| Aug 4, 2026 | -7.04% | Q2 FY2026 | Read transcript briefingQuarter in one view- Revenue $11.5B, +50% y/y, +13% q/q — above the $11.2B ±$300M guide (beat the midpoint by ~$300M, landed inside the range). Sixth consecutive quarter of >30% y/y growth. Comparable EPS +82% y/y; GM 56% (+200+ bps y/y, +80 bps q/q, in line with guide); opex $3.4B vs ~$3.3B guided — a fifth consecutive overrun. FCF $1.6B (down from $2.6B in Q1); CFO $2.4B; cash $13.1B; inventory up to ~$8.5B to support DC demand.
- Segments: Data Center record $6.7B (+107% y/y, +16% q/q; op income $2.1B / 31% — recovered from 28% in Q1). Server CPU: fifth consecutive record, cloud and enterprise each +70%+ y/y, exceeding the +70% guide. Instinct more than doubled y/y. Client & Gaming $3.8B (+6% y/y; op margin 15% vs 21% a year ago): Client $3.1B (+23%, record mobile), Gaming $779M (-31% y/y on semi-custom, +8% q/q). Embedded $977M (+19% y/y, best growth in 3+ years; op margin 40%).
- Q3 FY26 guide: ~$13B ±$300M (+41% y/y, +13% q/q); DC and Embedded both strong double-digit q/q growth; C&G down modestly (slight client growth offset by strong double-digit gaming decline); GM ~56%; opex ~$3.65B; tax 13%; ~1.66B shares; ~$55M other income.
- Headline strategic moves: Anthropic announced as third gigawatt-scale customer (up to 2GW MI450/Helios, first GW deploying 1H27, plus Claude-ROCm co-engineering); Microsoft will deploy Helios at scale on Azure. TAMs raised again: AI accelerator >$1.4T by 2030 (>45% CAGR) and server CPU ~$220B by 2030 (>50% CAGR — up from >$120B last quarter, which was itself up from ~$60B at Analyst Day). Server CPU now guided +80% y/y in 2H26 and +70%+ for FY27; total DC segment guided to "well over" double in 2027. Prior >35% revenue CAGR and $20 EPS targets now framed as materially conservative.
What management is focused on- Reframing 2027 around a much bigger TAM and steeper trajectory: Su now says AMD is "tracking materially ahead" of the November Analyst Day model, will grow "substantially above" the >35% revenue target, and will "significantly exceed" the $20 EPS target. The overall compute TAM is now ~$2T by 2030 at ~40% CAGR.
- Server CPU as a supply story: demand was "unforecasted" in 1H26, supply chain is "tight right now," and capacity additions (wafers, back-end, substrates) are the mechanism behind the raised 2H26 (+80% y/y) and FY27 (+70%+) guides. Moore extracted that 2027 supply "should be better than '26" and growth "may... go higher."
- Broadening the AI anchor base: Anthropic (up to 2GW, first GW 1H27) joins OpenAI and Meta; Microsoft Azure added as a Helios deployment channel for frontier-model inferencing. Su emphasized additional sub-gigawatt Helios customers for diversity (Schneider question).
- Helios ramp mechanics: in production now, initial shipments later this quarter (Q3), step-up in Q4, further step-up in Q1 2027, ramping through 2027. Customer pull "tracking ahead of our initial forecast" — clarified under Buchalter's pressure as a 2027 volume/demand comment, not yields.
- Annual rack-scale cadence: MI500 + Verano + Pensando platform for 2027, with expanded copper and optical scale-up domains; claimed >2,000x inference performance gain over 4 years; "multiple customers" already engaged on MI500.
- ROCm as a maturity claim: 3M+ models out of the box, 10x open-source contribution growth, ROCm.ai agentic dev platform (Claude, Codex, Cursor), co-optimization with OpenAI/Anthropic/Meta.
- Managing consumer downcycle: softer 2H PC market on memory/component costs, but client still expected to grow y/y in 2026 and outperform; gaming guided to strong double-digit q/q decline in Q3.
Key numbers and quarter mechanics- GM 56% (+80 bps q/q) on DC mix; Q3 guided 56%. Hu's 2027 GM framework: server CPU accretive, Embedded accretive, DC AI "slightly below corporate average" but adds large gross profit dollars; mix determines the path; no 2027 GM guide given (Muse deflected to "we'll give you more color when we get there").
- DC op margin 31% (vs 28% in Q1, 33% in Q4 with the reserve release) — underlying DC profitability is recovering on volume.
- Server CPU mechanics (Buchalter): Q2's +70% was double-digit growth in both units and ASP, "more unit" driven — consistent with last quarter's framing. ASP growth tied to core-count mix. Venice expected to expand workloads vs Turin.
- Opex $3.4B vs ~$3.3B guided (fifth straight miss); Q3 guided ~$3.65B. New framing from Hu (Malik): opex growth will be "less than top-line revenue growth" — an operating-leverage commitment, but no near-term ratio or timeline.
- FCF $1.6B vs $2.6B in Q1; inventory +$500M q/q to ~$8.5B (ramp build). No buyback disclosure in the transcript — capital return went unmentioned.
- Gaming: -31% y/y in Q2 (semi-custom cycle + component-cost-hit Radeon demand); Q3 guided strong double-digit q/q decline — the 2H gaming warning from last quarter is now landing.
- Revenue per gigawatt: Su explicitly pushed back on Arya's ~$15B/GW math — "we've said double-digit billions. We're still in that range." That implies a lower per-GW monetization than the Street's working assumption.
- Instinct dollars: still undisclosed. Muse's implied ~$30B 2027 Instinct estimate was called "probably too low" by Su, with DC AI guided "well over 100%" growth in 2027 — the closest thing to a number management has ever given, delivered verbally and without a dollar figure.
Product and launch scorecard- Helios/MI450: in production; initial shipments on track for later this quarter (Q3), ramp Q4 and into 2027 — consistent with prior phasing. Claimed up to 15% more throughput at same rack power and up to 30% more tokens/dollar vs competition. Demand "ahead of initial forecast" = 2027 volumes, per Su; yields expected to improve over the first few quarters (Buchalter) — an implicit acknowledgment that early-ramp yields/margins are not yet mature.
- Anthropic partnership: up to 2GW MI450 in Helios, first GW 1H27, multiyear Claude-ROCm joint engineering — third named gigawatt-scale anchor; no revenue sizing, margin, or warrant/equity terms disclosed.
- Microsoft: Helios at scale on Azure for frontier inferencing — first hyperscaler deployment channel named for Helios; no size or timing detail.
- Venice (Zen 6, 2nm): in production now — pulled forward in specificity from "on track for later this year." >30 SKUs, >2x performance/watt vs leading x86, up to 3.3x vs leading ARM CPUs; every major OEM launching platforms, leading cloud providers deploying later this year; demand "stronger than any prior EPYC generation."
- MI500/Verano (2027): next annual rack-scale platform; copper + optical scale-up; "largest generational leap in Instinct history"; multiple customers engaged — roadmap credibility claim, no product detail.
- MI355X: adoption broadening across AI labs, cloud providers, startups, national labs, and a sovereign deployment.
- Client: record mobile revenue; Ryzen Pro +50% y/y; Ryzen AI Halo developer systems on sale; Gorgon Halo (192GB unified memory, 300B-parameter local models) introduced with Hugging Face Pro bundling.
- Cerebras partnership: disaggregated fast-inference solution (Helios + wafer-scale engine), available Q4 in Cerebras cloud, extending into 2027 — the first concrete answer to the low-latency inference gap Arcuri pressed on last quarter, delivered via partnership rather than owned silicon.
- Embedded: +19% y/y, broad-based; embedded x86 "grew significantly" on hyperscaler networking/control-plane adoption; >$18B in new design wins tracked for the year (record); Ryzen AI Embedded x100 and Kria robotics platform launched.
Sell-side read-through- O'Malley (Barclays): CPU/GPU crossover timing — Su declined a crossover date; both grow strongly, AI TAM larger, "substantial growth" in DC AI in 2027. Extracted Hu's 2027 GM puts-and-takes: server accretive, DC AI below corporate average but gross-profit additive, Embedded a tailwind.
- Arcuri (UBS): caught an apparent tension — +80% 2H server guide could imply a weak Q4; Su answered that both server and DC AI grow in Q3 and grow again in Q4, with more supply coming. Also got the company-level commitment to outgrow the >40% TAM CAGR.
- Arya (BofA): the most economically important exchange. His ~3GW / ~$15B-per-GW framework was partially rejected — Su held revenue/GW at "double-digit billions," implying downside to Street per-GW assumptions, while asserting supply to "more than meet" guidance with upside potential. HBM: ~50% more HBM than competition acknowledged as a design advantage; Su conceded memory-footprint flexibility for medium-sized models — a quiet admission that HBM content could be tuned down if TCO doesn't justify it.
- Buchalter (TD Cowen): units-vs-ASP (both double-digit, units higher); pinned down that "ahead of forecast" = 2027 Helios volumes, and extracted that yields will improve over the first few ramp quarters — i.e., early Helios margins carry yield immaturity.
- Rakers (Wells Fargo): agentic/sandbox servers are the largest and fastest-growing piece of the $220B CPU TAM but the smallest today — near-term growth is still general-purpose and head-node driven. Client: 2026 y/y growth reaffirmed; market "held up better than most people thought."
- Rasgon (Bernstein): caught the press release's "data center accelerates in 2H" language against a Q3 guide that implies decelerating y/y growth vs Q2's +107% — Hu's answer (2H vs 1H, not vs Q2) resolved it, but the wording was sloppy. His "how much work is 'more' doing" question extracted the call's most quotable guide: DC AI 2027 growth "well over 100%," total DC "well over" double.
- Schneider (Goldman): customer diversity — anchors consume through CSPs plus sub-gigawatt customers; good diversity by Q4/Q1. Data-center readiness: nothing "that would give us pause," but Su introduced a range caveat — the range depends on "the ability to bring on more capacity in a timely fashion," the first hedged language on the 2027 ramp.
- Muse (Cantor): implied ~$30B 2027 Instinct from the guide; Su said his number is "probably too low" — a rare verbal steer upward on the most important undisclosed number. GM-by-quarter for Instinct in 2027 was declined.
- Moore (Morgan Stanley): server CPU is supply-constrained now; 2027 supply better forecasted; chiplet architecture reduces leading-edge wafer needs — the cleanest supply-chain answer of the call.
- Malik (Citi): Cerebras fast-inference timing (Q4 availability); extracted Hu's opex-leverage commitment (opex growth < revenue growth) tied to the "significantly higher than $20" EPS claim.
- Repeated themes: 2027 sizing (O'Malley, Arya, Rasgon, Muse — four separate attempts to pin down DC AI dollars, all answered directionally), server CPU supply/units/ASP (Arcuri, Buchalter, Rakers, Moore), GM trajectory into the MI450 ramp (O'Malley, Buchalter, Muse — no quantification given).
Management credibility- Delivery vs guide: revenue beat ($11.5B vs $11.2B midpoint); GM 56% as guided; server CPU +70%+ delivered against the +70% guide; DC +16% q/q vs "double digits" guided; Embedded +19% vs double-digit guide; gaming decline as flagged. The Q2 acceleration guide from last quarter was fully delivered.
- Credibility upgrades: Venice moved from "on track" to "in production" on schedule; Helios initial shipments reaffirmed for Q3; Anthropic converts "additional multi-gigawatt opportunities" into a third named anchor within one quarter; the server CPU +80% 2H guide is specific and falsifiable.
- Slippage: opex missed for the fifth consecutive quarter ($3.4B vs ~$3.3B) and Q3 guides to ~$3.65B — the new "opex grows slower than revenue" commitment is the third framing in three quarters (2H26 leverage promise → "investment drives revenue" → operating leverage), with no demonstrated control yet. FCF fell to $1.6B from $2.6B on inventory build.
- Evasions: Instinct dollars refused for the twelfth quarter running — four analysts (O'Malley, Arya, Rasgon, Muse) circled the 2027 DC AI number and got "well over 100%" and "your number is probably too low" instead of dollars. Revenue-per-GW was actively walked back from Street math without a corrected figure. 2027 GM declined. Anthropic and Microsoft deal economics undisclosed.
- Consistency watch: the server CPU TAM has now been raised twice in ~nine months ($60B → $120B → $220B), and the AI accelerator TAM to $1.4T — each revision attributed to demand unfolding faster than expected. The pattern is directionally bullish but makes every long-range number management publishes suspect as a forecast. The press release's "accelerating 2H" language did not survive Rasgon's arithmetic.
- Under pressure: Su was substantive and consistent on supply, units/ASP, and ramp phasing; the Arya per-GW pushback was unusually candid (it lowered a Street assumption on a record quarter). Hu's GM answers remain framework-only. No analyst asked about China, buybacks, or the opex streak this quarter — notable absences given prior quarters.
What changed versus the prior quarter- A third gigawatt-scale customer exists: Anthropic (up to 2GW, first GW 1H27, Claude-ROCm co-engineering) — the concentration-risk mitigation accelerated from one new name per quarter (Meta in Q1, Anthropic in Q2). Microsoft added as a named Azure deployment channel.
- TAMs raised again, sharply: server CPU $120B → $220B by 2030 (CAGR >35% → >50%); AI accelerator now sized at ~$1.4T by 2030 (>45% CAGR). Long-term targets upgraded: revenue "substantially above" >35% CAGR, EPS "significantly" above $20.
- Server CPU guidance stepped up again: from +70% y/y in Q2 to +80% y/y for 2H26 and +70%+ for FY27 — and the constraint is now explicitly supply, not demand ("tight right now," demand was "unforecasted").
- 2027 DC AI framing hardened: from "tens of billions" to "well over 100%" segment growth with Muse's ~$30B Instinct estimate called "too low" — the strongest verbal steer yet, still without dollars.
- Revenue-per-GW walked down: Su held "double-digit billions" per GW against Arya's ~$15B — a quiet tempering of per-deal monetization assumptions even as GW counts rise.
- Venice is in production (was "on track"), and MI500/Verano 2027 platform was pre-announced with an annual rack-scale cadence commitment.
- Gaming deterioration landed: -31% y/y in Q2 and a strong double-digit q/q decline guided for Q3 — last quarter's >20% 2H decline warning is now in the numbers.
- FCF dropped to $1.6B (from $2.6B) on a $500M inventory build for the ramp; buybacks went unmentioned.
- China disappeared entirely from the call — no questions, no commentary, after being the sharpest exchange last quarter.
- Opex framing changed again: fifth miss, but a new explicit commitment that opex grows slower than revenue — replacing last quarter's "investment drives revenue" rationale.
Bull case- The beat-and-raise cadence is compounding: Q2 beat, Q3 guided +41% y/y / +13% q/q with DC and Embedded both up strong double digits, and management is now verbally steering 2027 above Street models ("your number is probably too low").
- Three named gigawatt-scale anchors (OpenAI, Meta, Anthropic) plus Microsoft Azure as a deployment channel, plus claimed sub-gigawatt breadth — the customer-concentration bear case weakens with each quarter, and every anchor is multi-generation.
- Server CPU is a second, independent growth engine with raised visibility: +80% y/y 2H26, +70%+ FY27, supply-constrained (demand-led), units-driven share gains, Venice in production with claimed demand above any prior generation, and a $220B TAM.
- Margin structure held through the transition: 56% GM delivered and guided, DC op margin recovered to 31%, and Hu's 2027 bridge (server + Embedded accretion vs below-corporate DC AI) is internally consistent with holding the 55–58% range.
- Execution milestones keep landing on time: MI450 sampling (Q1) → Helios in production with Q3 initial shipments (Q2) → Venice in production ahead of prior "later this year" language.
- The low-latency inference gap got a first answer (Cerebras, Q4 availability), and ROCm's maturity claims (3M models, agentic tooling, lab co-optimization) address the software moat argument.
- EPS power is being demonstrated, not just promised: +82% y/y comparable EPS on +50% revenue, with an explicit opex-leverage commitment now on record.
Bear case- The most important number in the story remains undisclosed: twelve quarters without Instinct dollars, and this call added a verbal "well over 100%" / "too low" steer that raises the stakes on a number investors still cannot verify. If the 2027 ramp disappoints, there is no disclosed baseline to measure the miss against.
- Per-GW economics were quietly lowered: "double-digit billions" against the Street's ~$15B working assumption compresses the revenue implied by the announced GW totals — the GW headlines and the monetization math now point in different directions.
- TAM inflation is becoming a pattern: server CPU TAM up ~3.7x in ~nine months ($60B → $220B). Either the market is inflecting at unprecedented speed or the TAM is being sized to the narrative; Buchalter's prior-quarter challenge was never answered with data, and this quarter no one pressed it.
- Opex is a five-quarter miss streak with a third successive framing; the new "slower than revenue" commitment is unfalsifiable until 2027 guides arrive, and Q3's $3.65B guide is another step up.
- Early Helios margins carry yield risk: Buchalter extracted that yields "will improve over the first few quarters" — meaning Q4/Q1 absorb an immature, below-corporate-margin, HBM-heavy product at scale, with HBM cost inflation a live exposure Su only partially addressed (footprint "modifiable" for some workloads).
- Consumer drag is now in the actuals: gaming -31% y/y with another strong double-digit q/q decline guided; C&G op margin fell to 15% from 21%; memory inflation is hitting both gaming and 2H PC demand.
- FCF quality dipped: $1.6B vs $2.6B, inventory building to $8.5B, buybacks unmentioned — cash generation is being consumed by the ramp, and capital return has gone silent.
- Su's one hedge — the 2027 range "depends on the ability to bring on more capacity in a timely fashion" — is the first admission that the ramp's binding constraint may be data-center and supply-chain buildout, not demand.
Next-quarter watchlist- Q3 delivery: $13B ±$300M; DC and Embedded strong double-digit q/q growth; gaming strong double-digit q/q decline; GM 56%; opex ~$3.65B (sixth-miss watch, and first test of the "opex < revenue growth" commitment).
- Helios first shipments: confirmation that initial shipments began in Q3 as promised, rack-builder/OEM evidence, Q4 step-up sizing, and any early yield or margin commentary (Buchalter's yield question sets the trap).
- 2027 DC AI sizing: whether Q3 disclosure or guidance forces any dollar framing after the "well over 100%" / "too low" steers; watch for a fourth named at-scale customer and any quantification of the sub-gigawatt pipeline.
- Server CPU: evidence for the +80% 2H guide (third-party share data, capacity additions coming online), whether supply tightness eases as Moore was told, Venice platform launch cadence, and any Intel/ARM response to the >50% share ambition.
- Gross margin: whether 56% holds as Helios volume begins mixing in during Q4; any 2027 GM framework beyond "mix-dependent."
- Deal economics: Anthropic and Microsoft terms (revenue phasing, warrants/equity, custom-silicon margins) — the OpenAI/Meta disclosure gap now has two more siblings; reconcile "double-digit billions" per GW against announced GW totals.
- HBM/memory: allocation security for 2027, cost pass-through, and whether the "modifiable memory footprint" option becomes actual SKU strategy.
- Consumer: depth of the Q3 gaming decline vs guide; PC market trajectory into Q4 on memory costs; whether client y/y growth holds in 2H as promised.
