| Aug 12, 2026 | -1.60% | Q2 FY2026 | Read transcript briefingQuarter in one view- Q2 FY2026 core revenue $209.9M, +103% YoY, beating guidance (~$194M guided); beat on all three guided metrics (revenue, GM, OM). Cloud/services $127.7M (+287% YoY, nearly 4x); hardware $82.1M (+17% YoY).
- Core gross margin 40.6% (vs. 36–38% guided; 46.5% in Q1; ~31% a year ago). Core operating loss -$33.6M (-16% margin) vs. -30% to -32% guided and -42% a year ago — a ~2,600bp YoY improvement.
- Guidance raised across the board for FY26: core revenue $880–890M (from $855–865M), GM 41–43% (from 38–41%), OM -19% to -17% (from -28% to -32%). Q3 guided: revenue $214–216M, GM 38–40% (now called the trough, shifting from Q2), OM -25% to -23%.
- RPO now $25.4B (vs. ">$20B" OpenAI deal cited last quarter); explicitly excludes any AWS or other hyperscaler backlog. Management now projects >3x core revenue growth in 2027.
- Capacity land-grab quantified: >600MW secured (live or contracted through end-2027) across 13 named sites; pipeline "measured in gigawatts." Manufacturing capacity up >10x in 2026, with 3–4x more contracted for 2027.
- New since last call: AMD disaggregation partnership (Helios prefill + CS decode, 5x throughput at Cerebras speed), GPT-5.6 Sol support at claimed 10x speed, AWS GA timing set for Q1 2027 via Bedrock, CS4 unveiling next week, CS5 on track for H2 2027. Cash position $8.6B post-IPO.
What management is focused on- The "capacity, capabilities, customers" framework: 2026 is explicitly a foundation-building year for "massive 2027, 2028 and 2029," anchored on the $25.4B RPO.
- Data centers remain the binding constraint, but the tone has shifted from scramble to competence: "at the end of 2025, we were on a steep learning curve, today... we're pretty good at data center build-out with a clear path to becoming excellent." Inference needs smaller footprints than training clusters, giving siting flexibility.
- Disaggregation is now the central strategic narrative, not a side project: two partnerships (AMD Helios, AWS Trainium), framed as expanding the market for both parties — GPUs get access to fast inference; Cerebras gets 5x throughput per system, better token economics, and more tokens per MW in a power-constrained world.
- Roadmap aggression: double speed every year for several years (from a claimed 15x industry lead), >20x throughput increase within 18 months, CS4 reveal at Supernova next week, CS5 in H2 2027.
- TAM expansion via new verticals: security (CrowdStrike win — inline LLM traffic inspection "only exists if AI is fast"), coding (Figma, Cognition, Lovable), agentic enterprise (Block, AlphaSense, GSK). Six deals >$30M signed in Q2 outside the anchor customers.
- Margin story: Q3 is now the GM trough; recovery in Q4 as owned systems replace rented capacity; path to 60%+ GM reiterated with multiple drivers (pricing, rent-back roll-off, 20x throughput, 5nm wafer cost advantage, no HBM exposure).
Key numbers and quarter mechanics- Core revenue $209.9M (+103% YoY) vs. ~$194M guided — a ~$16M beat. Cloud/services $127.7M (+287%); hardware $82.1M (+17%). Mix shift to cloud driven by OpenAI ramp, other cloud usage growth, and hardware customers waiting on their own data center capacity.
- Core GM 40.6%, +940bp YoY, -590bp QoQ. Cloud/services GM 41.8% (+1,600bp YoY, but down from 52.9% in Q1 — the rent-back drag). Hardware GM 38.8% (+510bp YoY, vs. low-30s normalization guided last quarter — holding up better than signaled).
- Rent-back impact quantified: absent higher rented-capacity costs, GM would have been ~500bp higher (~45.6%, "more similar to last quarter"). Q3 guided 38–40% is the new trough call; significant Q4 improvement expected as owned systems come online.
