Three Scaling Laws, One Supply Constraint: Data Center +112%, Blackwell Ramping Ahead of Plan, Q4 Guide $37.5B — Upgrading to Outperform

Rating Action: Walking the Q2 Initiation to Today

We initiated coverage on NVIDIA at Hold in our Q2 FY2025 note with a constructive bias. The Hold framing was deliberate: NVDA at Q2 was the most operationally validated AI compute platform on the planet, but we wanted execution evidence on three specific concerns before committing capital at the multiples the stock was already trading at. Those three concerns were:

1. Blackwell ramp execution. The mask change announced earlier in calendar 2024 was the highest-stakes transition NVIDIA had attempted at scale, with seven custom chips per system, multiple form factors (NVLink 8/36/72, x86/Grace, air- and liquid-cooled), and integration with the world's most demanding hyperscaler data center architectures. Execution slippage, even by one quarter, could meaningfully de-rate the FY26 trajectory.
2. Gross-margin trajectory. The Q2 print already showed early softening into the high-70s on Hopper-mix shift. The Blackwell ramp was expected to compress GM further during the steep portion of the ramp curve. The question was how deep, how durable, and whether management would frame it in a way investors could underwrite.
3. Demand-cycle durability. The "AI scaling has stalled" narrative was incipient at Q2 (and has since gained more sell-side oxygen). If pre-training scaling truly plateaued, the multi-year compute trajectory would compress and the Hopper installed base would face a digestion period rather than a continuous expansion cycle.

Q3 FY2025 delivered concrete, well-quantified evidence on all three:

• Blackwell ramp: 13,000 samples shipped (zero in Q2), Q4 Blackwell revenue now expected to exceed the previously stated "several billion dollars" estimate, system stand-ups underway at Microsoft, Oracle, Google, Dell, CoreWeave, Meta. The most aggressive supply-chain ramp NVIDIA has ever executed is on track. Concern (1): retired.
• GM trajectory: Q4 guide 73.5% (down from 75% in Q3); CFO explicitly framed the trough as "71%, maybe about 72%, 72.5%" and the recovery to mid-70s as "reasonable" in 2H calendar 2025. The dip is bounded, telegraphed, and tied to a specific operational cause (Blackwell yield ramp + multi-config complexity). Concern (2): bounded and underwritable.
• Demand-cycle durability: Jensen's three-scaling-law framework plus the explicit "next generation starts at 100,000 Blackwells" comment vs. the 100,000-Hopper end-state of the prior generation reframes the demand cycle as multiplicatively expanding rather than digesting. Sovereign AI (India 10x GPU growth by year-end, Japan SoftBank/NTT/Fujitsu, US T-Mobile AI-RAN), Enterprise AI (1,000+ companies on NIM, Accenture 30,000-strong NVIDIA AI practice, AI Enterprise revenue 2x+ YoY exiting at $2B+ ARR), and physical AI/robotics layer on as orthogonal vectors. Concern (3): materially de-risked.

The Q2 initiation at Hold was calibrated against multi-billion-dollar execution risk on Blackwell and a non-trivial probability that the demand cycle would enter digestion. Q3 retires both. The thesis that justified caution at Q2 has been replaced with operational evidence that justifies conviction. We move to Outperform.

Results vs. Consensus

Q3 FY2025 cleared the company's own $32.5B outlook by ~$2.6B and beat the most bullish published Street estimates. Every line item printed above plan; the only material softness anywhere in the model was networking sequential decline, which management framed as a transient timing item with sequential growth resuming in Q4.

