| Aug 26, 2026 | +8.74% | Q2 FY2027 | Read transcript briefingQuarter in one view- Revenue $96B, >100% y/y (fourth consecutive quarter of y/y acceleration), vs the $91B ±2% guide (~5% above midpoint). Data Center $89B, +18% q/q: Hyperscale $49B, +13% q/q; ACIE $40B, +25% q/q, +138% y/y.
- Q3 FY27 guide: $108B ±2% (+12.5% q/q at midpoint), again zero China DC compute assumed. Vera Rubin expected to be ~20% of DC revenue in Q3 — the first quantified ramp contribution.
- First-ever full-year-ahead revenue guide: FY28 revenue growth ~70% y/y, explicitly supply-constrained — management says demand implies ~100% growth ("customers' forecasts point to our growth doubling next year"). Rasgon backed out this is roughly a $200B uplift vs the prior $1T three-year framework.
- Gross margin reset — the quarter's biggest negative: Q2 held at 75% GAAP/non-GAAP, but Q3 guided to 74% ±50bps, bottoming in Q4 at 71–72%, settling at 72–73% in FY28 as price increases take effect in Q1. Driver: "extreme pricing conditions in memory," exceeding prior expectations and "headed even higher into next year." This ends the multi-quarter mid-70s framework.
- FY27 opex raised again: low-50s % growth (from upper-40s last quarter, low-40s the quarter before) — third consecutive upward revision.
- Vera Rubin production shipments commenced earlier this month (August), with POs "from every major hyperscaler, AI cloud and system OEM"; claimed to be "the fastest product ramp in NVIDIA's history."
- AWS expansion: +2M GPUs (Blackwell/Rubin) deploying from this quarter through Q2 FY29, plus Vera CPUs (integrated and standalone), Nemotron on Bedrock/SageMaker, and full physical-AI stack for warehouse robots.
- Frontier-lab financing disclosed at scale: ~$50B invested in frontier AI labs; $500B+ third-party capital via six financing platforms (Apollo, BlackRock, Blackstone, Brookfield, Goldman, KKR); OpenAI commitments ~12GW of NVIDIA compute through 2030; SoftBank Energy Portsmouth site, initial 4.25GW exclusively for NVIDIA compute, to be used by OpenAI. Management pre-emptively addressed "circular financing" criticism.
- Record $26B returned in Q2 ($20B buyback + $6B dividend); 60% of FCF returned YTD vs the 50% pledge.
What management is focused on- The FY28 ~70% guide as an alignment tool: Jensen framed the unprecedented year-ahead guide as giving customers, suppliers, and shareholders "the same set of information" because "everybody is putting a lot of resources at play." The constraint is supply, not demand — repeated at least four times.
- Memory as the margin story, reframed as a demand symptom: Kress was deliberately direct ("we want to be direct about this rather than let it linger") — memory scarcity is "driven in large part by the AI build-out itself," and price increases take effect Q1. This is the first time memory cost has been given a full P&L trajectory after two quarters of unanswered margin-mechanics questions.
- Revenue per gigawatt as the new master metric: Hopper $18B/GW → Blackwell $25B/GW → Vera Rubin $40B/GW, with Jensen telling Rakers the trajectory continues higher beyond Vera Rubin ("infinity per gigawatt" directionally). Vera Rubin claims re-based again: 30x throughput per megawatt and 35x lower token cost vs Grace Blackwell Ultra (last quarter: 35x inference throughput / 10x AI factory revenue vs Blackwell).
- Financing the demand chain: three distinct mechanisms now disclosed — (1) ~$50B direct equity in frontier labs, (2) $500B+ third-party infrastructure financing platforms, (3) a new NeoCloud revenue-sharing model: NVIDIA gives take-or-pay commitments/minimum revenue guarantees on part of a facility's capacity so lenders will underwrite it, and shares in revenue above the floor. "We get paid twice." Also "selective credit enhancement for nearly 2GW" for one unnamed frontier lab.
- ACIE as the growth engine: guided to drive Q3 sequential growth; Jensen says the non-hyperscaler half is "growing 100% a year" and will be "roughly half of our data center business" and likely larger than hyperscale over time. NeoCloud partners expected to exit the year with 8GW installed, up from ~3GW at end of 2025.
- CPU as a confirmed pillar: Grace CPU >$5B TTM revenue; Vera CPU in full production, shipping to OCI, SpaceX AI, and now AWS; ~$20B total server CPU demand reiterated; preliminary expectation for CPU revenue to more than double in FY28.
- Groq 3 LPX in full production — first rack-scale LPU system, ~4x tokens/sec vs "next best alternative" (artificial analysis benchmark), volume shipments later this quarter, Nebius first. Jensen still frames it as a bolt-on for high-interactivity services: "the vast majority of the world's data centers will just be Vera Rubin and NVLink 72."
Key numbers and quarter mechanics- Revenue $96B vs $91B ±2% guide; Q3 guide $108B ±2%; FY28 preliminary ~+70% y/y, supply-constrained (demand framed as ~100%).
- DC $89B, +18% q/q: Hyperscale $49B (+13% q/q); ACIE $40B (+25% q/q, +138% y/y). Hyperscale growth expected to reaccelerate in Q4 and into FY28 as Vera Rubin supply grows — i.e., Q3 is ACIE-led by design.
- Networking: record quarter, +18% q/q; Spectrum-X +2.6x y/y (decelerating from ~3x y/y last quarter on a larger base). Still no absolute networking dollar figure this quarter.
- Gross margin 75.0% Q2 → 74% Q3 guide → 71–72% Q4 trough → 72–73% FY28. China H200 shipments (<1% of DC revenue) are dilutive to gross margin.
- Opex +10% GAAP / +11% non-GAAP q/q; Q3 guided ~$9.2B GAAP / ~$9.0B non-GAAP; FY27 growth now low-50s % (third raise: low-40s → upper-40s → low-50s).
- Tax: Q2 non-GAAP 16%; FY27 16–18% reiterated.
- Inventory $32B (build for Vera Rubin launch). DSO jumped to 60 days (from 45 in Q1, guided mid-50s) — attributed to extended payment terms for large multi-quarter purchases by certain investment-grade customers. This is a new, structurally different working-capital explanation, not just timing.
- Capital returns: $26B in Q2 ($20B buyback, $6B dividend at $0.25); 60% of FCF YTD; intent to "increase and return excess free cash flow net of strategic uses" — note the new "net of strategic uses" qualifier alongside the $50B lab investments.
- Demand-side data points: cloud industry backlog >$2T; top-5 hyperscaler capex ~$800B in 2026, $1.3T in 2027 (up from ">$1T by 2027" last quarter); global AI VC funding >$400B in H1 2026 (~70% spent on compute) vs $265B in all of 2025; ~20 AI-native companies now >$1B ARR (up from 13 in Q4).
- Enterprise verticals (TTM): automotive on-prem $8B; financial services + manufacturing + healthcare $7B combined. Sovereign: +35% q/q, >3x y/y (re-accelerated from +80% y/y last quarter).
- China: <1% of Q2 DC revenue in H200 under licenses; zero in forward outlook.
Product and launch scorecard- Vera Rubin — shipping, quantified, claims re-based again: production shipments began August; POs from every major hyperscaler/AI cloud/system OEM; ~20% of DC revenue in Q3 is the first hard ramp number — checkable next quarter. Performance framing changed for the third time: now 30x throughput/MW and 35x lower token cost vs Grace Blackwell Ultra (prior: 35x inference throughput / 10x factory revenue; before that: 1/4 GPUs for MoE / 10x token cost). The comparison baseline also shifted (GB Ultra vs Blackwell). Independent validation still absent.
- Vera CPU — converting from "visibility" to shipments: full production; lead partners named (OCI, SpaceX AI, AWS from this quarter); ~$20B server CPU demand reiterated; CPU revenue guided to more than double in FY28; Grace CPU >$5B TTM gives the first actual revenue baseline. Spec claims: 1.8x faster agentic tasks, 5x bandwidth/watt.
- Groq 3 LPX — from niche framing to production in one quarter: full production, volume shipments this quarter, Nebius first customer, ~4x tokens/sec claim. Last quarter Jensen capped LPX at "<20% of the market, niche for some time" — the tone is warmer now, though he still says the vast majority of data centers will be Vera Rubin/NVLink 72. Note: LPX is excluded from the $1T framework, so this is unmodeled upside per management's own definition.
- Blackwell — still the hyperscale workhorse: sustained strength drove Hyperscale +13% q/q; AWS's 2M-GPU commitment spans Blackwell and Rubin through FY29.
- Networking — record but decelerating growth rate: +18% q/q, Spectrum-X +2.6x y/y (vs ~3x last quarter); "largest and fastest-growing network company" claim repeated. Run-rate number still never given (eighth quarter).
- Nemotron/open models — new distribution proof point: AWS will serve Nemotron on Bedrock/SageMaker; Jensen's open-model answer (Arcuri) positioned open models as additive, "nearly all open models run on NVIDIA."
- Physical AI — first mega-customer: Amazon adopting Omniverse/Cosmos/Isaac/Jetson for warehouse robots. No updated LTM revenue figure this quarter (was >$9B).
- Software monetization — still no metric (eighth quarter), though the NeoCloud revenue-share model is a new recurring-revenue adjacency ("potential to drive billions over the medium to long term").
Sell-side read-through- Eight questions again; call ended with "no further questions." The margin reset and FY28 guide absorbed the oxygen; zero questions on China (fourth straight quarter), zero on the OpenAI equity agreement status, zero on gaming/consumer, zero on the revenue-share model's risks.
- Moore (Morgan Stanley) asked the obvious: why guide a full year out for the first time, and what separates 70% from 100%? Jensen's answer: supply, and a desire to align the whole ecosystem on one number. No quantified constraint breakdown offered.
- Muse (Cantor) on inference share got the four-phase AI-lifecycle framing and the fungibility argument — but notably no repeat of last quarter's "growing share very, very quickly" claim; the answer was architectural, not share-quantified.
- Rasgon (Bernstein) did the math live: FY28 ~70% implies roughly a $200B uplift vs the $1T framework, and asked how much is pricing (given the Q1 price increases). Jensen did not deny the uplift math and did not separate price vs volume — a gap to press on. He also asked the unconstrained number; answer: "a lot higher," no figure.
- Arya (BofA) totaled ~$500B of commitments/guarantees from the CFO commentary and asked the sharpest strategic question: OpenAI and Anthropic are both designing custom silicon (OpenAI's "Jalapeno" claims vs Blackwell) — how does NVIDIA reconcile funding its own competition? Jensen: XPUs are inference-specific single-cloud chips; NVIDIA is a full platform; "100% confidence" they remain customers. Kress added that supply commitments are concentrated in the first three years and underpin the revenue confidence. The $500B total was not confirmed or denied as the right ballpark.
- Arcuri (UBS) pivoted from capital returns (now answered) to open-source models — a soft question; Jensen: both open and closed are skyrocketing, NVIDIA runs all of them.
- Reitzes (Melius) asked about recursive self-improvement/AGI as a demand catalyst — Jensen: "it's going to inflect further," agents already crossed 50% of AI usage "this last month," and milestones are "senseless"; what matters is profitable tokens.
- Schneider (Goldman) asked for a rank-ordering of supply constraints (power/shell, DRAM, wafers) — Jensen declined to rank ("our entire supply chain is challenged"), deflecting with an anecdote. Given the memory margin reset, the refusal to rank DRAM is conspicuous.
- Rakers (Wells Fargo) on $/GW trajectory beyond Vera Rubin: Jensen confirmed directionally higher ($40B → more), and cited ROIC on AI data centers now under one year — a new third-party-sounding data point supporting token economics.
Management credibility- Beat-and-raise extended: $96B vs $91B guide; Q3 guided $108B (+$12B q/q) — sixth consecutive large sequential add, still ex-China.
- The margin reset is a credibility event in both directions: they broke the news proactively and completely (Q3, Q4 trough, FY28 settle, price-increase timing) — good practice. But "mid-70s FY27" was reiterated just last quarter, and memory mechanics questions went unanswered for two quarters before this. The magnitude "exceeded our prior expectations" is an admission the prior guide was set without full visibility into their largest input cost.
- Opex guide raised for the third consecutive quarter (low-40s → upper-40s → low-50s). The pattern now looks like systematic under-guiding of opex, not one-off revision.
- FY28 ~70% guide is a major credibility bet: first year-ahead revenue guide in company history, delivered with an explicit supply-constrained caveat and a demand-vs-supply gap (~100% vs 70%). If supply improves, this becomes a low bar; if memory/power worsens, it's exposed. Rasgon's $200B-uplift math went uncorrected.
- OpenAI: the agreement question was answered by disclosure, not by statement: no "definitive agreement" language, but the call disclosed OpenAI's ~12GW of committed NVIDIA compute through 2030, the 4.25GW SoftBank Portsmouth site for OpenAI, and ~$50B of frontier-lab investments. The substance arrived; the contractual status of the equity agreement itself remains undisclosed.
- Vera Rubin claims re-based for the third time (now 30x throughput/MW, 35x lower token cost vs GB Ultra). Each re-basing is favorable and none are independently validated; the moving baseline (Blackwell → GB Ultra) makes quarter-over-quarter claim tracking unreliable. However, the ~20%-of-DC-revenue-in-Q3 figure is the first falsifiable Rubin commitment — a credibility positive.
- DSO explanation changed character: from "collection timing" (Q1, 45 days) to "extended payment terms for large multi-quarter purchases by investment-grade customers" (60 days). Extending terms to large customers while simultaneously providing take-or-pay guarantees, credit enhancement, and equity investments is a coherent but escalating pattern of using the balance sheet to support demand — disclosed, but worth tracking as a quality-of-revenue signal.
- "Circular financing" addressed head-on before any analyst asked: fungible/redeployable compute, investment-grade ultimate offtake, labs ~a quarter of next year's business. The pre-emption suggests management knows this is the emerging bear thesis.
- Persistent gaps: software monetization (8 quarters), Spectrum-X run rate (never given), price-vs-volume split of the FY28 guide, constraint rank-ordering (declined), OpenAI equity agreement status.
What changed versus the prior quarter- Guidance horizon extended dramatically: from one quarter ahead to a preliminary FY28 ~70% growth guide — unprecedented for this company, and framed as supply-capped with demand at ~100%.
- Gross margin framework broken: mid-70s reiterated for a year → explicit reset to 74% (Q3), 71–72% trough (Q4), 72–73% (FY28) on memory costs, with price increases taking effect Q1. The first margin downcycle in the AI buildout era.
- Vera Rubin went from "POs in hand, Q3 start" to shipping, with a quantified Q3 contribution (~20% of DC revenue) — the ramp is now measurable.
- China went from zero to <1%: first H200 shipments under licenses, disclosed as margin-dilutive, still excluded from outlook.
- Balance-sheet support for demand escalated and was disclosed: ~$50B lab investments, $500B third-party financing platforms, take-or-pay/revenue-share NeoCloud model, ~2GW credit enhancement for an unnamed lab, 4.25GW SoftBank site for OpenAI. Last quarter none of this structure was visible.
- Opex raised again (upper-40s → low-50s); tax reiterated.
- DSO deteriorated to 60 days with a new extended-terms explanation; inventory up to $32B for the Rubin launch.
- AWS relationship stepped up: from ">1M GPUs starting this year" to +2M GPUs through Q2 FY29 plus Vera CPUs, Nemotron distribution, and physical AI.
- Groq LPX tone shifted: from "niche, <20%" to full production, first customer (Nebius), volume this quarter — though still framed as a bolt-on.
- CPU story firmed: Grace >$5B TTM baseline disclosed; FY28 CPU revenue guided to >2x; named lead customers (OCI, SpaceX AI, AWS).
- Capital return language qualified: "increase and return excess FCF net of strategic uses" — new carve-out alongside the $50B lab-investment disclosure.
Bull case- $108B Q3 guide (+$12B q/q) ex-China, with Vera Rubin already shipping and quantified at ~20% of DC revenue in Q3 — the fastest-ramp claim is now testable and the early evidence (POs from every major customer, AWS 2M GPUs, Google XGS up to 960K Rubin) supports it.
- FY28 ~70% growth with demand at ~100% means the guide is supply-capped, not demand-capped — every incremental gigawatt of supply (memory, wafers, power/shell) is upside to a number management says it can already see. Hyperscale reacceleration is explicitly guided for Q4/FY28 on top of ACIE's 100% growth.
- Revenue per gigawatt expansion ($18B → $25B → $40B) is a structural ASP/mix tailwind independent of unit growth — and Jensen confirmed the trajectory continues beyond Vera Rubin. Combined with <1-year data-center ROIC claims, the customer's economic incentive to keep buying is quantified.
- The margin reset is fully scoped and has a recovery mechanism: trough identified (Q4, 71–72%), recovery path stated (price increases Q1, 72–73% FY28), and the cause (memory) is framed as a demand symptom. If memory stabilizes or price increases over-deliver, FY28 margins are a beat candidate.
- Demand breadth is now undeniable in the numbers: ACIE +138% y/y, sovereign >3x y/y, NeoCloud capacity 3GW → 8GW in a year, ~20 AI-native companies >$1B ARR, enterprise verticals at $15B TTM combined. Hyperscale is ~55% of DC and the rest is compounding faster.
- CPU is a real second business: $5B Grace TTM baseline, $20B Vera demand, >2x FY28 growth guided, named hyperscaler customers — incremental to GPU and excluded from nothing.
- Groq LPX is excluded from the $1T framework and is now in production with a first customer — defined upside per management's own accounting.
- Capital returns running ahead of pledge (60% of FCF YTD vs 50% commitment) even while funding $50B of ecosystem investment.
Bear case- The mid-70s margin era is over: 71–72% trough in Q4, 72–73% FY28 — a ~300bp structural step-down guided in one quarter, on an input cost (memory) management admits is "headed even higher into next year." If memory worsens beyond plan or price increases lag, the trough deepens; the reset's credibility rests on supplier relationships and unproven pass-through.
- NVIDIA is now financing its own demand at multiple layers: $50B equity in labs, take-or-pay guarantees and revenue shares with NeoClouds, ~2GW credit enhancement for an unnamed lab, extended payment terms to large customers (DSO 60 days), and $500B of third-party capital it helped organize. Management pre-emptively rebutted "circular financing" — which confirms the critique has traction. Labs supported by NVIDIA's balance sheet are ~a quarter of next year's business. If any frontier lab's economics crack, NVIDIA wears it on equity, guarantees, receivables, and revenue simultaneously.
- Opex guided up for the third straight quarter (low-40s → low-50s) while margins fall — operating leverage is compressing from both sides for the first time in the AI cycle.