- Core operating loss -$33.6M (-16%) vs. -$3.5M (-2%) in Q1 — the planned investment step-down, but far better than the -30% to -32% guide. Implied opex roughly $119M (vs. $92.6M in Q1), consistent with the flagged G&A step-up and capacity investment.
- RPO $25.4B at 6/30/26 — up from the >$20B OpenAI figure; excludes AWS/hyperscaler backlog entirely.
- Cash/securities/restricted cash $8.6B (vs. $3.3B pre-IPO) plus undrawn $850M revolver. CapEx advantage claimed: lower BOM cost (no third-party margin) plus partial reimbursement of data center fit-out via OpenAI pass-through.
- FY26 raise mechanics: revenue midpoint +$22.5M, GM midpoint +250bp, OM midpoint +1,000bp — the Q2 beat plus better H2 visibility. Q3 guide implies only ~2–3% sequential revenue growth, with growth still back-end loaded.
Product and launch scorecard- OpenAI / GPT-5.6 Sol: supported in Q2 at claimed 10x faster speed; management says this "lays to rest any remaining concerns" about frontier-model support. Note the version cadence — last quarter GPT 5.4 was live and 5.5 in progress; 5.5 was not mentioned this call. "Fast tokens with frontier intelligence are only available through OpenAI Cerebras partnership" — the exclusivity claim is now explicit.
- AMD disaggregation (new): announced "a few weeks ago"; Helios racks (prefill) + CS systems (decode), claimed 5x throughput at full Cerebras speed. Feldman confirmed the economics: Cerebras buys the Helios racks, installs in its own cloud, and keeps all service revenue — no revenue share. Claims "we have buyers for it already." Generic GPU disaggregation running in labs now, deployable Q4 2026.
- AWS: timing firmed — GA in Q1 2027 via Bedrock, deployed in Amazon data centers, deployments being organized now. Still no MW scale or economics disclosed. Feldman confirmed interest in eventually hosting Trainium+CS in Cerebras' own cloud or other clouds.
- CS-3/CS-4/CS-5: CS4 unveiling next week at Supernova; CS5 on track H2 2027. Claims: speed doubling annually, >20x throughput by end-2027, quadrupling of speed through 2027 (Komin). These are forward claims with no benchmarks disclosed this quarter (no Kimi-style demo this time).
- Customer breadth: six >$30M deals in Q2; named wins — Figma, Cognition, Lovable, Block, AlphaSense, GSK, CrowdStrike. Hardware pipeline described as "hundreds of millions of dollars" in late-stage deals plus significant 2027 cloud deals. Neocloud channel identified as a 2027 opportunity.
- Government R&D: partnerships for stacked memory and wafer-scale optical integration — long-dated, no revenue framing.
Sell-side read-through- Customer concentration was the lead question (Arcuri): he floated AWS at "$1 billion next year" and asked whether OpenAI + AWS = ~2/3 of 2027 revenue. Feldman did not dispute the magnitudes, said OpenAI stays "a meaningful portion" next year but shrinks as a percentage over time. No concentration numbers given — still deferred to filings.
- AMD economics pinned down (Bleustein for O'Malley): clean confirmation that Cerebras owns the racks and keeps all revenue. Bolton probed whether AMD-boosted throughput serving OpenAI creates incremental revenue — Feldman answered generally (more throughput = more revenue per system) and explicitly declined OpenAI specifics.
- Moore tested disaggregation readiness: answer — in labs now, deployed Q4. His follow-up extracted a useful nuance: partner silicon (Helios, Trainium 3) will outperform generic/NVIDIA installed-base pairings, but even N-1 generation GPUs are "vastly better" disaggregated than not — a lifecycle-extension pitch to GPU owners.
- Bolton's sharpest question: does the internal 20x throughput roadmap obviate disaggregation? Feldman's answer ("we're exploring all sorts of ways... that is what we're thinking about") was candid but non-committal — the tension between the proprietary roadmap and the partner strategy is unresolved.
- Buchalter's AWS economics question got structure (Bedrock delivery, Amazon data centers, Q1 live) but no revenue model detail — demand-based, "difficult to forecast" framing accepted.
