$39.3B Record Quarter, Data Center +93% YoY, Blackwell $11B First Full Quarter — the FY26 Framework Reset Confirms the Multi-Year Cycle: Maintaining Outperform

Rating Action: Maintaining Outperform

We initiated coverage at Hold in Q2 FY2025 with the framing that NVIDIA's operating story was unambiguously strong but that the valuation already discounted the bull case and that meaningful execution risk sat in the Blackwell transition (the architectural step from NVLink-8 Hopper systems to NVLink-72 Blackwell rack-scale systems was, at that point, untested at scale). We upgraded to Outperform at Q3 FY2025 once Blackwell sample shipments validated execution at scale (13,000 units in customer hands), the GM trough was quantified at 71-72.5% with mid-70s recovery committed for 2H calendar 2025, and Huang's three-scaling-laws framework (pretraining + post-training + test-time inference) materially expanded the multi-year TAM beyond the merchant-GPU framework that had anchored the Hold. The Q3 upgrade was the structural answer to the post-DeepSeek bear case as it was emerging.

The Q4 FY2025 print extends the Q3 upgrade thesis on every operating axis:

• Blackwell ramp execution: $11B of Blackwell revenue in its first full production quarter against expectations of a more measured sequential build, characterized by management as the fastest product ramp in NVIDIA's history. Multiple 100,000-GPU clusters are already shipping; CoreWeave has gone public with bring-up of a 100,000 GB200 cluster; Microsoft, OpenAI, and others are operational. The architectural transition risk is materially behind us.
• FY26 framework: Q1 guide of $43B implies a sequential acceleration from the $39.3B Q4 base (+9% QoQ) into a quarter that is normally seasonally lighter for hyperscaler spend. The "significant ramp of Blackwell in Q1" framing removes the "1H FY26 air pocket" risk that some sell-side bears had been flagging.
• Three-scaling-laws TAM expansion: Huang's reframe of the demand vector along pretraining + post-training + test-time inference scaling is the structural answer to the DeepSeek-efficiency-bear case. Reasoning models do not reduce compute consumption; they multiply it (100x today, with management framing future generations at potentially "millions of times more compute"). This is a bullish reframe of the long-cycle unit economics.
• Margin recovery committed: The low-70s trough is shallow vs the bear case (some had modeled high-60s as the Blackwell-ramp cost intensity surprised), and management committed to a return to the mid-70s before fiscal year-end. This is a deliberate, well-flagged ramp dynamic rather than a structural margin reset.

The valuation point that anchored Hold has not changed: NVDA still trades at a premium multiple that prices in significant continued execution. But the operating story has materially de-risked across the four axes that mattered, and the FY26 framework reset is the strongest framing the Street will see this cycle. We move to Outperform.

Results vs. Consensus

The Q4 print beats management's own outlook by $1.8B (4.8%) on the top line, with Data Center carrying the upside at +93% YoY against a base that was already the largest segment in the company's history. The Blackwell $11B contribution is the standout: it is materially above the sub-$10B range that most sell-side desks had been modeling for the architecture's first full production quarter, and it is the most concrete execution data point the Street has had on the Hopper-to-Blackwell transition.

Quality of Beat

• Revenue: The $1.8B beat against management's own outlook is unusually clean. NVIDIA has consistently outperformed its guides through the AI cycle, but the absolute scale of this beat — nearly 5% above the midpoint — against a base that itself had been guided up through the year is materially positive. The beat is concentrated in Data Center compute (+18% QoQ, +2x YoY), which is exactly the line the Street most cares about.
• Margins: The 73.5% non-GAAP gross margin is at the upper end of the "low 70s" framing management has held since the Blackwell-ramp dilution was first flagged. The trough is therefore more contained than the bear case had positioned, and the recovery commitment (mid-70s by year-end) is an explicit management line that we treat as high-conviction given Kress's track record on guidance.
• Mix: Large CSPs were "about half" of Data Center revenue with sales nearly 2x YoY; consumer Internet 3x YoY (XAI on GB200, Meta Andromeda on Grace Hopper); Enterprise nearly 2x YoY. The breadth of the Data Center beat is the single most important compositional signal — this is no longer a "two hyperscalers carrying the print" story.
• Capital return: $8.1B returned in Q4 alone (buybacks plus dividend) confirms NVIDIA is generating cash at a pace that comfortably funds both the ramp and meaningful capital return. This is a structural positive that bear-case framings tend to under-weight.