- Earnings quality flags are accumulating: DSO 60 days on extended terms, inventory $32B ahead of a ramp, China shipments margin-dilutive, and the FY28 guide's price-vs-volume split undisclosed (Rasgon's question unanswered).
- Custom-silicon risk is now named and proximate: Arya cited OpenAI's "Jalapeno" claims vs Blackwell days before the call. Jensen's "100% confidence" answer is assertion, not evidence — and NVIDIA is simultaneously funding the two companies (OpenAI, Anthropic) most motivated to displace it.
- Vera Rubin claims have been re-based three times with shifting baselines and no third-party validation; the 30x/35x figures are unverifiable as stated. If MLPerf or customer benchmarks disappoint, the $40B/GW expansion thesis loses its anchor.
- Supply constraint cuts both ways: a 70% guide against 100% demand means ~30 points of demand is being turned away or deferred — to competitors, custom silicon, or later quarters — and management declined to rank which constraint (memory? wafers? power?) binds hardest, limiting investors' ability to handicap the gap closing.
- China remains structurally zero with dilutive licensed trickle; the competitor-risk language from two quarters ago remains unresolved.
Next-quarter watchlist- Q3 delivery vs $108B ±2%: the ACIE-led mix (hyperscale reacceleration is a Q4 promise — check whether it materializes); Vera Rubin vs the ~20%-of-DC-revenue commitment — the single most checkable claim on the call.
- Gross margin trajectory: Q3 vs 74% guide; confirmation of the Q4 71–72% trough; details on the Q1 price increases (magnitude, scope, customer acceptance); memory-cost updates and any supplier capacity announcements.
- FY28 guide mechanics: price vs volume split; whether the 70% supply ceiling moves; any update to the $1T framework (now apparently ~$200B light); constraint rank-ordering if memory worsens.
- Financing structures: first revenue-share income from NeoCloud take-or-pay deals; size/terms of the ~2GW credit enhancement and the unnamed lab; total ecosystem commitments vs Arya's ~$500B tally; any impairment or lab-distress signals; OpenAI equity agreement status (still undisclosed contractually).
- Working capital: DSO vs 60 days (does extended-terms become structural?); inventory vs $32B as Rubin ramps; FCF vs the new margin/opex reality; capital returns vs the "net of strategic uses" qualifier.
- Opex: tracking vs low-50s FY27 growth — a fourth raise would confirm systematic under-guiding.
- CPU and LPX conversion: first standalone Vera CPU revenue evidence vs the >2x FY28 guide; Groq 3 LPX volume shipments and Nebius deployment; any additional LPX customers.
- Custom silicon: any customer-visible traction for OpenAI Jalapeno or Anthropic silicon; whether the 12GW OpenAI commitment holds as stated.
- China: H200 revenue trajectory above/below 1%; margin dilution quantification; policy shifts.
- Events: Goldman Communacopia fireside (Sept 10), GTC Berlin keynote (Oct 21); Q3 FY27 earnings November 17, 2026.
|
| May 20, 2026 | -1.77% | Q1 FY2027 | Read transcript briefingQuarter in one view- Revenue $82B, +85% y/y, +20% q/q — a +$13.5B record sequential add, beating the $78B ±2% guide (~5% above midpoint). Third consecutive quarter of y/y acceleration; 14th straight sequential increase. Data Center $75B, +92% y/y, +21% q/q.
- Q2 FY27 guide: $91B ±2% (+11% q/q at midpoint), again assuming zero China DC compute revenue. Gross margin guided 74.9% GAAP / 75.0% non-GAAP ±50bps; FY27 "mid-70s" reiterated.
- The $500B framework became $1 trillion: Kress stated "full confidence in $1 trillion in Blackwell and Rubin revenue from 2025 through calendar 2027" — the first new dollar figure since the $500B was verbally superseded last quarter. Schneider confirmed this was first flagged at GTC and excludes Vera standalone CPUs, LPX, and CPX.
- Reporting segmentation overhaul: two market platforms — Data Center (split into Hyperscale $38B, ~50% of DC, +12% q/q and ACIE $37B, +31% q/q — AI clouds, industrial, enterprise) and Edge Computing ($6.4B, +29% y/y, +10% q/q). Nine quarters of recast history posted on the IR site.
- Vera CPU elevated to a headline growth driver: claimed $200B new TAM, ~$20B of standalone CPU revenue visibility this year, and a closing "top 5 things" slot. Vera Rubin production shipments confirmed for 2H26 starting Q3; now 7 chips again (was six last quarter, seven before that), 35x inference throughput, 10x AI factory revenue vs Blackwell.
- Capital allocation stepped up hard: record $20B returned in Q1; dividend raised from $0.01 to $0.25/quarter (Kress said $0.20 in prepared remarks; Jensen corrected her live to $0.25 — a notable on-air slip); new $80B buyback authorization on top of $39B remaining; commitment to return ~50% of FCF this year. This directly answers Arcuri's prior-quarter mega-buyback push.
- Record FCF $49B (vs $35B in Q4). Supply position (inventory + purchase commitments + prepaids) now $145B.
What management is focused on- "Compute is revenues, compute is profit" — the master framing continues, now extended: Jensen's close declares "demand has gone parabolic… agentic AI has arrived… tokens are now profitable." Kress added a new supporting data point: H100 rental pricing +20% YTD, A100 +~15% — used to argue GPUs generate profitable revenue beyond depreciable life.
- The new segmentation as a strategic story: the Hyperscale/ACIE split exists to show NVIDIA is more than hyperscaler capex. Jensen repeatedly steered questions (Moore, Reitzes, Rasgon) to the "second category" — AI-native clouds, enterprise, industrial, sovereign — claiming it will grow faster than hyperscale and that NVIDIA serves it "almost uniquely." Rasgon confirmed neoclouds sit in ACIE.
- Vera CPU as the next act: positioned as "the world's first CPU purpose-built for agentic AI" — agents' harnesses, orchestration, and tool use run on CPUs; "billions of agents" each effectively needing a PC. Four Vera use cases named: in VeraRubin, standalone, Vera+CX9 for storage, Vera+CX9 for security/confidential computing. Jensen: "we will be supply constrained throughout the entire life of VeraRubin."
- Inference share offense: Muse's inference-share question got an unambiguous "we are growing share in inference very, very quickly," anchored on the Anthropic partnership (coverage "largely zero until just recently") and new frontier labs (Cursor, Perplexity, TML, Reflection). Claim: "every single frontier model company will jump on VeraRubin from the get go."
- Frontier-model roster as proof of platform lock: OpenAI GPT-5.5 "co-designed for, trained with, and served on Blackwell"; AWS to add >1M Blackwell and Rubin GPUs starting this year plus Spectrum networking collaboration; Google Blackwell cloud instances with confidential computing; Google XGS bare-metal instances supporting up to 960K Rubin GPUs; Microsoft Fairwater live "ahead of schedule."
- LPX deliberately downplayed: Arcuri's custom-merchant question got a firm niche framing — low latency/high token rate but limited throughput and context; "a lot less than 20%" of the market today, "niche product for some time to come." This is a notable tempering of the Grok-derived product line teased last quarter.
Key numbers and quarter mechanics- Revenue $82B vs $78B ±2% guide; Q2 guide $91B ±2%, zero China DC compute assumed.
- Data Center $75B, +92% y/y: compute $60B, +77% y/y; networking $15B, ~3x y/y (InfiniBand >4x y/y on XDR; Spectrum-X claimed "larger than all Ethernet network peers combined"). Hyperscale $38B (+12% q/q); ACIE $37B (+31% q/q), with AI-cloud revenue >3x y/y.
- Gross margin 74.9% GAAP / 75.0% non-GAAP — flat q/q, in line with guide; Blackwell systems still most shipments.
- Opex +12% q/q (GAAP and non-GAAP) on compensation and compute/infrastructure. Q2 guided ~$8.5B GAAP / ~$8.3B non-GAAP. FY27 opex growth raised to "upper forties" % y/y (from "low 40s" last quarter) — driven by R&D and internal AI-tool usage.
- Tax: Q1 non-GAAP 16% (below outlook, geographic mix); FY27 guided 16–18%, lowered from 17–19% — resolving last quarter's garbled range.
- FCF $49B (record, vs $35B Q4). DSO 45 days on collection timing; guided back to mid-50s in Q2 — a small working-capital quality flag.
- Total supply (inventory + purchase commitments + prepaids) $145B — no prior-quarter comparator given on the call, but the framing shifted from "commitments into CY2027" to a single aggregate figure.
- Edge Computing $6.4B, +29% y/y, +10% q/q: Blackwell workstations strong; consumer demand fell modestly on higher memory and system prices — the gaming supply/memory warning from last quarter confirmed in effect, though gaming is no longer broken out.
- Physical AI >$9B LTM revenue (vs >$6B FY26 last quarter). Uber robotaxi partnership: ~30 cities, 4 continents by 2028.
- Sovereign revenue +80% y/y (vs "more than tripled to >$30B FY26" last quarter — growth rate decelerating off a larger base); NVIDIA AI deployed in ~40 countries. Partner data centers >10MW nearly doubled in a year to 80+ sites.
- Analyst hyperscaler capex forecasts now cited as >$1T by 2027 (vs "approaching $700B in 2026" last quarter); $3–4T annual AI infrastructure spend by end of decade reiterated.
Product and launch scorecard- Blackwell/GB300 — "fastest product ramp in company history": frontier model builders and hyperscalers each cumulatively deployed "hundreds of thousands" of Blackwell GPUs. MLPerf inference sweep; GB300 throughput +2.7x and cost/token -60% vs six months ago via full-stack optimization. Installed-base pricing (H100 +20%, A100 +15% YTD) is the new monetization evidence.
- Vera Rubin — on schedule, claims escalated: production shipments 2H26 starting Q3; Kress says POs in hand, "almost all major customers ready to go," ramp building through Q4 with Q1 FY28 "very big as well" — but she declined to say whether it ramps faster than GB300 ("hard to say," "little early"). Claims moved from "1/4 GPUs for MoE training, 10x lower token cost" to 35x inference throughput and 10x AI factory revenue vs Blackwell — a different (and larger) framing; not directly comparable to last quarter's metrics. Chip count back to seven without explanation (six last quarter).
- Vera CPU — the quarter's biggest new claim: $200B TAM, ~$20B standalone CPU revenue visibility this year (Arya confirmed this is standalone, excluding Vera-in-VeraRubin), "every major hyperscaler and system maker partnering." This is a forward visibility statement, not booked revenue — treat as a pipeline claim to be tested.
- Networking — $15B quarter, ~3x y/y: Spectrum-X now claimed larger than all Ethernet peers combined (last quarter Jensen said "probably the largest Ethernet networking company"); AWS Spectrum collaboration added to the prior NVLink/custom-silicon win. Still no Spectrum-X run-rate number.
- LPX — explicitly niched: low-latency decode, SRAM-based, limited context/throughput; <20% of market today. Schneider confirmed LPX/CPX/Vera standalone are all excluded from the $1T — i.e., framed as upside.
- Gaming/consumer — confirmed soft: folded into Edge; consumer demand down modestly on memory/system prices. No severity/duration update; the segment-level warning from last quarter is now harder to track under the new segmentation.
- Auto/robotics — narrative progress: Uber robotaxi fleet (~30 cities by 2028) is the first named large-scale auto deployment in the packet; physical AI >$9B LTM. No auto revenue line disclosed under the new structure.
- Software — still no monetization metric (seventh quarter).
Sell-side read-through- Only eight questions; the call ended with "no further questions" — unusually short. Segmentation and the new CPU story dominated; zero questions on China (third straight quarter), zero on gross margin sustainability, zero on gaming/memory, zero on the OpenAI agreement status (which was not mentioned at all — see credibility).
- Moore (Morgan Stanley) asked why resegment and where the "surprising CPU number" lands. Jensen gave a long taxonomy answer; the practical takeaway: the split is designed to de-emphasize hyperscaler-capex dependence. No competitive-dynamics detail per segment was offered.
- Reitzes (Melius) asked whether NVIDIA should grow faster than hyperscaler capex (DC ex-China +~120% y/y vs capex +90–100%). Jensen: "we should be growing faster than hyperscale CapEx," citing ACIE diversity and Anthropic share gains — an explicit endorsement of outperformance, with the burden of proof shifted to the second segment.
- Muse (Cantor) on inference share into late '26/'27: got the most direct share claim of the call ("growing share very, very quickly") plus "every single frontier model company will jump on VeraRubin from the get go" — a checkable, aggressive prediction.
- Arcuri (UBS) on LPX/CPX traction: Jensen capped expectations — niche, <20% today. Notably, Arcuri did not follow up on capital returns since the $80B buyback and 50%-of-FCF pledge preempted his prior push.
- Arya (BofA) pinned down the $20B CPU figure as standalone Vera and asked cannibalization vs GPU. Jensen's answer: CPUs run agent harnesses/orchestration, GPUs run thinking — incremental, not cannibalistic. This was the most analytically useful exchange of the call.
- Rasgon (Bernstein) confirmed neoclouds are in ACIE and got Jensen to state ACIE should grow faster than hyperscale near-term and be larger long-term — the clearest forward mix signal given.
- Schneider (Goldman) extracted the $1T exclusions: upside ranked as (1) frontier-model share gains, (2) Vera standalone CPU ("second largest"), (3) LPX. This defines the bridge above $1T.
- Buchalter (TD Cowen) asked the key execution question — Rubin ramp slope vs GB300. Kress: Q3 start, ramp through Q4, big Q1 FY28, but refused to commit to a faster-or-slower slope — the most hedged answer of the call.
Management credibility- Beat-and-raise streak extended: $82B vs $78B guide; Q2 guided $91B (+$9B q/q) — fifth consecutive ~$9–13.5B sequential add, all ex-China.
- The $1T figure replaces the verbally-superseded $500B — a concrete, checkable number delivered one quarter after the ceiling-to-floor language shift. Credibility positive, though it bundles two years (2025–CY2027) and multiple platforms, making quarterly verification indirect.
- Margin delivery continues: 75.0% non-GAAP vs 75.0% guide; FY27 mid-70s reiterated through the Rubin transition. No memory-cost mechanics were discussed — the Reitzes question from last quarter remains unanswered, and no one asked again.
- Rubin timeline held (2H26, now specified as Q3 start) and POs claimed in hand — but the performance claims were re-framed (from training-GPU-reduction/token-cost to 35x throughput/10x factory revenue) and the chip count moved back to seven without explanation. The metrics keep shifting in NVIDIA's favor; independent validation (MLPerf for Rubin) is still ahead.
- OpenAI agreement: not mentioned once — after two quarters of "working toward" and "we believe we are close," the call is silent on whether a definitive agreement was signed. GPT-5.5 is described as trained/served on Blackwell, but the investment/agreement status is undisclosed. This is now a three-quarter-old open item and its disappearance from the script is itself a signal.
- Dividend slip: Kress stated $0.20 in prepared remarks; Jensen corrected to $0.25 live. Minor, but a prepared-remarks error on a headline capital-return number.
- FY27 opex guide raised (low-40s → upper-40s % growth) one quarter after being set — the framework delivered last quarter was revised upward almost immediately, partly attributed to internal AI-tool spend.
- Segmentation change mid-surge: recast history provided (9 quarters posted), which is good practice, but the change also retires the gaming/auto/proviz lines just as gaming warned and auto stalled — reducing visibility into the weakest segments. Skeptics will note the timing.
- Persistent gaps: software monetization (seven quarters), Spectrum-X run rate (never given), ecosystem-investment criteria (never given), OpenAI agreement status (now unaddressed).
What changed versus the prior quarter- Demand framework quantified upward: $500B "exceeded" language → $1T Blackwell+Rubin revenue 2025–CY2027, with named exclusions (Vera standalone, LPX, CPX) defining upside.
- Capital allocation flipped from defense to offense: from declining Arcuri's mega-buyback to a $80B new authorization, 25x dividend increase, and a ~50%-of-FCF return pledge — the largest capital-return escalation in the packet's history.
- Vera CPU went from a GTC-deferred teaser to a headline pillar: $200B TAM, ~$20B standalone revenue visibility this year, four use cases, "supply constrained throughout VeraRubin's life."
- Segmentation replaced: gaming, proviz, auto lines retired into Edge Computing; DC split into Hyperscale/ACIE. ACIE (+31% q/q) is now growing much faster than Hyperscale (+12% q/q) — the diversification thesis is now structurally embedded in reporting.
- China unchanged in substance, quieter in tone: H200 still zero revenue, imports uncertain, excluded from outlook — but the prior quarter's "Chinese competitors could disrupt the global AI industry" language was not repeated.
- Opex trajectory raised (low-40s → upper-40s % FY27 growth); tax lowered (17–19% → 16–18%).
- Rubin claims re-based: from 1/4-GPUs/10x-token-cost to 35x throughput/10x factory revenue; chip count back to seven; ramp timing specified (Q3 start, Q4 build, big Q1 FY28).
- Gaming warning confirmed indirectly: consumer demand fell on memory/system prices — but the segment is no longer separately reported.
- Physical AI scaled: >$6B FY26 → >$9B LTM; Uber robotaxi deal is the first concrete fleet deployment named.
- Supply disclosure consolidated: from inventory +8% q/q and extended commitments to a single $145B total-supply figure.
Bull case- $91B Q2 guide (+11% q/q) ex-China, with the sequential-add machine intact ($13.5B added in Q1) and a $1T multi-year revenue framework now stated with "full confidence" — plus defined upside (Vera CPU, LPX, share gains) explicitly excluded from it.
- Demand broadening is now structural in the numbers: ACIE at $37B and +31% q/q, AI-cloud revenue >3x y/y, sovereign +80% y/y, 80+ partner AI factories >10MW. Hyperscale is only ~50% of DC and the other half is growing more than twice as fast sequentially.
- Vera CPU is a genuinely new revenue vector: ~$20B standalone visibility this year against a claimed $200B TAM NVIDIA has "never addressed" — if real, this is incremental to everything previously modeled, and Arya's exchange confirmed it is not cannibalizing GPUs.
- Rubin de-risked further: POs in hand, nearly all major customers committed, Q3 production start, Google XGS publicly sized at up to 960K Rubin GPUs, AWS >1M Blackwell+Rubin GPUs. Jensen's "every frontier model company will jump on VeraRubin from the get go" is aggressive but checkable within two quarters.
- Inference economics keep improving: GB300 +2.7x throughput / -60% cost per token in six months; H100/A100 rental pricing rising — evidence the installed base appreciates rather than strands, directly supporting the "compute is revenue" thesis.
- Capital returns now match the cash machine: $49B quarterly FCF, $80B new buyback, 25x dividend, ~50% FCF return pledge — the bear case on cash deployment discipline weakened materially.
- Anthropic is converting: from $10B investment disclosure to active multi-cloud capacity build-out (AWS, Azure, CoreWeave, xAI) with coverage "largely zero until recently" — a pure share-gain vector in inference.