- Rakesh's neocloud question opened a new 2027 channel narrative; his capacity question confirmed the 600MW/10x figures support 2027 acceleration. Notably, no analyst challenged the >3x 2027 growth claim, the 15x performance-lead claim, or the GPT 5.5 omission.
Management credibility- Beat-and-raise in only the second public quarter: beat all three guided metrics and raised all three FY26 lines — the deliberate-conservatism posture from Q1 is so far validated.
- The margin-trough call moved: Q1 call said Q2 (36–38%) was the trough with rental costs "fully baked in"; actual Q2 came in at 40.6% (above guide), yet Q3 is now called the trough at 38–40%. The trough migrated one quarter later — explained by increased rent-back volumes, but the prior quarter's "Q2 is the trough" framing did not hold as stated.
- Delivery claims remain verifiable and met: AWS definitive agreement (promised Q1, delivered), GPT-5.6 Sol support shipped, 600MW secured vs. the Q1 "dog fight" description, manufacturing 10x on track.
- Specificity improved materially on capacity (13 named sites, 600MW, gigawatt pipeline, 3–4x contracted 2027 manufacturing) and on AMD economics (full revenue retention). Still evasive on: AWS deal size/economics, customer concentration percentages, OpenAI contract specifics, benchmark configurations.
- The >3x 2027 revenue claim is new, bold, and RPO-anchored ($25.4B covers it arithmetically), but depends on data center delivery — the one input management admits it doesn't fully control.
- Version-naming inconsistency (5.5 skipped, 5.6 Sol now the flagship claim) is minor but worth noting given how much of the narrative rests on OpenAI model support cadence.
What changed versus the prior quarter- Guidance: FY26 raised on all lines (revenue $880–890M from $855–865M; GM 41–43% from 38–41%; OM -19%/-17% from -32%/-28%). Q2 actuals beat the Q1 guide on every metric.
- Margin trough shifted from Q2 to Q3 (38–40% guided); Q2 GM of 40.6% came in above the 36–38% guide but cloud margin fell from 52.9% to 41.8% on rent-back costs (~500bp total GM impact disclosed).
- Hardware margin held at 38.8% vs. the "low 30s" normalization signaled last quarter — better than flagged.
- AWS moved from "definitive agreement, 2027 revenue" to a concrete GA date (Q1 2027, Bedrock, Amazon data centers). AMD partnership is entirely new. Disaggregation went from one partner to two, with claimed buyers already in hand.
- Capacity disclosure stepped up: from the Bell Canada 120MW data point and geography list to >600MW secured across 13 named sites and a gigawatt-scale pipeline; manufacturing capacity quantified (>10x in 2026, 3–4x more contracted for 2027).
- RPO disclosed at $25.4B (vs. >$20B OpenAI deal previously); cash up to $8.6B from $3.3B post-IPO.
- New 2027 framing: >3x core revenue growth, hyperscaler revenue starting mid-2027, neocloud channel, security vertical (CrowdStrike) — none of these were in the Q1 narrative.
- OpenAI model support moved from GPT 5.4 live / 5.5 in progress to GPT-5.6 Sol supported at 10x speed; 5.5 not mentioned.
Bull case- Beat-and-raise with a $25.4B RPO book that excludes all hyperscaler business — the >3x 2027 growth claim is contractually underpinned, and AWS/Bedrock (Q1 2027 GA) plus other hyperscalers (mid-2027) are pure upside to that backlog.
- Disaggregation is a genuine economic unlock if the 5x throughput claim holds: more tokens per system, per watt, and per MW directly attacks the binding constraint (data centers), improves cloud margins, and creates a partner funnel (AMD, AWS, and "any GPU" installed base) that positions Cerebras as complementary to GPU estates rather than purely competitive.
- Capacity execution is now evidenced, not aspirational: 600MW secured, 13 sites, repeatable build-out process, 10x manufacturing ramp with TSMC wafer supply secured through next year — on a 5nm node with no HBM/CoWoS/3nm exposure while competitors face cost pressure.