Segment Performance

Data Center: The Print Inside the Print

Data Center is the entire story. At $35.6B, the segment is +93% YoY and +16% QoQ, with compute revenue specifically +18% QoQ and more than doubling YoY. The composition is broad-based: large CSPs at ~50% of segment revenue (and nearly 2x YoY), consumer internet 3x YoY (XAI on GB200, Meta's Andromeda ad engine on Grace Hopper), and enterprise nearly 2x YoY. The Blackwell first-full-quarter contribution of $11B is the operational headline — this is materially ahead of the sub-$10B sell-side range and confirms the ramp is execution-on rather than execution-troubled.

"In Q4, Blackwell sales exceeded our expectations. We delivered $11 billion of Blackwell revenue to meet strong demand. This is the fastest product ramp in our company's history. Unprecedented in its speed and scale." — Colette Kress, CFO

Networking declined 3% QoQ as the segment transitions from NVLink 8 with InfiniBand to NVLink 72 architecture; Spectrum-X and NVLink switch revenue grew and management committed to a Q1 return to growth. The first Stargate data centers will use Spectrum-X; Cisco's announcement of integrating Spectrum-X into its enterprise networking portfolio adds a meaningful new merchant-distribution lever for the Ethernet AI franchise.

Gaming: Supply-Constrained Quarter, Q1 Inflection Coming

Gaming printed $2.5B (-22% QoQ, -11% YoY), a clear weak spot in an otherwise pristine quarter. Management explicitly attributed the decline to supply constraints rather than demand softness ("demand remains strong throughout the holiday"). The RTX 50 series Blackwell desktop and laptop GPUs are launching from March with up to 3,400 AI TOPS, DLSS 4 with transformer-based AI frame generation, and Max-Q technology delivering up to 40% better laptop battery life. We treat the Q4 dip as a transient supply-side issue with a clean catalyst path through Q1 and into the new launch cycle.

Automotive: $570M Record, FY26 to ~$5B (3x)

Automotive was the most surprising upside vector of the print. Q4 revenue of $570M is a record and the segment grew 103% YoY in the quarter, 55% YoY for the full year, and management framed FY26 expected revenue as ~$5B — roughly 3x the FY25 base. The driver is the convergence of (a) Toyota's commitment to NVIDIA Drive OS for next-generation vehicles, (b) Aurora + Continental's driverless-truck deployment on Drive Thor, and (c) the broader AV development ecosystem that "virtually every AV company is developing on NVIDIA in the data center." This is a real second-leg revenue line for the multi-year framework.

Key Topics & Management Commentary

Overall Tone: The most confident framing of the cycle. Huang repeatedly used long-cycle language — "we're really three years into that journey" of modernizing decades-built data centers, "we're at the beginning of this new transition," "we will grow strongly in 2025." Kress's discipline on the margin trough — framing it as a deliberate operational choice rather than a forced compression — was consistent and on-message. There is no defensive posture in this print.

The Three Scaling Laws

The most important conceptual framing of the call was Huang's articulation of three independent scaling laws driving compute demand. This is the structural answer to the DeepSeek-efficiency-bear case that has pressured AI semiconductor narratives since late January.