- Cash and capital: inventory build vs Helios revenue conversion, FCF recovery, and whether buybacks resume with $13.1B cash.
- China: complete silence this quarter — any MI308/MI325 license development or resumption is now an unmonitored variable.
- Events: KeyBanc (Aug 11), Citi TMT (Sep 8), Goldman Communacopia (Sep 11).
- Carried gaps: Instinct dollar levels, server-CPU revenue split, MI350/MI450 ASPs, all anchor-deal economics and dilution, MI450-ramp GM quantification, 2026/2027 GPU revenue sizing.
|
| May 5, 2026 | +18.61% | Q1 FY2026 | Read transcript briefingQuarter in one view- Revenue $10.3B, +38% y/y, flat q/q — above the high end of the ~$9.8B ±$300M guide. EPS $1.37 (+43%), record FCF $2.6B (25% of revenue; CFO $3.0B). GM 55% (+170 bps y/y), clean this quarter with no reserve-release distortion flagged. Opex $3.1B (+42% y/y) vs ~$3.05B guided — a fourth consecutive overrun.
- Segments: Data Center record $5.8B (+57% y/y, +7% q/q; op income $1.6B / 28% vs 33% in Q4 — margin fell back once the Q4 reserve release lapped). Server CPU revenue +50%+ y/y, fourth consecutive record; DC AI down modestly q/q on the China step-down. Client & Gaming $3.6B (+23% y/y, -9% q/q; op margin 16%): Client $2.9B (+26%), Gaming $720M (+11% y/y on Radeon 9000, semi-custom down as expected). Embedded $873M (+6% y/y; op margin 39%).
- Q2 FY26 guide: ~$11.2B ±$300M (+46% y/y, +9% q/q) — a large acceleration; DC and Embedded both up double digits q/q, C&G up modestly; GM ~56%; opex ~$3.3B; tax 13%; ~1.66B shares; ~$60M other income.
- Headline strategic moves: server CPU TAM doubled — from ~$60B/18% CAGR (Analyst Day, November) to >$120B by 2030 at >35% CAGR, driven by agentic AI CPU demand; server CPU guided +70% y/y in Q2. Meta announced as a second gigawatt-scale customer: up to 6GW of Instinct across multiple generations including a custom MI450-based GPU, shipments starting 2H26. "Tens of billions" 2027 DC AI revenue reaffirmed with "strong and increasing confidence," and Su now says AMD can exceed its >80% DC AI growth target.
What management is focused on- Server CPU as the newly dominant story: agentic AI is framed as structurally additive to CPU demand (orchestration, data movement, parallel execution), with CPU:GPU ratios moving from 1:4/1:8 host-node configurations toward 1:1. Su split the CPU TAM into three buckets — general purpose (low-double-digit growth), AI head nodes, and agentic AI (the largest growth piece). Capacity is being "meaningfully increased" on wafer and back-end; customers are now discussing 2027 and 2028 CPU demand.
- Broadening the AI customer base: Meta (up to 6GW, custom MI450-based accelerator, Helios rack-scale with Venice CPUs) joins OpenAI as a named anchor. Su claims lead-customer MI450 forecasts now exceed initial 2027 plans, with "additional multi-gigawatt opportunities" in engagement — the concentration-risk answer finally has a second name.
- MI450/Helios execution: MI450 sampling to lead customers has begun; Helios production shipments 2H26, initial volume Q3, significant ramp Q4, continuing into Q1 2027. Visibility claimed "down to which data centers the GPU is going to be installed in."
- Venice (Zen 6, 2nm) on track for later this year, including Verano — the first AI-infrastructure-purpose-built EPYC; claimed >2x throughput per socket vs leading ARM-based AI solutions; more customers validating at this stage than any prior EPYC generation.
- Supply assurance as a repeated theme: memory supply "secured... to meet and exceed our targets" (Arya); supply chain "tight" but "an area where we excel" (Curtis); confidence in supplying above guided growth levels (O'Malley).
- Managing the consumer downcycle: memory/component inflation expected to hit 2H PC shipments and gaming (2H gaming revenue guided down >20% vs 1H); AMD plans to grow client y/y and outperform the market anyway, leaning on commercial (Ryzen Pro sell-through +50% y/y).
Key numbers and quarter mechanics- GM 55% (+170 bps y/y) on favorable mix (higher DC contribution); Q2 guided 56%. Hu's 2H bridge: tailwinds from server CPU mix, richer client mix, accretive Embedded, and lower gaming; offset by MI450 ramping below corporate average in Q4 — she nonetheless feels "really good" about 2026 GM and reiterated the 55–58% long-term range.
- DC op margin 28% vs 33% in Q4 — the Q4 reserve-release benefit lapped; underlying DC margin trajectory is the honest read now.
- Server CPU mechanics (Arcuri): Q1 growth was "much more unit driven" though ASPs and units both rose y/y; Q2/back-half growth is majority units, with modest ASP increases explicitly framed as passing through supply-chain inflation, not price-taking. Hu added that generational core-count increases lift ASP via mix. Turin crossed >50% of server revenue; Genoa still strong; Milan declining.
- DC AI was down modestly q/q in Q1 due to China ($390M in Q4 vs "not material" in Q1 — Hu, pressed by Rasgon). Q2: both server and DC AI guided up double digits q/q.
- Opex $3.1B vs ~$3.05B guided (fourth straight miss); Q2 guided ~$3.3B. Hu's explanation: investment is driving the revenue acceleration, some opex is revenue-linked, and customer-engagement support requires resources. SG&A has been outpacing R&D recently (Curtis observation); Hu committed that for the full year R&D grows faster than SG&A. Su: SG&A is going into enterprise servers, commercial PCs, and SMB — "paying off."
- Cash $12.3B (up from $10.6B); inventory ~flat at $8B; buybacks only $221M (1.1M shares), $9.2B authorization remaining — repurchase pace still timid despite record FCF.
- Gaming: 2H26 revenue guided down >20% vs 1H on memory/component costs — a new, specific negative disclosure beyond the prior "significant double-digit FY decline" framing for semi-custom.
- GPU dollars: still undisclosed. Rasgon's attempt to pin down whether Q1 DC AI grew q/q ex-China was deflected twice ("China revenue in Q1 is not material" — a non-answer to the ex-China question).
Product and launch scorecard- MI450/Helios: sampling underway to lead customers; production ramp 2H26 (Q3 initial volume, Q4 significant, continuing into Q1 2027 — phasing consistent with last quarter's disclosure and now extended). Lead-customer forecasts above initial 2027 plans; largest deployments are inference. Meta custom MI450-based GPU is a new design-win proof point for the architecture.
- Meta partnership: up to 6GW, multi-generation, co-designed custom accelerator, Helios rack-scale with Venice CPUs, shipments 2H26 — the first second named gigawatt-scale customer, directly addressing the biggest prior-quarter gap.
- MI355X: MLPerf results cited — "strong competitive performance across the full suite with leadership results in multiple categories"; day-0 support for Gemma 4, Qwen, Kimi. ROCm cadence accelerated via increased investment and agent-based coding workflows.
- EPYC: fourth consecutive record server CPU quarter; cloud and enterprise each +50%+ y/y; record enterprise revenue and sell-through; >1,600 cloud instances (+~50% y/y). Venice on track for later this year with Verano AI-optimized variant; >50% share target reaffirmed (Buchalter).
- Client: Ryzen AI 400 / AI Pro 400 desktop CPUs launched; notebook/premium mix strengthening; desktop softer on memory costs; commercial sell-through +50% y/y (accelerating from +40%) with Dell/HP/Lenovo broadening offerings.
- Gaming: Radeon 9000 demand drove +11% y/y; FSR software updates; semi-custom down as expected; next-gen platform engagements "strong" (no new timing detail).
- Embedded: +6% y/y on test/measurement/emulation, aerospace/defense, communications, embedded x86; design wins up double digits y/y; semi-custom engagements expanding into data center and comms customers.
Sell-side read-through- Buchalter (TD Cowen): challenged the TAM doubling ($60B → $120B in ~six months) and the >50% share target against improving Intel supply and ARM momentum — Su attributed it to bottoms-up customer forecasts and workload analysis over "the last few months," and held the share target. Also extracted that MI450 forecast upside is broad-based across customers, not just OpenAI/Meta upsizing, and is a 2027-timeframe comment.
- O'Malley (Barclays): got the cleanest intra-year cadence — Q1 DC AI down modestly q/q (China), Q2 both server and DC AI up double digits, Helios Q3 initial/Q4 significant/continuing Q1. Supply question answered with data-center-level visibility and confidence in exceeding guided growth; Ramsay cut off his multi-part follow-up.
- Seymore (Deutsche Bank): x86/ARM differentiation — Su's answer: portfolio breadth across workload-optimized CPUs; hyperscalers will use both x86 and ARM. GM framework: Hu gave the full 2H puts-and-takes (CPU mix, client mix, Embedded accretion, gaming decline vs below-corporate MI450) — the most complete 2026 GM bridge to date, and it partially closes last quarter's open question.
- Arcuri (UBS): units-vs-ASP — majority unit-driven, ASP mostly inflation pass-through plus generational mix; also probed low-latency inference architectures (the competitor's 20%+ claim) — Su conceded architectural fragmentation is natural and pointed to the full compute portfolio plus semi-custom, still no discrete low-latency product answer.
- Arya (BofA): agentic CPU growth is "largely additive" to AI TAM, not cannibalizing GPUs; CPU:GPU ratios moving toward 1:1. Memory: supply secured to "meet and exceed" targets; the bigger watch item is consumer demand destruction, already embedded in 2H PC/gaming planning.
- Rakers (Wells Fargo): AI-optimized CPU ASP structure — no number given ("depends on the workload"), only direction (core counts up → ASP up). ARM framed as "point products relative to a portfolio."
- Muse (Cantor): client 2H seasonality and ASP implications — notebook/premium mix up, desktop softer, y/y growth expected despite memory-driven demand impact. On Instinct GM: Hu acknowledged compute is effectively sold out and margins could move toward corporate average over time via ASP and cost-at-scale, but the near-term priority is topline — no commitment.
- Rasgon (Bernstein): the sharpest exchange of the call. His ex-China q/q AI GPU growth question was deflected twice; Hu's "China in Q1 is not material" did not answer whether ex-China AI grew. His opex-forecasting challenge (fourth miss) got the "investment drives revenue" defense, not a forecasting-fix commitment.
- Curtis (Jefferies): older-node longevity (competitor 7nm restart) — Su says AMD sees the opposite: customers prefer newer products; Turin >50% of revenue, Milan declining. SG&A-vs-R&D growth question extracted the commitment that FY R&D grows faster than SG&A.
- Repeated themes: supply sufficiency (O'Malley, Arya, Curtis — uniformly confident), CPU TAM/share credibility (Buchalter, Seymore, Rakers), opex discipline (Rasgon, Curtis), GPU dollar disclosure (Rasgon — still refused).
Management credibility- Delivery vs guide: revenue beat the high end ($10.3B vs $9.8B midpoint); DC up sequentially as guided (+7%); server CPU delivered the anti-seasonal q/q increase that was last quarter's sharpest falsifiable claim; GM 55% as guided and "clean" as promised; Embedded grew y/y as guided; gaming down 15% q/q "consistent with expectations."
- Credibility upgrades: the anti-seasonal server CPU guide was met with +50% y/y growth and a +70% Q2 guide — the "multi-quarter phenomenon" claim is now strongly evidenced. The Meta 6GW deal converts "a number of other customers" from assertion into a named second anchor. MI450 sampling began on schedule. The 2H GM bridge (Seymore) and units/ASP split (Arcuri) were answered with unusual specificity.
- Slippage: opex missed guide for the fourth consecutive quarter ($3.1B vs ~$3.05B) and Q2 guides to ~$3.3B — the 2H26 leverage promise from last quarter was not reiterated this call; the framing has shifted to "investment drives revenue," which is a rationale, not a control mechanism. DC op margin fell to 28% from 33% once the reserve release lapped.
- Evasions: Rasgon's ex-China q/q AI GPU question deflected twice — management would not confirm the underlying GPU business grew sequentially in Q1, which is the single most important undisclosed mechanic in the quarter. FY25/Q1 Instinct dollars still refused. AI-CPU ASP premium declined (Rakers). Instinct margin trajectory acknowledged but not committed (Muse).
- Consistency watch: the server CPU TAM doubled in roughly six months ($60B → $120B) — Su explained it as demand unfolding faster than expected, but the speed of the revision cuts both ways for forecast reliability. The >80% DC AI growth target is now framed as something to "exceed," another upward drift in long-range language without re-anchored numbers.
- Under pressure: Su was substantive on TAM methodology, CPU/GPU ratios, and competitive positioning; Hu was specific on GM bridge but visibly evasive on the Rasgon exchange — the pattern of accepting directional framings while refusing GPU numbers is unchanged.
What changed versus the prior quarter- A second gigawatt-scale customer exists: Meta (up to 6GW, custom MI450-based GPU, 2H26 shipments) — the biggest single change; OpenAI concentration risk is now partially mitigated by a named, co-engineered, multi-generation deal.
- Server CPU went from growth engine to headline: TAM doubled to >$120B by 2030 (>35% CAGR vs 18%), Q2 guided +70% y/y, capacity being meaningfully increased, and agentic AI is now the central demand narrative. Last quarter's anti-seasonal Q1 CPU guide was delivered.
- Guidance accelerated sharply: Q2 at $11.2B (+46% y/y, +9% q/q) vs Q1's -5% q/q guide — the back-half-weighted year now has a strong first-half bridge, with both server and DC AI up double digits q/q in Q2.
- China faded to immaterial: Q1 China revenue "not material" (vs $100M guided), and DC AI declined q/q because of it — but management refused to confirm ex-China AI grew, leaving the underlying GPU trajectory ambiguous.
- GM quality normalized: 55% clean (no reserve release), guided to 56% in Q2, with a full 2H puts-and-takes bridge disclosed — but DC op margin dropped to 28% from 33%, revealing how much the Q4 print relied on one-time items.
- Gaming outlook worsened: 2H26 revenue now guided down >20% vs 1H on memory/component costs — a new, quantified consumer headwind beyond the semi-custom cycle decline.
- Opex ratchet continued: fourth straight miss, Q2 guided ~$3.3B, and the prior quarter's explicit 2H26 leverage promise was replaced by an investment-justification framing; a new commitment emerged that FY R&D grows faster than SG&A.
- Capital return stayed minimal: $221M buyback against $2.6B FCF and $12.3B cash — the gap between cash generation and repurchase widened further.
Bull case- The growth algorithm is accelerating and broadening: Q2 guided +46% y/y with both DC engines up double digits q/q; server CPU +70% y/y with capacity secured and customers planning into 2028; the CPU TAM doubling is grounded in claimed bottoms-up customer forecasts.
- Concentration risk materially reduced: Meta's up-to-6GW, multi-generation, custom-silicon commitment plus OpenAI gives AMD two named anchor customers, with management claiming lead-customer forecasts above plan and additional multi-gigawatt opportunities in pipeline — supporting "tens of billions" in 2027 and an "exceed" on the >80% DC AI CAGR target.
- MI450/Helios execution is on schedule with rising evidence: sampling underway, data-center-level deployment visibility, HPE/Lenovo racks, Meta custom part, Q3 initial volume / Q4 significant ramp reaffirmed.
- Margin structure is intact and improving: 55% clean GM, 56% guided, with server CPU mix, client mix, and Embedded accretion positioned to offset below-corporate MI450 margins in Q4; Hu reaffirmed the 55–58% long-term range and acknowledged room to move Instinct margins toward corporate average over time.
- Server CPU economics are attractive: growth is majority unit-driven (share, not price), ASPs rising on generational mix, and agentic AI is additive to the AI TAM rather than cannibalizing GPUs (CPU:GPU toward 1:1).
- Client keeps gaining share through a downcycle: commercial sell-through +50% y/y, premium/notebook mix improving, y/y growth expected despite memory-driven TAM pressure.
- Balance sheet: $12.3B cash, record $2.6B FCF, inventory flat — the 2H26 ramp is funded with supply claimed locked.
Bear case- The core GPU question went unanswered again: DC AI declined q/q in Q1, and management refused to confirm it grew ex-China — meaning the underlying Instinct trajectory entering the biggest ramp in company history is still unverifiable, with no dollar disclosure for the eleventh quarter running.
- Opex is an uncontrolled ratchet: four consecutive guide misses, Q2 guided to ~$3.3B, and the prior 2H26 leverage promise was quietly replaced by "investment drives revenue" — if the MI450 ramp slips, the cost base will not.
- The TAM doubling invites skepticism: raising the server CPU TAM from $60B to $120B within ~six months, and the >50% share target held against improving Intel supply and ARM momentum, is either a genuine demand inflection or forecast inflation — Buchalter's challenge was answered with narrative, not data.
- DC op margin fell to 28% (from 33%) — the Q4 margin recovery is now confirmed as substantially one-item-driven, and Q4 2026 will absorb below-corporate MI450 margins against a much larger revenue base.
- Consumer exposure is a growing drag: 2H gaming guided down >20% vs 1H, PC shipments planned lower in 2H on memory costs — two segments now explicitly impaired by component inflation in the second half.
- Supply confidence remains total and unhedged ("secured enough supply to meet and exceed our targets"; "confident in our ability to supply... and to exceed") on a complex rack-scale ramp — any 2H26 hiccup reads as execution failure against explicit assurances.
- Meta deal economics undisclosed: no revenue sizing, no margin commentary on the custom part, no warrant/dilution analog disclosed — the OpenAI warrant-economics gap now has a sibling.
- SG&A outpacing R&D in recent quarters is an odd pattern for a company claiming technology-led share gains; the fix (R&D growing faster for the FY) is only a commitment.
Next-quarter watchlist- Q2 delivery against the big guide: $11.2B ±$300M (+46% y/y); server CPU +70% y/y; both server and DC AI up double digits q/q; GM 56%; opex ~$3.3B (fifth-miss watch); Embedded double-digit growth.
- MI450/Helios milestones: Q3 initial-volume confirmation, rack-builder shipment evidence, first Meta/OpenAI deployment signals, any third named at-scale customer, and Advancing AI event (July) for next-gen Instinct/EPYC/Helios detail.
- GPU disclosure: whether Q2's double-digit DC AI growth forces any dollar or ex-China framing; Rasgon's ex-China q/q question remains open and is now a standing challenge.
- Server CPU: evidence for the +70% guide (third-party share data, capacity additions), units-vs-ASP split as inflation pass-through flows through, Venice launch timing specificity, and any sign of Intel supply recovery or ARM share impact on the >50% share target.
- Gross margin: whether 56% lands in Q2 and whether Hu's 2H bridge (CPU/client/Embedded tailwinds vs MI450 dilution) holds as Q4 approaches; any update to the 55–58% long-term range.
- Opex: first test of the R&D-faster-than-SG&A commitment; any re-anchoring of the abandoned 2H26 leverage promise.
- Consumer: memory-cost impact on Q2 client/gaming actuals vs plan; whether 2H PC shipment cuts deepen; gaming tracking toward the >20% 2H decline.