Bear case- The OpenAI agreement vanished from the script — unmentioned after two quarters of "close." If the largest named demand pillar's definitive agreement slipped or changed terms, the silence is the tell. No analyst asked.
- Opex is now guided to upper-40s % growth one quarter after the low-40s framework was set — operating leverage is being deliberately spent (R&D + internal AI tools), and the framework's credibility is already dented.
- The segmentation change reduces visibility exactly where weakness lives: gaming (supply-constrained, consumer demand now falling on memory prices) and auto (stalled last quarter) are buried in Edge Computing. Investors can no longer track the two soft segments independently.
- Memory pressure is now visible in results (consumer demand down on "higher memory and system prices") while margin mechanics for memory cost remain undisclosed — mid-70s is reiterated but unexplained against a known rising input cost.
- The $20B Vera CPU figure is "visibility," not orders disclosed — and Jensen simultaneously claims supply constraint "throughout the entire life of VeraRubin," which cuts both ways: it caps how much of the $20B can convert and signals allocation risk across four Vera use cases.
- Rubin claims keep changing shape: 1/4 GPUs → 35x throughput; 10x token cost → 10x factory revenue; seven chips → six → seven. Each version is favorable, but the moving definitions make independent verification hard, and Kress would not commit to the ramp slope vs GB300.
- China remains zero-revenue despite licenses — two quarters of approvals with no imports; the structural-competitor risk flagged last quarter is unresolved even if unmentioned.
- DSO at 45 days guided back to mid-50s — Q1's cash conversion flattered by collection timing; Q2 FCF will not repeat $49B mechanically.
- The entire thesis still rests on token economics holding: "tokens are now profitable" is asserted with rental-price evidence, but the flat-capex contingency question (Arya, two quarters ago) has never been answered.
Next-quarter watchlist- Q2 delivery vs $91B ±2%: whether the +$9B sequential add lands; first full quarter under new segmentation — Hyperscale vs ACIE growth rates (ACIE guided implicitly to keep outpacing); any update to the $1T framework or its exclusions.
- OpenAI definitive agreement: signed, restructured, or a fourth quarter of silence — plus any disclosure of investment size/conditions. The most important open item in the packet.
- Vera Rubin ramp: Q3 production start confirmation, early customer deployments (Google XGS, AWS), whether Kress's Q4-build/big-Q1-FY28 shape holds, and third-party validation of the 35x/10x claims.
- Vera CPU revenue evidence: first standalone Vera revenue disclosure, customer names beyond "every major hyperscaler and system maker," and whether the ~$20B visibility converts on schedule.
- Gross margin through the transition: Q2 vs 75.0% guide; any memory-cost pass-through language; whether mid-70s FY27 survives early Rubin yields.
- Opex and tax: tracking vs the new upper-40s opex framework and 16–18% tax; whether the opex guide moves again.
- China: any H200 import clearance or first revenue; any return of the Chinese-competitor language; policy changes.
- Supply and balance sheet: the $145B total-supply figure's trajectory; inventory vs commitments mix; DSO normalization to mid-50s; FCF vs the $49B timing-flattered print.
- Capital returns: pace of the $80B authorization; whether ~50%-of-FCF return holds; dividend review cadence.
- Edge/consumer: whether memory pricing worsens the consumer decline; any auto/robotaxi revenue evidence from the Uber deployment; physical AI trajectory vs the >$9B LTM base.
- Events: TD Cowen TMT (May 28), Computex/GTC Taipei keynote (June 1), BofA Global Technology (June 4); Q2 FY27 earnings August 26, 2026.
|
| Feb 25, 2026 | -5.46% | Q4 FY2026 | Read transcript briefingQuarter in one view- Revenue $68B, +73% y/y (accelerating from Q3's +62%), a record +$11B sequential in data center — beating the $65B ±2% guide (~4.6% above midpoint). Data Center $62B, +75% y/y, +22% q/q; FY26 DC revenue $194B, +68% y/y (~13x since FY23).
- Q1 FY27 guide: $78B ±2% (+15% q/q at midpoint), again assuming zero China DC compute revenue. Gross margin guided 74.9% GAAP / 75.0% non-GAAP ±50bps; FY27 margins "mid-70s."
- The $500B framework was raised in language, not in number: Kress said NVIDIA expects sequential growth through CY2026 "exceeding what was included in the $500 billion Blackwell and Rubin revenue opportunity," with inventory/supply commitments now covering shipments into calendar 2027. No updated dollar figure given.
- Gross margin 75.0% GAAP / 75.2% non-GAAP — the promised mid-70s exit delivered. Free cash flow $35B in Q4, $97B FY26; $41B (43% of FCF) returned via buybacks/dividends.
- Accounting change: starting Q1 FY27, stock-based comp is included in non-GAAP results — a structural comparability break; Q1 opex guide of ~$7.5B non-GAAP includes $1.9B SBC.
- China: small amounts of H200 approved by the USG, but zero revenue generated and "we do not know whether any imports will be allowed into China." Chinese competitors "bolstered by recent IPOs" flagged as a long-term structural threat — sharper language than last quarter.
What management is focused on- "Compute equals revenues" — Jensen's new master framing, repeated at least four times: tokens are monetized, inference performance per watt = CSP revenue per watt, therefore customer capex is self-funding. This is the direct answer to the capex-sustainability question (Arya) and the $3–4T TAM question (Schneider).
- Agentic AI inflection as the demand proof: Claude Code, Claude Cowork, OpenAI Codex named as having "achieved useful intelligence"; Anthropic revenue cited as "10x in a year" and "severely capacity constrained." Jensen dates the public inflection to "the last 2–3 months."
- Frontier-model partnerships as strategy: OpenAI GPT-5.3-Codex trained and inferencing on GB NVL72, partnership agreement "we believe we are close" (still not definitive); Meta deploying "millions of Blackwells and Rubin GPUs" plus NVIDIA CPUs and Spectrum-X (announced last week); Anthropic partnership + $10B investment (first time the Anthropic investment size is disclosed); Grok nonexclusive licensing deal for low-latency inference technology, Mellanox-style integration teased for GTC.
- Rubin: first Vera Rubin samples shipped to customers this week; production shipments on track for 2H 2026; quantified for the first time — trains MoE models with 1/4 the GPUs, up to 10x lower inference token cost vs Blackwell; six chips named (Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6).
- Customer diversification narrative: top-5 CSPs/hyperscalers now "a little over 50%" of DC revenue; sovereign AI more than tripled to >$30B in FY26 (Canada, France, Netherlands, Singapore, UK); non-hyperscaler growth explicitly confirmed as faster (Lipacis question).
- Capital allocation defense: ecosystem investment framed as the priority use of cash; buybacks continue but no large new authorization despite Arcuri's direct prod.
Key numbers and quarter mechanics- Revenue $68B vs $65B ±2% guide; Q1 guide $78B ±2%, zero China DC compute assumed.
- Data Center $62B, +75% y/y, +22% q/q; Grace Blackwell systems ~2/3 of DC revenue. ~9 gigawatts of Blackwell infrastructure deployed nearly a year after NVL72 launch.
- Networking $11B, >3.5x y/y, record; FY26 networking >$31B (>10x vs FY21 Mellanox year). NVLink 72 scale-up switches the y/y driver; scale-up and scale-out both grew double digits q/q. AWS NVLink integration with custom silicon announced — a landmark Fusion-style win.
- Gross margin 75.0% / 75.2% vs ~74.8%/75.0% guide; Q1 guided 74.9%/75.0%; FY27 "mid-70s" reiterated.
- Opex: GAAP +16% q/q, non-GAAP +21% q/q (new product intros, compute/infrastructure). Q1 guided ~$7.7B GAAP / ~$7.5B non-GAAP including $1.9B SBC — not comparable to prior non-GAAP opex due to the SBC inclusion. FY27 non-GAAP opex guided to grow "low 40s" % y/y — the FY27 opex framework refused last quarter was delivered.
- Tax: Q4 non-GAAP 15.4% (below 17% ±1% guide, one-time benefit); FY27 guided 17–19% (transcript renders "7% and 19%" — almost certainly 17–19%; treat exact range as garbled).
- Inventory +8% q/q (vs +32% in Q3 — the build decelerated); purchase commitments "increased significantly," now extending beyond the next several quarters into CY2027 — "further out in time than usual."
- Gaming $3.7B, +47% y/y (down from $4.3B q/q); supply constraints flagged as a Q1-and-beyond headwind — Kress: "very tight" for "a couple of quarters," FY27 y/y growth uncertain (memory implied via Rasgon's question).
- ProViz $1.3B, +159% y/y, +74% q/q — crossed $1B for the first time; RTX PRO 5000 Blackwell workstation launched.
- Auto $604M, +6% y/y — growth decelerated sharply from +32%; Alpamayo open VLA model portfolio launched, first car in Mercedes-Benz CLA "soon." Physical AI claimed >$6B FY26 revenue.
- Top-5 hyperscaler 2026 capex expectations now approaching $700B, +~$120B since start of year (vs ~$600B / +$200B cited last quarter — note the base shifted; treat as analyst-expectation figures, not NVIDIA commitments).
- R&D budget "approaching $20 billion annual."
Product and launch scorecard- Blackwell/GB300 — still the engine: 50x perf/watt and 35x lower cost/token vs Hopper (InferenceX/SemiAnalysis "Inference King"); 5x GB200 improvement in 4 months via CUDA optimization. Even Hopper and 6-year-old Ampere are "sold out" in cloud with pricing rising — a striking installed-base data point.
- Rubin — first quantified claims, on schedule: 1/4 GPUs for MoE training, 10x lower token cost, samples shipped this week, 2H26 production, "every cloud model builder" expected to deploy. This answers the quantified-uplift question open since Q2 FY26. Note chip count is now six (last quarter said seven) — the count moved again without explanation.
- Networking — the standout: $11B quarter, AWS NVLink/custom-silicon integration, Spectrum-6 102T teased; Jensen claimed NVIDIA is "probably the largest Ethernet networking company in the world." Sur's Spectrum-X run-rate question (~$11–12B annualized estimate) got no number — deflected into platform narrative.
- Gaming — first negative forward signal in the packet: supply-constrained into Q1 and beyond; FY27 growth explicitly uncertain. Memory cost/availability is the implied cause.
- ProViz — breakout: $1.3B, +159% y/y; the "multibillion-dollar product line" thesis now has a $5B+ annualized run rate.
- Auto — stalled: +6% y/y, roughly flat q/q for a third quarter; the old ~$5B FY26 vertical claim is now conclusively dead (FY26 ended at ~$2.4B implied run rate). Alpamayo/Mercedes CLA is the new hope.
- Vera CPU standalone — Rakers asked about the push to sell Vera CPUs discretely; Jensen positioned Vera as a post-training/data-processing CPU (LPDDR5, single-thread performance), details deferred to GTC.
- Grok licensing — low-latency decode technology, team joining NVIDIA; positioned as Mellanox-like architectural extension. No financial terms.
- Software — still no monetization metric (sixth quarter).
Sell-side read-through- Eleven questions; capex durability, margins, and capital deployment dominated. Again zero questions on China despite the H200-approval-but-zero-revenue disclosure — two straight quarters of silence on what was once the top topic.
- Arya (BofA) asked the cycle's central question — can customers keep growing capex into 2027 given compressed cash flows, and can NVIDIA grow if they don't? Jensen answered with conviction ("I am confident in their cash flow growing") but only via the compute-equals-revenues logic — no contingency answer for the flat-capex scenario.
- Moore (Morgan Stanley) on ecosystem investments (Anthropic, OpenAI, CoreWeave, Intel, Nokia, Synopsys): Jensen gave philosophy ("expanding and deepening ecosystem reach") — still no criteria, caps, or sizing framework, though the $10B Anthropic number emerged in prepared remarks.
- Sur (JPMorgan) pressed for a Spectrum-X run-rate number — refused; Jensen instead claimed Ethernet leadership. A notable non-answer on a directly modelable metric.
- Muse (Cantor) on custom silicon/CPX/Grok and chiplet strategy: Jensen defended monolithic reticle-limited dies, deferred Grok details to GTC. No roadmap specificity.
- Rasgon (Bernstein) got two answers: (1) Blackwell and Rubin will co-ship through the transition — no clean crossover promised, Rubin ramp contribution "too early to determine"; (2) gaming FY27 growth explicitly uncertain — the most hedged answer of the call.
- Reitzes (Melius) on mid-70s margin sustainability into CY27: Jensen's answer — generational perf/watt leadership is "the single most important lever" — is a strategy statement, not a commitment; no memory-cost mitigation specifics despite the direct prompt.
- Arcuri (UBS) noted the stock "hasn't really gone up much" and pushed for a mega-buyback — Kress declined, prioritizing ecosystem/supply investment. Management is choosing ecosystem deployment over shareholder-return escalation at ~$100B annual FCF.
- Schneider (Goldman) on the $3–4T 2030 TAM: Jensen reaffirmed ("the world needs a lot more than $700B") and named the next inflection as physical AI after agentic.
- Chkaiban (New Street) on space data centers — Jensen engaged seriously ("economics are poor today"); notable only as a sign of where the narrative frontier has moved.
- Notable absences: no questions on the SBC accounting change's comparability impact, the gaming supply warning, China/H200, or the OpenAI agreement status.
Management credibility- Beat-and-raise streak extended: $68B vs $65B guide; Q1 guided $78B (+$10B q/q) — the fourth consecutive quarter of ~$8–11B sequential adds, all ex-China.
- Margin promise fully kept: mid-70s exit delivered at 75.2% non-GAAP; FY27 "mid-70s" reiterated. But Reitzes' sustainability question got philosophy, not mechanics — the memory-cost hedge from last quarter stands unresolved.
- The $500B claim was managed upward verbally ("exceeding what was included") without a new number — credibility-enhancing if Q1–Q2 deliver, but it also quietly converts a ceiling into a floor without a checkable figure.
- Rubin quantification finally delivered (1/4 GPUs, 10x token cost) after two quarters of refusal — a credibility positive; samples shipped on the promised schedule.
- OpenAI agreement slipped again: from "working toward a definitive agreement" (Q3) to "we believe we are close" (Q4) — still not signed after a full quarter. This is now a two-quarter-old commitment.
- FY27 opex framework delivered (low-40s % growth) after last quarter's refusal — responsive to Rasgon's prior challenge.
- New credibility risks: gaming supply warning landed with no prior hint (Q3 gaming commentary was benign); the SBC-in-non-GAAP change, while praised by Reitzes, arrives with no restated history on the call; the Rubin chip count moved from seven to six without explanation; the FY27 tax range is garbled in the transcript.
- Persistent gaps: software monetization (six quarters), Spectrum-X run rate (refused), investment criteria (refused twice running), auto $5B claim (never addressed, now moot).
What changed versus the prior quarter- The $500B anchor was superseded in language: growth through CY2026 will "exceed" it, and visibility now extends into CY2027 via purchase commitments — the demand horizon formally lengthened.
- Anthropic quantified: $10B investment disclosed (previously undisclosed); Meta added as a "millions of GPUs" Blackwell+Rubin+CPU+Spectrum-X customer — the two largest new demand disclosures of the quarter.
- China worsened rhetorically: from "POs never materialized" to H200 approved but zero revenue and unknown import permission, plus explicit acknowledgment that IPO-funded Chinese competitors could "disrupt the structure of the global AI industry." The competitive-loss framing hardened.
- Gaming flipped from stable to constrained: +47% y/y quarter but a forward supply warning — the first non-DC segment caution in the packet's history.
- Auto decelerated to +6% y/y; the narrative shifted entirely to Alpamayo/robotaxi optionality.
- SBC enters non-GAAP — a permanent reporting change that raises reported non-GAAP opex by ~$1.9B/quarter and breaks y/y comparability.
- FY27 frameworks arrived: opex low-40s growth, tax 17–19%, margins mid-70s — last quarter's refusals were answered.
- Rubin went from "x-factor" to quantified (1/4 GPUs, 10x token cost) and from "silicon back" to samples shipped.
- Networking stepped up again: $8.2B → $11B; AWS custom-silicon NVLink integration is the biggest Fusion-ecosystem win yet.
- Inventory build decelerated (+8% vs +32%) while commitments extended further out — the balance-sheet posture shifted from building stock to locking CY2027 capacity.
Bull case- $78B Q1 guide (+15% q/q) with zero China — the sequential-add machine is accelerating, not decelerating, and management now claims visibility into CY2027 with supply secured to match.
- Demand claims broadened and hardened: Meta "millions of GPUs," Anthropic $10B + capacity-constrained 10x revenue growth, OpenAI "close," 9GW of Blackwell deployed, sovereign >$30B and tripled, top-5 capex expectations approaching $700B. Non-hyperscaler customers growing faster than hyperscalers — concentration risk is diluting.
- Rubin is de-risked and quantified: samples shipped on schedule, 2H26 production, 10x token-cost improvement — the annual cadence now has three generations of execution evidence, and Blackwell+Rubin co-shipment smooths the transition risk.
- Networking is a $44B annualized business growing >3.5x, with AWS adopting NVLink for its own custom silicon — the strongest possible anti-ASIC/ecosystem-lock-in signal.
- Margins at 75%+ with FY27 mid-70s reiterated, and the margin defense (generational perf/watt leadership) is backed by 50x/35x inference benchmarks.
- $97B FY26 FCF funds both the ecosystem-investment strategy and 43% capital return without leverage.
- Agentic AI is presented as a step-function in token demand (agents running minutes-to-hours, spawning sub-agents) — if true, inference demand is structurally underappreciated, and even 6-year-old GPUs being sold out supports it.
Bear case- The entire forward story now rests on "compute equals revenues" — a circular claim (customers' capex is safe because it generates revenue, which funds more capex) that was asserted, not evidenced, when Arya pressed on compressed hyperscaler cash flows. No answer exists for the flat-capex scenario.
- China is now a stated structural threat: management itself said IPO-funded Chinese competitors could disrupt the global AI industry — the first time the China risk was framed as global, not just domestic. Zero H200 revenue despite approvals shows policy movement doesn't convert.
- Gaming supply constraints and memory tightness are the first concrete signs that input markets (memory) are biting — the same input-cost pressure management declined to quantify for margins.
- OpenAI remains unsigned after two quarters of "close"/"working toward" — the largest single named demand pillar lacks a definitive agreement, and the Grok licensing deal suggests NVIDIA is also hedging its inference technology position.
- Ecosystem investment has no disclosed discipline: $10B into Anthropic, stakes across CoreWeave/Intel/Nokia/Synopsys, no criteria or caps after two quarters of direct questions — and Arcuri's buyback push was declined, meaning cash is preferentially going into the same customers who drive the revenue.