Quality of Beat

• Revenue: The $2.6B beat versus the company's own outlook is fully organic and broad-based across every market platform. Data Center alone delivered ~$1.6B above the implicit pace embedded in the outlook; Gaming, ProViz, and Automotive each contributed incremental upside. The +17% sequential is the largest in-cycle dollar increase since the original Hopper ramp and demonstrates that the demand backlog is not yet "filling" — it continues to expand faster than supply can be ramped.
• Data Center composition: CSPs ~50% of Data Center, with CSP revenue more than doubling YoY. Regional GPU cloud revenue 2x YoY across North America, EMEA, and Asia-Pacific. Consumer Internet revenue more than doubled YoY on next-gen multimodal and agentic AI training plus generative AI inference. Networking declined sequentially — flagged as transient with InfiniBand, Ethernet switches, SmartNICs, and BlueField DPUs all growing sequentially individually; the aggregate networking number was timing-driven and sequential growth resumes Q4. Spectrum-X Ethernet for AI revenue was up 3x+ YoY with a building pipeline.
• Margins: 75.0% non-GAAP GM is down sequentially from Q2's 75.7%, driven primarily by mix-shift toward more complex and higher-cost H100 systems within Data Center. This is the early phase of the natural product-cycle GM compression we underwrite for the Blackwell transition. The Q4 guide of 73.5% extends the trajectory in the expected direction; management was clear that the trough is in the low-70s during the steep ramp and the recovery to mid-70s is "reasonable" by 2H calendar 2025.
• Capital return: $11.2B returned to shareholders in Q3 between buybacks and dividends. NVIDIA continues to demonstrate that it can simultaneously fund the most capital-intensive product transition in its history and return cash at scale — a function of the cash-flow generation that the demand-supply imbalance has produced.
• Segment breadth: The fact that Gaming (+15% YoY), ProViz (+17%), and Automotive (+72%, record) all printed solid prints alongside the Data Center +112% means this is not a single-segment story. The non-Data-Center segments collectively annualize at ~$17B with double-digit composite growth and provide a non-trivial base of revenue durability under the Data Center cycle.

Segment Performance

Data Center: H200 in Double-Digit Billions, Blackwell On Deck

Data Center at $30.8B (+112% YoY, +17% QoQ) is the most operationally important number on the print. Three things matter: (1) H200 sales scaled "significantly to double-digit billions" in a single quarter — management called it "the fastest product ramp in our company's history," with the chip delivering up to 2x faster inference and 50% improved TCO vs. H100; (2) CSPs were ~50% of Data Center with CSP revenue more than doubling YoY, demonstrating that the largest-buyer segment is still in expansion mode rather than digestion; (3) the "tail-end of the last generation" was 100,000-Hopper clusters; the next generation starts at 100,000-Blackwell — an order-of-magnitude floor expansion that materially raises the multi-year demand trajectory.

"NVIDIA Hopper demand is exceptional and sequentially, NVIDIA H200 sales increased significantly to double-digit billions, the fastest product ramp in our company's history." — Colette Kress, CFO

Blackwell is in full production after a successfully executed mask change, with 13,000 GPU samples shipped to customers in Q3 (including the first Blackwell DGX engineering sample to OpenAI). Microsoft will be the first CSP to offer Blackwell-based instances in private preview powered by GB200 + Quantum InfiniBand. Oracle announced the world's first Zettascale AI Cloud computing clusters scaling to over 131,000 Blackwell GPUs. Dell, CoreWeave, Google, Meta systems are all standing up. The most operationally significant call comment on Q4: management explicitly said Blackwell will exceed the previously stated "several billion dollars" Q4 estimate.

"Blackwell production is in full steam... we will deliver this quarter more Blackwells than we had previously estimated. And so the supply chain team is doing an incredible job working with our supply partners to increase Blackwell." — Jensen Huang, CEO

The MLPerf Training benchmark sweep is the third-party performance validation: Blackwell delivered 2.2x per-GPU performance vs. Hopper, and on the GPT-3 training benchmark just 64 Blackwells matched the throughput of 256 H100s — a 4x reduction in compute footprint. NVLink Switch enables up to 30x faster inference, especially valuable for reasoning models like OpenAI's o1.

Networking footnote: Networking revenue declined sequentially, but management was emphatic that this is transient. InfiniBand, Ethernet switches, SmartNICs, and BlueField DPUs all grew sequentially individually; the aggregate decline was timing-driven, and sequential growth resumes Q4. Spectrum-X Ethernet for AI revenue grew 3x+ YoY; xAI's Colossus 100,000-Hopper supercomputer using Spectrum-X achieved zero application latency degradation and 95% data throughput vs. 60% on traditional Ethernet — the operational case for Spectrum-X is now production-validated at the largest training cluster in the world.

Assessment: Data Center is in the strongest position we have observed in this cycle. H200 is ramping faster than any product in NVIDIA history; Blackwell ships in volume in Q4 ahead of plan; the next-gen scale floor has expanded by an order of magnitude; networking is on track to resume sequential growth. The demand-supply imbalance is structural and persistent.

Gaming: Solid Quarter, Q4 Supply-Constrained

Gaming at $3.3B (+15% YoY, +14% QoQ) was a strong back-to-school quarter with notebook, console, and desktop all up sequentially and YoY. Channel inventory is healthy. RTX AI PCs began shipping with up to 321 AI tops from ASUS and MSI, with Microsoft Copilot+ capabilities anticipated in Q4. The 25th anniversary of the GeForce 256 (the world's first GPU) was noted in the call — a reminder of the institutional positioning the franchise still occupies.