- Meta deal economics: revenue phasing, custom-part margin commentary, any warrant or equity component.
- China: MI325 license status and any resumption of China GPU revenue beyond the immaterial Q1 level.
- Capital allocation: buyback pace vs $9.2B authorization with $12.3B cash; Hu presents at BofA Global TMT (June 2).
- Carried gaps: Instinct dollar levels, server-CPU revenue split, MI350/MI450 ASPs, OpenAI/Meta deal economics and dilution, MI450-ramp GM impact quantification, 2026 GPU revenue sizing.
|
| Feb 3, 2026 | -17.31% | Q4 FY2025 | Read transcript briefingQuarter in one view- Revenue $10.3B, +34% y/y, +11% q/q — above the ~$9.6B ±$300M guide, with ~$390M of MI308 China revenue that was explicitly excluded from guidance. Net income $2.5B (+42%), record EPS $1.53 (+40% y/y), record FCF $2.1B (CFO $2.3B). FY2025: revenue $34.6B (+34%), EPS $4.17 (+26%), GM 52%.
- Non-GAAP GM 57% (+290 bps y/y) — but flattered by a $306M release of previously reserved MI308 inventory. Ex-reserve-release and ex-China, GM ~55% (+80 bps y/y). Opex $3.0B (+42% y/y, vs ~$2.8B guided — a third consecutive overrun). Op income record $2.9B / 28% margin.
- Segments: Data Center record $5.4B (+39% y/y, +24% q/q; op income $1.8B / 33% vs 30% — margin recovered sharply, aided by the reserve release). Client & Gaming $3.9B (+37% y/y, -3% q/q; op income $725M / 18%): Client record $3.1B (+34% y/y, +13% q/q), Gaming $843M (+50% y/y, -35% q/q — the pre-flagged console cliff landed). Embedded $950M (+3% y/y, +11% q/q; op margin 38% vs 39%) — the promised return to y/y growth delivered.
- Q1 FY26 guide: ~$9.8B ±$300M (+32% y/y, -5% q/q), including ~$100M MI308; DC up sequentially (both CPU and GPU incl. China), C&G and Embedded down seasonally; GM ~55%; opex ~$3.05B; tax 13%; ~1.65B shares; ~$35M other net income.
- Headline strategic claims: DC segment revenue ">60% annually over the next three to five years," "tens of billions" AI revenue in 2027 reaffirmed, and Analyst Day targets reiterated (>35% revenue CAGR 3-5yr, >$20 EPS "in the strategic time frame"). Su said >60% DC growth is "certainly possible in 2026."
What management is focused on- MI450/Helios as the 2026 inflection: development "right on track" for 2H26 launch; revenue starts Q3 2026 with "significant volume in Q4" (Su, to Arcuri) — the first intra-year phasing disclosure for the MI450 ramp. OpenAI ramp "on schedule" for 2H26 (to Reitzes), with "deep co-development" across OpenAI and CSP partners.
- Broadening beyond OpenAI: repeated emphasis on "a number of other customers" in active discussions for at-scale multi-year MI450/Helios deployments — the concentration-risk answer is now a standing talking point, still without a second named gigawatt-scale customer.
- Server CPU as a co-equal growth engine: Turin >50% of server revenue (up from ~half last quarter); record cloud and enterprise CPU sales; CPU order book strengthened "especially over the last sixty days"; Q1 server CPU guided up sequentially against a normal high-single-digit seasonal decline — an unusual anti-seasonal signal. Server CPU TAM framed as "strong double digits" growth in 2026.
- Portfolio widening around MI400: MI455X + Helios (superclusters), MI430X (HPC/sovereign — new wins: Genc/France, HLRS/Germany), MI440X (8-GPU enterprise). MI500 (CDNA6, 2nm, HBM4e) confirmed for 2027.
- Supply chain pre-emption: multiyear HBM/wafer agreements in place (O'Malley); Su flatly denied supply limitation on the 2H26 ramp ("I do not believe that we will be supply limited" — Schneider) and cited increased server CPU capacity as enabling the Q1 guide raise (Muse).
- China deliberately de-emphasized: Q4 MI308 revenue came from early-2025 orders on an approved license; only $100M forecast for Q1; nothing beyond ("dynamic situation"); MI325 licenses submitted, pending.
Key numbers and quarter mechanics- GM bridge: 57% reported = ~55% underlying + $306M MI308 reserve release + $390M low-cost-basis China revenue. Hu confirmed the reserve release also covers the Q1 $100M China shipment, making the Q1 55% guide "very clean" (Rasgon exchange). Note: Hu said "$360 million" once in the Rasgon answer vs $306M in prepared remarks — likely a transcript/misspeak discrepancy worth checking against the press release.
- Opex: $3.0B actual vs ~$2.8B guided — third straight miss; Q1 guided ~$3.05B. Su's new framing: opex grows slower than revenue in 2026, "especially... second half" — leverage is now a 2H26 promise, not a current fact.
- DC op margin 33% (vs 25% in Q3, 30% a year ago) — a big sequential jump, but the reserve release inflates it; underlying margin quality needs adjustment.
- Client: record $3.1B, +13% q/q (guided up, delivered); desktop record for the fourth consecutive quarter; commercial sell-through +40% y/y (accelerating from +30%). But Su now models PC TAM "down a bit" in 2026 on memory-cost inflation, with 2H "subseasonal" vs 1H — AMD expects to grow share in a down market.
- Gaming: $843M, -35% q/q as pre-flagged; FY26 semi-custom guided down "significant double-digit" for the full year (seventh year of console cycle); next-gen Xbox (AMD SoC) confirmed for 2027 launch; Valve Steam Machine shipping early 2026.
- Embedded: $950M, +3% y/y — first y/y growth of the recovery; design wins $17B in 2025 (+~20% y/y), >$50B cumulative since Xilinx. Q1 guided down seasonally / "modest growth" y/y.
- Cash: $10.6B cash + STI (up from $7.2B); inventory $7.9B (+$70M q/q) "to support strong data center demand"; FY buybacks $1.3B total (12.4M shares), $9.4B authorization remaining — repurchase pace stayed timid even with record FCF.
- GPU dollars: still undisclosed. Rasgon asked directly for the FY2025 Instinct number and was refused; Su offered only that ex-China DC AI revenue grew Q3→Q4. Arcuri floated a ~$14B 2026 DC GPU street number; management neither endorsed nor corrected it.
Product and launch scorecard- MI350/MI355: record Instinct revenue in Q4; hyperscaler availability expanded, new cloud providers launched MI350 instances; ramp continues through 1H26 — the no-air-pocket commitment from last quarter is holding.
- MI308 (China): first revenue since the ban — $390M in Q4 from early-2025 orders, $100M guided Q1, zero beyond. Treated as non-recurring; MI325 license applications submitted. The China optionality is now explicitly de-risked from guidance.
- MI450/Helios: on track for 2H26; revenue Q3, volume Q4; "vast majority" of 2026 MI450 is rack-scale, revenue recognized on shipment to rack builder (Moore); 8-way variant exists but minor in 2026. HPE (with Juniper Ethernet) and Lenovo announced as Helios rack OEMs. Rack-level and silicon testing "so far, so good" with parallel customer-driven testing — Moore's Nvidia-ramp-failure question got a process answer, not new evidence.
- OpenAI: 6GW ramp on schedule for 2H26; Su reaffirmed "tens of billions" 2027 and did not dispute Rakers' "high $20 billions" 2027 street framing — she answered around it ("we feel very good about that").
- EPYC: Turin >50% of server revenue; >230 new cloud instances in Q4 (>500 in 2025; ~1,600 total, +50% y/y); on-prem large-business deployments more than doubled in 2025; >3,000 OEM solutions. Venice launching "later this year" with claimed large-scale cloud engagements already underway.
- ROCm/software: vLLM upstream integration, day-zero support for leading models, enterprise AI suite launched, TCS partnership, healthcare imaging framework support — incremental ecosystem proof points, no new performance claims this quarter.
- Client: Ryzen AI 400 launched at CES (notebooks already shipping); Ryzen AI Halo dev platform (128GB unified memory, 200B-parameter local models). Commercial is the stated 2026 growth engine.
- Embedded: Versal AI Edge Gen2 in production, Spartan UltraScale+ shipping, three new embedded CPU families (EPYC 2005, Ryzen P100, X100) — cadence strong, revenue inflection just beginning.
Sell-side read-through- Rakers (Wells Fargo): tested the "high $20B" 2027 AI street number — Su reaffirmed "tens of billions" without disputing the figure; engagement commentary stayed qualitative ("proceeding very well").
- Arcuri (UBS): got the most useful new disclosure of the call — MI450 revenue starts Q3, significant volume Q4; Q1 DC up with both CPU and GPU (incl. China) up. His ~$14B 2026 GPU framing went unanswered.
- Arya (BofA): extracted the China clarification (nothing beyond $100M; MI325 licenses pending) and got Su to say >60% DC growth is "certainly possible in 2026" — a notable in-year endorsement of the long-term rate.
- Muse (Cantor): server CPU supply — Su admitted increasing TSMC capacity and raising the Q1 CPU guide on supply; pricing question ignored. GM framework: Hu gave mix tailwinds (Turin, MI355, commercial client, embedded recovery) but deferred MI450-ramp GM impact to "when we get there" — the 2026 GPU margin question remains open for a third quarter.
- Moore (Morgan Stanley): rack-vs-discrete (vast majority rack in 2026), revenue recognition at rack-builder shipment, and ramp-risk probing — management confident, no new mitigation specifics.
- Rasgon (Bernstein): the opex-leverage challenge ("every quarter you guide it up and it comes in higher") got a 2H26 leverage promise; his one-liners extracted the clean-Q1-GM confirmation and another refusal on FY25 Instinct dollars (softened only by "ex-China AI grew q/q").
- Buchalter (TD Cowen): memory-inflation/pull-in risk — Su conceded PC TAM down, 2H subseasonal; on SRAM/ASIC inference competition, Su leaned on chiplet flexibility and "no one size fits all" — no discrete inference-product answer.
- Reitzes (Melius): OpenAI on-track confirmation; x86-vs-ARM (incl. Nvidia's standalone ARM CPU) answered with agentic-workload x86 incumbency argument.
- O'Malley (Barclays): HBM procurement is multiyear and already locked for the ramp; architecture-evolution answer stayed at "flexible chiplet platform" level.
- Seymore (Deutsche Bank): best GM answer of the call — Hu: each GPU generation should carry higher GM long-term, but early-ramp GM is lower, improving with yield/scale — the first real generational margin framework. Gaming: FY26 semi-custom down significant double-digit, reversal with next-gen ramp.
- Schneider (Goldman): flat supply-limitation denial for 2H26; opex detail — 2025 spend went to AI hardware roadmap acceleration, software, ZT, go-to-market; 2026 continues "aggressively" but slower than revenue.
- Repeated themes: GPU dollar disclosure (Arcuri, Rasgon — still zero), 2026 GM shape through the MI450 ramp (Muse, Seymore — partially answered), supply constraints (Muse, O'Malley, Schneider — uniformly denied), OpenAI status (Rakers, Reitzes).
Management credibility- Delivery vs guide: revenue beat ($10.3B vs ~$9.6B midpoint, even ex-China ~$9.9B above midpoint); DC up double-digit q/q as guided (+24%); client up as guided (+13%); gaming down strong double digits exactly as pre-called (-35%); Embedded returned to y/y growth as promised; GM ~54.5% guided vs ~55% underlying — clean on an adjusted basis.
- Credibility upgrades: the Q4 console cliff and Embedded inflection both landed on script; MI450 phasing (Q3 start, Q4 volume) is now specific and checkable; the anti-seasonal Q1 server CPU guide is a falsifiable claim; the 2H26 opex-leverage promise is now on record (Su, to Rasgon).
- Slippage: opex missed guide for the third consecutive quarter ($3.0B vs ~$2.8B) and Q1 guides higher (~$3.05B) — "disciplined execution" (Hu) still does not apply to cost. The 57% headline GM required a reserve release; the underlying 55% is the honest number.
- Evasions: FY2025 Instinct revenue refused despite the year being closed (Rasgon); 2026 GPU sizing deflected (Arcuri's $14B); MI450-ramp GM impact deferred (Muse); server CPU pricing question ignored (Muse); no second named gigawatt-scale customer despite repeated "multiple customers" claims.
- Consistency watch: Hu's "$360M" vs "$306M" reserve-release figures; Su's "tens of billions 2027" now coexists with an unchallenged "high $20Bs" street framing — the commitment is drifting upward without explicit re-anchoring.
- Under pressure: Su was direct and specific on MI450 phasing, China, and supply; Hu gave the most substantive GM framework to date (Seymore). The pattern of accepting directional framings while refusing numbers continues, but the phasing disclosures this quarter were genuinely new.
What changed versus the prior quarter- China flipped from excluded to included — modestly: $390M recognized in Q4 (from early-2025 orders), $100M in the Q1 guide, and a $306M reserve release boosting reported GM. But management capped the forward contribution at $100M and called the situation "dynamic" — optionality realized once, then re-frozen.
- MI450 ramp got a calendar: Q3 revenue start, Q4 significant volume, 2H26 OpenAI first gigawatt on schedule, HPE and Lenovo named as Helios OEMs. Last quarter's "2H26" is now an intra-year bridge analysts can model.
- Server CPU evidence strengthened: Turin crossed >50% of server revenue, order book "strengthened over the last sixty days," and Q1 CPU guided up against seasonality — the "multi-quarter phenomenon" claim now has an anti-seasonal guide behind it, plus an admitted capacity increase.
- Opex went from step-function to ratchet: third straight guide miss, Q1 guided to ~$3.05B, and management responded by promising 2H26 leverage rather than near-term control.
- PC outlook turned cautious: memory inflation now expected to shrink the 2026 PC TAM with a subseasonal 2H — a new macro caveat on the segment that has been the most consistent beat.
- DC op margin recovered to 33% (from 25%) — but aided by the reserve release; underlying trajectory less clear.
- Cash position jumped to $10.6B; buybacks remained minimal ($1.3B FY) — capital return still not a story despite record FCF.
- Long-range targets restated (>35% revenue CAGR, >60% DC CAGR, >$20 EPS) with Su endorsing >60% DC growth as possible in 2026 — the Analyst Day framework is now being applied in-year.
Bull case- The 2026 setup is now phased and checkable: MI355 ramps 1H, MI450 revenue starts Q3 with significant Q4 volume, OpenAI first gigawatt on schedule, Venice launches 2H — and Su says >60% DC growth is "certainly possible" this year, not just long-term.
- Server CPU is inflecting independently: anti-seasonal Q1 guide, order book strengthening, Turin >50% of revenue, capacity being added, TAM growing strong double digits on agentic-AI CPU demand — a second engine with supply secured.
- Underlying margin structure improved: ~55% ex-one-items GM (+80 bps y/y) with GPU mix rising; Q1 guided 55% "clean"; Hu's generational framework (each GPU generation higher GM at scale) supports margin expansion through MI450/MI500.
- Client share gains are structural: record $3.1B, commercial sell-through +40%, and management expects to grow even in a down 2026 PC market.
- Embedded finally inflected (+3% y/y, 38% margin) with $17B record design wins — the recovery is now evidenced, not promised.
- Balance sheet: $10.6B cash, record $2.1B FCF, inventory built deliberately for the DC ramp — the 2H26 ramp is funded and supply is claimed locked via multiyear agreements.
- Ecosystem breadth: 8 of top 10 AI companies on Instinct, HPE/Lenovo Helios racks, MI430X sovereign wins (France, Germany), MI500 on track for 2027 — the franchise is wider than OpenAI.
Bear case- GPU P&L remains a black box entering the biggest ramp in company history: FY25 Instinct dollars refused, 2026 sizing deflected, MI450-ramp GM impact deferred — investors must underwrite Q4 volume with no disclosed base and no margin framework for the transition quarter.
- Opex is a ratchet: three consecutive guide misses, ~$3.05B guided for Q1, and the leverage promise is explicitly back-half-loaded — if MI450 slips, the cost base doesn't.
- Reported quality of Q4 beat is inflated: $390M China revenue at released-reserve cost basis and a $306M reserve release drove the 57% GM and part of the DC margin jump; none of it recurs beyond Q1's $100M.
- OpenAI concentration unresolved: still no second named gigawatt-scale customer; "a number of other customers" remains assertion; warrant economics still undisclosed.
- PC headwind is new: memory inflation shrinks the 2026 TAM, 2H guided subseasonal — the most reliable beat segment now faces macro pressure, and pull-in risk (Buchalter) wasn't fully addressed.
- Gaming is a 2026 drag: semi-custom down significant double-digit for the full year, with the Xbox recovery not until 2027.
- Supply confidence is total and unhedged: "I do not believe we will be supply limited" (Su) on a rack-scale ramp where the lead competitor stumbled — any 2H26 hiccup now reads as execution failure against an explicit denial.
- Q1 guide embeds a sequential decline (-5%) with C&G and Embedded down — 2026's growth is entirely back-half-weighted on the least-disclosed business.
Next-quarter watchlist- MI450 execution milestones: silicon/rack testing progress, first CSP deployment evidence for the OpenAI gigawatt, any second named at-scale MI450 customer, HPE/Lenovo Helios order signals, confirmation of Q3 revenue start.
- Q1 delivery: DC up sequentially (CPU and GPU both, incl. $100M China); server CPU up vs normal high-single-digit seasonal decline — the anti-seasonal claim is the quarter's sharpest test; GM 55% "clean"; opex ~$3.05B (fourth-miss watch).
- Opex leverage: whether 1H26 shows any deceleration toward the promised 2H26 leverage; Rasgon's challenge is now a standing question.
- GPU disclosure: tenth-quarter watch — FY25 Instinct sizing, MI350 ASP/margin commentary, any 2026 GPU revenue framework; reconciliation of the $306M vs $360M reserve-release discrepancy.
- China: MI325 license outcome, any Q2+ MI308 forecast resumption, demand re-assessment conclusion.
- Server CPU: third-party share data vs "record share" claim; Turin mix beyond 50%; evidence of the claimed capacity increase; pricing commentary (Muse's unanswered question); Venice launch timing specificity.
- Client: memory-inflation impact on TAM and any pull-in distortion in Q1; commercial sell-through trajectory vs +40%; 2H subseasonality modeling.
- Gaming: magnitude of the semi-custom decline through the year; Steam Machine launch traction; next-gen Xbox 2027 confirmation.
- Embedded: whether y/y growth sustains against seasonal Q1 decline; margin holding ~38%; design-win pace vs $17B.
- Capital allocation: buyback acceleration vs $9.4B authorization with $10.6B cash; Papermaster at Morgan Stanley TMT (March 3) for roadmap color.
- Carried gaps: GPU dollar levels, server-CPU revenue, MI350/MI450 ASPs, OpenAI warrant economics and dilution, MI450-ramp GM impact, 2026 GPU revenue sizing.
|
| Nov 4, 2025 | +2.51% | Q3 FY2025 | Read transcript briefingQuarter in one view- Revenue $9.2B, +36% y/y, +20% q/q — above the high end of the $8.7B ±$300M guide, with zero MI308 China revenue again. Non-GAAP GM 54% (+40 bps y/y), opex ~$2.8B (+42% y/y, vs ~$2.55B guided — a large overrun), op income $2.2B / 24% margin, EPS $1.20 vs $0.92. Record FCF $1.5B (CFO $1.8B).