- Auto is confirmed ex-growth (+6% y/y) and gaming is guided to constraint — the non-DC segments offer no cushion if DC ever pauses.
- The SBC change obscures comparability at exactly the moment opex is guided to grow low-40s % — reported non-GAAP profitability will look different from the trend investors have modeled.
- Purchase commitments extending into CY2027 mean NVIDIA is long capacity further out than ever — if the agentic-inflection thesis is early or wrong, the balance sheet owns the overbuild.
Next-quarter watchlist- Q1 delivery vs $78B ±2%: whether the +$10B sequential add lands; Blackwell/Rubin co-shipment mix; any updated dollar figure replacing the now-"exceeded" $500B.
- OpenAI definitive agreement: signed or a third quarter of "close" — plus size, structure, and conditions of the investment.
- Rubin ramp specifics: 2H26 production timing (early vs late), first customer deployments, whether the 10x token-cost claim gets third-party validation; GTC (March 16) is the venue.
- Gaming supply: severity and duration of the constraint; whether memory costs are quantified; FY27 gaming growth framework.
- Gross margin: Q1 vs 75.0% guide; any memory-cost pass-through language; whether "mid-70s" FY27 survives the Rubin transition's early-yield phase.
- SBC restatement: whether NVIDIA provides recast historical non-GAAP figures; how the Street normalizes the ~$1.9B/quarter add-back removal.
- China: any H200 import clearance and first revenue; further language on Chinese competitor progress; any policy shift under the new export framework.
- Purchase commitments/inventory: 10-K detail on the CY2027-extended commitments; whether inventory growth stays at +8% or re-accelerates.
- Anthropic and Meta ramps: evidence of the $10B/1GW and "millions of GPUs" commitments converting to revenue; networking attach on these builds.
- Grok integration: what the low-latency decode technology becomes at GTC; any revenue or product timeline.
- Capital allocation: buyback pace vs ecosystem investment; any response to Arcuri's public mega-buyback push.
- Events: Morgan Stanley TMT fireside (March 4); GTC keynote (March 16); Q1 FY27 earnings May 20, 2026.
|
| Nov 19, 2025 | -3.15% | Q3 FY2026 | Read transcript briefingQuarter in one view- Revenue $57B, +62% y/y, +$10B / +22% q/q — a record sequential dollar increase, landing well above the $54B ±2% guide (~5.6% above midpoint). Data Center $51B, +66% y/y; compute +56% y/y on the GB300 ramp; networking $8.2B, +162% y/y ("more than doubled" sequentially per Kress).
- Q4 FY26 guide: $65B ±2% (+14% q/q at midpoint), again assuming zero data-center compute revenue from China. Gross margin guided to roughly 74.8% GAAP / 75.0% non-GAAP ±50bps (transcript renders this as "74.875% respectively" — garbled; treat the exact split as unconfirmed, but the guide is clearly ~75% non-GAAP, i.e., the promised mid-70s exit).
- New headline disclosure: "visibility to a half a trillion dollars in Blackwell and Rubin revenue from the start of this year through the end of calendar year 2026." Kress confirmed the $500B framework is on track, ~$50B shipped this quarter, and said the number "will grow" — citing the same-day KSA/Humane deal (400–600K GPUs over three years) and Anthropic as additive.
- China went from conditional reopening to dead stop: H20 sales ~$50M in Q3 (vs the $2–5B upside scenario floated last quarter). "Sizable purchase orders never materialized due to geopolitical issues and the increasingly competitive market in China."
- Gross margin 73.4% GAAP / 73.6% non-GAAP, above the 73.3%/73.5% guide. Inventory +32% q/q; supply commitments +63% q/q — the balance sheet is being levered hard into the ramp.
- Hopper, in its 13th quarter, still did ~$2B (down sharply from Rasgon's ~$6–7B estimate for Q2 — the fade finally arrived).
What management is focused on- Killing the "AI bubble" narrative: Jensen opened his remarks with a direct rebuttal — three simultaneous platform shifts (CPU→GPU accelerated computing, classical ML→generative AI, agentic/physical AI), each independently driving infrastructure spend. Meta's GEM model cited as proof of revenue ROI (+5% Instagram ad conversions, +3% Facebook feed).
- The $500B Blackwell+Rubin revenue visibility through CY2026 as the new anchor metric — replacing the vaguer $3–4T-by-2030 TAM as the near-term demand proof point (though the $3–4T claim was repeated).
- Ecosystem/equity investments as strategy: OpenAI (10GW partnership, "working toward a definitive agreement," investment opportunity), Anthropic (first-time NVIDIA adoption, ~1GW initial commitment on GB/Rubin, deep co-optimization), plus Mistral, Reflection, Thinking Machines, xAI. Jensen's framing: investments "expand the reach of CUDA" and "rather than giving up a share of our company, we get a share of theirs."
- Rubin execution: silicon back from supply chain, bring-up "beautifully," ramp in 2026, "seven chips," "x-factor improvement" vs Blackwell, third-gen rack-scale with redesigned manufacturability, drop-in compatible with Grace Blackwell infrastructure.
- Perf-per-watt = customer revenue, now with generational content escalation: Jensen gave content-per-gigawatt numbers for the first time — Hopper ~$20–25B/GW, Grace Blackwell ~$30B/GW ±, Rubin "higher than that."
- Supply chain pre-emption: first Blackwell wafer on US soil (TSMC), US manufacturing expansion with Foxconn/Wistron/Amkor/SPIL over four years; Jensen emphasized the balance sheet as a supply-securing weapon ("when we make purchases, our suppliers can take it to the bank").
Key numbers and quarter mechanics- Revenue $57B vs $54B ±2% guide. Q4 guide $65B ±2%, zero China DC compute assumed.
- Data Center $51B, +66% y/y; compute +56% y/y (GB300-driven); networking $8.2B, +162% y/y (NVLink scale-up onset plus double-digit growth in Spectrum-X and Quantum-X InfiniBand).
- GB300 crossed over GB200, ~2/3 of total Blackwell revenue; transition "seamless," shipping to all major CSPs/hyperscalers/GPU clouds.
- Hopper ~$2B in Q3 (13th quarter). H20 ~$50M — the $2–5B scenario from last quarter did not happen; POs "never materialized."
- Gross margin 73.4% / 73.6% vs 73.3%/73.5% guide; sequential improvement attributed to DC mix, cycle time, cost structure. FY27: "input costs are on the rise but we are working to hold gross margins in the mid-seventies."
- Opex: GAAP +8% q/q, non-GAAP +11% q/q (infrastructure compute, comp, engineering). Q4 guided ~$6.7B GAAP / ~$5.0B non-GAAP (vs ~$5.9B/~$4.2B in Q3 — a large step up).
- Tax: Q3 non-GAAP ~17% (above 16.5% guide, "strong US revenue"); Q4 guided 17% ±1%.
- Inventory +32% q/q (from $15B → ~$20B implied); supply commitments +63% q/q. Kress: "preparing for significant growth ahead."
- Gaming $4.3B, +30% y/y (flat q/q). ProViz $760M, +56% y/y, record, driven by DGX Spark. Auto $592M, +32% y/y (roughly flat q/q), Uber L4 fleet partnership on new Hyperion robotaxi reference architecture.
- Announced AI factory/infrastructure projects in the quarter totaling 5M GPUs; xAI Colossus 2 (first gigawatt-scale DC); AWS–Humane up to 150K accelerators incl. GB300; xAI–Humane 500MW flagship; KSA agreement 400–600K GPUs over 3 years.
- Top CSP/hyperscaler 2026 aggregate capex expectations now ~$600B, +$200B since start of year (analyst expectations cited by Kress).
Product and launch scorecard- GB300 — full crossover delivered: ~2/3 of Blackwell revenue, seamless transition claim repeated, now the growth engine. MLPerf training: Blackwell Ultra 5x faster time-to-train than Hopper, swept every benchmark, only platform running FP4 within accuracy standards. SemiAnalysis inference: best perf and lowest TCO across every model; 10x perf/watt and 10x lower cost/token vs H200 on DeepSeek R1.
- Rubin — silicon back, bring-up underway, 2026 ramp on track: now described as seven chips (prior quarter said six taped out — count discrepancy likely reflects platform scope; not explained). "X-factor" performance claim repeated but still no quantified uplift; CPX variant described for the first time in detail (long-context workloads, strong perf/$ and perf/W).
- Networking — third consecutive record, $8.2B (+162% y/y): NVLink scale-up onset is the new driver; Ethernet GPU attach rates "roughly on par with InfiniBand"; Meta, Microsoft, Oracle, xAI building gigawatt factories on Spectrum-X. NVLink Fusion momentum: Fujitsu (October), Intel custom DC/PC collaboration, Arm integrating NVLink IP (announced at SC25) — the Fusion ecosystem expanded from one design win to three named partners in a quarter.
- Gaming — $4.3B, +30% y/y but flat q/q: no new product news of note; 25-year GeForce anniversary color.
- ProViz — record $760M (+56% y/y) on DGX Spark ("world's smallest AI supercomputer," small Grace Blackwell config). Second straight strong quarter; the "multibillion-dollar product line" claim is gaining evidence.
- Auto — $592M, +32% y/y, flat q/q again: Uber L4 fleet partnership and Hyperion robotaxi reference architecture announced. The old ~$5B FY26 vertical claim remains unmentioned — third straight quarter of silence.
- Software — still no run-rate disclosure (fifth quarter). Dynamo now adopted by every major CSP (AWS, Google Cloud, Azure, OCI) with quantified inference gains claimed, but no monetization metric.
Sell-side read-through- Seven questions; demand durability, financing, and margins dominated. Nobody asked about China despite the H20 collapse to $50M — a notable absence given it was the top topic for three quarters.
- Moore (Morgan Stanley) pinned down the $500B math: Kress confirmed on track, ~$50B shipped this quarter, "the number will grow," with KSA (400–600K GPUs/3yrs) and Anthropic explicitly additive. This is the quarter's most important confirmation.
- Muse (Cantor) asked whether supply catches demand in 12–18 months: Jensen gave the three-transitions essay, no direct answer — no commitment on when supply/demand balances.
- Arya (BofA) extracted the content-per-gigawatt ladder (Hopper ~$20–25B → Blackwell ~$30B → Rubin higher) — new, modelable disclosure. His vendor-financing question was deflected: "financing is up to them," with Jensen arguing hyperscaler spend is "fully cash flow funded" and agentic AI is net-new revenue.
- Reitzes (Melius) asked the circularity question directly (cash use, ecosystem investment criteria): Jensen's answer — balance sheet secures supply, buybacks continue, investments expand CUDA reach and "we get a share of their company... I fully expect that investment to translate to extraordinary returns." No criteria, no caps, no sizing disclosed.
- Schneider (Goldman) on inference mix (~40% historically): Jensen declined to update the percentage — "hard to know exactly" — but said "our hope is that inference is a very large part of the market." First substantive CPX description.
- Arcuri (UBS) on the biggest bottleneck: Jensen listed power/land/shell/financing as "all constraints... all solvable," expressed most confidence in supply chain. No single binding constraint named.
- Rasgon (Bernstein) on FY27 margins and opex: Kress confirmed mid-70s is now a "working to hold" target against rising input costs (memory implied, not named), levers = cost improvements, cycle time, mix. No FY27 opex growth guide given — deflected to "investing in innovation."
- Rakers (Wells Fargo) on ASICs post-Anthropic: Jensen's five-point moat answer (every transition phase, every AI phase, every model, every cloud, offtake diversity). Notably conceded nothing on ASIC traction; the Anthropic win itself is the anti-ASIC evidence.
- Notable absences: no questions on China/H20, the 15%-to-USG construct (never mentioned on the call), inventory/supply-commitment spike, or the Q4 opex jump.
Management credibility- Guide beat again, and the Q3 ramp landed bigger than guided: $57B vs $54B ±2% — the +$7B sequential guide became +$10B actual. The ex-China compounding thesis continues to deliver.
- Mid-70s margin exit promise kept: "exit the year in the mid-seventies" (said early FY26) → 73.6% in Q3, ~75% guided for Q4. Kress explicitly claimed the win. But note the FY27 language is a hedge: "working to hold" mid-70s against rising input costs — the first margin-ceiling language, replacing the prior upward trajectory.
- China credibility took the hit this quarter: last quarter's "$2–5B shippable in Q3" scenario produced $50M. Management had flagged it as conditional, but the gap between the floated scenario and reality is large, and the 15%-to-USG framework has vanished from disclosure entirely — never mentioned on this call.
- $500B visibility claim is the new credibility anchor: introduced at GTC, reaffirmed with specifics ($50B shipped this quarter, more orders expected). This is checkable over the next five quarters and management is now staking the narrative on it.
- Rubin execution evidence is real so far: silicon back, bring-up reported smooth — consistent with last quarter's tape-out claims. Still no quantified performance uplift (Schneider's question from Q2 remains unanswered; CPX detail was the substitute).
- Persistent gaps: auto $5B claim absent a third quarter; software run rate absent a fifth; inference mix percentage refused; FY27 opex guide refused; ecosystem investment criteria/sizing undisclosed despite direct questioning (Reitzes). The pattern holds: specific where the story is good, opaque where it isn't.
- Hopper honesty: management gave a clean ~$2B Hopper number, implicitly confirming the long fade Rasgon modeled — after last quarter's "sold out" framing, the decline was disclosed plainly.
What changed versus the prior quarter- China: conditional reopening → effective zero. H20 went from "$2–5B shippable" to $50M actual; POs never materialized; the 15%-to-USG construct disappeared from discussion; Q4 again assumes zero China DC compute. Management now blames both geopolitics and "the increasingly competitive market in China" — the first acknowledgment of Chinese domestic competition as a factor.
- New demand anchor: the $500B Blackwell+Rubin visibility through CY2026 is now the central metric, with Kress confirming it's tracking and growing. Last quarter's framing was the $3–4T TAM; this quarter the near-term number got concrete.
- Anthropic flipped: from non-customer to ~1GW commitment + deep technical partnership + equity investment — the single largest competitive-development of the quarter (an Anthropic-scale win had been ASIC territory).
- OpenAI relationship formalized further: 10GW build partnership, investment opportunity, "working toward a definitive agreement" — still not definitive.
- Content-per-gigawatt disclosed for the first time: ~$20–25B (Hopper) → ~$30B (Blackwell) → higher (Rubin) — replaces last quarter's ~$35B-of-$50–60B-factory framing with a cleaner, rising ladder.
- Hopper finally rolled over: ~$2B vs Rasgon's ~$6–7B Q2 estimate; the "H100/H200 sold out" language is gone.
- Margin story pivoted from recovery to defense: mid-70s achieved; FY27 framed as "working to hold" against rising input costs — first mention of cost pressure as a margin risk.
- Balance sheet intensity stepped up: inventory +32% q/q, supply commitments +63% q/q (vs the $11B→$15B build last quarter). Q4 opex guided up ~$800M non-GAAP sequentially.
- NVLink Fusion went from one win to an ecosystem: Fujitsu, Intel, Arm all announced in one quarter.
- Networking mix shifted: NVLink scale-up called out as the doubling driver, alongside Spectrum-X/InfiniBand growth.
Bull case- $65B Q4 guide (+14% q/q) with zero China — the ex-China business is now compounding $8–10B per quarter, and China is again a free option rather than a dependency.
- $500B of visible Blackwell+Rubin revenue through CY2026, confirmed on track with ~$50B shipped this quarter and management saying the number will grow (KSA, Anthropic additive). This is the hardest forward-demand disclosure NVIDIA has ever given.
- The demand base broadened structurally this quarter: Anthropic (first-time, ~1GW), OpenAI (10GW path), KSA/Humane (400–600K GPUs), xAI Colossus 2, AWS–Humane 150K accelerators, 5M GPUs of announced projects. Hyperscaler 2026 capex expectations up $200B since January to ~$600B.
- The ASIC threat lost its best evidence: Anthropic — the marquee non-NVIDIA training workload — is coming onto CUDA. Jensen's five-point moat framing (every phase, every model, every cloud, offtake) now has a fresh proof point.
- Margins delivered to ~75% as promised, with FY27 "hold mid-70s" language implying no margin collapse despite input-cost inflation — pricing/mix levers intact.
- Rubin de-risked on schedule (silicon back, smooth bring-up, 2026 ramp, rising content/GW) — the annual cadence now has two clean transitions of evidence.
- Networking is a confirmed $30B+ run-rate engine growing triple digits, with NVLink Fusion converting Intel, Arm, and Fujitsu into the ecosystem.
Bear case- China is now a competitive-loss story, not just a policy story: "increasingly competitive market in China" is the first admission that even if policy opens, domestic alternatives may have taken share. The $50B China TAM claim from prior quarters looks impaired; the 15%-to-USG framework vanished without explanation.
- Circularity questions went unanswered: Reitzes asked directly about ecosystem investment criteria and got philosophy, not parameters. OpenAI's "definitive agreement" is still not definitive; the scale of equity commitments vs the $500B revenue they support is undisclosed. Bears will read vendor-financing risk into every gigawatt announcement.
- Balance sheet is absorbing the ramp risk: inventory +32% q/q and supply commitments +63% q/q into a demand environment management itself describes as sold out — if any leg pauses, NVIDIA now owns the pipeline. This is the largest pre-build in company history relative to revenue.
- Margin trajectory has peaked linguistically: "working to hold mid-seventies" against rising input costs (memory) is a ceiling, not a bridge upward — the first quarter without an improving margin narrative.
- Opex is stepping up sharply (Q4 guide ~$5.0B non-GAAP vs ~$4.2B in Q3) with no FY27 opex framework given despite Rasgon's direct question.
- The $500B claim concentrates the credibility bet: it spans five quarters and two product generations; any push-out (power, financing, customer digestion) now has a single public number to miss against.
- Disclosure gaps persist: auto $5B claim gone for three quarters, software run rate for five, inference mix refused, investment criteria refused.
Next-quarter watchlist- Q4 delivery vs $65B ±2%: whether the +$8B sequential ramp lands; Blackwell Ultra's share; whether Hopper's ~$2B holds or fades to zero.
- The $500B tracker: how much has shipped cumulatively by Q4; whether Kress's "the number will grow" produces a raised figure; KSA/Anthropic/OpenAI order conversion.
- OpenAI definitive agreement: signed or still "working toward" — and any disclosure of investment size, structure, or conditions.
- China: any H20 license movement, any Blackwell-in-China application, any codification of the 15% USG revenue share (and its margin treatment), and further language on Chinese domestic competition.
- Gross margin: Q4 vs ~75% non-GAAP guide; FY27 "hold mid-70s" — watch for memory-cost quantification and whether pricing is named as a lever.
- Opex: Q4 vs ~$5.0B non-GAAP; any FY27 opex growth framework after this quarter's refusal.