Q4 wrinkle: Management guided Gaming revenue to decline sequentially in Q4 due to supply constraints — the supply allocation toward Data Center is biting. Atif Malik (Citi) pressed Colette on whether this was Data Center prioritization; she framed it as a ramp-pacing issue ("how fast could we get that supply getting ready into the market") with normalization expected by the calendar year start. We treat this as a timing issue, not a demand-side concern.

Assessment: Gaming is annualizing at ~$13B and is structurally durable. The Q4 sequential dip is supply-driven and likely backfills in Q1 calendar 2025. The AI-PC vector adds a non-trivial upside option.

ProViz: AI as a New Demand Driver

ProViz at $486M (+17% YoY, +7% QoQ) is the smallest segment by revenue but is showing meaningful demand-mix evolution. RTX workstations remain the primary driver, but management called out AI specifically as "emerging as a powerful demand driver" — autonomous-vehicle simulation, generative-AI model prototyping, and gen-AI content creation in media and entertainment. The segment is functionally becoming an AI workstation business rather than a traditional CAD/design business.

Assessment: Modest revenue contribution but high optionality on the AI-augmented enterprise workflow vector.

Automotive: Record Quarter, +72% YoY

Automotive at $449M (+72% YoY, +30% QoQ) is the strongest YoY growth segment and printed an all-time record. Driven by Orin self-driving ramps and robust NEV end-market demand. Volvo's fully-electric SUV (the EX90) launched on Orin + DriveOS in the quarter. The segment is annualizing at ~$1.8B and is the most credible non-Data-Center growth franchise NVIDIA operates.

Assessment: Automotive is now growing at a rate that materially supports the FY26+ thesis even if Data Center growth normalizes. With Drive Thor on the roadmap and additional OEM design wins continuing to accumulate, this segment is increasingly a real second-leg of the growth profile.

Key Topics & Management Commentary

Overall management tone: Confident bordering on euphoric on the demand side, measured and operationally specific on the supply-chain and margin sides. Jensen's framing of three concurrent scaling laws and the "AI factories" metaphor was the most expansive multi-year demand framework he has articulated in this cycle. Colette's GM trough specificity ("71%, maybe about 72%, 72.5%") was the most concrete margin-trajectory guidance management has offered through the Blackwell transition.

Three Scaling Laws: Pre-training Intact, Post-training and Test-time Additive

C.J. Muse (Cantor) opened the Q&A with the most pointed question: is foundation-model pre-training scaling stalling? Jensen's response was the most intellectually substantive part of the call. Pre-training scaling, he said, is empirical rather than physical-law, and the evidence is that it continues. But two additional scaling vectors have emerged: post-training scaling (RLHF, RLAIF, synthetic-data generation) and test-time scaling (catalyzed by OpenAI's o1/Strawberry, where longer model "thinking time" produces higher-quality outputs). All three scaling vectors independently consume compute and operate concurrently.

"Foundation model pre-training scaling is intact and it's continuing... we now have three ways of scaling and we're seeing all three ways of scaling. And as a result of that, the demand for our infrastructure is really great." — Jensen Huang, CEO

"At the tail-end of the last generation of foundation models were at about 100,000 Hoppers. The next generation starts at 100,000 Blackwells." — Jensen Huang, CEO

Assessment: The three-scaling-law framework is the most direct rebuttal of the "AI scaling has stalled" sell-side narrative we have seen articulated by an operating CEO. The 100,000-Hopper-to-100,000-Blackwell floor expansion is operationally specific and difficult to dismiss as marketing — it implies an order-of-magnitude expansion in compute consumption per next-gen-model project. Combined with the inference-side scaling that the o1 reasoning paradigm catalyzes, the multi-year compute-demand TAM is structurally larger than what the digestion-cycle narrative assumes.

Blackwell Supply Chain: Seven Custom Chips, "Almost Every Company in the World" Involved

Toshiya Hari (Goldman) asked about reported heating issues, the GTC roadmap (Ultra in 2025, Ruben in 2026), and supply constraints (HBM specifically). Jensen's response was operationally exhaustive: seven custom chips per Blackwell system, four form factors (NVLink 8/36/72 + x86/Grace), air- and liquid-cooled options, and a supply-chain partner list including TSMC, Amphenol, Vertiv, SK Hynix, Micron, Amkor, KYEC, Foxconn, Quanta, Wiwynn, Dell, HP, Super Micro, Lenovo. Q3 Blackwell shipments: zero. Q4 Blackwell shipments: measured in billions. The annual roadmap (Ultra 2025, Ruben 2026) is on track.