- Customer diversification is showing early proof points: six >$30M deals, named wins across coding (Figma, Cognition, Lovable), enterprise agents (Block, GSK), and a new security category (CrowdStrike) that is structurally speed-dependent.
- Balance sheet ($8.6B + undrawn revolver) plus claimed CapEx-per-MW advantage (own BOM, OpenAI fit-out reimbursement) funds the land-grab without obvious financing risk; operating leverage demonstrated (-16% OM vs. -42% a year ago while doubling revenue).
Bear case- Concentration is intensifying before it diversifies: the Q2 beat and the 2027 triple both rest substantially on the OpenAI ramp, and Feldman conceded OpenAI remains "a meaningful portion" of 2027 revenue. Arcuri's 2/3-of-revenue framing (OpenAI + AWS) went uncontradicted. OpenAI's cloud-vs-hardware election still swings mix and pass-through.
- The margin trough slipped a quarter: Q1's "Q2 is the trough" became Q2's "Q3 is the trough." Rent-back volumes are growing, not shrinking, near term — the Q4 recovery and 2027 margin expansion depend on owned data centers arriving on schedule, the admitted bottleneck.
- The 20x internal throughput roadmap and the disaggregation strategy are in unresolved tension (Bolton's question); if internal throughput scales as claimed, the partner-rack economics could cannibalize or strand the Helios/Trainium integrations Cerebras is buying into with its own capital.
- AWS economics remain undisclosed with GA only in Q1 2027 and demand-based revenue; AMD deal has "buyers" but no disclosed commitments, timing, or revenue. Both are pre-revenue narratives carrying significant 2027–28 expectations.
- Sequential growth is stalling near term: Q3 guide of $214–216M is only ~2–3% QoQ, meaning the FY26 raise still requires a large Q4 ramp — execution risk concentrated in data center delivery.
- Performance claims (10x on GPT-5.6 Sol, 15x industry lead, 5x disaggregated throughput) are management-asserted without disclosed configurations; the "core" non-GAAP framework still excludes pass-through and warrant amortization, requiring continued reliance on GAAP reconciliations as OpenAI pass-through grows.
Next-quarter watchlist- Q3 delivery vs. guide: $214–216M revenue, 38–40% GM, -25% to -23% OM — and critically, whether Q3 is finally the GM trough as now promised, with Q4 recovery evidence (owned systems replacing rented capacity, cloud GM stepping back up).
- CS4 launch at Supernova (next week): specifications, speed/throughput benchmarks with disclosed configurations, pricing, availability — and whether the 20x throughput / annual speed-doubling roadmap gets concrete milestones.
- Data center execution: MW actually delivered vs. the 600MW contracted, Q4 capacity coming online on schedule, and pipeline conversion from the "gigawatts" of opportunities.
- AWS: deployment progress toward Q1 2027 Bedrock GA, any MW/economics disclosure, and whether AWS backlog begins to enter RPO. AMD: named buyers, deployment timing, revenue recognition, and any response to AMD's inference startup acquisition.
- OpenAI: revenue trajectory within the ramp, model support cadence (what happened to 5.5; 5.6 Sol adoption), and any signal on the cloud-vs-hardware election affecting pass-through.
- Disaggregation commercialization: the promised Q4 2026 generic-GPU disaggregated deployment — on time, with whom, and at what economics.
- Diversification evidence: conversion of the "hundreds of millions" late-stage hardware pipeline, additional >$30M deals, neocloud traction, and security-vertical follow-on beyond CrowdStrike.
- Filings to check: 10%+ customer concentration disclosure, GAAP-to-core reconciliation magnitudes (pass-through, warrant amortization), RPO composition, and rent-back cost trajectory vs. the ~500bp Q2 impact.
|
| Jun 23, 2026 | -18.79% | Q1 FY2026 | Read transcript briefingQuarter in one view- Cerebras' first earnings call as a public company (Q1 FY2026, reported 2026-06-23), weeks after what management calls the largest semiconductor IPO in history ($6.4B raised).