- Segments: Data Center record $4.3B (+22% y/y, +34% q/q on the MI350 ramp; op income $1.1B / 25% vs 29% a year ago — R&D investment diluting). Client & Gaming record $4.0B (+73% y/y, +12% q/q; op income $867M / 21% vs 12%) — Client record $2.8B (+46% y/y, +10% q/q), Gaming $1.3B (+181% y/y, +16% q/q on holiday semi-custom + Radeon 9000). Embedded $857M (-8% y/y, +4% q/q; op margin 33% vs 40% — mix erosion persists).
- The quarter's defining event is the OpenAI deal: 6 gigawatts of Instinct GPUs, first gigawatt of MI450 in 2H26, "well over $100 billion in revenue over the next few years," plus a warrant structure Su calls "unique" with "extremely aligned incentives." Oracle named as MI450 lead launch partner (tens of thousands of GPUs from 2026). Su put a date on the superlative: "tens of billions in annual [AI] revenue in 2027."
- Q4 guide: ~$9.6B ±$300M (+25% y/y, +4% q/q), again excluding MI308; DC up double-digit q/q (server + MI350), C&G down (client up, gaming down strong double digits as pre-flagged), Embedded up double-digit; GM ~54.5%; opex ~$2.8B; tax 13%; ~1.65B shares; ~$37M net interest gain.
- MI308 update: some licenses received ("appreciative of the administration"), but still excluded from guide; demand environment still being assessed; WIP inventory position unchanged.
What management is focused on- OpenAI as the franchise-defining validation: Su framed it as multi-gigawatt, multiyear, multigeneration co-development (hardware, software, networking, roadmaps — "deep in conversations on MI500 and beyond"). She deflected concentration risk (Arcuri: OpenAI could be ~half of DC GPU revenue in 2027-28) by asserting the supply chain is being dimensioned for "multiple customers at similar scale."
- Server CPU demand inflection: Turin nearly half of EPYC revenue already; hyperscalers "planning substantially larger CPU build-outs over the coming quarters"; Su called the AI-driven CPU demand "durable... a multi-quarter phenomenon" (to Buchalter) and denied supply constraints ("we have supplies to support our growth" — Hu). Agentic-AI CPU thesis now has claimed customer forecast evidence behind it.
- Venice (2nm, 2026): silicon in labs "performing very well," "strongest customer pull we have seen," multiple cloud OEM partners already have first platforms online — unusually early ecosystem claims for a 2026 part.
- Helios/MI450 execution: Sanmina sale closed (last week), now "lead manufacturing partner for Helios"; Meta open rack wide standard support; OCP reception "phenomenal" with customers bringing engineering teams; early MI450 customers "really around rack-scale solutions" (to O'Malley).
- Supply chain as the gating question: O'Malley pressed on power/component constraints; Su's answer was ecosystem-planning language ("we feel very good that we have a strong supply chain") — confident but unquantified.
- Software: ROCm 7 launched (up to 4.6x inference, 3x training vs ROCm 6 — note the inference claim stepped up from ">3x" last quarter); OpenAI collaboration on Triton; AI-assisted ROCm kernel development teased for Analyst Day.
Key numbers and quarter mechanics- DC GPU: still no dollar figure. Rasgon tried twice — y/y growth ranking (Hu: "directionally they are similar, but server is a little bit better") and a direct "$6.5B for the year" anchor, which Hu explicitly rejected ("I don't think what you just mentioned was what we guided"). Notably, Rasgon attributed that framing to last quarter; management disowned it. Su then clarified "strong double-digit" applied to y/y, not sequential — the Q4 sequential guide is DC up double-digit with both server and GPU up, no relative sizing.
- Server CPU: all-time record revenue, Turin ~half of EPYC revenue, >160 new cloud instances (>1,350 total, +~50% y/y), cloud adoption by large businesses "more than tripled y/y," enterprise sell-through up sharply. Still no absolute dollar figure — eighth consecutive quarter.
- Client: record $2.8B, +46% y/y, +10% q/q (guided single-digit — beat); record desktop channel sell-in *and* sell-out; commercial sell-through +30% y/y. Q4: client revenue guided up sequentially even as C&G segment declines.
- Gaming: $1.3B +181% y/y, +16% q/q; Q4 down strong double digits exactly as pre-disclosed last quarter — the console cliff arrived on script.
- Embedded: $857M, -8% y/y (worse than Q2's -4%), +4% q/q; op margin 33% for a second straight quarter; Q4 guided double-digit sequential growth and "return to growth" y/y at the segment level. Design wins >$14B YTD, second straight record year claimed.
- Opex: ~$2.8B actual vs ~$2.55B guided — a ~$250M overrun, and Q4 guided flat at ~$2.8B. The investment-supercycle posture is now a step-function, not a trend.
- Cash: FCF record $1.5B; cash + STI $7.2B (rebuilt from $5.9B); buybacks only $89M in Q3 ($1.3B YTD) despite $9.4B remaining authorization — repurchase pace is notably timid.
- ZT manufacturing now in discontinued operations; Sanmina close completed.
Product and launch scorecard- MI350/MI355: delivered the committed y/y growth (Su: "ramped really nicely... proceeded well"). Oracle first hyperscaler with public MI355X instances; neoclouds Crusoe, DigitalOcean, TensorWave, Vultr ramping; MI300X broadened (IBM/Zyphra training cluster, Cohere on OCI, Character.AI and Luma AI inference). Su guided continued MI355 ramp in 1H26 — explicitly no air pocket before MI450 (to Arya).
- MI450/Helios: transformed from roadmap to contracted business — OpenAI 6GW (first GW 2H26), Oracle tens of thousands of MI450s from 2026 "expanding through 2027 and beyond," Meta rack standard, DOE Discovery (MI430X + Venice) and Lux AI (early 2026, with OCI/HPE). Rack-scale is the lead form factor. This is the first quarter with named, sized, dated GPU commitments.
- OpenAI warrant structure: acknowledged by Su as "unique," vesting tied to price milestones; she left the door open to "conceptually similar creative" structures with others (to Seymore) — a new, dilutive-adjacent deal-making tool now on the table.
- ROCm 7: launched with quantified gains (4.6x inference / 3x training vs ROCm 6), Hugging Face/vLLM/SGLang contributing directly; Su claims new-customer onboarding is now "very smooth" — a deliberate contrast to the historical friction narrative.
- EPYC: Turin ramp "extremely fast," Genoa demand persisting (Rakers extracted the ASP-vs-unit answer: Turin lifts ASP, Genoa holds volume); Venice pull described as strongest ever.
- Client: Ryzen 9000 desktop records; commercial +30% sell-through; no Intel 18A mention anywhere on the call.
- Gaming: FSR 4 supported games doubled to 85+; Radeon 9000 sell-out growth; semi-custom holiday build delivered the pre-announced peak.
- Embedded: Versal Prime Gen 2 shipping, first Versal RF dev platforms, Ryzen Embedded 9000 — product cadence fine; revenue inflection still pending.
Sell-side read-through- Rasgon (Bernstein): the most consequential exchange of the call. Forced the first directional CPU-vs-GPU growth ranking (server slightly better y/y) and got his own "$6.5B FY GPU" framing publicly disowned by Hu — the informal street anchor from last quarter is now contested, not just unconfirmed. Analysts must re-derive GPU sizing from scratch.
- Arya (BofA): extracted the 1H26 continuity commitment (MI355 keeps ramping, no digestion pause) and pressed OpenAI visibility/allocation — Su's answer was planning-depth language ("multiple quarters out," power and supply chain secured), not contractual specifics.
- Arcuri (UBS): pivoted from GPU dollars to OpenAI halo and concentration; got Su to accept his "~half of DC GPU revenue in 2027-28" framing implicitly (she answered the risk question without disputing the sizing) and the admission that customer interest "accelerated... at higher scale" post-announcement.
- O'Malley (Barclays): rack-vs-discrete crossover answer (early MI450 = rack-scale-led) and the constraint question — Su would not rank power vs components, only "all of this is going to be tight."
- Buchalter (TD Cowen): CPU durability ("multi-quarter phenomenon," not seasonal reset) plus a flat supply-constraint denial; ROCm competitive gap answer was honest-but-vague ("always more work to do").
- Muse (Cantor): 2026 GPU GM framework refused beyond "transition period then normalize" — Hu reiterated the priority is revenue and GM dollars over GM percentage, a margin-quality warning embedded in guidance language.
- Moore (Morgan Stanley): MI308 — licenses received, demand still being sized, WIP intact; no swing-factor quantification offered.
- Seymore (Deutsche Bank): differentiation answer was TCO/time-to-market boilerplate; the warrant question got the more interesting answer — Su defended the structure and signaled openness to similar constructs.
- Repeated themes: GPU dollar disclosure (Rasgon, Arcuri — still zero), 2026 margin shape (Muse, Arya), supply/power constraints (O'Malley, Buchalter, Arya — three angles), OpenAI concentration (Arcuri, Seymore, Arya).
Management credibility- Delivery vs guide: revenue beat the high end; GM 54% as guided; DC up strong double-digit q/q as guided; client beat the single-digit guide (+10%); gaming up as guided; Embedded returned to sequential growth as promised (third attempt, delivered); the Q4 console cliff pre-call from last quarter is now embedded in guidance. The Q3 GPU y/y growth commitment (made ex-China against a China-containing base) was met.
- Major credibility upgrade: "tens of billions" got a date (2027) after Moore was refused one last quarter — and it arrived with named anchor contracts (OpenAI, Oracle) rather than aspiration. The 8-9-month lead-time disclosure from last quarter is now visibly consistent with 2H26 first-gigawatt timing.
- Slippage: opex missed badly (~$2.8B vs ~$2.55B guided, second consecutive overrun) and the Q4 guide ratifies the higher level — the "disciplined execution" claim (Su's words) does not apply to cost. Buyback pace ($89M) is trivial against $9.4B authorization despite record FCF.
- Evasions: GPU dollar levels refused for the eighth-plus quarter, now with an active disavowal of a street anchor; server CPU dollars refused (eighth quarter); Rasgon's sequential server-vs-GPU ranking for Q4 was deflected into a y/y clarification; OpenAI revenue phasing beyond "well over $100B over the next few years" deferred to Analyst Day.
- New disclosure risk: the warrant structure is acknowledged but its terms were not discussed on the call (vesting "according to a price that would be very accretive" per Seymore's framing, unchallenged) — dilution mechanics are a 10-Q/Analyst Day item.
- Under pressure: Su was direct on 1H26 continuity, CPU durability, and rack-scale-first MI450 mix; Hu was disciplined in refusing 2026 GM guidance and the $6.5B anchor. The pattern of refusing numbers while accepting directional framings continues.
What changed versus the prior quarter- The AI story went from ramp-evidence to contracted scale: last quarter's best proof was Oracle 27K nodes and "7 of top 10"; this quarter added OpenAI 6GW / >$100B potential, Oracle MI450 launch partnership, two DOE systems, and a dated 2027 "tens of billions" claim. The burden of proof shifted from "is MI350 real" to "can they execute 2H26."
- GPU disclosure got worse, not better, amid bigger numbers: the informal ~$1B+/quarter floor from last quarter was neither confirmed nor updated, and Rasgon's $6.5B FY framing was actively disowned — the street is flying blind on the current GPU base while underwriting a massive 2026-27 ramp.
- Server CPU narrative graduated from "agentic AI multiplies CPU work" (theory) to "hyperscalers forecasting significant CPU builds into 2026" (claimed customer visibility), with Turin at ~half of EPYC revenue as the mix proof.
- China moved from "under review" to "some licenses received" — but the guide still excludes MI308, demand is being re-assessed, and WIP conversion timing is unchanged. Optionality preserved, urgency removed.
- Opex discipline broke: two consecutive guide misses, now structurally ~$2.8B/quarter (+42% y/y) — the investment supercycle is absorbing the gross-margin gains.
- Embedded's recovery delivered sequentially but the y/y decline deepened (-8% vs -4%) and margin stayed at 33% — the Q4 double-digit sequential guide is the next test.
- Capital return went quiet: record FCF but minimal buyback — a shift from last quarter's aggressive authorization posture.
- The warrant deal introduces a new variable: equity-linked customer incentives are now part of AMD's playbook, with Su explicitly open to repeats.
Bull case- The AI business is now underwritten by contracts, not claims: OpenAI 6GW with first silicon 2H26, Oracle as MI450 lead partner, DOE Lux/Discovery, UAE G42/Cisco — plus Su's assertion of "numerous other customers" at potentially similar scale and a supply chain dimensioned for breadth. "Tens of billions in 2027" is now a dated, checkable commitment with named anchors.
- No air pocket: MI355 guided to keep ramping through 1H26, then MI450 accelerates 2H26 — the transition-risk question (Arya) got an unambiguous continuity answer.
- Server CPU is a second, independent growth engine: record revenue, Turin ~half of EPYC, hyperscaler CPU build-outs forecast into 2026, Venice pull "strongest ever" — AI is driving CPU demand, and AMD holds the share-gain position with no supply constraint admitted.
- Client keeps compounding: record $2.8B with commercial sell-through +30% and Q4 client guided up even in a down segment quarter — share gains with mix, not just ASP.
- Margin structure intact through the ramp: 54% GM delivered with GPU mix up, 54.5% guided for Q4; DC op margin 25% despite heavy R&D; operating leverage visible in EPS +30% y/y.
- Balance sheet: $7.2B cash, record $1.5B FCF, Sanmina cash in, $9.4B buyback capacity — the 2026 ramp is self-funded.
Bear case- The GPU base remains undisclosed while 2027 commitments multiply: management disowned the street's $6.5B FY anchor, refused every sizing attempt, and gave no 2026 GPU GM framework beyond "transition then normalize" — investors are being asked to underwrite a 6GW ramp with no current-quarter GPU P&L visibility.
- OpenAI concentration and structure risk: Arcuri's "~half of DC GPU revenue in 2027-28" went undisputed; the warrant structure ties customer economics to AMD's share price in ways not yet fully disclosed; and OpenAI's ability to fund simultaneous multi-vendor commitments (Arya's question) was answered with planning language, not contracted-revenue mechanics.
- Opex is outrunning the model: +42% y/y, two straight guide misses, $2.8B now the run-rate — if the MI450 ramp slips, the cost base doesn't.
- DC op margin fell y/y (25% vs 29%) despite record revenue — R&D intensity is structurally diluting the segment's profitability even at scale.
- Embedded still hasn't inflected: -8% y/y, 33% margin for two quarters; the Q4 double-digit sequential guide is aggressive against a two-year slippage history.
- Gaming cliff is now: Q4 C&G declines with gaming down strong double digits — Q4 growth rests entirely on DC, the least-disclosed segment.
- China is drifting: licenses received but demand "still being assessed" — the post-license revenue opportunity may be smaller or slower than the pre-ban ~$700M/quarter run-rate implied.
- Supply chain confidence is asserted, not evidenced: "all of this is going to be tight" (Su) on power, memory, packaging — execution risk into 2H26 is real and unquantified.
Next-quarter watchlist- Financial Analyst Day (Nov 11, one week out): long-range targets, updated AI TAM (Su teased "larger than $500B"), OpenAI revenue phasing, warrant terms/dilution, 2026 GPU margin framework, and whether "tens of billions 2027" gets a bridge.
- GPU dollar disclosure: ninth-quarter watch. Whether Q4 results or Analyst Day finally size the Instinct base; reconciliation of the disowned $6.5B framing; MI350 ASP/margin commentary.
- OpenAI mechanics: warrant vesting schedule and share count impact (10-Q), CSP deployment partners for the first gigawatt, power/site evidence, any additional gigawatt-scale customers announced to validate the "multiple customers at similar scale" claim.
- Q4 delivery: DC up double-digit q/q; client up sequentially; gaming down strong double digits vs guide; Embedded double-digit sequential growth and y/y return to growth; GM 54.5%; opex held at ~$2.8B (third-miss watch).
- Server CPU: third-party share data vs the claimed gains; Turin mix progression past ~50%; whether hyperscaler "CPU build-out" commentary shows up in competitor/capex disclosures; Venice launch timing specificity.
- MI308: license scope expansion, demand assessment conclusion, any Q1 2026 inclusion in guidance, WIP conversion timing.
- Helios/MI450: first silicon milestones, additional named rack-scale commitments, Sanmina manufacturing ramp evidence, Meta rack-standard ecosystem adoption.
- Embedded: margin recovery from 33%; design-win total vs the $14B pace; whether Q4's guided growth finally sticks.
- Capital allocation: buyback acceleration vs the $9.4B authorization; cash deployment priorities stated at Analyst Day.
- Carried gaps: GPU dollar levels, server-CPU revenue (eight quarters), MI350/MI450 ASPs, OpenAI warrant economics, 2026 GPU GM trajectory, MI308 revenue sizing.
|
| Aug 5, 2025 | -6.42% | Q2 FY2025 | Read transcript briefingQuarter in one view- Revenue $7.7B, +32% y/y, +3% q/q — came in at the top of the ~$7.4B ±$300M guide (management says "exceeded the midpoint"; the print sits at the high end of the band). Reported GM 43% (vs 53% a year ago), ~54% ex the $800M MI308 inventory charge — the charge landed exactly as pre-quantified. Opex ~$2.4B (+32% y/y, above the ~$2.3B guide). Op income $897M / 12% margin (vs 22% a year ago; the charge is the swing). EPS $0.48 vs $0.69; Hu quantified the charge impact at ~$0.43, implying ~$0.91 ex-charge (arithmetic on disclosed figures, not a management-stated number).
- Segments: Data Center $3.2B (+14% y/y, -12% q/q; operating *loss* of $155M vs +$743M/26% a year ago — the $800M charge concentrated here; ex-charge implies roughly ~$645M / ~20% margin, inference only). Client & Gaming $3.6B (+69% y/y, +20% q/q vs the guided "double-digit"; op income $767M / 21% vs 8% a year ago) — Client a record $2.5B (+67% y/y), Gaming $1.1B (+73% y/y on semi-custom holiday builds + Radeon 9000). Embedded $824M (-4% y/y, flat q/q; op income $275M / 33% vs 40% — mix-driven erosion).
- The quarter's defining shift is regulatory tone, not operations: Commerce is "moving forward with the review" of MI308 license applications; AMD excludes all MI308 revenue from the Q3 guide anyway and commits to Instinct y/y growth in Q3 purely on MI350. Inventory is mostly WIP — a couple of quarters to convert even once licensed.
- Q3 guide: revenue ~$8.7B ±$300M (+28% y/y, +13% q/q) excluding any MI308; DC up strong double-digit sequentially on MI350, Client up single-digit (per Su to Seymore), Gaming flattish, Embedded returning to growth; GM ~54%; opex ~$2.55B; tax 13%; ~1.63B shares; net interest flips to a ~$10M gain.
- Cash mechanics: record FCF $1.2B (CFO $1.5B); cash + STI $5.9B (down from $7.3B); repaid the $950M commercial paper from the ZT funding; long-term debt $3.2B; new $6B buyback authorization, $9.5B total remaining.