"There are seven different chips, seven custom chips that we built in order for us to deliver the Blackwell systems... almost every company in the world seems to be involved in our supply chain." — Jensen Huang, CEO

Assessment: The supply-chain breadth Jensen described is itself a moat — it would take a competitor years to replicate the multi-vendor coordination NVIDIA has institutionalized. The annual-cadence roadmap (Hopper-Blackwell-Ultra-Ruben) is now the industry's de facto pace-setter; competitors race against NVIDIA's release schedule. The reported heating issues were not addressed directly but the operational framing ("Blackwell production is in full steam... we will deliver this quarter more Blackwells than we had previously estimated") is the strongest possible rebuttal.

Inference: Largest Installed Base, Test-Time Scaling Tailwind

Joseph Moore (Morgan Stanley) probed inference vs. training mix and whether aging Hopper clusters could service inference workloads. Jensen's response: NVIDIA is the largest inference platform in the world today; inference is "super hard" because it requires high accuracy + high throughput + low latency simultaneously, with growing context lengths and multimodality. Hopper inference throughput improved 5x in one year through software (CUDA, NIM); the upcoming NIM release adds another 2.4x.

"We are the largest inference platform in the world today because our installed base is so large and everything that was trained on Amperes and Hoppers inference incredibly on Amperes and Hoppers." — Jensen Huang, CEO

Assessment: The inference-side framing is structurally bullish. As Hopper trains-then-inferences cycles play out, the installed base becomes a recurring inference cash-flow source. Test-time scaling on o1-class reasoning models is incremental compute consumption layered on top. The CUDA software-stack moat is the central reason this dynamic accrues to NVIDIA rather than to merchant competitors.

Sovereign AI: Non-trivial New Vertical

Atif Malik (Citi) asked for an update on sovereign AI, which Colette had previously framed as low-double-digit billions. Q3 update: India's CSPs (Tata Communications, Yotta) building AI factories with tens of thousands of NVIDIA GPUs, with year-end deployments boosting India NVIDIA GPU footprint by ~10x. Infosys, TCS, Wipro adopting NVIDIA AI Enterprise and upskilling ~500,000 developers. Japan: SoftBank building the nation's most powerful AI supercomputer with DGX Blackwell + Quantum InfiniBand, plus AI-RAN partnership for telco. T-Mobile rolling out the same AI-RAN architecture in the US. Fujitsu, NEC, NTT all adopting NVIDIA AI Enterprise.

Assessment: Sovereign AI is now a structurally meaningful demand vector. The geographic breadth (India, Japan, US telco, Europe, Asia-Pac regional clouds) materially diversifies the customer base away from US hyperscaler concentration. The pipeline framing ("absolutely intact") suggests this is multi-year incremental rather than one-time.

Enterprise AI: Agentic Adoption in Full Throttle

Colette's prepared remarks were unusually specific on enterprise AI: nearly 1,000 companies using NVIDIA NIM; AI Enterprise revenue 2x+ YoY this year, exiting at $2B+ ARR; software/service/support business annualizing at $1.5B and exiting at $2B+. Cadence, Cloudera, Cohesity, NetApp, Nutanix, Salesforce, SAP, ServiceNow listed as building agentic AI on NVIDIA. Accenture launching a 30,000-strong NVIDIA AI practice; using agentic AI internally to cut marketing-campaign manual steps by 25-35%. Foxconn using Omniverse digital twins for Blackwell factory bring-up, projecting 30% kilowatt-hour reduction in Mexico facility.

Assessment: The enterprise software business is no longer a rounding error. $2B+ exiting ARR is real revenue with high incremental margins. The Accenture practice is the most credible enterprise distribution channel NVIDIA has constructed. Combined with Omniverse for industrial AI, the enterprise vertical is increasingly a secondary growth engine alongside Data Center hyperscaler infrastructure.

China: Sequentially Up, Structurally Below Pre-Controls

Data Center revenue in China grew sequentially due to shipments of export-compliant products. As a percentage of total Data Center revenue, China remains "well below levels prior to the onset of export controls." Management characterized the China market as "very competitive" and committed to continuing compliance with export controls.