- Core (non-GAAP) revenue of $191.3M, +92% YoY: hardware $111.6M (+60%), cloud/services $79.8M (+167%). Core gross margin 46.5% (vs. 42.1% a year ago, 41% last quarter). Core operating loss near breakeven at -$3.5M (-2% margin).
- Two anchor partnerships dominate the story: a >$20B multi-year OpenAI compute agreement (signed 12/24/25, in production by 2/1 — 35 days signature-to-production) and a newly completed definitive agreement with AWS for disaggregated inference (Trainium 3 prefill + CS-3 decode) in AWS data centers.
- Guidance: Q2 core revenue ~$194M (+88% YoY) but gross margin dropping to 36–38% and operating margin to -30% to -32%; FY26 core revenue $855–865M (+69% at midpoint), GM 38–41%, OM -28% to -32%. Profitability deliberately regresses as the company rents back capacity to serve demand ahead of its own data center build-out.
- Cash position of $3.3B plus an $850M revolver (April '26) and the IPO proceeds — liquidity is not the issue; data center capacity is the stated binding constraint.
What management is focused on- Speed as the core thesis: Feldman argues fast inference is categorically more valuable ("fast tokens are the most valuable tokens"), claims Cerebras is >10x faster than GPUs "bar none," and frames the entire inference market as addressable — explicitly rebutting the "fast inference is only ~25% of the market" framing attributed to NVIDIA's CEO.
- Wafer-scale as a durable moat: 58x larger chips, SRAM instead of HBM, no CoWoS, no 3nm dependence — positioned as both a performance advantage and a supply-chain advantage. Feldman also floated longer-dated optionality (memory stacking, optical integration, data centers in space) that is vision, not near-term revenue.
- OpenAI and AWS execution: GPT 5.4 already running on Cerebras (OpenAI engineers + select customers); GPT 5.5 port in progress; AWS definitive agreement completed "as of this week," with revenue impact expected in 2027.
- Capacity land-grab: data centers added across US, Canada, France, Nordics; early discussions in Israel, UAE, Australia, Singapore, India, Indonesia; a 120MW Bell Canada deal cited. Feldman was blunt: "Demand is not the constraint. Supply is not the constraint. The constraint is data centers."
- New "core" non-GAAP reporting framework that strips out OpenAI data center pass-through (3% markup, reported gross) and warrant amortization — investors must track GAAP-to-core reconciliations carefully.
Key numbers and quarter mechanics- Core revenue $191.3M (+92% YoY); hardware $111.6M (+60%); cloud/services $79.8M (+167%).
- Core gross margin 46.5% (vs. 42.1% Q1'25, 41% Q4'25). Cloud/services margin 52.9% (up sharply on premium pricing for fast inference and better utilization); hardware margin 42% (vs. 30.6% Q1'25), flattered by performance-based incentive pricing recognized prospectively.
- Non-GAAP opex $92.6M (+51% YoY): R&D $69.8M, S&M $12.9M, G&A $9.9M (G&A steps up "significantly" next quarter on public-company costs).
- Core operating loss -$3.5M (-2% margin) vs. -$19.3M (-19%) a year ago and -10% in Q4'25; core net loss $2.5M.
- Cash/securities $3.3B; funding stack: $1B Series G (Sep '25), $1B Series H (Feb '26), $850M revolver (Apr '26), $6.4B IPO.
- Guidance mechanics to note:
- Hardware revenue planned to decline sequentially for the next few quarters as production shifts to internal cloud deployment.
- Hardware margin guided back to low 30s as incentive pricing normalizes.
- Cloud margin to drop 10–15 points temporarily due to renting systems back from an existing customer (G42, per analyst reference); Q2 rental costs are fully baked into guidance, and management indicated Q2 is the margin trough.
- FY26 growth is back-end loaded: OpenAI cloud ramp plus data centers coming online largely in H2; management expects increasing YoY growth rates each quarter of 2026.
- Long-term targets reiterated: ~60% gross margin, ~40% operating margin.