What management is focused on- Reframing China from damage to optionality: Su thanked the Trump administration by name, emphasized licenses "under review," and stressed the Q3 guide carries zero MI308 revenue — any approval is upside, albeit slow ("couple of quarters" to run WIP through). O'Malley, Rakers, and Reitzes all probed mechanics; Hu killed any charge-reversal hope (majority WIP, no finished goods to ship).
- MI350 as the proof point, now early: volume production began in June "ahead of schedule," "steep" 2H ramp, adoption "a bit faster than we might have expected" (Su, to Seymore). Competitive claims escalated — MI355 "matches or exceeds B200," comparable to GB200 "at significantly lower cost and complexity," 40% more tokens/dollar in inference.
- Rack-scale / MI400 / Helios: far more specific than last quarter — 40 PFLOPS FP4, 50% more memory/bandwidth/scale-out vs competition, 72 GPUs per rack as one accelerator, "10x generational" for frontier models, "highest-performance AI system in the world when it launches," and first revenue framing: "significant revenue contribution from Helios in 2026" (hyperscaler/neocloud-led, sovereign later, per Su to Muse).
- Sovereign AI as a new disclosed category: HUMAIN (Saudi) "multibillion-dollar" all-AMD collaboration with quarterly expansions; ">40 active engagements globally"; Su calls sovereign "additive," not substitutive (to Arya).
- Agentic-AI CPU narrative: each GPU token triggers multiple CPU-intensive tasks — management's answer to GPU-attached server demand, plus records in cloud and enterprise CPU, 33rd straight y/y share-gain quarter, >100 new cloud instances (vs >30 last quarter), ~1,200 total.
- Client quality defense: Su again denied pull-forwards (O'Malley) but voluntarily plans "a little bit less than seasonal" 2H; ASP/mix still doing the work; commercial (Dell ramp, +25% OEM consumption) is the forward story. She did not touch the Intel 18A portion of O'Malley's question.
- Capital/inorganic machine: Sanmina buying ZT US manufacturing for $3B cash + stock (incl. contingent), close near end-2025; Brium and Lamini added "last quarter" atop Nod.ai/Mipsology/Silo AI; opex guided up again to $2.55B — an explicit investment-supercycle posture.
Key numbers and quarter mechanics- DC GPU: still no dollar figure — Arcuri ("a little more than $1 billion?" "can you get to $7 billion for the year?") and Reitzes ("$2 billion run rate?") both failed; Su answered with commitments instead: Instinct y/y growth in Q3 (ex-MI308) and growth into Q4. Hu conceded the majority of the ~$1.0B Q3 sequential revenue increase is MI355-driven (to Reitzes) — the closest thing to sizing all call.
- GPU y/y mechanics: Q2 GPU declined y/y as Su pre-admitted last quarter; Q3-25 y/y growth is being committed against a base period that still contained China AI revenue — a stricter claim than the headline suggests, and checkable in one quarter.
- GPU GM: first quantification of any kind — Hu: "a little bit below corporate average," and she explicitly decoupled improvement from any revenue threshold (operational efficiency-driven). Q3 GM guided ~54%, flat vs Q2 ex-charge; Buchalter's challenge forced the mix bridge: Gaming-heavy mix is dilutive, offset by server margin, commercial PC, and ops efficiency.
- Server CPU: records in both cloud and enterprise; Hu claimed share up vs Q1 (no third-party report yet). Still no absolute revenue or growth figure — seventh consecutive quarter. With DC +14% y/y and GPU down y/y, the segment growth is arithmetically EPYC-driven (inference from two management statements).
- Client: record $2.5B, +67% y/y; H1 +68% y/y; record desktop channel; mobile sellout "large double-digit" y/y; ASP/mix remains the engine. Q3 guided single-digit sequential growth; 2H planned below seasonal — Muse's extraction from last quarter is now Su's own words.
- Gaming: $1.1B +73% y/y, roughly doubled sequentially against holiday semi-custom builds (segment +20% q/q with client ~+9%). Su pre-disclosed the Q4 shape: consoles "down substantially — strong double digits" q/q, C&G segment likely down in Q4.
- Embedded: flat q/q at $824M as guided; op margin 33% (-7 pts y/y) on mix; Q3 return to sequential growth guided; design wins "tracking ahead" of last year's record $14B pace.
- Charge mechanics closed out: $800M in DC, ~$0.43 EPS; Hu confirms majority WIP → no near-term reversal, and Su says minimal Q3 contribution even if licenses land ("we're sitting already in early August").
- Opex: ~$2.4B actual vs ~$2.3B guided — a small overrun; Q3 steps to ~$2.55B (+32% y/y trend). Operating leverage remains paused: ex-charge op margin still well below year-ago 22% (inference).
Product and launch scorecard- MI350 (MI355/MI355X): ahead of plan — June volume production vs "mid-year" commitment; Oracle now sized at 27,000+ nodes (MI355X + Turin + Pollara 400 NIC — the in-house NIC's flagship deployment); "7 of the top 10" model builders/AI companies on Instinct; developer cloud launched (Su: not meaningful to 2H revenue — an expectations-management answer to Arcuri). First actual revenue evidence (Q3 y/y growth) arrives next quarter.
- MI400/Helios (2026): specification-level disclosure for the first time (40 PFLOPS FP4, 72-GPU rack, 10x generational claim); "significant revenue contribution in 2026" is a new, checkable commitment; lead times 8-9 months disclosed (Buchalter) — implies orders/co-engineering being worked now; ZT team credited with customer data-center compatibility work. Still no named Helios commitments beyond Oracle.
- ROCm 7: >3x inferencing/training performance vs prior gen, large-scale training and distributed inference support; nightly builds, first developer cloud, native vLLM/SGLang, day-0 Llama 4 / Gemma 3 / DeepSeek-R1; ROCm Enterprise AI full-stack launched — software cadence is now a systematic proof point, not a promise.
- EPYC: Turin ramping "significantly"; >100 new cloud instances in the quarter; Google/Oracle Turin instances ~2x prior-gen; 28 new Turin enterprise platforms; EPYC 4005 SMB launch; KDDI 5G and Nokia wins; >1/3 of fastest supercomputers; El Capitan/Frontier #1/#2 Top500.
- Client: Threadripper (up to 96 cores, ~2x competition in content creation) launched; Ryzen AI 300 mix up; Dell commercial ramp started in Q2 with more in 2H.
- Gaming: Microsoft multiyear next-gen Xbox (consoles, PCs, handhelds) — the strategic win of the quarter, extending semi-custom into the next console cycle; Sony Project Amethyst co-engineering; Radeon 9600 XT launched; Radeon AI Pro R9700 (local inference/fine-tuning, full ROCm).
- Embedded: Spartan UltraScale+ first production shipments; Bosch robotaxi Versal win — adoption evidence but not yet revenue inflection.
Sell-side read-through- Arcuri (UBS): third-plus consecutive failure on GPU dollars, but he anchored the debate — "a little more than $1 billion" Q2 GPU and "$7 billion for the year" went uncorrected by Su; analysts will treat the un-rebutted $1B+ as a soft floor. His "$7B FY" ask was deflected into qualitative ramp language.
- Reitzes (Melius): extracted the call's only GPU sizing admission — Hu accepted "the majority of the [Q3 sequential] increase is really driven by MI355's strong ramp." His inference ($1B+ GPU doubling into Q3) is analyst math, not management confirmation.
- Arya (BofA): got the first GPU margin disclosure in six-plus quarters of asking (below corporate average, ops-driven improvement) — partial closure of a carried gap. Sovereign sizing for 2026 refused; "additive" and "starts with MI355" are the usable takeaways.
- O'Malley (Barclays): pull-in question returned; Su's posture softened from denial-plus-sub-seasonal-planning to denial-plus-"a little bit less than seasonal" — directionally validating the concern. His Intel 18A follow-up was simply not answered — the clearest evasion of the call.
- Seymore (Deutsche Bank): "better or worse than 90 days ago" framing produced "a bit faster than expected" on MI350 adoption; he also extracted the most concrete new forward data — Q3 client single-digit, gaming flattish, Q4 consoles down strong double digits.
- Buchalter (TD Cowen): lead-time quantification (8-9 months) plus the GM mix-bridge press — Hu's answer (gaming dilutive in Q3, server/commercial/ops offsets) is the first real decomposition of how 54% holds through the GPU ramp.
- Moore (Morgan Stanley): failed to time-bound "tens of billions" — Su explicitly declined his 2027/$20B+ anchor ("I won't necessarily speculate on the exact time"). Disciplined non-answer; the superlative entered the script anyway.
- Rakers (Wells Fargo): confirmed server CPU growing double-digit sequentially alongside GPU; settled the charge-reversal question (WIP, no finished goods, no quick reversal).
- Muse (Cantor): got the 350→400 handoff architecture: 2H25 = MI355 ramp, Helios significant in 2026, hyperscalers/neoclouds first, sovereigns later. His use-of-proceeds question got boilerplate back.
- Repeated themes worth tracking: GPU dollar disclosure (Arcuri, Reitzes — two attempts), China phasing mechanics (O'Malley, Rakers, Reitzes — three separate probes), pull-ins (O'Malley again), GPU GM (Arya, Buchalter).
Management credibility- Delivery vs guide was near-complete: revenue at the high end; reported GM 43% and ex-charge ~54% both as pre-quantified; the $800M charge and ~$0.43 EPS impact matched; C&G +20% q/q vs "double-digit" guide; Embedded flat as guided; DC down as guided; MI350 mid-year committed → June production (ahead); "strategic partner shortly" for ZT manufacturing → Sanmina in May at $3B; Q2 GPU y/y decline conceded last quarter → confirmed ("revenue declined year-over-year").
- Specificity improved where it was pressured: GPU GM vs corporate average, WIP composition of the charge, 8-9-month lead times, Q4 console cliff, Helios 2026 revenue significance — each closed a carried gap, partially.
- Specificity withheld where it matters most: GPU dollar levels (two analysts, no number), FY GPU sizing ($7B ask declined), server-CPU revenue (seventh quarter), ZT purchase price (still undisclosed even as the manufacturing exit price is public), MI350 ASPs, and the "tens of billions" timeframe.
- Consistency flags: Su denied pull-forwards while adopting sub-seasonal 2H planning — the position is coherent but has migrated in her direction over two quarters; opex guided ~$2.3B and printed ~$2.4B — investment framing covers it, but the guide missed; "highest-performance AI system in the world" and "tens of billions" are new superlatives introduced without supporting disclosure — raises the bar for 2026 delivery.
- Under pressure: Su gave Seymore a direct better-than-expected admission and boxed herself into two checkable claims (Q3 GPU y/y growth ex-China; Q4 GPU growth); Hu declined to promise charge reversals or margin scale thresholds — both credibility-preserving rather than promotional.
- Unresolved gaps to carry: GPU dollar levels, server-CPU quantification (now seven quarters), ZT purchase price, MI350 ASPs, Helios named customers, "tens of billions" timing.
What changed versus the prior quarter- China flipped from bounded damage ($700M Q2 / $1.5B FY / $800M charge) to pending restoration: license review progressing, guide conservatively excludes MI308, and the inventory is WIP — the recovery is now a timing question, and Su made clear it will not rescue Q3 even if approved.
- MI350 moved from schedule risk to early delivery: June production, an adoption claim of "faster than expected," a sized anchor (Oracle 27K+ nodes), and a hard y/y growth commitment for Q3 — the air-pocket scenario from two quarters ago did not fully close, but the burden of proof tightened.
- The GPU base question intensified: last quarter management admitted a y/y decline without a base; this quarter two analysts created an informal "a little more than $1B" floor that management chose not to correct — the opacity now interacts with street anchors rather than a vacuum.
- MI400/Helios went from road-map bullet to specified system with a revenue claim (significant in 2026) and a disclosed supply lead time — materially more underwritable, still unnamed beyond Oracle.
- Server CPU strengthened its evidence (records, 33rd share quarter, share up q/q claimed) while the disclosure drought hit a seventh quarter; the agentic-AI CPU demand thesis is new narrative scaffolding for the same unquantified line.
- Client's migration completed: last quarter ASP/mix did all the work while management planned sub-seasonal; this quarter Su said it out loud (Q3 single-digit, 2H "less than seasonal") while Intel 18A entered the question set unanswered.
- Gaming delivered the inflection *and* disclosed its own cliff: +73% y/y with the Microsoft next-gen Xbox win, while Q4 console revenue is pre-guided down strong double digits — the 2025 DC-only dependency in Q4 is now explicit.
- Embedded's recovery slipped in quality even as timing held: revenue flat as promised, but op margin fell to 33% from 40% — the 2H25 return-to-growth call now carries a margin-recovery question too.
- Financial posture shifted from funding to repayment: the $950M ZT commercial paper is repaid, FCF hit a record $1.2B, and a $6B authorization was added — balance-sheet concern from last quarter is largely retired; opex discipline is the new open item.
Bull case- A clean beat-and-raise with China as free optionality: $7.7B at the top of the band with zero MI308, and a $8.7B / +13% q/q guide that still excludes it — if licenses land, 2H has an unguided tailwind; if they don't, the commitments stand on MI350 alone.
- MI350 de-risked on schedule and claims: early June production, "faster than expected" adoption, a sized hyperscale anchor (Oracle 27K+ nodes including CPU + NIC content), 7-of-10 top AI companies, and an explicit y/y Q3 growth commit — the 2H ramp thesis now has named, dated, countable evidence.
- The competitive framing upgraded: MI355 vs B200/GB200 claims plus Helios' specified 72-GPU rack and 2026 revenue significance — AMD is now selling rack-scale against Nvidia's roadmap rather than conceding the category, with ZT/Sanmina resolving the manufacturing question.
- Server CPU is the quiet compounding engine: dual records, 33 share quarters, >100 new instances in one quarter, and a structural argument (agentic AI multiplies CPU work per GPU token) that ties server growth to GPU deployment, not just to share.
- Client is a margin story now: 21% C&G op margin vs 8% a year ago, commercial OEM consumption +25%, Dell only just ramping — share gains with mix, and management is planning conservatively (below seasonal), which sets up upside if pull-ins are as absent as Su claims.
- Gaming's strategic value jumped: Microsoft next-gen Xbox multiyear extends semi-custom into the next cycle — the revenue stream just got a longer runway beyond the Q4 seasonal cliff.
- Cash engine intact: record $1.2B FCF in a charge-impaired quarter, CP repaid, $9.5B repurchase capacity, $3B inbound from Sanmina — the investment supercycle is self-funded.
Bear case- The reported P&L genuinely deteriorated: GM 43%, op margin 12%, EPS down y/y, and Data Center — the flagship segment — posted a $155M operating loss (ex-charge ~20% margin, down from 26% a year ago by inference), showing how much of the segment's profitability had been GPU-China-supported.
- The GPU ramp is still unverifiable: no dollar base (two failed attempts), no FY sizing ($7B declined), GPU GM admitted below corporate with no scale threshold, and a Q3 y/y commitment whose base period contained China revenue — skeptics will note every proof point arrives next quarter, not this one.
- Margin mechanics are fragile at exactly the wrong time: Q3 GM guided flat at ~54% despite a dilutive GPU ramp only because server/commercial/ops offsets are assumed; opex steps to $2.55B with a +32% y/y trend — any MI350 slip hits revenue and leverage simultaneously.
- Client quality-of-growth questions compounded: ASP/mix-driven records, explicit below-seasonal 2H planning, a dodged Intel 18A question, and the desktop-channel/X3D strength that may include channel restocking AMD refuses to acknowledge.
- Q4 is now explicitly a one-segment story: consoles pre-guided down strong double digits, C&G segment down — the entire Q4 growth case rests on GPU, the one line management won't quantify.
- Embedded keeps deferring: revenue flat as promised but margin down 7 points y/y on mix; the 2H recovery call has survived multiple quarters of slippage and now carries a profitability question.
- China upside is slower than the headline: WIP-heavy inventory means a couple of quarters post-license to convert; the $800M charge is structurally unlikely to reverse; and Su conceded — unsolicited — that approval would barely touch Q3.
- Aspirational language outran disclosure: "tens of billions," "highest-performance AI system in the world," 10x generational Helios claims — introduced with a refusal to be timed or named, which concentrates credibility risk into 2026.
Next-quarter watchlist- The single checkable claim: Q3 Instinct y/y growth ex-MI308 (Su's commitment to Seymore/Arcuri) and GPU growth into Q4 — vs the uncorrected Arcuri/Reitzes "~$1B+ base / ~doubling" framing. Any dollar disclosure at quarter's end matters; the segment-only policy has now survived two soft GPU quarters.
- MI350 ramp evidence: additional named customers beyond Oracle, first ASP/margin commentary, whether "below corporate average" GPU GM gets a number, and whether Hu's 54% flat-GM bridge holds with GPU mix up (gaming dilutive, server/commercial/ops offsets).
- China mechanics: license approvals; Su's minimal-Q3 framing tested; whether the ~$700M/quarter run-rate assumption survives post-license commentary; any change to WIP-conversion timing ("couple of quarters").
- Helios/MI400: first named Helios commitment beyond Oracle; co-engineering milestones from the 8-9-month lead-time window now disclosed; whether "significant 2026 revenue" gets any sizing; hyperscaler vs neocloud vs sovereign mix development.
- Server CPU: eighth-quarter watch for a revenue number; verification of Hu's "share up vs Q1" when third-party data publishes; Turin ramp pace; whether the agentic-AI CPU narrative shows up in cloud capex commentary.
- Client: Q3 single-digit sequential growth vs below-seasonal 2H framing; ASP/mix durability; Dell commercial ramp progress; any response to Intel 18A now that it was publicly ignored.
- Gaming: Radeon 9000 supply catch-up ("demand outpaced supply"); Microsoft/Sony program milestones; Q4 console decline vs the pre-disclosed strong-double-digit guide.
- Embedded: Q3 sequential growth delivery (third attempt at a return-to-growth call); op margin recovery from 33%; design-win pacing vs the $14B record.
- ZT/Sanmina: closing near end-2025; cash/stock mix and contingent-payment terms; ZT purchase price (still undisclosed — 10-Q watch); first observable Helios design work attributable to the ZT team.
- Financial mechanics: opex ~$2.55B delivery vs the +32% investment trend; GM 54% ex-charge trajectory into Q4; FCF durability from the record $1.2B; buyback acceleration against $9.5B capacity; cash rebuild from $5.9B.
- Carried gaps: GPU dollar levels, server-CPU quantification (seven quarters), ZT purchase price, MI350 ASPs, "tens of billions" timeframe, charge-reversal treatment if licenses arrive.
|
| May 6, 2025 | +1.76% | Q1 FY2025 | Read transcript briefingQuarter in one view- Revenue $7.4B, +36% y/y, -3% q/q — above the high end of the ~$7.1B ±$300M guide; GM 54% (+140 bps y/y, fifth straight y/y expansion); opex $2.2B (+28% y/y, above the ~$2.1B guide); op income $1.8B (24% margin, down from 26% in Q4); EPS $0.96, +55% y/y.