"From a geographic perspective, our Data Center revenue in China grew sequentially due to shipments of export-compliant copper products to industries. As a percentage of total Data Center revenue, it remains well below levels prior to the onset of export controls." — Colette Kress, CFO

Ben Reitzes (Melius) asked about the incoming US administration and tariffs/China implications. Jensen's response was deliberately measured: NVIDIA will support whatever the administration decides, comply with regulations, and compete. No specific tariff or policy color was offered.

Assessment: China is a managed variable rather than an existential thesis driver. The export-compliant product line is producing sequential growth at lower margins; the structural reduction from pre-controls levels is the cost. The post-election regulatory environment is the relevant near-term policy variable to monitor.

Gross Margin Trajectory: Trough Quantified, Recovery Specified

Tim Arcuri (UBS) and Vivek Arya (BofA) both pressed Colette on the GM trajectory through the Blackwell ramp. Colette's response was the most operationally specific GM commentary management has offered: low-70s during the steep ramp (specifically "71%, maybe about 72%, 72.5%"), with the recovery to mid-70s "reasonable" by 2H calendar 2025 as Blackwell yields and mix mature.

"Low, of course, is below the mid, and let's say we might be at 71%, maybe about 72%, 72.5%, we're going to be in that range. We could be higher than that as well... we'll get up to the mid-70s by that point." — Colette Kress, CFO

Assessment: The GM trough is bounded and the recovery cadence is specified. This is the most underwritable margin-trajectory guidance NVIDIA has provided through any major product transition. Investors should model FY26 GM in the 71-75% range with the trough in the steep-ramp quarters and the recovery layered on as yields mature.

Q4 FY2025 Guidance

The $37.5B midpoint is meaningfully above the pre-print Street consensus and is the most concrete operational commitment yet on the Blackwell ramp. Two specific framings matter: (1) Blackwell revenue in Q4 will exceed the previously-stated "several billion dollar" estimate — management has gained supply visibility and is upsizing the trajectory; (2) Hopper demand "will continue through next year, surely the first several quarters of the next year" — meaning Hopper is not stepping off the ramp as Blackwell ramps on, but the two are running concurrently. This is an additive-not-substitutive transition pattern, which is the most cash-flow-favorable architecture possible for a major product cycle.

FY26 framing implied: Management did not provide an FY26 revenue framework on this call (full FY guide typically arrives at Q4 print). But the Q4 guide and the Blackwell trajectory framing implies FY25 lands near $130B (vs. the $60B FY24 base) and FY26 sets up at materially higher levels with Blackwell at scale plus Hopper still running plus Sovereign AI plus Enterprise AI plus Automotive. We model FY26 in the $185-210B range pending the formal guide.

Analyst Q&A Highlights

Has AI Pre-Training Scaling Stalled?

• C.J. Muse, Cantor Fitzgerald: Opened with the question that has dominated sell-side debate — whether LLM scaling has stalled, and whether Blackwell demand is being driven by clusters that have yet to benefit from new architecture. Jensen's response was the most intellectually substantive of the call, articulating three concurrent scaling laws and the 100,000-Blackwell next-gen floor.
  Assessment: This was the most important exchange in the Q&A. Jensen's three-scaling-law framing reframes the "scaling has stalled" debate from a binary (stalled/not stalled) to a multi-vector (pre-train + post-train + test-time, all scaling concurrently). The 100,000-Blackwell floor on the next-gen model project is operationally specific and difficult to dismiss as narrative. Bear desks running the digestion-cycle thesis now have a higher bar to clear.

Blackwell Execution Risk and Roadmap

• Toshiya Hari, Goldman Sachs: Asked about reported Blackwell heating issues, ability to execute the GTC roadmap (Ultra 2025, Ruben 2026), and supply-chain constraints (HBM specifically). Jensen confirmed the annual roadmap is on track and described the seven-custom-chip Blackwell architecture and the multi-vendor supply chain in operational detail.
  Assessment: The supply-chain breadth Jensen described is itself a competitive moat. The annual roadmap cadence (Hopper-Blackwell-Ultra-Ruben) sets the industry pace. The reported heating issues were not addressed directly but the framing of "production in full steam" and Q4 shipments exceeding prior estimates is the strongest possible rebuttal.

Blackwell-Hopper Crossover and Margin Trajectory

• Timothy Arcuri, UBS: Asked when Blackwell crosses over Hopper (April quarter previously indicated) and whether April is the worst of the gross margin pressure. Colette confirmed low-70s GM trough, with mid-70s recovery as the ramp matures. Jensen confirmed Hopper continues through next year, with Blackwell ramping each subsequent quarter.
  Assessment: The additive Hopper-Blackwell concurrent ramp pattern is a cash-flow-positive transition architecture. The GM trough is bounded. This is the cleanest answer to the "FY25 transition will mechanically compress GM and earnings" bear concern.