- Inventory and supply commitments: whether the +32%/+63% build converts to Q4/Q1 revenue; any purchase-commitment detail in the 10-Q.
- Rubin: ramp timing specificity (1H vs 2H 2026), quantified performance uplift (still owed since Q2), CPX positioning and customers.
- Anthropic ramp: pace of the ~1GW commitment; whether Anthropic workloads show up in networking attach.
- Auto: whether the abandoned ~$5B claim is ever addressed; Uber/Hyperion revenue timing.
- Software: any monetization metric after five quarters of silence.
- Events: UBS Global Tech & AI Conference (Dec 2); Q4 FY26 earnings February 25, 2026.
|
| Aug 27, 2025 | -0.79% | Q2 FY2026 | Read transcript briefingQuarter in one view- Revenue $46.7B, above the $45B ±2% guide (~2% above midpoint), with sequential growth across all platforms despite a $4B sequential decline in H20 revenue. Data Center +56% y/y and up sequentially; Blackwell platform revenue +17% q/q to record levels.
- Q3 FY26 guide: $54B ±2% — over $7B sequential growth, with zero H20-to-China assumed. Kress: if geopolitics resolve, $2–5B of H20 could ship in Q3 on top of that ("if we had more orders, we can bill more").
- China policy shifted from "closed" to "conditionally reopening": licenses began being granted in late July; a select number of China customers have licenses but no H20 has shipped yet; USG has expressed an expectation of 15% of licensed H20 revenue to the US government — not yet codified in regulation. Management is now openly advocating for Blackwell approval in China; Jensen called it "a real possibility."
- Gross margin: 72.4% GAAP / 72.7% non-GAAP, including a $180M (~40bps) benefit from releasing previously reserved H20 inventory; ex-benefit 72.3% non-GAAP, still above the 72.0% guide. Q3 guided 73.3% / 73.5% ±50bps; mid-70s exit reiterated.
- Networking record $7.3B, +46% q/q, +98% y/y; Spectrum-X now >$10B annualized (up from >$8B last quarter). Gaming record $4.3B (+14% q/q, +49% y/y). ProViz $601M (+32% y/y). Auto $586M (+69% y/y).
- $10B returned in Q2; board approved a new $60B buyback authorization (on top of $14.7B remaining).
- FY26 opex growth guide raised to high-30s% from mid-30s% — "accelerating investments."
What management is focused on- The $3–4 trillion TAM claim: Kress and Jensen both framed $3–4T of AI infrastructure spend by end of decade — a major escalation from the prior ~$1T-by-2028 compute framing (Reitzes pressed this directly; see sell-side). Jensen sized NVIDIA's content at ~$35B of a $50–60B gigawatt AI factory (~35%, "plus or minus").
- China re-engagement as the live policy battle: advocacy has shifted from lamenting the ban to actively lobbying for Blackwell in China. Jensen re-ran the argument (50% of world's AI researchers in China, Chinese open-source models — DeepSeek, Qwen, Kimi — fueling global enterprise adoption) and restated the $50B China TAM growing ~50%/year.
- GB300 ramp execution: production shipments began in Q2; factory builds converted in late July/early August; ~1,000 racks/week run rate restored, accelerating further in Q3; CoreWeave bringing GB300 instances to market with claimed 10x inference performance on reasoning models vs H100.
- Rubin de-risking narrative: all six Rubin-platform chips (Vera CPU, Rubin GPU, CX9 SuperNIC, NVLink 144 switch, Spectrum-X scale-out/scale-across, silicon photonics processor) taped out and in fab at TSMC; volume production next year; "third-generation NVLink rack-scale" with a "fully scaled-up supply chain."
- Perf-per-watt = revenue economics: repeated framing that power-limited data centers make NVIDIA's efficiency directly translate to customer revenue — $3M GB200 investment → $30M token revenue (10x); GB300 NVFP4 + NVLink 72 = 50x energy efficiency per token vs Hopper; NVFP4 training 7x faster than H100 (FP8).
- Sovereign AI quantified for the first time: >$20B sovereign revenue this year, more than double last year; EU €20B / 20 AI factories plan, UK Isambard (21 exaflops).
- Demand-is-sold-out messaging: Jensen closed the Arcuri question with "H100s sold out, H200s sold out," CSPs renting capacity from each other, AI-native startup funding $100B last year → $180B this year, AI-native revenue $2B → $20B.
Key numbers and quarter mechanics- Revenue $46.7B vs $45B ±2% guide. Q3 guide $54B ±2% (+$7B+ q/q), excluding any H20-to-China; upside scenario +$2–5B H20 if licenses/geopolitics resolve.
- H20 mechanics: ~$4B sequential H20 revenue decline absorbed; ~$650M of H20 sold to an unrestricted customer outside China; $180M reserve release added ~40bps to gross margin. China fell to low single digits % of DC revenue (from "meaningful decrease" guided last quarter).
- Hopper surprise: H100/H200 shipments increased sequentially — management framed it as breadth of workloads and CUDA value; Rasgon estimated Hopper still $6–7B/quarter (his math, not confirmed).
- Blackwell: +17% q/q; Rasgon estimated ~$27B in Q2 vs ~$23B in Q1 (analyst inference; Kress declined to confirm splits, said only Blackwell is "the lion's share" of Q3 growth across compute and networking).
- Networking $7.3B (+46% q/q, +98% y/y): Spectrum-X >$10B annualized (double-digit sequential and y/y growth); InfiniBand revenue nearly doubled sequentially (XDR adoption); NVLink "strong growth"; new Spectrum-XGS scale-across announced at Hot Chips (CoreWeave initial adopter, ~2x GPU-to-GPU communication claim).
- Gross margin 72.4% / 72.7% (72.3% ex-H20-release) vs 72.0% guide; Q3 guided 73.3% / 73.5% ±50bps; mid-70s year-end exit reiterated.
- Opex: Q3 guide ~$5.9B GAAP / ~$4.2B non-GAAP; FY26 opex growth raised to high-30s% (from mid-30s%). Transcript states GAAP opex "rose 86% on a non-GAAP basis sequentially" — internally garbled; treat the sequential opex increase as driven by compute/infrastructure and comp costs, exact figure unclear from transcript.
- Inventory up $11B → $15B sequentially to support Blackwell/Blackwell Ultra ramp — a large build worth monitoring.
- Other income ~$500M guided; tax 16.5% ±1%.
- Capital return: $10B in Q2 (down from $14.3B in Q1) plus new $60B authorization.
- Geography: Singapore 22% of billed revenue; >99% of DC compute billed to Singapore was for US-based customers — the defensive transshipment disclosure repeated.
- Sovereign: >$20B this year, >2x last year — first hard number on this leg.
Product and launch scorecard- GB300/Blackwell Ultra — delivered on the "later this quarter" promise from Q1: production shipments began in Q2, "tens of billions in revenue" claimed for the Blackwell Ultra platform in the quarter, transition "seamless" for CSPs (shared architecture/footprint), factories converted late July/early August, ~1,000 racks/week with further Q3 acceleration. The drop-in-compatibility de-risking thesis from last quarter is now evidenced.
- Rubin — on schedule, all six chips taped out/in fab, volume production next year. Jensen refused to detail Rubin performance uplift vs Blackwell when Schneider asked directly ("I cannot right now... save it for GTC") — annual cadence rationale given (cost reduction, customer revenue generation) but no quantified step-up.
- Networking — second consecutive delivered quarter, now the clearest second engine: $7.3B record, Spectrum-X >$10B annualized (from >$8B), InfiniBand ~2x sequentially, NVLink Fusion landed Fugaku NEXT (Fujitsu CPUs) — first named Fusion design win. Spectrum-XGS extends the story to data-center interconnect.
- Gaming — another record: $4.3B (+14% q/q, +49% y/y) on Blackwell GeForce ramp and improved supply; RTX 5060 desktop shipped; GeForce NOW Blackwell upgrade in September (RTX 5080-class, 5K/120fps, catalog doubling to 4,500+ titles); OpenAI GPT open-model optimization for RTX PCs.
- Auto — Thor shipping, but the $5B question still unanswered: $586M (+69% y/y, roughly flat q/q from $567M); Thor SoC shipments begun ("most successful robotics and AV computer we have ever created"); segment definition now explicitly "in-car compute only." The ~$5B FY26 vertical claim from two quarters ago was again not mentioned or reconciled — second straight quarter of silence.
- ProViz — tariff recovery delivered: $601M, +32% y/y (from $509M flat quarter); RTX Pro servers in full production, "90 companies" adopting (Hitachi, Lilly, Hyundai, Disney); management now calls RTX Pro a future "multibillion-dollar product line."
- Robotics/Thor edge platform: Jetson Thor available; adopters named (Agility, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic, Meta); 2M+ developers, 1,000+ partners; Siemens Omniverse expansion.
- Software — still no run-rate disclosure: fourth straight quarter without the $1.5B→$2B framework being updated. Only qualitative CUDA/Dynamo/TensorRT-LLM claims (Blackwell perf +2x since launch via software).
Sell-side read-through- Seven questions; China and TAM dominated. No one pressed the auto $5B claim, software run rate, inventory build, or the opex raise.
- Muse (Cantor) on 2026 growth and networking vs compute split: got Jensen's agentic-AI essay (100x–1000x compute vs one-shot) and the $3–4T framing; no quantified split.
- Arya (BofA) extracted the key China mechanics: licenses received, supply ready, $2–5B shippable this quarter, more can be built if orders/licenses arrive — but "still waiting on geopolitical issues," amount uncertain. His ASIC question (Broadcom's ~55–60% AI growth signal) drew Jensen's full-stack/ubiquity rebuttal: "a lot of projects are started... very few products go into production," perf-per-watt and perf-per-dollar as the structural answer. No direct engagement with merchant-share-loss scenarios.
- Reitzes (Melius) did the TAM math: prior ~$1T compute-by-2028 vs new $3–4T infrastructure — implying $2T+ compute. Jensen confirmed the framing: top-4 hyperscaler capex doubled to $600B/yr, US ~60% of world compute, NVIDIA ~$35B of a $50–60B gigawatt factory. Bottleneck answer: power is the limiter, which he reframed as NVIDIA's advantage.
- Moore (Morgan Stanley) on China long-term: Jensen restated $50B TAM growing ~50%/yr and said Blackwell-in-China is "a real possibility" — the most constructive China language since the ban.
- Rakers (Wells Fargo) on Spectrum-XGS sizing: Jensen's three-layer answer (scale-up NVLink / scale-out InfiniBand+Spectrum-X / scale-across XGS) with the utilization argument (65% → 85–90% "makes networking free," $10–20B effective benefit on a $50B factory). No XGS revenue sizing given.
- Rasgon (Bernstein) asked the sharpest modeling question — how to apportion the +$7B Q3 guide across Blackwell/Hopper/networking, with his own estimates (Blackwell ~$27B vs ~$23B; Hopper $6–7B). Kress declined to confirm any split, said only Blackwell is "the lion's share" and spans compute and networking. Hopper durability left open.
- Arcuri (UBS) pressed whether the 50% AI-market CAGR is a bogey for NVIDIA's DC growth next year: Jensen pointed to "very, very significant" customer forecasts for next year, sold-out H100/H200, CSPs renting from CSPs, startup funding/revenue ramp — no explicit endorsement or denial of the 50% bogey.
- Notable absences: no challenge on the $15B inventory build, the high-30s opex raise, the 15%-to-USG H20 mechanics (margin treatment, legality), auto's missing $5B claim, or software monetization.
Management credibility- Guide beat with the China hole fully absorbed: $46.7B vs $45B guide despite the $4B H20 sequential decline — the ex-China outperformance thesis from last quarter (Reitzes' math) played out. The $54B Q3 guide again excludes China, keeping the same conservative structure.
- GB300 promise kept precisely: "production shipments later this quarter" (said in Q1) → production shipments began in Q2, seamless transition, ~1,000 racks/week restored. The dual-ramp execution risk flagged two quarters ago is now retired.
- Margin trajectory on plan: 72.3% ex-benefit vs 72.0% guide; 73.5% guided for Q3; mid-70s exit reiterated. Note the $180M H20 reserve release flattered the headline by 40bps — disclosed transparently, but the underlying beat was smaller than the print.
- Networking credibility fully rebuilt: $5.0B → $7.3B, Spectrum-X annualized $8B → $10B; two consecutive delivered quarters after the two prior misses.
- TAM inflation is a new credibility risk: $1T-by-2028 compute → $3–4T-by-2030 infrastructure in one quarter. Jensen's bridge ($600B top-4 capex, $35B/GW content) is arithmetic, not evidence; Reitzes' question exposed that the number is a construct. Checkable only over years.
- China messaging whipsawed with policy: Q1 — "$50B TAM effectively closed," "end of the road for Hopper"; Q2 — licenses arriving, $2–5B shippable, Blackwell-in-China "a real possibility." Management didn't control the policy, but the prior quarter's terminal language aged poorly within 90 days. The 15% revenue-to-USG expectation is uncodified — management disclosed this plainly, which is to their credit, but it means the H20 recovery has an unpriced margin/legal overhang.
- Persistent disclosure gaps: auto $5B claim missing for a second quarter; software run rate missing for a fourth; Blackwell/Hopper revenue split refused under direct questioning (Rasgon); Rubin performance uplift refused (Schneider); next-year growth bogey deflected (Arcuri). The pattern: management is specific where the story is good, opaque where it isn't.
- Opex raise buried in the outlook section: FY26 opex growth moved mid-30s → high-30s% with one sentence of justification ("accelerating investments"). No analyst asked about it.
What changed versus the prior quarter- China: structural zero → conditional reopening. From "no compliant product, nothing to announce" to licenses granted, $2–5B Q3 upside scenario, ~$650M H20 sold to a non-China unrestricted customer, a $180M reserve release, and open advocacy for Blackwell in China. The 15%-to-USG construct is entirely new and uncodified.
- TAM language escalated an order of magnitude: ~$1T compute by 2028 → $3–4T AI infrastructure by end of decade, with a new per-gigawatt content claim (~$35B of $50–60B).
- GB300 went from sampling to production revenue ("tens of billions"), with the seamless-transition promise validated.
- Rubin went from roadmap to silicon: all six chips taped out, in fab — a year ahead of volume production.
- Hopper refused to die: H100/H200 shipments grew sequentially and Jensen says both are "sold out" — contradicts the "Hopper transition nearly complete" framing from Q1.
- Sovereign AI got its first hard number: >$20B this year, >2x y/y (previously named deals only).
- Networking stepped up again: $5.0B → $7.3B; Spectrum-X annualized $8B → $10B; first NVLink Fusion design win (Fugaku NEXT); new scale-across product (XGS).
- Opex trajectory raised: FY26 growth mid-30s% → high-30s%.
- Inventory built sharply: $11B → $15B.
- Capital return restructured: $14.3B → $10B quarterly, but a new $60B authorization signals sustained elevated buybacks.
- China revenue now de minimis: low single digits % of DC revenue (vs ~$4.6B H20 in Q1).
Bull case- $54B guide with zero China implies the ex-China business is compounding at ~$7B+/quarter; the $2–5B H20 scenario is pure upside to the guide, with supply already built and more buildable. China has shifted from a written-off $8B/quarter hole to a free option.
- GB300 ramp is executing exactly as promised — production shipments on time, seamless CSP transition, ~1,000 racks/week accelerating, "tens of billions" already in revenue. The annual cadence (Blackwell → Ultra → Rubin, all taped out) now has a delivery track record.
- Demand evidence is broadening beyond hyperscalers: sovereign >$20B (>2x y/y), enterprise RTX Pro in full production with 90 adopters, AI-native startup revenue $2B → $20B, H100/H200 sold out, CSPs renting capacity from each other. Four demand legs (hyperscaler, sovereign, enterprise, AI-native) are all described as growing simultaneously.
- Networking is a confirmed second engine at scale: $7.3B quarter, Spectrum-X >$10B annualized in ~18 months, InfiniBand doubling sequentially, NVLink Fusion converting the ASIC threat into ecosystem wins (Fugaku NEXT), and XGS opening the data-center-interconnect layer.
- Margin recovery on schedule (72.3% underlying → 73.5% guided → mid-70s exit) with the H20 charge now partially recaptured via reserve release; $60B buyback authorization signals management's confidence in cash generation.
- Perf-per-watt economics (50x per-token efficiency vs Hopper, 10x token-revenue return on GB200) give customers a revenue justification for capex, not just a capability one — management's answer to the digestion question.
Bear case- The China reopening is fragile and encumbered: licenses exist but nothing has shipped; the 15% revenue-to-USG expectation is uncodified — its legality, margin treatment, and Chinese-customer acceptance are all unknown. The $2–5B is a scenario, not a guide. Beijing's posture (customers "trying to determine their purchases") is a second gate.
- TAM claims are outrunning disclosure: $3–4T by 2030 and 50% China CAGR are assertion, not evidence; management refused to confirm Blackwell/Hopper splits, next-year growth bogeys, or Rubin uplift when pressed by Rasgon, Arcuri, and Schneider respectively. The market is being asked to underwrite numbers management won't decompose.
- Inventory up $4B sequentially to $15B alongside an opex guide raised to high-30s% — the cost structure is being built for the $3–4T world before that world is contracted. If any demand leg pauses, both lines work against margins simultaneously.
- Hopper's sequential growth cuts both ways: it evidences demand breadth, but it also means the "nearly complete" transition narrative was premature, and Rasgon's $6–7B/quarter Hopper estimate (unrefuted) is revenue that must eventually be converted or lost.
- Disclosure discipline continues to slip: auto's ~$5B FY26 claim now missing for two quarters; software run rate absent for four; the 15% H20 economics unexplained; Singapore at 22% of billings still requiring defensive transshipment disclosures.
- Policy dependence is now total in both directions: the upside (H20 licenses, Blackwell-in-China, Middle East/sovereign deals) and the downside (export rules, tariffs, the uncodified 15% levy) are all government-set variables outside management's control.
- Gross margin quality: the headline 72.7% included a 40bps one-time reserve release; underlying beat was modest, and the path to mid-70s still lacks a quantified bridge.
Next-quarter watchlist- Q3 delivery vs $54B ±2%: whether the +$7B sequential ramp lands; Blackwell's share of growth; whether Hopper strength persists or fades as Rasgon's question implied.
- H20 shipments: do any licensed sales actually ship? Does the $2–5B scenario materialize? Is the 15% USG revenue share codified, and how is it treated in revenue/margin? Any Chinese-government pushback on purchases.
- Blackwell-in-China: any license application, approval, or spec disclosure — Jensen called it "a real possibility"; watch for it becoming a plan.
- Gross margin: Q3 vs 73.5% non-GAAP ±50bps; whether mid-70s year-end exit holds; any further H20 reserve releases flattering the print; tariff cost disclosure.