- Segments (new three-segment structure; four revenue lines still disclosed): Data Center $3.7B (+57% y/y, -5% q/q; op income $932M, 25% vs 23% a year ago, 30% in Q4); Client & Gaming $2.9B (+28% y/y, +2% q/q; op income $496M, 17% vs 10% a year ago) — Client $2.3B (+68% y/y), Gaming -30% y/y; Embedded $823M (-3% y/y, -11% q/q; op income $328M, 40% vs 41%).
- The quarter's defining event is regulatory, not operational: April export license requirement on MI308 China shipments — $700M Q2 revenue headwind, ~$1.5B FY25 impact, and an ~$800M inventory/reserve charge that drops Q2 guided GM to 43% (54% ex-charge).
- Q2 guide: revenue ~$7.4B ±$300M (+27% y/y at midpoint despite the $700M removal); Client & Gaming up double-digit q/q, Embedded flattish, DC down on MI308 exclusion; opex ~$2.3B (incl. ~$50M ZT); tax 13%; ~1.64B shares (incl. 9M ZT shares).
- ZT Systems closed March 31 — funded with $1.5B new debt plus $950M commercial paper; cash + STI $7.3B; CFO $939M, FCF $727M (down from Q4's record $1.1B); buyback tripled to $749M; $4B authorization remaining. ZT manufacturing goes to discontinued operations from Q2; strategic partner announcement "shortly."
- FY25 framing: "strong double-digit-percentage revenue growth" reiterated; DC GPU guided to strong double-digit y/y growth for the year, second-half weighted on the MI350 ramp.
What management is focused on- Containing the China export-control narrative: Su opened with "evolving dynamics related to tariffs and the regulatory environment" and closed by arguing headwinds are "more than offset" by product tailwinds. Hu quantified everything ($700M Q2, $1.5B FY, $800M charge) — a deliberate effort to bound the damage in one call. Notably, this is the first China AI SKU disclosure after six quarters of silence.
- MI350 as the 2H25 hinge: sampling with multiple customers confirmed in Q1 (as promised last quarter), mid-year accelerated production reiterated, 35x vs MI300X claim repeated, and the first named large-scale commitment — Oracle deploying a "multi-billion-dollar" MI355X + Turin + Polara 400 AI NIC cluster. June 12 Advancing AI event set as the next proof point.
- Rack-scale and ZT integration: ZT framed as co-designing MI400 rack-level solutions with key customers and accelerating MI350 time-to-market; Su tied the acquisition directly to the MI400 rack-scale challenge raised by Sur.
- Server CPU momentum: >30 new cloud instances (Alibaba, AWS, Google, Oracle, Tencent, incl. first Turin instances); Forbes 2000 enterprise EPYC activations more than doubled y/y; on-prem EPYC up "large double-digit" for the seventh straight quarter; >150 Turin platforms coming from Dell, Cisco, HPE, Lenovo, Supermicro; TSMC Arizona EPYC production 2H25; Venice (N2) silicon in labs for 2026.
- Client mix story: record client ASP, more than half of client growth from ASP/mix rather than units (units actually declined double-digit q/q per Hu); desktop channel sellout +50% y/y; AI PC sales +50% q/q; Ryzen Pro sell-through +30% y/y with an 80% increase in commercial systems from HP, Lenovo, Dell, Asus.
- Inference positioning: a "frontier model developer" now serving "a significant portion" of daily inference traffic on Instinct; distributed inferencing clusters; day-zero support for Llama 4, Gemma 3, DeepSeek-R1; >2M Hugging Face models out-of-the-box (doubled from >1M last quarter).
Key numbers and quarter mechanics- DC GPU: grew "significant double-digit" y/y; down "very modestly" q/q (Su, to Arya) — consistent with last quarter's flattish-1H framing. MI300/MI325X were the majority of Q1 GPU revenue (Hu). No dollar figure disclosed — the segment-level-only policy held through direct attempts (Arcuri, Arya, Rasgon) to extract one.
- GPU y/y trajectory explicitly conceded: Q2 GPU will not grow y/y (Su, to Rasgon) given the $700M MI308 removal; y/y growth resumes Q3/Q4; full-year strong double-digit growth requires the 2H MI350 ramp. The $1.5B FY China impact lands mostly in Q2/Q3, "very little in Q4" — Q4 was always planned as non-China MI350 revenue.
- Server CPU: share gains again, but still no absolute server-CPU revenue or growth number — sixth straight quarter. DC segment -5% q/q vs the guided "down ~7%"; with GPU down modestly, server CPU was down somewhat more than 5% sequentially (inference from the two statements; management did not quantify).
- Client: $2.3B +68% y/y, flat-ish q/q (revenue roughly flat vs Q4's $2.3B while units fell double digits — ASP/mix did all the work). Management again denied tariff pull-ins (Buchalter): "we have not seen a lot of tariff-related activity."
- Gaming: -30% y/y but grew sequentially (segment +2% q/q with client flat); Radeon 9070 first-week sellout >10x the prior best Radeon launch; semi-custom down y/y but consoles normalized and FY25 semi-custom growth guided.
- Embedded: -3% y/y, -11% q/q ("consistent with expectations"); return to y/y growth pushed to 2H25 on test/measurement, comms, A&D recovery.
- GM bridge: 54% in Q1; Q2 guided 43% including the ~$800M MI308 charge, ~54% ex-charge; Hu guided 2H GM "improve slightly" as DC becomes the #1 growth driver, partially offset by client/gaming strength — softer than last quarter's explicit "2H step-up."
- Balance sheet: cash + STI $7.3B (from $5.1B) but now carrying $1.5B new debt + $950M CP for ZT; net interest flips to ~$5M expense; FCF $727M vs $1.1B in Q4; inventory commentary shifted to deliberate builds for client/server and 2H GPU ramp (no dollar figure given on the call).
Product and launch scorecard- MI350 (CDNA 4): on track — Q1 multi-customer sampling delivered as promised, mid-year production reiterated, Oracle named as the first multi-billion-dollar MI355X deployment (with Turin and the Polara 400 NIC — first named deployment of the in-house NIC). 1.5x memory capacity/bandwidth, new data types, 35x throughput vs MI300X. This is now the single most important 2H25 deliverable.
- MI325X: ramping; majority of Q1 GPU revenue with MI300; a "large foundational model company" adopted MI325X significantly (unnamed); >35 MI300-series platforms in production (up from >25).
- MI400: on track for 2026; rack-scale, ZT co-design underway; "early customer feedback very positive" — still no named commitments; June 12 event is the next disclosure window.
- ROCm: 6.4 released (PyTorch/JAX/vLLM performance, cluster management tools); biweekly container cadence now in effect; day-zero Llama 4 / Gemma 3 / DeepSeek-R1; claimed MI300 "leadership inferencing throughput" on DeepSeek-R1 after post-launch optimizations.
- EPYC: Turin instances launching at all major CSPs; Oracle Exadata X11m optimized for Turin (up to 25% faster); CrowdStrike multi-cloud deployment; Siemens software-defined vehicle on Azure; EPYC Embedded 9005 won Cisco high-end firewalls and IBM Storage Scale 6000; Arizona fab milestone passed, 2H25 first shipments; Venice lead HPC product on TSMC N2, 2026.
- Client: Ryzen 9 9950X3D launched; Ryzen AI Max Plus and AI 7/5 300 series notebooks shipping; commercial momentum quantified for the first time (80% more AMD commercial systems, +30% Pro sell-through).
- Gaming: Radeon 9070 launch executed — record first-week sellout, >10x prior best, weekly replenishment underway; FSR 4 (first ML-based) in 30+ games, 75 targeted by year-end.
- Embedded: Spartan UltraScale Plus and Versal AI Edge Gen 2 initial shipments; Vitis AI suite update.
Sell-side read-through- Rasgon (Bernstein): got the cleanest admission of the call — Q2 GPU will decline y/y (the $700M MI308 hole), with y/y growth resuming in Q3. This confirms his flattish-to-down 1H25 GPU math from last quarter was directionally right.
- Arcuri (UBS): again failed to get a GPU dollar figure; extracted the China-impact phasing instead (majority Q2/Q3, very little Q4) and confirmation Q4 was always planned non-China/MI350.
- Arya (BofA): got GPU sequential direction (down modestly in Q1), MI300/MI325X as majority of GPU revenue, and Su's TAM defense — China restrictions were already factored into the $500B TAM; AI diffusion rules engagement ongoing, no quantified impact.
- Buchalter (TD Cowen): pressed the client pull-in question again; Su denied tariff-related ordering and attributed strength to ASP/mix and above-seasonal desktop. Hu's disclosure that units fell double-digit q/q while revenue held flat is the key data point for the pull-in debate — it cuts both ways.
- Sur (JPMorgan): MI400 rack-scale question produced the clearest statement that ZT was bought precisely for this; Su acknowledged learning from "challenges that have occurred with some of the recent deployments" (industry rack-scale stumbles, unnamed).
- Rakers (Wells Fargo): confirmed the 1H DC shape ex-China and got Hu's 2H GM framing — "improve slightly," a walk-back from last quarter's more explicit step-up language.
- O'Malley (Barclays): interconnect/UALink question — Su claimed "all the pieces required" plus ecosystem partnerships, no internal interconnect gap acknowledged.
- Muse (Cantor): extracted that AMD is planning 2H client sub-seasonal — management is not extrapolating the 1H client run-rate, an implicit acknowledgment of pull-in risk even while denying observed pull-ins.
- Seymore (Deutsche Bank): embedded recovery timing (Q3/Q4 y/y growth, Q4 strongest) and ZT opex run-rate (~$50M/quarter, fully in the $2.3B Q2 guide).
- Moore (Morgan Stanley): inference/custom-silicon — Su's answer stayed at portfolio/TCO level; "ASICs have a place, GPUs have a larger piece," no share or pricing commitment.
Management credibility- Delivered: revenue above the high end of guidance (vs ±$300M band); GM 54% as guided; MI350 multi-customer sampling in Q1 as promised; ZT closed in 1H25 as promised (March 31); Radeon 9070 early-March launch on time with record sellout; DC down ~5% q/q vs the ~7% guide — modestly better.
- New disclosure discipline on China: after six quarters of silence on China AI SKUs, management quantified the MI308 impact comprehensively ($700M Q2, $1.5B FY, $800M charge, phasing) in one call — the right move, but it also reveals MI308 was a material revenue line that was never previously disclosed as such.
- GPU opacity persists: no Q1 GPU dollar figure despite three direct attempts; "significant double-digit y/y" and "down very modestly q/q" are the only data points. The guide-withdrawal timing from last quarter now looks worse in hindsight — the first quarter under the new regime included a $700M quarterly China revenue removal that the old per-GPU guide would have forced into the open earlier.
- Language shifts: 2H GM "step-up" became "improve slightly"; "consistent with 2H24" framing replaced by explicit Q2 GPU y/y decline admission; training language stabilized at "training engagements ramped" with tier-one customers scaling clusters — firmer than last quarter's "customers want to see us as a strong training solution," but still no balance commitment.
- Consistency checks: client strength attributed to sell-out while units declined double-digit q/q — the ASP/mix explanation is internally consistent but means the +68% y/y is price/mix-driven, not volume; embedded recovery timing slipped again (now 2H25 y/y growth); "strong double digits" FY25 revenue guide maintained despite the $1.5B China removal, implying underlying ex-China guidance effectively rose.
- Under pressure: Su gave Rasgon a direct, unhedged answer on Q2 GPU y/y decline — a credibility positive; Hu's GM walk-back to "improve slightly" was delivered without being pressed — mixed signal worth tracking.
- Unresolved: GPU dollar levels (Q4 and Q1), server-CPU quantification (sixth quarter), GPU GM level/timeline, ZT purchase price (still undisclosed post-close), ZT manufacturing buyer, MI350 ASPs.
What changed versus the prior quarter- China went from unmentionable to the dominant variable: MI308 export controls remove $700M in Q2 and ~$1.5B for FY25, trigger an ~$800M charge (Q2 GM 43% reported), and confirm a China AI revenue stream AMD had never quantified. The Q4 plan was always non-China/MI350 — per Su, the FY impact is front-loaded.
- The flattish-1H25 GPU guide from last quarter is now confirmed in fact: Q1 GPU down modestly q/q, Q2 down y/y — Rasgon's implied sequential declines materialized. The entire FY25 GPU growth claim now rests on the 2H MI350 ramp, exactly the air-pocket scenario Buchalter and Curtis probed last quarter.
- First named multi-billion-dollar GPU commitment: Oracle MI355X cluster — the first time a hyperscale-scale MI350 deployment has been named with a dollar magnitude, and the first named Polara NIC design win.
- ZT moved from regulatory process to integration: closed, debt-funded, manufacturing headed to discontinued ops, ~$50M/quarter opex, 9M shares issued — but purchase price still undisclosed.
- Client narrative shifted from volume to price: +68% y/y with record ASPs and double-digit unit decline q/q; management now planning sub-seasonal 2H — an implicit hedge on the pull-in question they continue to deny.
- Gaming inflected as predicted: Radeon 9070 launch outperformed (>10x prior best first-week sellout), semi-custom guided to FY growth; the segment merger means Gaming op income is no longer visible — the transparency concern from last quarter is now live.
- GM trajectory softened: 2H "step-up" became "improve slightly"; Q2 reported GM drops to 43% on the charge.
- Server CPU story extended (Arizona production, Venice on N2, Exadata win) but the growth-rate disclosure drought hit a sixth quarter.
- Embedded recovery pushed out again: from "slower than expected" to a specific 2H25 y/y growth target with Q2 flattish.
Bull case- Underlying demand absorbed a $700M quarterly removal: Q2 guide of $7.4B flat q/q including the MI308 hole implies ex-China sequential growth of ~$700M — the business ex-China accelerated into Q2, with Client & Gaming guided up double-digit.
- MI350 is executing to the pulled-in schedule with the first named anchor customer (Oracle, multi-billion-dollar, full AMD rack: GPU + Turin + Polara NIC) — if mid-year production holds, AMD enters 2H25 with a higher-ASP part, a 35x generational claim, and a fast ramp path (MI350 systems environment similar to MI300).
- The FY25 "strong double-digit" revenue guide was maintained after removing $1.5B of China revenue — management is effectively guiding the ex-China business higher than the prior plan.
- Server CPU keeps compounding: seventh straight quarter of large double-digit on-prem growth, enterprise activations doubling, Turin instances at every major CSP, Arizona supply diversification, Venice on N2 for 2026 — the durable share-gain engine is intact regardless of GPU noise.
- Client is gaining share with margin: 17% segment op margin (vs 10% a year ago), record ASPs, commercial finally opening (80% more systems, Dell/HP/Lenovo/Asus), and management guiding client "well ahead of the market" for 2025.
- Gaming bottomed exactly as called: normalized inventories, record RDNA 4 launch, FY semi-custom growth guided — the 2025 tailwind thesis from last quarter is on track.
- ZT closed on schedule and is already co-designing MI400 racks — the rack-scale capability gap (the clearest competitive deficit vs Nvidia) is being addressed ahead of the 2026 MI400 launch.
- Software momentum is measurable: >2M Hugging Face models (2x in a quarter), day-zero Llama 4/Gemma 3/DeepSeek-R1, biweekly containers — the ROCm gap keeps narrowing at the inference layer where AMD's memory advantage is strongest.
Bear case- The GPU business just printed its first y/y decline quarter (guided Q2) since the AI ramp began, and management still won't disclose the dollar base — investors are being asked to underwrite a 2H MI350 ramp to "strong double-digit" FY growth without knowing the starting level.
- The $1.5B FY China hole was a revenue stream never previously disclosed; its removal exposes how much of the 2024-25 GPU trajectory was China-supported, and further export tightening remains an open risk with diffusion rules pending (May 15 implementation date referenced by Arya, unquantified by Su).
- Everything now hinges on one event: mid-year MI350 production and a 2H ramp. Any slip re-opens the air pocket with no China cushion, and Q4 was explicitly described as the MI350 quarter — there is no plan B in the guide.
- Client quality of growth is deteriorating: +68% y/y is ASP/mix-driven with units down double-digit q/q; management is planning sub-seasonal 2H while denying pull-ins — if tariff-driven demand did pull forward, 2H client could undershoot even the sub-seasonal plan.
- Margin story muddied: Q2 reported GM of 43% (even if ex-charge 54%), 2H GM language softened to "improve slightly," and opex stepped up to ~$2.3B with ZT — operating leverage is pausing (24% op margin vs 26% in Q4).
- FCF fell to $727M from $1.1B while inventory is being deliberately built for the 2H GPU ramp — if MI350 slips, AMD holds AI inventory against a restricted China market.
- Server CPU remains unquantified for a sixth quarter, and the segment's -5% q/q with GPU down "modestly" implies server CPU declined more than the segment — the strongest narrative line had a sequentially soft quarter.
- ZT integration risk is now real: debt-funded, price undisclosed, manufacturing divestiture pending, $50M/quarter opex — execution proof points are all ahead.
Next-quarter watchlist- MI350: confirmation that mid-year production started; first revenue contribution timing; whether Oracle's "multi-billion-dollar" deployment gets sizing/phasing; any additional named customers from the June 12 Advancing AI event; ASP and margin commentary.
- China: whether the $700M Q2 / $1.5B FY estimates hold or worsen; any license approvals; AI diffusion rule outcomes post-May 15; whether the $800M charge is final.
- GPU disclosure: does the segment-level-only policy survive a second quarter of declines, and does any dollar figure (or 2025 sizing) emerge at the June event?
- Q2 GPU y/y decline magnitude vs the implied ~$700M; Q3 return to y/y growth as Su committed to Rasgon — the single most checkable forward claim.
- Server CPU: seventh-quarter watch for a growth number; Turin instance ramp pace; any Intel pricing commentary; Arizona production start in 2H.
- Client: whether the guided double-digit Q2 sequential segment growth materializes; ASP sustainability; any admission of tariff pull-ins as 2H sub-seasonality plays out; Dell commercial ramp progress.
- Gaming: Radeon 9070 replenishment and sustained sell-through vs the record launch; FSR 4 title count toward 75; semi-custom order patterns for the FY growth guide.
- Embedded: Q2 flattish delivery; evidence for the 2H25 y/y growth call (test/measurement, comms orders); design-win metric stability.
- ZT: purchase price disclosure (10-Q), manufacturing divestiture buyer and terms, integration costs vs the ~$50M/quarter opex guide, first rack-scale design-win evidence.
- Financial mechanics: Q2 GM 43% reported / ~54% ex-charge; whether 2H GM "improve slightly" firms up; opex ~$2.3B; inventory build vs the 2H ramp; FCF recovery from $727M; buyback pace after the $749M step-up.
- Carried gaps: GPU dollar levels, server-CPU quantification (six quarters), GPU GM level/timeline, ZT price, MI350 ASPs.
|
| Feb 4, 2025 | -6.27% | Q4 FY2024 | Read transcript briefingQuarter in one view- Revenue $7.7B, +24% y/y, +12% q/q — above the ~$7.5B ±$300M guide midpoint; GM 54% (+330 bps y/y, in line with ~54% guide); opex $2.1B (+23% y/y, slightly above the ~$2.05B guide); op income record $2.0B (26% op margin); EPS $1.09, +42% y/y.