Digestion Risk and Hardware Cycle Pattern

• Vivek Arya, BofA: Probed historical hardware cycle digestion patterns and when this cycle reaches that phase. Jensen's response: "no digestion until we modernize a trillion dollars" of installed-base data center infrastructure, with the modernization runway extending several years and a parallel new-industry creation in AI factories.
  Assessment: The "trillion-dollar IT modernization + AI factory new-industry creation" framing is the most expansive multi-year demand narrative we have heard from an operating CEO. Whether literally accurate or directionally right, it argues that the digestion phase is multiple years away — not a calendar 2025/2026 concern.

Hopper-Blackwell Transition into Q4

• Stacy Rasgon, Bernstein: Pressed on the precise Hopper-vs-Blackwell mix into Q4, given Blackwell several-billion-plus and aggregate $37.5B guide. Colette indicated H200 substantial growth continues, with Hopper potentially up sequentially Q3-to-Q4 even as Blackwell ramps.
  Assessment: The Hopper-up-sequentially-while-Blackwell-ramps framing is the operationally cleanest version of an additive product transition. Both products contributing to growth concurrently is structurally rare and very cash-flow-favorable.

Inference Outlook and Hopper Repurposing

• Joseph Moore, Morgan Stanley: Asked about inference market dynamics and whether aging Hopper clusters could service inference loads. Jensen described inference as "super hard" (high accuracy + throughput + low latency simultaneously), with Hopper inference throughput already improved 5x in one year and 2.4x more from upcoming NIM.
  Assessment: The inference framing is structurally bullish. As Hoppers cascade from training to inference duty over time, the installed base becomes a recurring revenue source via NIM and CUDA software optimizations. The CUDA moat is the central reason this dynamic accrues to NVIDIA rather than to merchant competitors.

Networking Dip and Spectrum-X Pipeline

• Aaron Rakers, Wells Fargo: Asked about the sequential networking decline against the strong demand framing. Colette: timing-driven dip, multiple CSP design wins, sequential growth resumes Q4.
  Assessment: Networking is a watch item, not a thesis problem. The timing-driven framing is plausible given the Blackwell prep dynamic at customers. Spectrum-X 3x YoY revenue and the xAI Colossus production validation are the operational evidence that the architecture is working.

Sovereign AI Pipeline Update

• Atif Malik, Citi: Asked for an update on sovereign AI (previously framed as low-double-digit billions) and the gaming supply situation. Colette confirmed sovereign pipeline "absolutely intact" with India 10x GPU expansion and Japan SoftBank/Fujitsu/NEC/NTT deployments; gaming Q4 dip is supply-driven and resolves into calendar 2025.
  Assessment: Sovereign AI is no longer a story; it is a measurable demand vector with multi-year runway. Gaming Q4 supply constraint is benign timing.

Q1 Sequential Trajectory and Tariffs/China

• Ben Reitzes, Melius: Asked about Q1 sequential trajectory implied by Blackwell catch-up and incoming US administration tariff/China policy. Jensen declined to extend guide beyond one quarter and gave a measured response on regulatory compliance.
  Assessment: The one-quarter guide discipline is consistent with NVIDIA's historical posture. The tariff/China response is appropriately deferred until policy specifics emerge.

Compute Mix Across Pre-train, Post-train, Inference

• Pierre Ferragu, New Street: Asked for a high-level compute-mix split between pre-training, post-training (RL), and inference today. Jensen: vastly pre-training today, with post-training and inference both scaling concurrently and multimodality foundation models trained on petabytes of video continuing the pre-training compute trajectory.
  Assessment: The "vastly pre-training today" framing rebuts the implicit assumption that pre-training is at end-of-life. Multimodality + petabytes-of-video training implies the compute-per-model trajectory continues to expand even as post-training and inference layers add incremental demand.