- GB300 ramp: rack output beyond ~1,000/week; CoreWeave GB300 instance launch; MLPerf inference results (September) with Blackwell Ultra benchmarks.
- Rubin: any performance detail ahead of GTC DC (Oct 27–28, keynote Oct 28); supply-chain readiness vs the GB200 ramp's early friction.
- Networking: does $7.3B hold/grow; Spectrum-XGS early adopters beyond CoreWeave; NVLink Fusion pipeline beyond Fugaku NEXT.
- Inventory: whether the $15B build converts to revenue or becomes a risk line; any purchase-commitment disclosure.
- Opex: Q3 vs ~$4.2B non-GAAP; evidence for what the high-30s% FY26 growth is buying.
- Auto: whether the ~$5B FY26 vertical claim is ever reinstated or formally abandoned; Thor revenue ramp; segment now defined as in-car compute only.
- Software: any run-rate metric after four quarters of silence.
- Events: Goldman Sachs Tech Conference (Sept 8), GTC DC keynote (Oct 28), Q3 FY26 earnings November 19, 2025.
|
| May 28, 2025 | +3.25% | Q1 FY2026 | Read transcript briefingQuarter in one view- Revenue $44.0B, +69% y/y, +12% q/q — above the $43B ±2% guide (~2% above midpoint), despite a $2.5B H20 revenue hole from the April 9 export controls. Q2 FY26 guide: $45B ±2%, which explicitly absorbs an ~$8B H20 revenue loss — implying the ex-China business is running well ahead of prior expectations.
- The quarter's defining event: the H20 ban. New export controls (April 9, no grace period) forced a $4.5B charge (inventory + purchase obligations; lower than initially feared due to material reuse). Q1 still recognized $4.6B of H20 revenue pre-ban. Jensen: "the end of the road for Hopper" in China; the ~$50B China AI accelerator TAM is "effectively closed" with no compliant product today.
- Data Center $39.1B, +73% y/y. Blackwell now ~70% of DC compute revenue; Hopper transition "nearly complete." Hyperscalers each deploying ~1,000 NVL72 racks (~72,000 GPUs) per week.
- Gross margin: 60.5% GAAP / 61.0% non-GAAP as reported; ex-charge, non-GAAP would have been 71.3% — slightly above the 71.0% guide. Q2 guided to 71.8% / 72.0% ±50bps; mid-70s now framed as "late this year."
- Networking delivered on the third promise: +64% q/q to $5.0B. NVLink switch shipments >$1B in Q1; Spectrum-X annualizing >$8B, added Google Cloud and Meta.
- Gaming record $3.8B, +48% q/q, +42% y/y (guided recovery delivered, bigger than expected). Auto $567M, -1% q/q, +72% y/y. ProViz $509M, flat q/q (tariff uncertainty cited).
- Record $14.3B returned to shareholders (vs $8.1B in Q4).
What management is focused on- Reframing the China loss as a policy failure, not a demand problem: Jensen's prepared remarks are an extended argument against export controls — "shielding Chinese chipmakers from US competition only strengthens them," the assumption that China can't make AI chips is "clearly wrong," and the platform that wins China's developers wins globally. Notably political: "The president has a plan. He has a vision. I trust him."
- Declaring the inference inflection arrived: Microsoft processed >100 trillion tokens in Q1, 5x y/y; inference startups on B200 tripling token generation; reasoning models need "hundreds to thousands more tokens per task." GB200 NVL72 claimed 30x inference throughput vs 8-GPU H200 (MLPerf Llama 3.1) and ~40x vs Hopper (Jensen, Q&A).
- Sovereign AI as the new growth engine: Saudi Arabia 500MW project, UAE 5GW campus, Taiwan's first AI factory (Foxconn), Sweden national AI infrastructure; "nearly 100 AI factories in flight this quarter, 2x y/y" (Jensen also referenced "some 800 AI factories being built" — inconsistent numbers, see credibility). Line of sight to "tens of gigawatts" of projects.
- Onshoring narrative aligned with the administration: TSMC Arizona (6 fabs, volume production by year-end), Foxconn Houston (1M sq ft), Wistron Fort Worth, "chip to supercomputer built in America within a year," backed by "substantial long-term purchase commitments."
- Blackwell Ultra (GB300) transition teed up: sampling began earlier this month at major CSPs; production shipments "later this quarter"; drop-in compatibility with GB200 (same architecture, footprint, electrical/mechanical specs); B300 = +50% HBM, +50% dense FP4 inference vs B200.
- NVLink Fusion as the ASIC-co-option move: opening NVLink to semi-custom CPUs/accelerators with MediaTek, Alchip, Astera, Qualcomm, Fujitsu(?) ("Pizizzo" — transcript garbled) — turning the custom-silicon threat into ecosystem attachment.
Key numbers and quarter mechanics- Revenue $44.0B (+69% y/y) vs $43B ±2% guide. Q2 guide $45B ±2% with "modest sequential growth across all platforms"; DC Blackwell ramp "partially offset by a decline in China revenue."
- H20 bridge: $4.6B recognized in Q1 + $2.5B unshipped = $7.0B would-be Q1 China run rate; Q2 orders foregone ~$8B; $4.5B charge covers inventory and purchase commitments on future orders. Kress confirmed to Arya that the write-down captures the committed pipeline, but the ongoing loss is the ~$50B TAM with no product to serve it.
- Segments: DC $39.1B (+73% y/y); Gaming $3.8B (+48% q/q, +42% y/y); ProViz $509M (flat q/q, +19% y/y); Auto $567M (-1% q/q, +72% y/y).
- Blackwell ~70% of DC compute revenue — first mix disclosure since the $11B Q4 print; implies DC compute roughly $27B+ on Blackwell (inference from the 70% figure; exact compute split not given).
- Networking $5.0B, +64% q/q — the promised return to growth finally delivered. NVLink (scale-up) >$1B in its first disclosed quarter; Spectrum-X >$8B annualized; silicon photonics switches (Spectrum-X/Quantum-X Photonics) announced: 3.5x power efficiency, 10x resiliency, 1.3x faster time-to-market.
- Gross margin: 60.5% GAAP / 61.0% non-GAAP; 71.3% non-GAAP ex-charge (vs 71.0% guide). Q2 guide 71.8% / 72.0% ±50bps — "better Blackwell profitability" driving sequential improvement; mid-70s "late this year."
- Opex: +7% GAAP / +6% non-GAAP q/q. Q2 guide ~$5.7B GAAP / ~$4.0B non-GAAP; FY26 opex growth still mid-30s%.
- Q2 other items: other income ~$450M; tax 16.5% ±1%.
- $14.3B returned in Q1 (record; up from $8.1B in Q4).
- Geography: China % of DC revenue "slightly below expectations and down sequentially"; Q2 guided to a "meaningful decrease." Singapore ~20% of billed revenue, but >99% of H100/H200/Blackwell billed to Singapore was for US-based customers — a disclosure clearly aimed at transshipment scrutiny.
- Demand proof points: Microsoft >100T tokens Q1 (5x y/y); ~100 AI factories in flight (2x y/y, average GPUs per factory also doubling); Capital One 5x latency cut with Dynamo; Cisco 40% accuracy / 10x response with NIM; software optimizations improved Blackwell performance 1.5x in the last month alone; Hopper inference perf improved 4x over two years.
Product and launch scorecard- Blackwell — ramp compounding, transition from Hopper "nearly complete": ~70% of DC compute revenue; ~1,000 NVL72 racks/week per major hyperscaler; Microsoft "tens of thousands" of Blackwell GPUs deployed, ramping to "hundreds of thousands of GB200s" with OpenAI as key customer. Manufacturing yields "significantly improved." The ramp narrative is now fully about slope and margin, not feasibility.
- GB300 / Blackwell Ultra — on schedule, de-risked by design: sampling at major CSPs now, production shipments "later this quarter" (i.e., within fiscal Q2). Drop-in compatibility is the explicit answer to the dual-ramp execution concern from last quarter. B300: +50% HBM, +50% dense FP4 vs B200.
- Networking — the twice-broken promise finally kept, decisively: +64% q/q to $5B. NVLink scale-up is a new disclosed revenue line (>$1B); Spectrum-X added Google Cloud and Meta (joining CoreWeave, Azure, OCI, xAI); NVLink Fusion opens the fabric to third-party silicon — strategically significant as an ASIC moat move.
- Gaming — recovery delivered at record level: $3.8B (+48% q/q) vs guided "strong sequential recovery"; RTX 5060/5060 Ti launched ($299 desktop, $1,099 laptop); Nintendo Switch 2 confirmed on custom RTX/DLSS silicon (150M+ installed base predecessor).
- Auto — first sequential stumble (-1% q/q): +72% y/y on self-driving ramps and NEV demand; GM partnership (vehicles, factories, robots); Mercedes CLA full-stack production "hitting roads in the next few months." The ~$5B FY26 vertical claim from last quarter was not reiterated or reconciled — still unresolved.
- ProViz — tariff-impacted: flat q/q; "tariff-related uncertainty temporarily impacted Q1 systems"; sequential growth guided for Q2. DGX Spark (1 PFLOP) calendar Q3, DGX Station (20 PFLOP, GB300) later this year.
- Enterprise/robotics stack: RTX Pro enterprise server with "every major IT company"; Llama Nemotron open reasoning models (+20% accuracy, 5x inference speed); Isaac GR00T N1 humanoid foundation model; Cosmos WFM adopters (1X, Agility, Figure, Uber, Waabi); Yum! Brands deal — 500 restaurants this year, 61,000 over time.
- Software: still no run-rate disclosure — third straight quarter without the $1.5B→$2B framework being updated. NIM/Nemo traction anecdotes only.
Sell-side read-through- Only six analyst questions — a short Q&A for a quarter with an $8B guidance hole; several obvious topics went unpressed.
- Moore (Morgan Stanley) on inference sizing and whether NVL72 rack-scale is required for reasoning: Jensen — "we would like to serve all of it... on track to serve most of it"; GB200 ~40x Hopper throughput. Still no quantified inference mix — the "vast majority" claim from last quarter was not repeated or substantiated.
- Arya (BofA) extracted the cleanest H20 bridge: $7B Q1 run rate, $8B Q2 foregone, write-down covers committed orders, but the structural loss is the ~$50B TAM. His second question (path to ~$1T AI spend, digestion risk) got Jensen's "very beginning of the build-out" essay — no cyclicality acknowledgment.
- Muse (Cantor) on unannounced mega-clusters and lead times: "more orders today than at GTC," supply chain "quite busy for several more years," Europe trip next week with more announcements expected. No lead-time quantification.
- Reitzes (Melius) did the key guidance math: the $8B H20 hole is ~$3B worse than consensus assumed, so the ex-China business is doing $2–3B better than the Street expected to still guide $45B. Kress confirmed the framing without disputing it — the most important analytical takeaway of the call. Jensen then listed "four positive surprises" vs GTC: reasoning-AI inference step-function, AI diffusion rule rescission, enterprise AI (RTX Pro), industrial AI/Omniverse.
- Arcuri (UBS) pressed the China SKU question directly: no approved replacement, "we don't have anything at the moment... nothing to announce today," limits make further Hopper reduction impossible. No commitment that $7–8B quarterly China run rates are recoverable — a clear negative for models that add China back.
- Rakers' associate (Wells Fargo) on networking attach/Ethernet: Jensen's four-platform answer (NVLink scale-up, InfiniBand, Spectrum-X, BlueField) with the claim that Spectrum-X lifts cluster utilization from ~50% to 85–90% — "worth $4B on a $10B cluster."
- Striking absences: nobody asked about the ~$5B auto vertical claim, the software run rate, HBM supply, gross-margin bridge detail beyond "better Blackwell profitability," or the Singapore/transshipment question (management volunteered the >99% US-customer stat unprompted — suggesting they expected the question). No challenge on the "100 vs 800 AI factories" discrepancy.
Management credibility- Guidance beat with a hole in it: $44B vs $43B guide while absorbing a $2.5B H20 shortfall — underlying ex-China performance was materially above plan. The Q2 guide's credibility rests on the same math Reitzes surfaced: ex-China strength covering an $8B hole.
- Networking promise finally kept: after two consecutive misses with identical "growth resumes next quarter" commitments, +64% q/q to $5B restores the guidance track record — with new disclosure (NVLink >$1B, Spectrum-X >$8B annualized) to support durability.
- Margin trajectory on track, language tightened: ex-charge 71.3% vs 71.0% guide; Q2 guided 72.0%; "mid-seventies late this year" is now slightly firmer than last quarter's "later this fiscal year." The charge itself was smaller than initially anticipated ($4.5B, due to material reuse) — conservative initial reserving.
- Gaming recovery delivered at record level as guided; ProViz tariff impact disclosed plainly.
- Inconsistencies and gaps: Kress said "nearly 100 AI factories in flight"; Jensen said "some 800 AI factories being built" — an 8x discrepancy never reconciled (likely "in flight" vs "planned," but sloppy). The ~$5B FY26 auto vertical claim vanished without explanation. Software run rate absent for a third quarter. Inference mix still unquantified despite being the core demand thesis.
- China messaging is candid on the loss, evasive on the path back: unusually specific numbers ($4.6B / $2.5B / $4.5B / $8B / $50B TAM) paired with "we're thinking about it" on a compliant SKU and "I trust him" on policy. Arcuri's run-rate-recovery question was not answered.
- Jensen's "four positive surprises" framing is a notable rhetorical upgrade — management is now claiming demand is ahead of where they believed it was at GTC (late March), which is checkable against future guides.
What changed versus the prior quarter- China went from a capped percentage to a structural loss: last quarter "roughly half of pre-control share, stable" → this quarter H20 banned, $4.5B charge, $8B Q2 hole, ~$50B TAM "effectively closed," Hopper "end of the road," no compliant product. The single biggest negative change in the series.
- Ex-China demand accelerated beyond plan: Q2 guided to $45B despite the $8B hole — implying ~$53B ex-China run rate vs $44B total in Q1. Reitzes' $2–3B-above-Street math was confirmed, not disputed.
- Networking: broken promise → delivered: -3% q/q → +64% q/q to $5B, with NVLink (>$1B) and Spectrum-X (>$8B annualized) newly quantified.
- Gaming: trough → record: $2.5B → $3.8B (+48% q/q), plus Switch 2 confirmation.
- Margins: trough confirmed and recovery started: 71.3% ex-charge (slight beat) → 72.0% guided → mid-70s "late this year." The Q1-as-bottom question Kress dodged last quarter is now implicitly answered yes.
- Blackwell mix disclosed: ~70% of DC compute; Hopper transition "nearly complete" — the Hopper bridge narrative is now officially over (and Hopper is dead in China specifically).
- Sovereign AI went from unquantified narrative to named gigawatt deals: Saudi 500MW, UAE 5GW, Taiwan, Sweden — the AI diffusion rule rescission is the catalyst management credits.
- Auto momentum broke: +27% q/q → -1% q/q; the ~$5B FY26 vertical claim was dropped rather than explained.
- Capital return stepped up sharply: $8.1B → record $14.3B.
- Tariffs moved from "unknown" to visible impact (ProViz systems) while the onshoring build-out (TSMC Arizona, Foxconn Houston, Wistron Fort Worth) became a core strategic pillar.
Bull case- $45B Q2 guide with an $8B China hole implies ex-China revenue of ~$53B — roughly +20% q/q ex-China — with management confirming the rest of the business is running ahead of Street expectations. Demand, not policy, is setting the ceiling.
- Inference inflection is now evidenced, not just asserted: Microsoft 100T+ tokens (5x y/y), inference startups tripling token revenue on B200, GB200 MLPerf 30x vs H200, software alone adding 1.5x in a month. Reasoning AI is a third, compounding compute-scaling vector.
- Blackwell at ~70% of DC compute with yields improving and GB300 already sampling — drop-in compatibility removes the dual-ramp execution risk that defined the last transition; annual cadence "through 2028" aligns with customer planning cycles.
- Networking is now a second confirmed growth engine: $5B quarter, NVLink >$1B, Spectrum-X >$8B annualized with Google and Meta added, and NVLink Fusion co-opts the ASIC threat into the ecosystem.
- Sovereign AI converted from narrative to contracts: Saudi 500MW, UAE 5GW, "tens of gigawatts" in line of sight, diffusion-rule rescission opening allied markets — a demand leg that didn't exist in guidance a quarter ago.
- Margin recovery is underway on schedule (71.3% → 72.0% → mid-70s "late this year") with the charge absorbed in one quarter; record $14.3B buyback signals management's own view of the China hit as manageable.
- Gaming at a record with Switch 2 as a multi-year annuity; enterprise AI (RTX Pro, DGX Spark/Station) and robotics (GR00T, Cosmos) seeding the next legs.
Bear case- China is now a structural zero, not a capped percentage: ~$7–8B/quarter run rate gone, no compliant SKU, "nothing to announce," and Arcuri's recovery question unanswered. Jensen's own framing — the ban "strengthens" Chinese competitors abroad — concedes long-term share creation for rivals in a ~$50B TAM. Any model adding China back is speculative.
- The guide's quality depends on ex-China outperformance continuing: Q2 embeds a ~$2–3B beat-and-raise dynamic in the non-China business; if hyperscaler digestion ever arrives, there is no China cushion left to absorb it.
- Reported margins took a 10-point hit (61.0% non-GAAP) and the recovery to mid-70s still has no quantified bridge — "better Blackwell profitability" is the entire explanation. Tariff exposure is now visibly hitting ProViz and remains unquantified elsewhere.
- Disclosure discipline is slipping: 100 vs 800 AI factories contradiction, the ~$5B auto vertical claim silently dropped, software run rate missing for a third quarter, inference mix still unquantified while being the central demand claim. Singapore at ~20% of billings required a defensive disclosure (>99% US customers) that invites further regulatory scrutiny of transshipment.
- Auto decelerated sequentially (-1% q/q) right after management floated a ~3x growth vertical number — the one segment with a specific ambitious target is now off-script.
- Policy dependence is now explicit and bilateral: the bull case leans on the diffusion-rule rescission and Middle East deals (politically reversible), while the bear case (China) is also policy-set. Jensen's "I trust him" ties the equity story to a single administration's decisions.
- Opex still growing mid-30s% against a margin recovery that must reach mid-70s in two quarters to hit "late this year" — little room for a second surprise charge.
Next-quarter watchlist- Q2 delivery vs $45B ±2%: the ex-China implied ~$53B run rate; whether Blackwell ramp fully offsets the China decline as guided; any residual H20 revenue (should be ~zero).
- China SKU: any approved H20 successor, its specs/timing, and whether management restores any China revenue to the back half; further export-control or tariff action; Singapore billing scrutiny.