- Segments: Data Center record $3.9B (+69% y/y, +9% q/q; op income $1.2B, 30% vs 29% a year ago and 29% in Q3); Client record $2.3B (+58% y/y; op income $446M, 19% vs 4% a year ago and 15% in Q3); Gaming $563M (-59% y/y, +22% q/q; op income $50M, 9% vs 16% a year ago, 2% in Q3); Embedded $923M (-13% y/y, flat q/q; op income $362M, 39% vs 44% a year ago, 40% in Q3).
- FY2024: revenue $25.8B (+14%); DC +94%, Client +52%; GM +300 bps; EPS +25%; FCF more than doubled y/y; DC ~50% of annual revenue.
- Headline: FY2024 DC AI revenue confirmed "greater than $5B" — but no Q4 GPU number given; Su guided 1H25 DC segment (both server and GPU) "consistent with" 2H24, i.e., roughly flattish halves, with 2H25 stronger on the MI350 ramp. MI350 production pulled in from 2H25 to mid-year; lead-customer sampling this quarter.
- Q1 FY25 guide: revenue ~$7.1B ±$300M (+30% y/y, ~-7% q/q); GM ~54%; opex ~$2.1B; 13% tax; ~1.64B shares. Sequential split: DC down about the corporate average (~7%), Client and Embedded down more, Gaming down less (atypical, post-inventory normalization).
- Cash: CFO $1.3B, record FCF $1.1B; inventory +$360M to $5.7B (fourth straight build); cash + STI $5.1B (up from $4.5B); $256M buyback (1.8M shares), $4.7B authorization remaining.
- Segment reporting change: starting Q1 FY25, Client and Gaming combine into one reportable segment (three segments total), with continued distinct revenue disclosure for all four businesses.
What management is focused on- Reframing GPU guidance: Su explicitly retired the quarterly GPU dollar guide now that the business is "at scale" (>$5B) — guidance shifts to segment level with qualitative color. This is a deliberate disclosure reduction, presented as a maturity milestone (Arya challenge).
- MI350 pull-in as the proof of execution: silicon bring-up "better than expected" — large-scale LLMs running within 24 hours of first silicon, validation ahead of schedule — enabling production shipments at mid-year instead of 2H25. The 35x vs CDNA 3 claim (dropped last quarter) is back.
- "Tens of billions" long-term framing: Su reiterated the >$500B 2028 accelerator TAM and guided the AI franchise from >$5B (2024) to "tens of billions of dollars of annual revenue over the coming years" — no 2025 number attached.
- Server CPU share narrative escalated: "well over 50% share at the majority of our largest hyperscale customers"; >1,000 public cloud instances (+27% in 2024); >450 EPYC platforms (>120 Turin in production in Q4); enterprise cloud consumption nearly tripled y/y; named on-prem wins (Akamai, Hitachi, LG, ServiceNow, Verizon, Visa).
- DeepSeek/Stargate as demand validation: Su argues model-efficiency innovation expands adoption and TAM rather than shrinking it; ASICs remain "the smaller part" of TAM given algorithm churn; AMD positioned across CPUs, GPUs, and ASIC conversations.
- Dell commercial PC partnership: first full Ryzen Pro commercial portfolio from Dell — notebooks this spring, full portfolio 2H25 — the spearhead for commercial share gains.
- ZT Systems: unconditional approvals in Japan, Singapore, Taiwan; close on track for 1H25; "significant interest" in the US manufacturing divestiture, expected shortly after close. Purchase price still undisclosed.
Key numbers and quarter mechanics- DC GPU: no Q4 number disclosed. Su's only math: DC segment +9% q/q, server "a bit more than that," GPU "a little less than that" — implying GPU grew sequentially but low-single-digits. She told Rasgon his Q3 GPU estimate "might be a little bit light," i.e., Q3 was above the ~$1.5B+ he modeled, without giving a figure. Arcuri's ~$2B Q4 GPU guess was not confirmed or denied (Hu deflected to "server did better than data center GPU").
- 1H25 shape: DC segment (server and GPU both) "consistent with the second half of '24" — roughly flattish halves, "plus or minus," deployment-timing dependent. Su committed the 2025 GPU exit rate will be higher than the 2024 exit rate ("hard to grow strong double digits otherwise").
- Server CPU: double-digit sequential growth in Q4 (Su, to Sur), in both cloud and enterprise; enterprise grew sequentially again. Still no absolute server-CPU revenue or growth-rate number — fifth straight quarter.
- Client: $2.3B (+58% y/y) — well above the "grows sequentially" guide; record desktop channel sell-out (>70% share at Amazon, Newegg and others over the holidays); claimed record OEM notebook sell-through share; desktop constrained and still shipping strongly through January. Op margin 19% (from 15% in Q3, 4% a year ago).
- Gaming: $563M (-59% y/y but +22% q/q) — better than the "modest" framing; console cumulative shipments surpassed 100M units; channel inventories "now normalized," semi-custom to return to "historical patterns" in 2025. Graphics declined y/y again on RDNA 4 transition sellout.
- Embedded: $923M (-13% y/y, flat q/q) — recovery "slower than expected"; A&D and test/emulation strong, industrial/comms soft. Record $14B design wins closed in 2024, +25% y/y (metric shifted again: +40% 1H framing in Q2, +20% tracking in Q3, now +25%/$14B closed).
- GM bridge: 54% delivered on DC/Client mix up and Gaming down, partially offset by lower Embedded; Q1 guided ~54%; Hu guided 1H25 GM consistent with 54% and a 2H25 step-up as DC mix rises.
- Balance sheet: inventory $5.7B (+$360M, fourth build); cash recovered to $5.1B (from $4.5B) on record FCF; buyback steady at ~$256M.
Product and launch scorecard- MI300X: production deployments expanded at Meta (exclusive Llama 405B serving on meta.ai; Instinct added to the OCP Grand Teton platform for DLRM/inference) and Microsoft (GPT-4 Copilot services; flagship instances scaling to thousands of GPUs). New CSP deployers: IBM (plus watsonx enablement planned), DigitalOcean, Vultr; >12 CSPs deploying Instinct; >25 MI300-series enterprise platforms in production; Dell AI Factory integration with Hugging Face containers. HPC: El Capitan #1 on Top500 (>44,000 MI300A, >1.7 exaflops); Stuttgart's Hunter on MI300A.
- MI325X: volume production began in Q4 as promised; ramp "progressing very well"; new MI300/MI325 wins at "Lighthouse AI" customers deploying at scale across inference and training for the first time. OEM availability timing not re-stated this call.
- MI350 (CDNA 4): the quarter's biggest product news — production pulled from 2H25 to mid-year; lead-customer sampling this quarter; 35x AI compute vs CDNA 3 claim reinstated; "strong" customer feedback with deeper engagements at existing and net-new hyperscalers. Su frames it as addressing broader workloads including training, with higher ASPs and larger deployments.
- MI400 (CDNA Next): on track for 2026; rack-scale integration of networking/CPU/GPU at silicon level; "strong customer interest" — still no named commitments.
- ROCm: 6.3 released (flash attention up to 3x faster, SGLang runtime, day-zero DeepSeek V3 support); MI300X inference performance now +2.7x since launch (vs 2.4x last quarter); >1M out-of-box Hugging Face models; biweekly container releases started in January.
- EPYC/Turin: >540 performance records (vs >130 at launch); >120 Turin platforms in production in Q4; custom HBM-equipped EPYC powering new Azure HPC instances (claimed 8x memory bandwidth vs competition); three generations still selling.
- Client: 22 new mobile processors at CES; claimed only vendor with full Copilot+ CPU portfolio across segments; Dell Ryzen Pro commercial deal (spring notebooks, 2H25 full ramp); 2025 PC TAM assumed mid-single-digit growth, with Client revenue growing "well ahead of the market."
- Gaming: RDNA 4 / Radeon 9070 series on sale early March, targeting the highest-volume enthusiast segment with AI upscaling; PS5 Pro not discussed beyond the 100M cumulative console milestone.
- Embedded: Versal RF series launched (A&D), Versal Premium Gen 2 (first adaptive SoCs with CXL 3.1 / PCIe Gen 6), next-gen Alveo shipping for ultra-low-latency trading.
Sell-side read-through- Rakers (Wells Fargo): got the 2025 GPU shape — no number, but 2H stronger than 1H on MI350, and "strong double digits" for both server and GPU in 2025. His follow-up extracted the Q1 sequential split: DC down ~the corporate average (-7%), Client/Embedded down more, Gaming down less.
- Arcuri (UBS): pressed on Intel server pricing aggression — Hu acknowledged a competitive market but claimed TCO leadership across three generations, no concession on the mid-to-high-teens core-count growth framework. His attempt to pin Q4 GPU at ~$2B and get a Q1 number was deflected; Su substituted the "1H25 consistent with 2H24" framing.
- Arya (BofA): directly challenged the guidance retreat (explicit targets last year vs none this year) and the 60% CAGR applicability to 2025 — Su answered with "tens of billions over the coming years," no 2025 commitment. His DeepSeek question produced Su's adoption-expansion argument and the ASICs-stay-smaller view.
- Buchalter (TD Cowen): probed whether the flat 1H25 GPU shape reflects an MI350 air pocket — Su said the shape is unchanged and the pull-in is "incrementally positive," citing 1H MI300/MI325 design-win deployments. His client-inventory question got a firm denial: no tariff pull-ins seen, strength attributed to sell-out and desktop constraint.
- Sur (JPMorgan): extracted confirmation of double-digit sequential server growth in Q4 (cloud and enterprise both) and sequential enterprise growth. His networking question surfaced the in-house Pensando AI NIC and full rack solutions at MI350 and MI400 levels.
- Curtis (Jefferies): named the core tension — sequential GPU growth "slows for the next three quarters" — Su attributed it to a product-transition window, not demand. On training vs inference, she acknowledged customers "want to see us as a strong training solution" — softer than last quarter's "fairly balanced within a year" commitment, which was not repeated.
- Rasgon (Bernstein): ran the mezzanine math again — server up double digits with DC +9% implies GPU could be down sequentially, and flattish halves imply Q1 and Q2 GPU declines. Su corrected "strong double digits" to "double digits," said his Q3 GPU estimate may be "a little light," and held the flattish-halves line — the implied 1H25 sequential GPU declines were not refuted. She did commit to a higher 2025 GPU exit rate than 2024.
- Hari (Goldman): got the >50% hyperscaler share claim probed against custom silicon — Su's answer was product-portfolio breadth, not a share defense. Hu/Jean confirmed 1H25 GM ~54% with a 2H25 step-up on DC mix.
Management credibility- Delivered: revenue above guide midpoint; GM 54% as guided; DC and Client both records as guided; Gaming better than the "modest" framing; MI325X volume production started in Q4 as promised; FY24 GPU >$5B as guided (four raises, then landed); ZT regulatory milestones passing on schedule.
- Pulled forward, not just met: MI350 mid-year production vs prior 2H25 — the first schedule acceleration on the AI roadmap, supported by specific bring-up detail (LLMs in 24 hours, validation ahead of schedule). This is the strongest execution evidence on the call.
- Disclosure retreat: the quarterly GPU dollar guide — the scoreboard management itself created and raised four times — was withdrawn precisely when the sequential trajectory flattened. Su's framing ("business is now at scale") is plausible, but the timing aligns with Rasgon's math showing flattish-to-down 1H25 GPU quarters. Q4 GPU revenue was not disclosed despite direct asks (Arcuri, Rasgon).
- Language shifts: last quarter's "lumpy" was replaced by "consistent with the second half of '24" — more specific but still non-numeric; the "fairly balanced training/inference within a year" commitment from Q3 was not repeated, replaced by "customers want to see us as a strong training solution"; the 35x MI350 claim returned after being dropped last quarter.
- Consistency checks: Embedded design-win metric changed a third time (+40% 1H → +20% FY tracking → $14B closed, +25%); "strong double digits" vs "double digits" correction to Rasgon shows sensitivity to the GPU-down-sequentially inference; Hu's answer to Arcuri's Q4 GPU question ("server did better than data center GPU") was a non-answer.
- Unresolved: no 2025 GPU quantification (asked by Rakers, Arya, Rasgon); no Q4 GPU number; no server-CPU growth number (fifth quarter); GPU GM level/timeline not raised this call but still undisclosed; China AI SKUs unmentioned (sixth quarter); ZT purchase price still undisclosed.
What changed versus the prior quarter- GPU guidance regime ended: four consecutive raises ($2B → >$5B) are done; 2025 guidance is segment-level with qualitative color. Simultaneously, the sequential GPU growth streak effectively stalled — Q4 GPU grew only "a little less" than the segment's +9%, and 1H25 is guided flattish vs 2H24, implying Q1/Q2 sequential declines Rasgon's math surfaced and Su did not refute.
- MI350 accelerated: 2H25 → mid-year production, sampling this quarter — the biggest positive roadmap change; the 35x claim reinstated.
- Rasgon's Q4 flat-to-down GPU math from last quarter was partially resolved: Su said Q3 GPU was higher than his ~$1.5B estimate, so Q4 likely grew modestly — but the same math now applies to 1H25.
- Client delivered a second straight blowout ($2.3B, +58% y/y, 19% op margin) with a strategic Dell commercial win; management denied tariff pull-ins and inventory build, claiming desktop constraint into January.
- Gaming inflected off the bottom: +22% q/q, inventories declared normalized, 2025 framed as a return to historical patterns; RDNA 4 dated (early March).
- Embedded recovery pushed out: "slower than expected," Q1 guided to a modest decline; design-win metric restated again ($14B closed, +25%).
- Server CPU narrative strengthened: >50% share at most large hyperscalers claimed, double-digit sequential growth confirmed, enterprise momentum continuing — but Intel pricing pressure was acknowledged as a live question (Arcuri).
- Cash position improved ($4.5B → $5.1B) on record FCF, though inventory built a fourth straight quarter to $5.7B.
- Reporting change: Client + Gaming merge into one segment from Q1 FY25 — reduces standalone Gaming margin visibility (distinct revenue disclosure continues, op income does not).
- New disclosures: El Capitan Top500 win; IBM/DigitalOcean/Vultr deployments; Dell AI Factory; ROCm 6.3 with DeepSeek V3 day-zero support; >1,000 cloud instances; >450 EPYC platforms; 100M cumulative consoles; Pensando in-house AI NIC; ZT approvals (Japan, Singapore, Taiwan).
Bull case- MI350 pulled to mid-year with clean bring-up and claimed strong customer demand — if it holds, AMD closes the cadence gap and enters 2H25 with a higher-ASP, training-capable part against Blackwell, with management committing to a higher 2025 GPU exit rate than 2024.
- The >$5B 2024 AI franchise is real and broadening: >12 CSPs, >25 enterprise platforms, IBM/watsonx, Dell AI Factory, first at-scale training-plus-inference "Lighthouse" wins, and the world's fastest supercomputer — deployment evidence keeps accumulating.
- Server CPU is the quiet engine: >50% share at most large hyperscalers, double-digit sequential growth, enterprise nearly tripling cloud consumption, Turin's >540 records — and 1H25 server is guided flat-to-2H24 levels even in the "down" Q1, implying resilience.
- Client has delivered two consecutive record quarters with margin inflection (4% → 15% → 19%), record sell-out, desktop constraint into January, and the Dell commercial deal opening a historically weak segment — management guiding Client "well ahead of the market" in 2025.
- Gaming bottomed with inventories normalized; Embedded holds ~39% margins with record $14B design wins; both are positioned as 2025 tailwinds rather than drags.
- Record FCF ($1.1B), cash back up to $5.1B, and a 2H25 GM step-up guided on DC mix — the financial model is absorbing the AI ramp without the margin dilution bears predicted; DC op margin hit 30%.
- DeepSeek-style efficiency gains, if Su is right, expand the inference TAM where MI300X is strongest — and ROCm's day-zero DeepSeek V3 support shows the software gap narrowing in practice.
Bear case- The GPU guide was withdrawn exactly when growth stalled: Q4 GPU grew less than the segment's +9%, 1H25 is guided flattish vs 2H24 (implying sequential declines in Q1 and Q2 per Rasgon's unrefuted math), and management refused every request for a 2025 number. "Tens of billions someday" replaced a near-term target.
- The entire 2025 GPU story now rests on one event: the mid-year MI350 ramp. Any slip re-opens the air-pocket question (Curtis, Buchalter), and the competitive claim vs Blackwell remains management assertion.
- Training credibility softened: last quarter's "fairly balanced within a year" commitment was not repeated; Su now says customers "want to see us as a strong training solution" — an admission the proof is still pending.
- Server CPU faces acknowledged Intel pricing aggression (Arcuri) and hyperscaler custom-silicon questions (Hari) that Su answered with portfolio breadth rather than share or pricing commitments; the server-CPU growth number remains undisclosed for a fifth quarter.
- Client's 58% y/y quarter invites the inventory/pull-in question (Buchalter); management denied it, but Q1 is guided down more than the corporate average, and the claim rests on sell-out assertions, not disclosed channel data.
- Embedded recovery slipped again ("slower than expected," Q1 declining) with the design-win metric restated a third time; Gaming's recovery is a 2025 promise, not a Q4 fact.
- Inventory built a fourth straight quarter to $5.7B while GPU revenue flattens — the ramp-justification argument weakens if 1H25 GPU is flat; ZT's price and integration remain undisclosed with close in 1H25.
- Segment consolidation (Client + Gaming) reduces margin transparency on the two consumer businesses just as Gaming recovers — disclosed as an alignment move, but it obscures Gaming op income going forward.
Next-quarter watchlist- Q1 GPU trajectory vs the flattish-halves guide: does management confirm or deny sequential GPU decline, and does any 2025 GPU number emerge (or does the segment-level-only policy hold)?
- MI350: confirm lead-customer sampling happened in Q1 as promised; mid-year production timing specificity; any named customer commitments or early ASP/deployment-size evidence.
- MI325X: ramp progress and OEM availability; whether the "Lighthouse" training-plus-inference wins convert to disclosed deployments.
- Server CPU: evidence on Intel pricing impact (ASP commentary, share claims); whether the >50% hyperscaler share claim is updated; sixth-quarter watch for a server-CPU growth number.
- Client: magnitude of the Q1 seasonal decline vs guide (down more than corporate average); desktop constraint resolution; Dell commercial notebook spring launch on time; any tariff pull-in admission for Q1.
- Gaming: Q1 "down less than average" delivery; Radeon 9070 early-March launch execution, pricing, and sell-through; semi-custom order patterns post-normalization.
- Embedded: Q1 modest decline as guided; comms/industrial inflection timing; whether the design-win metric stabilizes.
- ZT Systems: close within 1H25, purchase price disclosure, manufacturing divestiture buyer and terms, and any rack-scale design-win evidence tied to MI350/MI400.
- Financial mechanics: GM ~54% in Q1 and the promised 2H25 step-up; opex ~$2.1B; inventory ($5.7B) vs the flat-GPU 1H; first report under the new three-segment structure — check what Gaming disclosure survives.