What They're NOT Saying

1. Explicit FY26 revenue framework: Management did not provide a forward-year framework. Understandable timing (formal FY guide arrives at Q4 print in February), but the implication of the Blackwell trajectory and three-scaling-law framing is materially higher than current sell-side modeling. The market will fill the gap with bullish bottom-up models; the asymmetry is that consensus may end up tracking below where management ultimately formalizes.
2. Hyperscaler-specific Blackwell pre-orders: The "exceeds several billion dollars" framing is intentionally non-specific. We know Microsoft is the first CSP in private preview, Oracle has the 131,000-Blackwell zettascale design announced, and Google/Dell/CoreWeave/Meta systems are standing up. But unit-level commitment by customer is not disclosed and likely won't be.
3. Customer concentration disclosure: The "CSPs ~50% of Data Center" framing leaves the other 50% (regional cloud + consumer internet + sovereign + enterprise) opaque on a per-customer basis. The 10-K will provide the formal customer concentration disclosure for FY25.
4. China revenue specificity: "Well below pre-controls levels" is qualitative. We do not know the exact China contribution to Q3 Data Center, nor the trajectory implied for Q4. Sequential growth was acknowledged but not quantified.
5. Heating/yield issues with Blackwell: The reported heating issues from press accounts were not addressed directly. Management's "production in full steam" and "exceeds prior estimates" framing implies any issues are bounded, but explicit acknowledgment was avoided.
6. Inference vs. training revenue mix: Jensen described inference as a structurally large and growing portion of the demand, but no specific mix percentage was offered. As inference scales with the o1 reasoning paradigm and the AI-native company proliferation, this mix will become an increasingly important variable.
7. Capex / R&D trajectory: Q4 OpEx guide implies +9% sequential, consistent with continued R&D scale-up. But no multi-year framework was offered. As NVIDIA increasingly takes on full-stack data-center scope, the capital intensity of the model may evolve.

Market Reaction

• After-hours move (Nov 20 evening): NVDA traded modestly lower in the immediate after-hours session despite the ~8% beat versus company guide. The pattern is consistent with the "in-line is a disappointment" dynamic that has characterized AI complex prints during periods of elevated expectations — with NVDA at all-time-high market cap heading into the print, the bar required for upside reaction was a guide materially above $37.5B, not at $37.5B.
• Pre-print context: NVDA had rallied substantially through the calendar year and entered the print as one of the largest market-cap companies in the world. Forward P/E was elevated relative to historical NVDA standards. The Blackwell-mix GM compression to 73.5% in Q4 was within Street range but at the lower end — some sell-side desks had been modeling 74-74.5%.
• Cited reasons for muted post-print response: (1) Q4 GM guide of 73.5% at lower end of Street range; (2) Blackwell "exceeds previously stated several billion dollars" was operationally meaningful but not as quantitatively specific as some bulls had hoped; (3) reported Blackwell heating issues remained an overhang absent direct management address; (4) lofty pre-print valuation amplifying the "sell the news" dynamic; (5) broader AI complex sentiment digestion concerns from sell-side note flow.
• Volume: Materially elevated on the AH session.

The muted post-print reaction is structurally typical for NVDA in this cycle and is not a thesis signal. The print operationally clears the bar on every concern we listed at our Q2 initiation. The price action over the next several weeks will be driven by positioning dynamics and macro-AI-narrative flow rather than by any change in NVIDIA's operating trajectory. This is the kind of "fundamental beat, muted price reaction" setup that historically produces favorable risk/reward when the underlying operating story is intact and visibility is expanding — both of which are clearly the case here.

Street Perspective

Debate: Is Pre-Training Scaling Actually Stalling?

Bull view: Pre-training scaling is empirical, not physical-law, and the empirical evidence (per Jensen) is that it continues. Multimodality foundation models trained on petabytes of video maintain the per-model compute trajectory; post-training and test-time scaling are additive vectors. The 100,000-Hopper-to-100,000-Blackwell next-gen-model floor expansion is operationally specific and difficult to dismiss. Bear desks running the "scaling has stalled" thesis are essentially betting against the data Jensen is seeing in production with his largest customers.

Bear view: Recent reporting on the leading frontier-model labs has suggested that incremental capability gains from pre-training scaling may be saturating, and that the o1 reasoning paradigm may be a substitute for rather than an additive layer to pre-training scaling. If true, the per-model compute trajectory could compress as model labs reallocate spend from pre-training to inference-time compute — a shift that may favor inference-optimized chips over training-optimized chips and could compress NVIDIA's training-revenue trajectory.

Our take: The bull is meaningfully more right than the bear. Three concurrent scaling vectors are demonstrably growing in absolute compute terms; substitution risk is real but bounded. The o1 paradigm consumes its own compute (test-time scaling) and does not eliminate the pre-training requirement for the underlying model. We treat the "scaling has stalled" narrative as a sentiment overhang rather than a fundamental thesis risk for the FY25-FY26 cycle.