"There are now multiple scaling laws. There's the pretrained scaling laws... the second is post-training scaling law... the amount of computation you use for post-training is actually higher than pretraining... And the third part is test time compute or reasoning. Long thinking, inference scaling... The amount of tokens generated, the amount of inference compute needed, is already a hundred times more than the one-shot examples... The next generation could have thousands of times and even hopefully extremely thoughtful and simulation-based and search-based models that could be hundreds of thousands, millions of times more compute than today." — Jensen Huang, CEO

Assessment: This is a TAM-expanding reframe, not a TAM-defending one. Management is not arguing that pretraining scaling continues unabated; it is arguing that pretraining + post-training + test-time inference combine to produce a multi-axis demand curve that is structurally larger than the single-axis pretraining curve the Street had been modeling. Even if pretraining scaling decelerates, post-training (RLHF, distillation, synthetic-data generation) and test-time inference (reasoning models like O3, R1, Grok 3) more than offset the slowdown.

Blackwell as Reasoning-AI Architecture

Kress's prepared remarks explicitly characterized Blackwell as architected for the reasoning-AI workload: "Blackwell supercharges reasoning AI models with up to 25x higher token throughput and 20x lower cost versus Hopper 100." The Transformer Engine plus the NVLink 72 fabric (delivering "14x the throughput of PCIe Gen 5") create an architecture-level fit for the test-time-compute workload that, by management's framing, is the fastest-growing segment of AI demand. "Many of the early GB200 deployments are earmarked for inference. A first for a new architecture."

"The vast majority of our compute today is actually inference, and Blackwell takes all of that to a new level. We designed Blackwell with the idea of reasoning models in mind." — Jensen Huang, CEO

Geographic Composition and Sovereign AI

US sequential growth in Data Center revenue was the strongest geography in Q4, driven by the initial Blackwell ramp. Management explicitly cited France's €200B AI investment and the EU's €200B Invest AI initiative as glimpses into the sovereign-AI build-out that "will redefine global AI infrastructure in the coming years." Sovereign AI continues to be framed as a multi-year, multi-region demand vector rather than an isolated geography play.

China: Stable at Half of Pre-Export-Control Levels

Kress noted that China data center revenue remains "well below levels seen onset of export controls" and that "absent any change in regulations, we believe that will remain roughly at the current percentage." Huang reinforced the point: China is "approximately the same percentage as Q4. And as in as previous quarters. It's about half of what it was before the export control. But it's approximately the same in percentage." This converts China from an open downside variable into a stable baseline contribution.

ASIC vs Merchant GPU

Huang's response to Tim Arcuri's custom-ASIC question was the most assertive defense of the merchant-GPU franchise we have heard him make. Four arguments stacked: (1) NVIDIA architecture is general (autoregressive, diffusion, vision, multimodal); (2) end-to-end coverage from data processing through training, post-training, and inference; (3) performance per watt of "2x to 4x to 8x" translates directly to AI-factory revenue because tokens are directly monetizable; (4) the software-stack ecosystem on top of NVIDIA architecture "is ten times more complex today than it was two years ago," and "just because the chip is designed doesn't mean it gets deployed." This is the argumentative spine that pushes back on the hyperscaler-ASIC-displacement bear case.

Capital Allocation

NVIDIA returned $8.1B to shareholders in Q4 in the form of share repurchases and cash dividends. The capital return cadence continues at a pace that comfortably accommodates both the Blackwell ramp's working capital intensity and meaningful equity retirement. With OpEx growth guided to "mid-thirties" YoY for FY26, the company is investing aggressively into the cycle while still generating sufficient FCF to support the buyback. This is an unusual operating-leverage position.

Guidance: The FY26 Framework Reset

The Q4 call is the FY framework-reset event of the year for NVIDIA, and the framing was unambiguously positive. Q1 FY26 specific guide: revenue ~$43B ±2%, with sequential growth in both Data Center and Gaming, and within Data Center sequential growth in both compute and networking. GAAP gross margin 70.6% / non-GAAP 71.0% ±50bps. GAAP OpEx ~$5.2B / non-GAAP OpEx ~$3.6B. Full-year FY26 OpEx growth in the mid-thirties. Other income +$400M. Tax rate 17% ±1%.