- Gross margin: Q2 vs 72.0% non-GAAP; a quantified bridge to mid-70s "late this year"; confirmation the $4.5B charge was one-time; tariff cost disclosure.
- GB300/Blackwell Ultra: production shipments commencing "later this quarter" — confirm on-time start, early customer uptake, and that the drop-in promise holds without the NVLink 8→72-style disruption.
- Networking durability: does $5B hold or grow? NVLink Fusion design-win traction (MediaTek, Alchip, Astera, Qualcomm); Spectrum-X photonics switch timeline.
- Sovereign conversion: Saudi 500MW / UAE 5GW order timing and revenue recognition; Europe announcements from Jensen's early-June tour and GTC Paris keynote (June 11).
- Auto: whether the ~$5B FY26 vertical claim is reinstated, redefined, or dead; Mercedes CLA launch; GM program detail; sequential growth resumption.
- Software: any reinstatement of a run-rate metric; NIM/Nemotron monetization evidence beyond anecdotes.
- ProViz: tariff impact resolution and the guided sequential recovery; DGX Spark availability (calendar Q3).
- Opex: Q2 vs ~$4.0B non-GAAP; FY26 mid-30s% framework holding.
- Events: BofA Global Tech (June 4), GTC Paris/VivaTech keynote (June 11), Q2 FY26 earnings August 27, 2025.
|
| Feb 26, 2025 | -8.48% | Q4 FY2025 | Read transcript briefingQuarter in one view- Revenue $39.3B, +12% q/q, +78% y/y — above the $37.5B ±2% guide (~5% above midpoint). FY2025 revenue $130.5B, +114% y/y. Q1 FY26 guide: $43B ±2% (+~9% q/q), with sequential growth guided in both DC and gaming, and within DC both compute and networking.
- The quarter's headline: Blackwell delivered $11B of revenue in its first ramp quarter — "fastest product ramp in company history," exceeding management's own expectations (prior commitment was only "exceed several billion"). DC compute +18% q/q, >2x y/y.
- Data Center record $35.6B, +16% q/q, +93% y/y. CSPs still ~half of DC, CSP sales ~2x y/y; consumer Internet 3x y/y; enterprise ~2x y/y; regional clouds grew as a % of DC.
- Gross margin landed on guide: 73.0% GAAP / 73.5% non-GAAP (from 74.6/75.0). Q1 guide steps down again to 70.6% / 71.0% ±50bps — the low-70s trough is now arriving; mid-70s recovery now framed as "later this fiscal year" (i.e., FY26, ending Jan 2026).
- Networking declined 3% q/q — a second consecutive sequential decline, against last quarter's explicit commitment to resume sequential growth in Q4. Management again guides Q1 networking growth; attach rate to GPU systems disclosed at >75%.
- Gaming $2.5B, -22% q/q, -11% y/y on the pre-flagged supply constraints; strong sequential recovery guided for Q1. Auto record $570M, +27% q/q, +103% y/y; FY26 auto vertical revenue expected at ~$5B (a striking new number vs $1.7B FY25 segment revenue — appears to be a broader vertical framing).
- Capital: $8.1B returned in Q4 (vs $11.2B in Q3). FY26 opex guided to grow mid-30s% — a new full-year opex framework, absent last quarter.
What management is focused on- Declaring the Blackwell ramp won: $11B in Q4, "full gear across multiple configurations," 350 manufacturing sites / 1.5M components per NVL72 rack / ~100,000 factory operators, named bring-ups (CoreWeave, Microsoft, OpenAI). Jensen on the prior mask hiccup: "cost us a couple of months. We're fully recovered."
- Reasoning AI as the new demand engine: test-time scaling (OpenAI o3, DeepSeek R1, Grok 3) framed as 100x more compute per task today, with future models "hundreds of thousands, millions of times more." Blackwell positioned as purpose-built for it: 25x token throughput, 20x lower cost vs H100. Jensen now claims "the vast majority of our compute today is actually inference" — a notable, unquantified assertion that replaces the dropped ">40% inference" metric with something much larger and even less verifiable.
- DeepSeek R1 embraced, not feared: "ignited global enthusiasm... open-sourced a world-class reasoning AI model" — the efficiency-threat narrative answered with Jevons-style compute expansion.
- Pre-empting the ASIC question (Arcuri): generality, end-to-end coverage, ubiquity, perf/watt → revenue per fixed-power data center, software-stack complexity, and "just because the chip is designed doesn't mean it gets deployed." Pure argumentation, no win/loss data.
- Roadmap cadence reassurance: Blackwell Ultra confirmed H2 CY2025, "slots right in" (same system architecture, unlike the hard Hopper→Blackwell NVLink 8→72 transition); Vera Rubin teased for GTC with "one click after that."
- Managing the margin trough as temporary and chosen: "we are focused on expediting manufacture" over cost optimization; cost/margin improvement work begins once ramp completes; mid-70s "later this fiscal year."
Key numbers and quarter mechanics- Revenue $39.3B (+12% q/q, +78% y/y) vs $37.5B ±2% guide. Q1 FY26 guide $43B ±2%.
- Segments: DC $35.6B (+16% q/q, +93% y/y); Gaming $2.5B (-22% q/q, -11% y/y); ProViz $511M (+5% q/q, +10% y/y); Auto record $570M (+27% q/q, +103% y/y; FY $1.7B, +55%).
- Blackwell revenue: $11B in Q4 — first disclosure of the actual ramp number; vs prior "exceed several billion" commitment. "Significant ramp of Blackwell in Q1" guided.
- DC mix: CSPs ~half (flat vs Q3), CSP ~2x y/y; consumer Internet 3x y/y (accelerating from >2x); enterprise ~2x y/y; regional clouds up as % of DC. US drove the sequential growth (~$5B q/q per Reitzes' question, not disputed).
- China: "well below" pre-control levels, roughly half of pre-control share, approximately the same percentage as prior quarters — Jensen's most specific China framing to date, though still no absolute number. Guided to remain "roughly at the current percentage" absent regulatory change.
- Networking: -3% q/q (second straight sequential decline); Spectrum-X and NVLink Switch revenue grew; attach rate >75%; Q1 return to growth guided. Cisco integrating Spectrum-X; first Stargate data centers will use Spectrum-X.
- Gross margin: 73.0% GAAP / 73.5% non-GAAP; Q1 guide 70.6% / 71.0% ±50bps; low-70s during ramp, mid-70s "later this fiscal year" (FY26).
- Opex: +9% GAAP / +11% non-GAAP q/q. Q1 guide ~$5.2B GAAP / ~$3.6B non-GAAP; FY26 opex growth guided mid-30s% (new framework).
- Other Q1 items: other income ~$400M; tax 17% ±1% (back up from 16.5%).
- $8.1B returned in Q4 (down from $11.2B in Q3).
- Auto vertical: ~$5B expected this fiscal year (vs $1.7B FY25 segment revenue — definitional scope unclear, likely broader than the reported segment; needs clarification).
- Inference proof points: Perplexity 435M monthly queries, 3x cost reduction; Microsoft Bing 5x visual-search speedup; Snap 3x throughput / 66% cost cut (transcript garbled on the Snap name).
Product and launch scorecard- Blackwell — ramp delivered at scale, ahead of commitment: $11B in the first quarter vs "exceed several billion" guided; multiple 100,000+ GPU clusters already shipping; early GB200 deployments "earmarked for inference — a first for a new architecture." The overheating issue from last quarter's Q&A was not raised by anyone this quarter; Jensen's "350 plants / 1.5M components" framing acknowledges complexity while asserting control. Own read: this is the strongest possible answer to the delay narrative; the open question shifts entirely to margin cost and Q1 ramp slope.
- Blackwell Ultra — on track, H2 CY2025: key de-risking claim is architectural continuity ("slots right in," same NVLink 72 chassis) vs the painful Hopper→Blackwell transition. Simultaneous ramp management with current Blackwell acknowledged but not detailed.
- Hopper — still contributing, but clearly the secondary engine now: "Hopper 200 continued to contribute growth"; no H200-specific numbers this quarter (last quarter: double-digit-billions sequential increase). The Hopper bridge narrative is quietly receding as Blackwell carries the print.
- Networking — second consecutive sequential miss (-3% q/q): management attributes it to the NVLink 8 + InfiniBand → NVLink 72 transition; Spectrum-X momentum cited (Azure, OCI, Stargate, Cisco integration). Q1 growth guided again — this is now a twice-made commitment after a miss; credibility on the line.
- Gaming — trough quarter as flagged: -22% q/q on supply constraints; RTX 50-series (Blackwell, DLSS 4, up to 3,400 AI TOPS) launched, laptops from March; strong Q1 sequential growth guided.
- Auto — breakout continues: +103% y/y; Toyota (world's largest automaker) on Orin/DriveOS; Aurora/Continental driverless trucks; Drive Hyperion passed third-party safety assessments. The ~$5B FY26 auto vertical statement, if it refers to the segment, implies ~3x growth — a major unverified claim to pin down.
- Cosmos / physical AI: world foundation model platform announced at CES (transcript says "TDX"); Uber among first adopters. Early, no revenue framing.
- Software: no run-rate update this quarter — the $1.5B annualized / >$2B exit framework from last quarter was not reiterated. A disclosure gap.
Sell-side read-through- Muse (Cantor) asked about inference-dedicated clusters as training/inference blur; Jensen's answer — fungible unified architecture, "vast majority of our compute today is actually inference" — is the new mix claim, delivered with zero data.
- Moore (JPMorgan) probed GB200 rack-level bottlenecks and NVLink 72 enthusiasm; Jensen: "more enthusiastic today than at CES... we shipped a lot more," 350 plants, "nothing is easy... but we're doing great." No component-level constraint disclosure (HBM still unaddressed).
- Arya (BofA) asked the two sharpest questions: is Q1 the margin bottom (Kress declined to confirm directly — repeated "low seventies during ramp, mid-seventies later this year"), and what dashboard supports demand durability post-DeepSeek (Jensen: near-term POs/forecasts, mid-term CapEx, long-term software transformation — framework, not data).
- Ramsay (Cowen) on Ultra demand dynamics and dual-ramp execution; got the "hiccup cost us a couple of months, fully recovered" admission and the architectural-continuity argument.
- Arcuri (UBS) on custom ASIC vs merchant GPU and heterogeneous clusters; Jensen's five-point rebuttal is the company's standard ASIC defense — rhetorically complete, no customer-specific evidence.
- Reitzes (Melius) on US-concentration risk (~$5B q/q US surge) and China; Jensen confirmed China ~half of pre-control share, stable as a percentage — the most concrete China disclosure in several quarters, extracted by the question, not volunteered.
- Lipacis/Pappas (Evercore) confirmed enterprise grew ~2x y/y (similar to CSPs) and asked for the CSP internal/external workload split — Jensen declined to split it, pivoted to "enterprise long-term is by far larger" via the three-computers (agentic/physical/robotic) thesis.
- Rakers (Wells Fargo) on replacement/refresh cycles; Jensen's answer — Voltas, Pascals, Amperes still employed on data curation via CUDA compatibility — is a direct rebuttal to the obsolescence/refresh bear point.
- Malik (Citi) pressed the margin-recovery math: H2 would need ~200bps/quarter to reach mid-70s, plus tariff exposure. Kress's answer was the weakest of the call: "complex... many configurations... start as soon as possible," and on tariffs "a little bit of an unknown... awaiting" the government's plan. No quantified bridge to mid-70s was offered.
- Striking absences: nobody asked about the overheating reports (last quarter's live issue), the software run rate, sovereign AI quantification, or the networking miss directly. Only nine analyst questions; call ended early-ish ("time for one more").
Management credibility- Beat the raised Blackwell bar decisively: "exceed several billion" → $11B actual. The mask-change episode is now fully closed with Jensen's own retrospective ("cost us a couple of months"). This is the strongest delivery-versus-promise data point in the series.
- Margin guidance precision held: Q4 landed exactly on the 73.0/73.5 guide; the low-70s trough predicted last quarter is arriving on schedule (Q1 guide 71.0% non-GAAP). However, the recovery timeline slipped in framing: last quarter "H2 CY2025 reasonable" → this quarter "mid-seventies later this fiscal year" (FY26 ends Jan 2026 — roughly consistent but looser), and Kress would not confirm Q1 as the bottom nor provide the recovery bridge under Malik's 200bps/quarter math.
- Networking guidance missed again: Q4 sequential growth was explicitly committed last quarter; delivered -3% q/q. The NVLink 72 transition explanation is plausible, but this is the second consecutive networking miss and the second identical "growth resumes next quarter" commitment.
- New disclosure, both directions: positive — actual Blackwell revenue ($11B), networking attach rate (>75%), China share framing (~half of pre-control, stable); negative — no software run-rate update, no sovereign number (second straight quarter), no H200 detail, inference mix replaced by an unverifiable "vast majority" claim.
- Tariff exposure acknowledged as unknown: Kress's "awaiting the US government's plan" is honest but leaves a live cost risk unquantified against a guided margin recovery.
- Jensen's demand answers remain essay-length and metric-free; the financial guidance remains precise. The asymmetry persists, but the $11B Blackwell print shifts the burden of proof to the bears on execution.
What changed versus the prior quarter- Blackwell: from "will exceed several billion" to $11B delivered — the ramp question is settled; the questions are now slope, margin, and Ultra transition.
- Margin trough arrived on schedule: Q4 73.5% non-GAAP → Q1 guided 71.0%; recovery language shifted from "H2 CY2025 reasonable" to "mid-70s later this fiscal year," and Kress declined to call Q1 the bottom.
- Networking missed again: -15% q/q → -3% q/q, against a guided return to growth; attach rate >75% disclosed; Q1 growth re-promised.
- Gaming trough delivered as flagged (-22% q/q); RTX 50-series launched; Q1 recovery guided.
- Auto accelerated again: +72% y/y → +103% y/y; new ~$5B FY26 auto vertical expectation introduced; Toyota win announced.
- China quantified (partially) for the first time in quarters: ~half of pre-control share, stable percentage, guided to stay there.
- Hopper de-emphasized: no H200 metrics after last quarter's "fastest ramp ever"; Blackwell now carries the narrative.
- Inference narrative escalated: from ">40% inference" (dropped last quarter) to "vast majority of our compute today is actually inference" — bigger claim, less verifiable.
- FY26 opex framework added: mid-30s% growth — new disclosure that was absent last quarter.
- Capital return moderated: $11.2B → $8.1B.
- Tariffs emerged as a named unknown in margin Q&A; DeepSeek R1 absorbed into the demand narrative rather than treated as a threat.
Bull case- $11B Blackwell in the first ramp quarter — roughly 3-5x the original "several billion" framing — with a $43B Q1 guide (+9% q/q) embedding a "significant" further Blackwell ramp and growth in every DC sub-line. The execution-risk discount should compress.
- Demand evidence broadened again: consumer Internet 3x y/y, enterprise ~2x, CSP ~2x at ~half of DC, regional clouds up, multiple 100k+ GPU clusters already shipping, and reasoning models (o3, R1, Grok 3) adding a third compute-scaling vector with 100x-per-task intensity.
- Blackwell's inference economics (25x throughput, 20x lower cost vs H100) plus early GB200 inference earmarking position NVIDIA for the reasoning-AI wave rather than just the training wave.
- Margin trough is guided, bounded (71% Q1), and management has now hit both margin guides in the descent; recovery levers (cost-down, configurations, yields) are identified even if unquantified.
- Auto inflecting hard (+103% y/y, Toyota, ~$5B FY26 vertical target) and physical AI (Cosmos, Uber) opening the next TAM leg.
- CUDA installed-base argument (Volta→Ampere still employed) directly counters the refresh/obsolescence bear case; ASIC threat answered with deployment-speed and ecosystem logic while no ASIC share loss is visible in the numbers.
Bear case- Gross margin guided to 71% with no confirmed bottom: Kress would not affirm Q1 as the trough, offered no quantified bridge to mid-70s, and Malik's math (~200bps/quarter H2 recovery) went unanswered. Tariffs are an acknowledged unknown sitting on top of the recovery path.
- Networking missed twice in a row (-15%, then -3% vs guided growth) — the "growth resumes next quarter" commitment now has a broken track record, and the NVLink 72 transition explanation implies the lumpiness is structural to the ramp, not a one-off.
- Disclosure is thinning where it matters: software run rate dropped, sovereign unquantified for a second quarter, H200 metrics gone, inference mix now an unverifiable "vast majority" claim. The trend is toward narrative over verifiable mix data.
- US concentration is rising (~$5B of the ~$4B... the US drove essentially all sequential growth per Reitzes' framing) while China is structurally capped at ~half its prior share and tariff/export policy is in flux — geographic diversification is narrowing, not broadening.
- Opex guided to grow mid-30s% in FY26 while gross margin is in a trough — operating leverage pauses precisely when the multiple needs margin recovery.
- Dual-ramp execution risk: Blackwell still ramping while Ultra launches H2 — management asserts architectural continuity, but the last transition (NVLink 8→72) "cost a couple of months" and produced overheating reports that were never substantively addressed on the record.
- Expectations: the stock now needs Q1's $43B, the 71% margin floor to hold, networking's third-time promise to deliver, and Ultra to launch cleanly — a long chain of guided-but-undelivered items.
Next-quarter watchlist- Q1 delivery vs $43B ±2%: Blackwell ramp slope off the $11B base; whether Hopper declines are fully offset; DC compute and networking both growing as guided.
- Gross margin: Q1 vs 71.0% non-GAAP — is it the trough? Any quantified bridge to mid-70s "later this fiscal year"; tariff developments and any quantification of tariff exposure; Blackwell cost-down/yield commentary.
- Networking: does the third "return to growth" commitment land? Spectrum-X traction (Cisco integration, Stargate build-out); NVLink 72 transition completion.
- Blackwell Ultra: GTC (March 17–19, Jensen keynote March 18, analyst Q&A March 19) — Ultra specs/timing reaffirmation, Vera Rubin detail, and the teased "one click after that"; any sign of dual-ramp strain.
- Auto: reconciliation of the ~$5B FY26 vertical claim vs the $1.7B FY25 segment — scope, definition, and segment-level trajectory; Toyota/Aurora/Continental program timing.
- China: whether the ~half-of-pre-control share holds; any new export-control or tariff action; H20-class product traction.
- Software: whether the $1.5B→$2B run-rate framework is reinstated or quietly retired; AI Enterprise growth.
- Gaming: magnitude of the guided Q1 sequential recovery; RTX 50-series supply normalization.
- Sovereign AI: France/EU €200B initiatives — any conversion into quantified NVIDIA pipeline.
- Opex: Q1 vs ~$3.6B non-GAAP; FY26 mid-30s% growth framework holding.