Debate: Is the Blackwell GM Trough a Buying Opportunity or a Multi-Quarter Drag?

Bull view: Colette specified the trough at 71-72.5% with mid-70s recovery by 2H calendar 2025. The arithmetic of NVIDIA's revenue scale means even at trough GM, gross profit dollars expand sharply (because the revenue base is growing 60-90% YoY). Operating income and EPS both expand through the trough. The GM rate compression is a transition feature, not a thesis problem.

Bear view: Multi-quarter GM compression below 75% is a structural shift if Blackwell complexity (seven custom chips, four form factors) results in persistent yield/cost headwinds. If competitive dynamics or hyperscaler ASIC pressure prevent GM recovery to mid-70s, the long-term margin profile may anchor in low-70s rather than reverting.

Our take: The bull is roughly right. NVIDIA's pricing power on Blackwell at the steep ramp phase is functionally unlimited (demand >> supply), which gives the company optionality to defend margin even as cost intensity peaks. The trough is bounded and the recovery is specified. We model FY26 GM in the 73-75% range with quarterly variability around the Blackwell-Hopper mix.

Debate: Does Hyperscaler ASIC Erode NVIDIA's Merchant GPU TAM?

Bull view: CSPs' own ASIC programs (TPU, MTIA, Trainium) have existed for years and have not materially compressed NVIDIA's hyperscaler revenue. The pace of model-architecture innovation favors flexible programmable platforms (NVIDIA + CUDA) over purpose-built ASICs that lock to specific architectures. CSPs continue to diversify supply rather than substitute, and the OpenAI deal (announced post-Q3 with AMD, but the principle generalizes) demonstrates that even merchant alternatives do not displace NVIDIA at the leading edge.

Bear view: Hyperscaler ASIC programs are accelerating in maturity. TPU v5/v6 and Trainium 2 have meaningfully closed the performance gap on specific workloads. As inference-side compute consumption grows (and inference is more workload-specific than training), ASIC economics may become more favorable for hyperscalers in 2026+.

Our take: The bull is right for the FY25-FY26 cycle; the bear has a valid 2027+ concern. ASIC competition is a watch item, not a near-term thesis driver. NVIDIA's CUDA software-stack moat plus annual-cadence roadmap make it structurally hard for any ASIC competitor to hold a sustained performance-per-dollar lead at the leading edge.

Model Implications

Valuation framework: Net of the changes, the multi-year revenue and earnings trajectory is meaningfully higher than where Street consensus sits. The combination of (a) Blackwell ramp ahead of plan, (b) three concurrent scaling laws expanding the compute TAM, (c) sovereign + enterprise AI as orthogonal demand vectors, (d) Hopper persistence through 1H FY26 supporting concurrent revenue, and (e) the GM trough being bounded and specified produces a fundamentally better risk/reward than the Q2 initiation framework reflected. We move to Outperform with conviction on the operating story; valuation is full but justifiable given the multi-year visibility expansion the print delivered.

Thesis Scorecard Post-Earnings

Overall: Eight bull points strongly confirmed or confirmed; five bear points either contained, bounded, managed, or neutralized. This is the cleanest thesis-favoring print we have observed in the cycle. The combination of execution evidence (Blackwell on track) plus demand visibility expansion (three scaling laws + sovereign + enterprise) plus margin trajectory specification (trough bounded, recovery specified) is the rare configuration that supports an upgrade to Outperform from a Hold initiation.

Action: Upgrading from Hold to Outperform. Our Q2 initiation framed NVDA as a Hold pending evidence on three concerns: (a) Blackwell ramp execution, (b) gross-margin trajectory, and (c) demand-cycle durability. Q3 retired all three with operationally specific evidence: 13,000 Blackwell samples shipped with system stand-ups at every major CSP and Q4 exceeding prior estimates; GM trough quantified at 71-72.5% with mid-70s recovery specified for 2H calendar 2025; and three concurrent scaling laws plus the 100,000-Blackwell next-gen-model floor materially expanding the multi-year compute TAM. Sovereign AI, Enterprise AI, and Automotive layered on as orthogonal demand vectors that materially diversify the revenue base. The setup post-print is the strongest operational position we have observed for NVIDIA in this cycle: execution on track, demand visibility expanding, margin dip bounded and underwritable, customer base diversifying. We move to Outperform with conviction on the multi-year operating story and calibrated humility on near-term price-action volatility — the AI complex remains positioning-sensitive and digestion-narrative flow is unlikely to fully fade through the next several quarters.