Margin trajectory: Q1 71% non-GAAP is the trough. Kress explicitly committed to a return to "mid-seventies later this fiscal year" as Blackwell production yields and configuration mix optimize. Atif Malik (Citi) pressed her on whether the back-half ramp from 71% to mid-70s implies ~200bps per quarter of expansion — a steep recovery. Kress did not retreat: "we have many different configurations as well on Blackwell that will be able to help us do that... If we can improve it in the short term, we will also do that. Tariffs, at this point, it's a little bit of an unknown."

OpEx mid-thirties FY26 growth: Reflecting "higher engineering development costs and higher compute and infrastructure costs for new product introductions." The OpEx profile signals continued aggressive investment into the multi-year cycle — consistent with Blackwell Ultra (2H FY26), Vera Rubin (the next architecture after Ultra, previewed for GTC), and the broader software / Cosmos / agentic-AI roadmap.

Blackwell Ultra and Vera Rubin: Huang confirmed Blackwell Ultra remains on track for second-half FY26 launch on the annual product cadence. Critically, the transition from Blackwell to Blackwell Ultra is described as a "slot in" rather than a chassis-architecture change — "the system architecture is exactly the same. It's a lot harder going from Hopper to Blackwell because we went from an NVLink 8 system to a NVLink 72 base system." This is a meaningful execution-risk reduction for the next generational transition. Vera Rubin is the architecture beyond Ultra and was previewed for the upcoming GTC keynote.

Analyst Q&A Highlights

C.J. Muse (Cantor Fitzgerald): Test-time compute and inference-dedicated clusters

• Asked Huang about the implications of test-time compute and reinforcement learning blurring the line between training and inference, and what that means for inference-dedicated clusters and NVIDIA's positioning.
  Assessment: Huang used the question to deliver the three-scaling-laws framework that anchors the entire call's thesis. This was the most important conceptual exchange of the Q&A.

Joseph Moore (JPMorgan): GB200 ramp and rack-level bottlenecks

• Asked about complexity of rack-level GB200 systems, where NVIDIA is in the ramp, and whether systems-level bottlenecks remain. Huang noted "350 plants manufacturing the one and a half million components that go into each one of the Blackwell racks" and confirmed CoreWeave has gone public on bring-up of its 100,000 GB200 cluster, with Microsoft and OpenAI also operational.
  Assessment: The 350-plant / 1.5M-component supply-chain disclosure is a useful complexity data point that frames the magnitude of the operational achievement. The "$11B last quarter" framing signals confidence in continued scale.

Vivek Arya (Bank of America): Q1 gross-margin trough, DeepSeek implications

• Asked Kress to confirm whether Q1 is the gross-margin bottom and asked Huang what gives him confidence on FY26 demand sustainability post-DeepSeek. Kress confirmed the low-70s ramp framing with a return to mid-70s "later this year." Huang articulated "near-term, mid-term, long-term" demand signals: POs and forecasts (near-term), CapEx scale-out vs prior years (mid-term), and the structural shift to AI-based software (long-term). He also explicitly noted that "we've really only touched consumer AI" and that agentic enterprise AI, physical AI for robotics, and sovereign AI are "barely off the ground."
  Assessment: Q1 is operationally the trough. The DeepSeek question got the structural answer: efficiency innovations expand rather than contract aggregate compute demand because more workloads become economically viable.

Matt Ramsay (TD Cowen): Blackwell Ultra ramp dynamics

• Asked about simultaneous ramp of Blackwell and Blackwell Ultra in 2H. Huang confirmed Ultra is on track for 2H, with the chassis-level architecture identical to Blackwell — making the transition a "slot in" rather than a redesign.
  Assessment: This is the most important execution data point on the next-generation transition. Lower transition risk for Ultra than for Blackwell itself materially de-risks the FY26 H2 trajectory.