- Carried gaps: 2025 GPU sizing, Q4/Q1 GPU revenue levels, server-CPU quantification (five quarters), China AI SKUs (six quarters), GPU GM level/timeline, ZT price.
|
| Oct 29, 2024 | -10.62% | Q3 FY2024 | Read transcript briefingQuarter in one view- Revenue $6.8B, +18% y/y, +17% q/q — above the ~$6.7B ±$300M guide midpoint; GM 54% (Hu later specified 53.6% delivered, +250 bps y/y, vs ~53.5% guide); opex $1.96B (+15% y/y, slightly above the ~$1.9B guide); op income $1.7B (25% op margin); EPS $0.92, +31% y/y, +33% q/q.
- Segments: Data Center record $3.5B (+122% y/y, +25% q/q; op income $1.0B, 29% vs 19% a year ago and 26% in Q2); Client $1.9B (+29% y/y, +26% q/q; op income $276M, 15% vs 10% a year ago and 6% in Q2); Gaming $462M (-69% y/y, -29% q/q; op income $12M, 2% vs 14% a year ago); Embedded $927M (-25% y/y, +8% q/q; op income $372M, 40% vs 49% a year ago).
- Headline: FY2024 Data Center GPU guidance raised a fourth time, >$4.5B → >$5B; Su confirmed Q3 GPU revenue was "greater than $1.5B" and "approaching the scale of our CPU business" (Arcuri/Rasgon extraction). DC now 52% of revenue.
- Q4 guide: revenue ~$7.5B ±$300M (+22% y/y, ~+10% q/q); GM ~54%; opex ~$2.05B; 13% tax; ~1.64B shares. Sequential: DC largest contributor, Client grows, Gaming/Embedded "more modest" (Su, to Rakers).
- Cash: CFO $628M, FCF $496M ($619M excluding acquisition-related payments); inventory +$383M to $5.4B (DC ramp); cash + STI $4.5B (down from $5.3B); $250M buyback (1.8M shares), $4.9B authorization remaining.
What management is focused on- Fourth GPU guide raise framed as milestone-driven: Su attributes the >$5B raise to "completed some important customer milestones" (at-scale deployment, reliability, workload optimization) plus supply-chain improvement — the same milestone framework she described last quarter.
- ZT Systems acquisition as the rack-scale answer: definitive agreement announced; Su positions ZT as delivering "rack and cluster level solutions," parallel silicon/systems design, and faster deployment at scale. HSR waiting period passed; close expected 1H25; plan to divest ZT's US manufacturing business at close with "significant interest" from buyers. This is the concrete response to two straight quarters of rack-scale questions answered with "investment intent."
- Customer proof points escalated: Meta now running all live Llama 405B traffic exclusively on MI300X and publicly discussing expansion into training; Microsoft using MI300X "broadly" for multiple Copilot services on GPT-4 models; Oracle Cloud added to public MI300X instances; named AI-native adopters (Essential AI, Fireworks AI, Luma AI, Databricks).
- Software maturity narrative: >1M models run out-of-the-box on Instinct; ROCm 6.2 delivered 2.4x inference and 80% training performance improvement since launch; claimed 30% higher performance vs competitive offerings on tuned inference workloads; foundational ROCm support added to Triton, Llama Stack, SGLang, vLLM, TensorFlow.
- TAM framing: AI accelerator TAM now claimed >60% CAGR to $500B by 2028 — a new, aggressive market-sizing anchor.
- Turin launched (early October): >130 performance records, >130 OEM/ODM platforms in development, Google and OCI fifth-gen instances planned for early next year; Oracle Turin instances cited at +35% performance/core, +33% memory speed, 2x networking bandwidth.
- x86 ecosystem advisory group formed with Intel — a notable competitive-détente signal on architecture consistency.
Key numbers and quarter mechanics- DC GPU: Q3 revenue "greater than $1.5B" (Su, to Arcuri) and "approaching the scale of our CPU business" (to Rasgon). Rasgon's math: with DC segment at $3.5B and non-CPU/GPU content, CPU is roughly ~$1.7B+, implying GPU near that level; he noted >$5B for the year could imply a flat-to-down Q4 GPU quarter unless the true figure is $5.2–5.3B+ — Su did not resolve this, only reiterated "exceeds $5B."
- DC segment: $3.5B (+25% q/q); op margin 29% (from 26% in Q2, 19% a year ago) — third straight quarter of DC margin expansion while GPU scales; Hu again attributes to revenue leverage.
- EPYC: enterprise sales up strong double-digit y/y for the fifth straight quarter; public cloud instances >950, +20% y/y (growth rate decelerating: +34% y/y last quarter); Meta disclosed at >1.5M EPYC CPUs deployed; OEM Genoa/Bergamo platforms +50% in a year, >200 solutions; Seymore's ~33% y/y server-CPU growth math was not endorsed or refuted — fourth straight quarter without a server-CPU number.
- Client: $1.9B (+26% q/q) — well above the "strong growth" guide; Su claims highest desktop sell-through and good Ryzen AI 300 activation rates (Arya pressed on units vs ASP; Su said "both units up," ASP mix-dependent). Op margin jumped to 15% from 6% — the segment's profitability problem visibly improved.
- Gaming: $462M (-29% q/q, -69% y/y) — worse than the guided double-digit decline; op margin collapsed to 2% (from 12%). Semi-custom inventory reduction by Microsoft and Sony; gaming graphics also declined y/y now (last quarter it grew) as AMD preps the RDNA 4 transition.
- Embedded: $927M (+8% q/q) — better than the guided single-digit; recovery led by test/emulation (Versal VP1902 at all three largest EDA vendors) and aerospace/defense (SpaceX satellite win); comms and industrial still soft. Margin 40%, flat q/q but down from 49% a year ago. Design wins tracking +20% y/y in 2024 (note: last quarter's framing was +40% y/y to >$7B in 1H — the growth rate cited has slowed).
- GM bridge: 53.6% delivered vs ~53.5% guide; Q4 guided ~54% — Hu's puts/takes (to Hari): DC mix and enterprise server expansion and Embedded recovery as tailwinds; consumer-weighted Client mix as headwind; operational efficiency/scale as the offset.
- Balance sheet: inventory +$383M to $5.4B (third straight build); cash fell to $4.5B from $5.3B — Silo AI closed in the quarter (acquisition-related payments visible in the FCF bridge) plus buyback.
Product and launch scorecard- MI300X: >$1.5B quarter; Meta exclusive live-traffic deployment for Llama 405B (the strongest single production proof point to date) plus stated intent to expand into training; Microsoft broad Copilot/GPT-4 usage; Oracle Cloud instances added; adoption evidence now includes named AI-native clouds.
- MI325X: launched (early October); claimed up to 20% higher inference performance vs H200 and "competitive training performance"; production shipments start this quarter (Q4); broad OEM availability (Dell, HPE, Lenovo, Supermicro) from Q1 2025. Last quarter's "small Q4 contribution" framing not updated.
- MI350 series: "silicon is looking very good," on track for 2H25 launch; Su now calls it "the largest generational increase in AI performance we have ever delivered" (last quarter's specific 35x vs CDNA 3 claim was not repeated); Su reiterated it "will compete very well with Blackwell" (Arya challenge).
- MI400/CDNA Next: "progressing very well" toward 2026 — no new detail.
- ROCm: 6.2 released; 2.4x inference / +80% training performance since launch; >1M out-of-box models; 30% tuned-workload advantage claim; open-source ecosystem additions (Triton, SGLang, vLLM, TensorFlow, Llama Stack).
- EPYC/Turin: launched with >130 records; Google/OCI instances early next year; Milan (Zen 3) still selling on price/performance — three generations shipping simultaneously.
- Client: Ryzen 9000 desktop strength; Ryzen 9000 X3D launching November; Ryzen AI Pro 300 launched (first enterprise-class Copilot+ CPU claim); HP/Lenovo tripling Ryzen AI Pro platforms in 2024; >100 commercial platforms in 2025; Windows 10 end-of-support (2025) now an explicit demand catalyst.
- Gaming: PS5 Pro announced with new AMD semi-custom SoC (extends the Sony partnership but doesn't reverse near-term declines); RDNA 4 on track for early 2025 with claimed big ray-tracing and AI gains.
- Embedded: Telluride taped out (first second-gen Versal product, claimed up to 10x compute); Versal AI Core in SpaceX's latest satellites.
- Corporate: ZT Systems agreement (close 1H25, manufacturing divestiture planned); Silo AI closed; x86 advisory group with Intel.
Sell-side read-through- Hari (Goldman): got the raise rationale (customer milestones + supply) and a 2025 framing — broadening workloads at existing cloud customers plus engaged new large cloud/enterprise customers, no numbers. His GM question produced the clearest 2025 puts-and-takes list (DC mix, enterprise server, Embedded recovery up; consumer Client mix down).
- Rakers (Wells Fargo): supply-chain probe — Su kept "tight environment" language but added "we've planned for significant growth going into 2025"; extracted the Q4 sequential split (DC largest, Client grows, Gaming/Embedded modest).
- Seymore (Deutsche Bank): caught the Embedded ambiguity in the Q4 guide (Hu omitted it; Su confirmed "modest improvement"); his ~33% server-CPU growth math went unanswered — fourth quarter running. His GPU-crowding-out-CPU question got a "market environment has gotten better" answer with three-generation stack strength.
- Reitzes (Melius): Turin-as-AI-head-node framing endorsed by Su; on PCs, extracted an explicit admission of normal 1H seasonality ("we would expect some level of seasonality going into the first half") — a tempering of last quarter's above-seasonality momentum.
- Buchalter (TD Cowen): runway-at-existing-customers question; Su's answer was qualitative (workload expansion at Meta/Microsoft, "working with all the large customers") — no new customer named.
- Arcuri (UBS): extracted the quarter's key number — Q3 GPU "greater than $1.5B" and "approaching the scale of our CPU business." His 2025-shape question (pause risk before rack scale in 2026) got Su's new "lumpy" language: "it might be lumpy… not always predictable exactly which quarters."
- Moore (Morgan Stanley): training/inference split — Su now commits to "a fairly balanced portfolio between training and inference" within a year, a firmer claim than last quarter's "largely inference… over time."
- Arya (BofA): directly challenged the one-year-behind-Nvidia spec gap; Su rejected the framing, claimed the gap has closed with the accelerated roadmap, and leaned on market constraint and easier data-center retrofit as the equalizers. His channel/sell-through challenge (Intel guided down; Q3 sell-through "not that great") was rebutted with claimed highest desktop sell-through and good AI 300 activation rates.
- Sur (JPMorgan): enterprise follow-through confirmed for Q4 (sequential server growth, strength in both enterprise and cloud); China answer — Hu: "underrepresented in China… opportunity" — still no China AI SKU discussion, fifth straight quarter.
- Rasgon (Bernstein): did the mezzanine math again — showed that >$5B with Q3 at ~$1.5B+ implies Q4 GPU could be flat-to-down unless the real number is $5.2–5.3B+; Su confirmed the >$1.5B Q3 figure but gave no Q4 GPU number, leaving the implied deceleration unresolved.
- Kumar (Piper Sandler): gross-margin framework for Instinct — Jean Hu confirmed MI300/325 GM is below corporate average, improvement tied to revenue ramp, long-term data-center margins "better than corporate average," still no timeline or level.
- O'Malley (Barclays): extracted Q4 mix detail — MI300 remains the majority of Q4 GPU revenue; MI325X production starts late in Q4, ramp is a Q1 event.
Management credibility- Delivered: revenue above guide midpoint; GM 53.6% vs ~53.5% guide; DC largest sequential contributor as guided; Client strong growth as guided (and then some, +26% q/q); Embedded single-digit-plus growth delivered (+8%); Turin launched on the "later this year" schedule with Google/OCI commitments; MI325X launched in the stated window with Q4 production start; Silo AI closed in Q3 as stated; fourth consecutive GPU guide raise ($2B → >$3.5B → >$4B → >$4.5B → >$5B) continues the milestone-based pattern.
- Extracted, not volunteered: the Q3 GPU revenue level (>$1.5B, near CPU scale) came only under Arcuri/Rasgon pressure; the implied Q4 GPU flat-to-down math (Rasgon) was left unresolved; MI325X being a Q1-ramp product with MI300 still the Q4 majority (O'Malley); Q1 PC seasonality admission (Reitzes).
- New hedging language: Su introduced "lumpy" for 2025 GPU revenue (Arcuri) — the first explicit acknowledgment of quarter-to-quarter unpredictability, paired with continued refusal to size 2025. Last quarter's "tight through 2025" supply language was maintained but softened with "planned for significant growth."
- Consistency checks: MI350's "35x vs CDNA 3" claim from last quarter was not repeated — replaced with "largest generational increase… ever"; Embedded design-win growth cited as +20% y/y for 2024 vs +40% y/y (1H, >$7B) last quarter — the metric and rate shifted without explanation; gaming graphics flipped from y/y growth (Q2) to decline (Q3).
- Unresolved: no 2025 GPU quantification despite direct asks (Hari, Buchalter, Arcuri); no server-CPU growth number (fourth quarter); no GPU GM level or accretion timeline (Kumar, again); China AI SKUs unmentioned (fifth quarter); ZT purchase price not disclosed on this call.
What changed versus the prior quarter- FY2024 DC GPU guide raised a fourth time: >$4.5B → >$5B; Q3 GPU confirmed >$1.5B and "approaching" CPU scale — but Rasgon's math shows the raise implies a flat-to-down Q4 GPU quarter unless the true figure is higher, and management would not clarify.
- Rack-scale strategy moved from "investment intent" to a signed deal: ZT Systems acquisition (close 1H25, HSR cleared, US manufacturing to be divested) — the biggest strategic change of the quarter.
- Client inflected hard: +26% q/q to $1.9B with op margin 6% → 15% — the segment's chronic margin problem showed its first real improvement; but Su now concedes normal 1H25 seasonality.
- Gaming deteriorated beyond guide: -29% q/q, op margin 12% → 2%, and graphics revenue flipped to y/y decline; PS5 Pro win announced but not near-term accretive.
- Embedded recovery confirmed (+8% q/q vs single-digit guided) with sub-segment detail (test/emulation and A&D strong; comms/industrial soft), but design-win growth language slowed (+40% 1H → +20% FY24 tracking).
- Training narrative firmed: from "largely inference, ramping over time" to a commitment of "fairly balanced" training/inference within a year, backed by Meta's stated intent to expand into training workloads.
- 2025 language introduced: "lumpy" (Su) — new caution alongside continued refusal to quantify; supply still "tight" but with "significant growth planned."
- Cash continued down ($5.3B → $4.5B) with Silo AI closed and inventory up again to $5.4B; buyback slowed ($352M → $250M).
- New disclosures: Meta's 1.5M+ EPYC CPUs and exclusive 405B live traffic on MI300X; ROCm performance deltas (2.4x/80%); >1M out-of-box models; x86 advisory group with Intel; AI accelerator TAM claim of $500B by 2028.
Bull case- Fourth guide raise with Q3 GPU already >$1.5B and approaching CPU scale; the milestone-based raise pattern has now repeated four times, and Su attributes this one to completed at-scale customer milestones plus supply improvement.
- Demand evidence hardened again: Meta running all live Llama 405B traffic exclusively on MI300X and publicly discussing training expansion; Microsoft broad Copilot deployment; Oracle Cloud instances; named AI-native adopters; ROCm now with foundational support across the major open-source inference/training stacks and claimed 30% tuned-workload advantage.
- The rack-scale gap — the clearest competitive deficiency — now has a signed answer in ZT Systems, with HSR cleared and hyperscaler feedback claimed "very positive."
- DC op margin expanded for a third straight quarter (19% → 23% → 26% → 29%) while GPU revenue roughly 5x'd y/y — the margin-dilution bear case keeps weakening; Hu reiterates long-term data-center margins above corporate average.
- Client delivered a genuine beat-and-margin-inflection quarter (+26% q/q, 15% op margin) with X3D, Ryzen AI Pro, >100 commercial platforms, and the Windows 10 refresh as 2025 catalysts.
- EPYC momentum broad-based: fifth straight double-digit enterprise growth quarter, Turin launched with Google/OCI commitments, three generations selling simultaneously, and management claiming server share gains.
- Training/inference mix commitment ("fairly balanced within a year") would materially widen the addressable workload base if delivered.
Bear case- The raise math doesn't obviously clear: Rasgon showed >$5B with Q3 at >$1.5B implies Q4 GPU flat-to-down unless the real number is $5.2–5.3B+ — and Su declined to resolve it. Combined with the new "lumpy" 2025 language, the sequential GPU growth streak may be ending.
- 2025 visibility was asked four ways (Hari, Buchalter, Arcuri, Moore-adjacent) and answered qualitatively every time; "lumpy" is a downgrade in predictability language even if directionally positive.
- Gaming is in freefall (-69% y/y, 2% op margin) with graphics now also declining; Embedded margin stuck at 40% (vs 49% a year ago) with design-win growth language slowing; two of four segments remain impaired.
- Cash keeps falling ($4.5B, from $6.0B three quarters ago) while inventory builds to $5.4B and ZT Systems adds an undisclosed purchase price and integration risk on top of Silo AI; the manufacturing divestiture adds execution complexity.
- The spec-gap question (Arya) was answered with market-constraint and retrofit arguments rather than a claim of parity — MI350 vs Blackwell competitiveness remains a management assertion, and the 35x claim was quietly not repeated.
- Client's strong Q3 was partly launch-driven with admitted 1H25 seasonality ahead; Arya's sell-through challenge was rebutted with claims, not data.
- China AI SKUs unmentioned for a fifth consecutive quarter; server-CPU growth unquantified for a fourth; GPU GM level and accretion timeline still refused.
Next-quarter watchlist- Q4 GPU revenue vs the implied ~$1.5B-or-less: does the print confirm Rasgon's flat-to-down math, and does management give a 2025 GPU number at all (or hold the "lumpy" line)?
- MI325X: confirm production shipments started in Q4 as promised and the Q1 2025 OEM availability/ramp; any update to the "small Q4 contribution" framing.
- MI350: 2H25 launch timing specificity; whether the 35x performance claim is re-anchored; early customer commitments.
- ZT Systems: purchase price disclosure, close progress toward 1H25, manufacturing divestiture buyer, and any early rack-scale design-win evidence.
- DC op margin: does 29% hold as GPU mix rises further; any quantification of GPU GM trajectory (Kumar's framework question will return).
- Client: Q4 growth as guided, then the admitted 1H25 seasonal decline — magnitude vs normal seasonality; X3D launch (November) and Ryzen AI Pro platform expansion; whether 15% op margin holds.
- Gaming: Q4 "modest" sequential performance per Su's framing; RDNA 4 early-2025 launch timing; PS5 Pro contribution timing.
- Embedded: continued modest Q4 growth; comms/industrial inflection; margin stabilization at 40%; reconciliation of the design-win growth-rate language (+20% vs +40%).
- Corporate mechanics: GM ~54% delivery; opex ~$2.05B; inventory trajectory ($5.4B) vs ramp justification; cash level after ZT; buyback pace.
- Carried gaps: 2025 GPU sizing, server-CPU growth quantification (four quarters), China AI SKUs (five quarters), GPU GM level/timeline, and any sub-segment GPU disclosure.
|
Corbanu