Product and launch scorecard- OpenAI (flagship): >$20B committed over several years; production 35 days after signature on a model Cerebras had never seen; GPT 5.4 live for OpenAI engineers and select customers; GPT 5.5 in active work. Feldman: only two hardware vendors currently serve OpenAI models, and Cerebras is one — a genuine validation proof point. OpenAI retains the option to take future commitments as hardware in its own data centers vs. Cerebras cloud, which creates real revenue-mix and pass-through variability.
- AWS (new, pre-revenue): binding term sheet in March, definitive agreement completed this week; disaggregated architecture (Trainium 3 prefill, CS-3 decode) claimed to be an order of magnitude faster. No revenue expected until 2027; scale (MW) and economics not disclosed; value split between Cerebras and Trainium explicitly unknown — Feldman conceded "nobody has yet deployed a true disaggregated solution, we have a lot to learn."
- CS-3 / wafer-scale platform: live demo of Kimi K2 (trillion-parameter open-source model) — 21 seconds vs. 4:37 on a "leading GPU" inference cloud (believed B300-based), claimed 13x. Demo conditions not independently specified (system count, batch, latency targets not disclosed).
- Cerebras Inference Cloud: demand "incredibly strong," premium pricing holding, utilization of late-2025 deployments ramped quickly; capacity being expanded via own builds plus the customer rent-back bridge.
- Additional disaggregated partners: management hinted at partnerships beyond AWS (decode-for-GPU-owners opportunity) but named none; Cerebras presence at Microsoft Build noted by an analyst, unconfirmed by management.
Sell-side read-through- Repeated pressure on TAM: three separate analysts (O'Malley, Bolton, Bryson) probed whether fast inference addresses the whole market or a premium niche, and whether premium token pricing limits adoption. Feldman's answer was philosophical (search/dial-up analogies) rather than quantitative — no TAM sizing, no willingness-to-pay data offered.
- Capacity and ramp scrutiny: Moore pinned down that data centers (not wafers, not demand) are the constraint; Buchalter extracted that H2 growth is mostly OpenAI cloud deployment and that management is "being conservative"; Rakesh's question about a "50MW/month ramp" was disowned by Feldman — that figure apparently came from elsewhere and was not confirmed.
- Disclosure gaps flagged in real time: customer concentration (10%+ customers) deferred to filings; AWS deal size (tens/hundreds of MW vs. GW) declined; Kimi benchmark configuration only partially answered.
- Margin quality questions: Bryson tested whether Q2 is the gross-margin trough (answer: yes, rental costs fully baked in); Moore's question revealed the GM beat drivers — higher pricing (including from existing customers), competitors' HBM-driven price floor rising, and more favorable capacity economics than modeled.
- Arcuri's supply question produced a clean answer: 2026 supply secured ("supply for our plan and beyond"), AWS impact is a 2027 event — setting a clear timeline expectation.
Management credibility- First call as a public company, so no public guidance track record; credibility rests on delivery claims that are at least partially verifiable: OpenAI production in 35 days, GPT 5.4 live, AWS definitive agreement completed when promised ("as of this week" after a March term sheet).
- Specificity was good on financials (margin bridge, rental impact quantified at 10–15 points, hardware margin normalization to low 30s, back-end-loaded year) and evasive on commercially sensitive items (AWS scale, customer concentration, benchmark specs).
- Feldman corrected the record rather than accept a friendly premise (Rakesh's 50MW/month question) — a small positive for candor.
- The "core" non-GAAP framework is aggressive in what it excludes (pass-through revenue/cost, warrant amortization contra-revenue); management says it matches internal decision-making and prior-period comparability, but investors should anchor on GAAP reconciliations, especially as pass-through grows.
- Guidance embeds deliberate conservatism (hardware declining, H2 conservative "as we're still pretty early"), and management pre-committed to updating as data center timing firms — a testable promise for coming quarters.
- Grandiose framing (TAM "thousands of times" larger, data centers in space) is promotional; the near-term claims, by contrast, were mostly concrete and internally consistent.