Tim Arcuri (UBS): Custom ASIC vs merchant GPU balance

• Asked about the balance between custom ASICs and merchant GPUs and whether heterogeneous super-clusters will mix both. Huang delivered the four-pillar defense (architecture generality, end-to-end coverage, performance per watt, software-stack complexity) and the "just because a chip is designed doesn't mean it gets deployed" framing.
  Assessment: This is the most assertive merchant-GPU defense Huang has delivered on a public call. The substance is consistent with NVIDIA's competitive position; the directness of the framing is new and is appropriately answering a Street concern that has firmed through 2H FY25.

Ben Reitzes (Melius): US sequential surge and geographic mix

• Asked about the sustainability of the US-driven Data Center surge. Huang confirmed China is "approximately the same percentage as Q4" and used the question to deliver the "AI has gone mainstream" prepared framing — AI as the largest software addressable-market expansion in history.
  Assessment: The China commentary is the most important takeaway. China is a stable baseline contribution rather than an open variable. The "AI mainstream" framing is the macro-level extension of the three-scaling-laws argument.

Marshall Pappas (Evercore ISI): Enterprise growth vs hyperscalers

• Asked Kress to confirm enterprise grew 2x YoY and Huang to discuss the long-term mix between hyperscaler internal/external workloads and enterprise direct. Kress confirmed enterprise grew 2x YoY, similar to large CSPs. Huang argued long-term enterprise will be "by far larger" than CSP, anchored by the three-computer framing for industrial AI (employee-productivity AI agents, AI factory training, in-vehicle/in-machine physical AI).
  Assessment: Enterprise 2x YoY is a critical breadth-confirmation data point. The three-computer industrial-AI framing is the long-cycle TAM extension argument and is consistent with the automotive-to-$5B framing in FY26.

Aaron Rakers (Wells Fargo): Replacement cycle and installed-base dynamics

• Asked about the two-year-anniversary replacement cycle from Hopper and whether GB300 or Rubin would be the refresh trigger. Huang's response noted that Voltas, Pascals, and Amperes are still in active production use for data-processing, data-curation, and machine-learning-search workloads, with the newer architectures dedicated to training. "All of our CPUs are very well employed."
  Assessment: This is a structurally important answer. NVIDIA's installed base does not simply depreciate to zero on a generational cycle — older architectures cascade into adjacent workloads. This pushes back against bear cases that model rapid Hopper depreciation as a headwind to net new GPU demand.

Atif Malik (Citi): Gross-margin ramp back to mid-70s, tariff exposure

• Pressed Kress on the back-half FY26 gross-margin ramp from 71% to mid-70s — implying ~200bps per quarter — and on tariff exposure. Kress reiterated the multiple-configuration optimization opportunity and "If we can improve it in the short term, we will also do that." On tariffs: "It's an unknown until we understand further what the US government's plan is, its timing, it's where, and how much."
  Assessment: The margin recovery commitment held under direct pressure. Tariff acknowledgement is appropriately conservative without retreating from the year-end framework.

What They're NOT Saying

1. Formal full-year FY26 revenue figure: Notably absent. Huang's "we will grow strongly in 2025" closing framing and Kress's segment-level guidance (Automotive ~$5B, OpEx mid-thirties YoY) are bullish but stop short of a full-year revenue number. The Q1 $43B guide is the only quantitative anchor; the Street will have to bridge from there.
2. Specific FY26 Data Center revenue trajectory: No segment-level full-year framework. With Data Center at ~$115B FY25 and Q4 alone at $35.6B (~$142B annualized), the implied FY26 Data Center range is wide ($150–200B+ depending on Blackwell ramp pace and Blackwell Ultra contribution). Sell-side will have to construct this, and the divergence in models will widen post-print.
3. Tariff impact quantification: Kress acknowledged the unknown but did not disclose any embedded tariff scenario in the Q1 guide. If tariffs are imposed on semiconductor imports during FY26, the gross-margin recovery to mid-70s could be at risk — not addressed.
4. Specific Blackwell yield trajectory: The "low 70s during ramp / mid-70s once fully ramped" framing implies meaningful yield improvement, but management did not provide specific yield curve, defect-rate trajectory, or HBM/CoWoS allocation commentary that some on the buy-side wanted.
5. Hyperscaler ASIC threat quantification: Huang's four-pillar defense is qualitative. There is no quantitative framing of how much of internal-hyperscaler-workload demand has migrated to TPU-style or Trainium-style ASICs vs merchant GPUs. Critical to the multi-year framework but unaddressed.
6. Sovereign AI revenue contribution: France €200B and EU €200B are cited as TAM signals, but no specific sovereign-AI revenue contribution disclosed for Q4 or FY25. The narrative is meaningful; the quantification is absent.
7. Vera Rubin specifications and timing: Mentioned only as "the click after Blackwell Ultra" with a GTC reveal. Specific architecture, networking, memory, and timing details are reserved for the keynote — an information vacuum for the period between earnings and GTC.