- Events: TD Cowen Healthcare (March 3), Morgan Stanley TMT (March 5), GTC (March 17–19), Q1 FY26 earnings May 28, 2025.
|
| Nov 20, 2024 | +0.53% | Q3 FY2025 | Read transcript briefingQuarter in one view- Revenue $35.1B, +17% q/q, +94% y/y — ~8% above the $32.5B ±2% guide midpoint; seventh consecutive multi-billion beat-and-raise in this series. Q4 guide: $37.5B ±2% (+~7% q/q), now explicitly including the initial Blackwell ramp plus continued Hopper demand.
- Data Center record $30.8B, +17% q/q, +112% y/y. CSPs now ~half of DC (up from ~45%), with CSP revenue >2x y/y; consumer Internet more than doubled y/y.
- The quarter's defining event: Blackwell is in full production and the Q4 revenue bar was raised. "On track to exceed our previous Blackwell revenue estimate of several billion dollars" as supply visibility improves; 13,000 Blackwell GPU samples shipped in Q3 (including an early DGX engineering sample to OpenAI); Microsoft first CSP to preview GB200 instances; Oracle announced a 131,000-GPU Blackwell cluster.
- H200 became the fastest product ramp in company history — sales "increased significantly to double-digit billions" sequentially; H200 instances live at AWS, CoreWeave, Azure, with Google Cloud and OCI coming.
- Gross margin landed on guide: 74.6% GAAP / 75.0% non-GAAP (vs 75.1/75.7), on mix-shift to more complex/higher-cost systems. Q4 guide steps down to 73.0% / 73.5% ±50bps, with explicit new language: margins moderate to low-70s as Blackwell ramps, mid-70s when fully ramped — Kress put numbers on "low-70s": 71%–72.5% range possible.
- Networking declined sequentially (~-15% q/q per Rakers) despite +20% y/y and Spectrum-X >3x y/y — a guidance miss versus last quarter's "sequential growth" expectation; management guides sequential growth resumes in Q4.
- Capital: $11.2B returned in Q3 (vs $7.4B in Q2). Opex +9% q/q; Q4 guide ~$4.8B GAAP / ~$3.4B non-GAAP; tax guide lowered to 16.5% ±1%.
- Software: AI Enterprise FY revenue guided to >2x y/y; total software/service/support annualizing at $1.5B now, >$2B exit reiterated — note the current run rate is $1.5B, below the "$2B exiting" framing from last quarter's summary context.
What management is focused on- Converting the Blackwell narrative from delay to upside: "full production," mask change "successfully executed," 13,000 samples shipped, Q4 revenue now above the prior "several billion" commitment, and a named-partner ramp litany (TSMC, SK Hynix, Micron, Amkor, Foxconn, Quanta, Wiwynn, Dell, HPE, Super Micro, Lenovo, Vertiv, Amphenol). Jensen's framing: "how much Blackwell we shipped last quarter... zero" to "billions" this quarter.
- A new demand narrative: three scaling laws. Pre-training scaling "intact," plus post-training scaling (RLHF → RL-AI feedback, synthetic data) and inference-time/test-time scaling (OpenAI o1/"Strawberry") — the direct rebuttal to the "scaling has stalled" debate (Muse's opening question). Next-gen clusters "start at 100,000 Blackwells" vs ~100,000 Hoppers at the tail of the last generation.
- Managing the margin descent proactively: for the first time, management pre-committed to a low-70s trough during the Blackwell ramp with mid-70s recovery "quite quickly" — Kress confirmed mid-70s in H2 calendar 2025 is "a reasonable assumption or a goal." This inverts last quarter's dynamic (where Kress had to push back on Rasgon's 71–72% extrapolation) into a guided, temporary trough.
- Hopper durability through the transition: H200 "exceptional" demand, double-digit-billions sequential increase, Hopper demand "will continue through next year, surely the first several quarters"; Kress conceded Hopper q/q growth in Q4 is only "possible" — the first hedging on Hopper sequential growth.
- Enterprise/agentic AI as the monetization story: ~1,000 companies using NIM; Accenture's 30,000-person NVIDIA practice and internal agentic AI use (25–35% fewer manual marketing steps); AI Enterprise revenue >2x y/y.
- Sovereign AI broadened geographically: India (Tata, Yotta — ~10x GPU deployment growth by year-end; Infosys/TCS/Wipro upskilling ~0.5M developers), Japan (SoftBank DGX Blackwell supercomputer + AI-RAN with T-Mobile US launch).
Key numbers and quarter mechanics- Revenue $35.1B (+17% q/q, +94% y/y) vs $32.5B ±2% guide. Q4 guide $37.5B ±2%.
- Segments: DC $30.8B (+17% q/q, +112% y/y); Gaming $3.3B (+14% q/q, +15% y/y) but guided down sequentially in Q4 on supply constraints; ProViz $486M (+7% q/q, +17% y/y); Auto record $449M (+30% q/q, +72% y/y — accelerating from +37%).
- DC mix: CSPs ~half (up from ~45%), CSP revenue >2x y/y; consumer Internet >2x y/y; regional GPU cloud revenue 2x y/y. China DC grew sequentially on export-compliant products, "well below" pre-control share, "very competitive" — still unquantified, third straight quarter.
- H200: sequential sales increase to double-digit billions; fastest ramp in company history; 2x inference performance, 50% TCO improvement claims; software optimizations drove 5x Hopper inference throughput in one year, upcoming NIM release adds 2.4x.
- Networking: +20% y/y but down ~15% q/q; Spectrum-X >3x y/y; xAI Colossus 100,000-GPU Hopper cluster cited with 95% throughput vs 60% for traditional Ethernet; Q4 sequential growth guided.
- Gross margin: 74.6% GAAP / 75.0% non-GAAP; Q4 guide 73.0% / 73.5% ±50bps; ramp trough low-70s (Kress: 71%–72.5% possible), mid-70s "when fully ramped," H2 CY2025 recovery "reasonable."
- Opex: +9% q/q both bases (compute, infrastructure, engineering for new products). Q4 guide ~$4.8B GAAP / ~$3.4B non-GAAP — no FY opex growth framework reiterated this quarter.
- Other Q4 guide items: other income ~$400M; tax 16.5% ±1% (down from 17%).
- $11.2B returned in Q3 (buybacks + dividends).
- Blackwell: 13,000 samples shipped in Q3; Q4 revenue to exceed "several billion dollars"; shipments to increase each quarter through next year; demand "staggering," exceeds supply.
- Software: $1.5B annualizing now; >$2B exit reiterated; AI Enterprise FY revenue >2x y/y; ~1,000 NIM customers.
Product and launch scorecard- Blackwell — ramp confirmed, bar raised, margin cost now explicit: full production post-mask-change; 13,000 Q3 samples; Q4 revenue above the prior "several billion"; MLPerf Training sweep with 2.2x per-GPU over Hopper and GPT-3 benchmark on 64 Blackwells vs 256 H100s (4x cost reduction); NVLink Switch enables up to 30x inference; Oracle 131,000-GPU cluster, Microsoft GB200 private preview, Dell/CoreWeave/Google systems standing up. Own read: the revenue raise and named deployments are real de-risking, but the margin trough (71–72.5%) is the quantified cost of the ramp, and "exceed several billion" remains unbounded — the actual Q4 Blackwell number is still undisclosed.
- Hopper/H200 — the strongest quarter of evidence yet: double-digit-billions sequential increase, fastest ramp in company history, broad CSP instance availability. But Kress's "possible" on Q4 Hopper sequential growth is the first crack in the "Hopper keeps growing" bridge narrative.
- Spectrum-X — growth but a sequential networking miss: >3x y/y, xAI Colossus metrics (95% vs 60% throughput), multiple CSP/consumer-Internet large-cluster pipeline — yet total networking fell ~15% q/q against last quarter's guidance of sequential growth. Kress called it "a slight dip" ahead of Blackwell attach; Q4 recovery is now a testable commitment.
- Gaming — supply-constrained into the holiday quarter: Q3 strong ($3.3B, +14% q/q) but Q4 guided down sequentially on supply; recovery expected "as we turn the corner into the new calendar year." New RTX AI PCs (up to 321 AI TOPS, ASUS/MSI) shipping with Copilot+ in Q4.
- Auto — breakout quarter: record $449M, +30% q/q, +72% y/y on Orin self-driving ramps and NAV demand; Volvo EX SUV on Orin/DriveOS. Now the fastest-growing segment.
- Software — run rate disclosed at $1.5B current / >$2B exit: AI Enterprise >2x y/y, ~1,000 NIM customers. Note: the current $1.5B annualized figure is new disclosure granularity; the exit target is unchanged.
- Omniverse/Industrial AI: Foxconn digital twins of Blackwell plants with a hard metric (>30% annual kWh reduction at Mexico facility).
Sell-side read-through- Muse (Cantor) opened with the scaling-stall debate; Jensen's "three ways of scaling" answer (pre-training, post-training, test-time) is now the company's standard demand-durability framework — rhetorically strong, still no customer-level data.
- Hari (Goldman) pressed the weekend overheating reports and roadmap execution (Blackwell Ultra next year, Rubin 2026), plus supply-constraint composition. Jensen did not directly address the heating reports — answered with ramp volume, partner list, and "everything is on track as far as I know." No component-specific constraint disclosure (HBM vs other); the seven-custom-chip complexity was acknowledged. The heating question was effectively evaded.
- Arcuri (UBS) extracted the two most useful guideposts: Blackwell crossing over Hopper was not confirmed for April — Jensen instead said Hopper continues "through next year, surely the first several quarters" and Blackwell shipments rise each quarter; and Kress confirmed the low-70s margin trough is front-loaded in the ramp with recovery "quite quickly" after.
- Arya (BofA) got Kress to endorse mid-70s gross margin in H2 CY2025 as "a reasonable assumption or a goal" — the closest thing to a margin recovery timeline. His digestion question drew Jensen's "no digestion until we modernize a trillion dollars of data centers" — the cycle-risk question answered with TAM, not evidence.
- Rasgon (Bernstein) again forced precision: Kress defined low-70s as "71%, maybe about 72%, 72.5%... could be higher" — a notable reversal from last quarter when she rejected his 71–72% Q4 scenario; that scenario is now the guided trough. His Hopper-down-sequentially probe got only "possible" growth — the most hedged Hopper comment to date.
- Moore (Morgan Stanley) on inference vs training mix: Jensen's answer was aspirational ("hopes and dreams") with no mix data — the prior ">40% inference" estimate was not repeated this quarter.
- Rakers (Wells Fargo) surfaced the networking sequential decline (~-15%) — the one clear guidance miss; Kress attributed it to pre-Blackwell timing and committed to Q4 sequential growth.
- Malik (Citi) asked for a sovereign AI update — Kress gave no number (pipeline "absolutely intact"), a step back from last quarter's low-double-digit-billions target; and extracted that gaming's Q4 decline is supply-driven, with recovery early in the new calendar year.
- Reitzes (Melius) on reacceleration got "we guide one quarter at a time"; his tariff/new-administration question got Jensen's full-compliance-and-compete non-answer — China policy risk acknowledged only procedurally.
- Ferragu (New Street) asked for the pre-training/post-training/inference compute split; Jensen said "vastly pre-training" today with all three scaling — directional, unquantified.
- Striking absences: nobody pressed the $1.5B vs $2B software run-rate gap, the sovereign target going unquantified, the dropped inference-share metric, or the overheating reports beyond Hari's question. Q&A remained friendly; ten analysts.
Management credibility- Delivered on the delayed ramp, then raised it: last quarter's "production ramp Q4, several billion dollars" became "full production, 13,000 samples shipped, will exceed several billion" — the mask-change episode now reads as a one-quarter slip fully recovered, and the raise restores some of the "a lot of Blackwell revenue" framing that had been resized.
- Margin guidance honesty improved: after pushing back on Rasgon's 71–72% extrapolation last quarter, management this quarter voluntarily guided to exactly that trough with a recovery path (mid-70s when fully ramped, H2 CY25 "reasonable"). Kress's specificity under Rasgon's clarification question (71–72.5%) is credibility-positive; the reversal itself shows last quarter's pushback was about timing, not level.
- First clear guidance miss in the series: networking guided to sequential growth last quarter, delivered ~-15% q/q. Kress's "slight dip" framing understates it; Q4 recovery is now a credibility test.
- Hedged where previously confident: Hopper Q4 sequential growth only "possible" (vs "Hopper shipments expected to increase in H2" last quarter); sovereign AI number not reiterated; inference share metric dropped; gaming Q4 decline disclosed in outlook rather than prepared with a number.
- Evasion pattern: the overheating-reports question was answered with ramp enthusiasm, not thermals; supply-constraint composition (HBM?) unanswered; tariff/China policy deflected to compliance boilerplate. Jensen's demand answers remain essay-length and metric-free, but the financial guidance (margins, opex, segment mechanics) remains unusually precise — the standing asymmetry persists.
- Consistency check on software: "$1.5B annualizing now, >$2B at exit" is internally consistent with last quarter's exit target but reveals the current base is lower than the prior framing implied.
What changed versus the prior quarter- Blackwell: from "ramp begins Q4, several billion" to "full production, will exceed several billion" — schedule commitment met and revenue bar raised; 13,000 samples already shipped in Q3.
- Margin trajectory re-anchored downward before recovery: Q4 guide 73.5% non-GAAP (vs 75.0% Q3 actual), with an explicit low-70s ramp trough (71–72.5%) and mid-70s recovery targeted for H2 CY2025 — last quarter's contested bear scenario is now the guided path.
- Hopper narrative peaked: H200 delivered the fastest ramp in company history (double-digit billions sequential), but Q4 sequential Hopper growth downgraded to "possible."
- Networking missed: +16% q/q last quarter → ~-15% q/q this quarter against guided growth; Spectrum-X still >3x y/y.
- CSP concentration rose: ~45% → ~half of DC revenue.
- Sovereign AI de-quantified: low-double-digit-billions target not reiterated; pipeline described qualitatively with new India/Japan programs.
- Software disclosure refined: current run rate revealed at $1.5B (vs >$2B exit target); AI Enterprise >2x y/y; ~1,000 NIM customers.
- Gaming flipped to a Q4 headwind: supply-constrained sequential decline guided for the holiday quarter.
- Auto accelerated sharply: +37% y/y → +72% y/y, record revenue.
- Inference-share metric (">40%") dropped after three repetitions; replaced by the three-scaling-laws framework and test-time-scaling narrative.
- Capital return stepped up: $7.4B → $11.2B.
- New external risk surfaced in Q&A: weekend overheating reports and tariff/administration-change questions — neither substantively addressed.
Bull case- Beat-and-raise sustained at $35B scale with a $37.5B Q4 guide that management says understates Blackwell ("on track to exceed" the prior several-billion estimate) — the ramp is now a revenue tailwind, not a promise.
- H200's double-digit-billions sequential ramp — fastest in company history — proves the Hopper bridge held with no air pocket, and Hopper demand is committed "through next year."
- Demand evidence broadened: CSPs >2x y/y at ~half of DC, consumer Internet >2x, regional clouds 2x, sovereign programs scaling (India ~10x GPU deployments by year-end), enterprise agentic AI monetizing (~1,000 NIM customers, AI Enterprise >2x y/y).
- Blackwell performance moat quantified: 2.2x per-GPU training, 4x cost reduction on GPT-3 benchmark, 30x inference with NVLink Switch — supports pricing power through the margin trough.
- Margin trough is now guided, bounded, and temporary: 71–72.5% low with mid-70s recovery targeted for H2 CY2025 — removes the open-ended margin-drift bear argument.
- Auto inflecting (+72% y/y) and software annualizing $1.5B→$2B provide second engines beyond DC compute.
- $11.2B returned in one quarter signals cash-generation confidence at the ramp's peak investment phase.
Bear case- Gross margin is now on a guided decline to 71–72.5% — from 78.9% three quarters ago — with recovery dependent on Blackwell yields, mix, and ramp execution; "quite quickly" and "reasonable assumption" are not commitments.
- Networking's sequential miss (~-15% q/q vs guided growth) breaks the "all platforms grow" pattern and suggests lumpiness or share/pacing issues in the attach business just as Blackwell systems (which bundle networking) ramp.
- Hopper hedging begins: Q4 sequential growth only "possible" — if Hopper rolls over before Blackwell fully scales, the transition math tightens; Q4's ~7% q/q guide already embeds both.
- Overheating reports went unaddressed on the record; seven custom chips, liquid-cooling complexity, and "nothing short of a miracle" integration language concede execution risk at scale.
- Key demand metrics are being retired as they age: inference share dropped, sovereign target unquantified, China share unquantified for a third quarter — the disclosure trend is toward narrative (three scaling laws) and away from verifiable mix data.
- Gaming supply-constrained into the holiday quarter — a rare operational stumble outside DC, with recovery only "as we turn the corner."
- Policy risk live and unpriced in commentary: tariff/export-control questions deflected; China "very competitive" and structurally below prior share.
- Expectations risk compounds: the stock now needs the Blackwell beat, the margin trough to hold at 71–72.5%, networking recovery, and H2 CY25 margin restoration — all guided, none yet delivered.
Next-quarter watchlist- Q4 delivery vs $37.5B ±2%: the actual Blackwell revenue number vs "exceed several billion" — the first hard ramp-slope data point; whether Hopper grew, held, or declined sequentially (Kress's "possible").
- Gross margin: Q4 vs 73.5% non-GAAP guide; whether the low-70s trough (71–72.5%) is Q4 or Q1 FY26; any update on the H2 CY2025 mid-70s recovery path; Blackwell yield commentary and whether inventory provisions recur.
- Blackwell execution: shipment growth quarter-over-quarter as committed ("more next quarter than this"); customer stand-up pace (Oracle 131k cluster, Microsoft GB200 preview → GA); any substantiation or recurrence of the overheating reports; Ultra/Rubin roadmap reaffirmation at CES (Jensen keynote January 6, analyst Q&A January 7).
- Networking: whether the guided Q4 sequential recovery materializes after the ~-15% miss; Spectrum-X progress toward the multibillion-dollar run rate; Blackwell attach effect.
- Hopper: H200 trajectory after the double-digit-billions quarter; supply allocation between Hopper and Blackwell.
- Gaming: magnitude of the Q4 supply-constrained decline; recovery timing into the new calendar year.
- China: any quantification of share; H20/export-compliant product traction; tariff or export-control developments under the new administration.
- Sovereign AI: whether the low-double-digit-billions FY25 target is reaffirmed or quietly dropped; India/Japan program conversion.
- Software: progress from $1.5B annualized toward the >$2B exit; AI Enterprise >2x growth pacing.
- Opex: Q4 vs ~$3.4B non-GAAP; whether an FY26 opex framework is given on the February 26 call.
- Events before the print: UBS Global Technology and AI Conference (December 3); CES keynote and analyst Q&A (January 6–7); Q4 earnings February 26, 2025.
|
Corbanu