What changed versus the prior quarter- No prior-quarter summary was provided; this transcript is the baseline. Sequential data points available within the call: core gross margin 46.5% vs. 41% in Q4'25; core operating margin -2% vs. -10% in Q4'25. The AWS relationship moved from binding term sheet (March) to definitive agreement (this week). Everything else should be treated as the starting baseline for future comparisons.
Bull case- >$20B OpenAI commitment plus an AWS definitive agreement gives multi-year revenue visibility rare for a newly public company; FY26 guide of +69% is back-end loaded with identifiable drivers (OpenAI cloud ramp, H2 data center capacity).
- Structural supply-chain advantage: no HBM, no CoWoS, no 3nm dependence — the three industry bottlenecks — while competitors face rising costs that are lifting the pricing floor; Cerebras is simultaneously raising prices.
- Demonstrated frontier-model credibility: one of only two hardware vendors serving OpenAI models; 35-day bring-up; GPT 5.5 in progress; disaggregated decode positions Cerebras as complementary (not just competitive) to GPU/Trainium estates, expanding the partner funnel.
- Massive liquidity ($3.3B cash + $850M revolver + IPO proceeds) to fund the capacity land-grab; operating leverage already visible (opex +51% vs. revenue +92%); near-breakeven core operating income despite hypergrowth.
- Margin regression is self-inflicted and temporary (rent-back bridge), with a defined path back to 60% GM / 40% OM targets.
Bear case- Customer concentration is extreme and undisclosed on the call: OpenAI is the growth engine, holds the cloud-vs-hardware option (which swings revenue mix and pass-through), and its decisions are "outside of our control."
- FY26 guidance implies a sharp profitability reversal (Q2 OM -30% to -32% vs. -2% in Q1) and depends on H2 data centers coming online on schedule — the one resource management admits is a "dog fight" and the stated binding constraint. Slippage directly hits the back-end-loaded guide.
- AWS contributes nothing until 2027, deal size and economics undisclosed, and the disaggregated architecture is unproven in production anywhere — Feldman admitted the value split is unknown and traffic-shape mismatch can strand compute.
- Hardware revenue and hardware margin are both guided down (revenue declining for several quarters, margin to low 30s), meaning near-term mix shifts toward a cloud business whose margin is about to drop 10–15 points.
- Premium-token pricing thesis is asserted, not evidenced; if fast-inference premiums compress as GPU inference improves, the pricing tailwind behind the 52.9% cloud margin reverses.
- Heavy reliance on non-GAAP "core" framing (excluding pass-through and warrant amortization) can obscure GAAP economics as OpenAI pass-through grows.
Next-quarter watchlist- Q2 delivery vs. guide: ~$194M core revenue, 36–38% GM, -30% to -32% OM — confirm Q2 is indeed the margin trough and that rent-back costs landed as modeled.
- Data center milestones: new capacity promised for Q3 and Q4; watch MW added, geographies, and whether the Bell Canada 120MW facility and other deals progress on schedule.
- OpenAI ramp: revenue contribution trajectory, GPT 5.5 timing on Cerebras, and any signal on OpenAI's cloud-vs-own-data-center election (drives pass-through and mix).
- AWS: technical collaboration milestones, deployment prep, MW scale disclosure, and any 2027 revenue framing; also whether Amazon selling Trainium externally becomes a Cerebras channel (Feldman said he views it as an opportunity).
- Margin path: confirmation that Q2 (36–38% GM) is the trough, rent-back volumes vs. plan, hardware margin glide to low 30s, and progress back toward 60% cloud margin.
- Disclosures to check in filings: 10%+ customer concentration, GAAP-to-core reconciliations (pass-through magnitude, warrant amortization), and any AWS commitment sizing.
- Pipeline breadth: evidence of customer diversification beyond OpenAI/AWS into 2027, additional disaggregated-inference partners (Microsoft Build signal), and data center deals closing in the discussed geographies (Israel, UAE, Australia, Singapore, India, Indonesia).
|
Corbanu