Market Reaction

• After-hours move (Feb 26 evening): The print was received initially positively in after-hours but action was choppy as the Street digested the gross-margin trough framing and the absence of a formal FY26 revenue figure. The Q1 $43B guide was meaningfully above the implied sequential trajectory but the magnitude of the beat was within the band that the most-bullish AI-momentum desks had been positioning for.
• Sentiment context: NVIDIA entered the print under the most pressure on AI-narrative sentiment of the cycle. The DeepSeek R1 release in late January had compressed the merchant-GPU TAM thesis on capex-efficiency arguments; AI-bubble framing was active in macro discourse; the Feb 26 print was the highest-profile event for the AI complex. Reaction is therefore being read as a referendum on the AI thesis broadly, not just on NVIDIA specifically.
• Multiple compression context: NVDA had de-rated from its mid-2024 peaks even as fundamentals continued to deliver. Forward P/E heading into the print had compressed materially from the late-2024 highs. The setup heading in was therefore "results need to clearly counter the DeepSeek bear case" — a setup that the print and the three-scaling-laws framing arguably accomplished, though the AI-momentum-positioned reaction may take days to fully process.
• Cited drivers (sell-side aggregation): (1) Q4 revenue beat the outlook by $1.8B and Blackwell delivered $11B in its first full quarter — clear positive; (2) gross-margin trough at 71% Q1 with a stated mid-70s recovery is a feature of the ramp, not a structural reset, but flagged as the principal short-term concern; (3) the three-scaling-laws framework is being received as a credible structural answer to DeepSeek-efficiency-bear positioning; (4) absence of a formal FY26 revenue figure leaves a wide bull-bear modeling gap that will keep the post-print consolidation noisy; (5) automotive at ~$5B for FY26 is a real second-leg vector that under-rated bull cases will start to add to models.

Our read of the post-print sentiment: the operating story is unambiguously bullish — this is the cleanest framework-reset quarter NVIDIA has delivered — but the absence of a formal FY26 revenue commitment and the gross-margin trough optics will keep the post-print price action choppy through GTC (March 17). The substantive framework reset (three scaling laws, Blackwell ramp execution, Blackwell Ultra slot-in, mid-70s margin recovery, ~$5B Automotive) is what matters for the multi-quarter framework, and this print is materially more bullish than the bear case had been positioning for.

Street Perspective

Debate: Does the Q4 print resolve the post-DeepSeek bear case?

Bull view: Yes, decisively. The three-scaling-laws framing argues that efficiency innovations like DeepSeek expand rather than contract compute demand because more workloads become economically viable. Test-time inference compute is "100x" today and management projects orders-of-magnitude higher with future reasoning models. Hyperscaler capex is ramping; sovereign AI is a multi-region demand vector; enterprise AI is "barely off the ground." Q1 $43B guide is structurally bullish.

Bear view: Partially. The three-scaling-laws framing is a coherent structural argument, but it is also management's argument — not yet validated by a Hyperscaler capex print or a major model-builder commitment that explicitly de-prioritizes pretraining scale. If hyperscaler capex deceleration emerges in calendar 2026 (post-Stargate buildout, post-Saudi/UAE FOMO peak), the merchant-GPU bull case has to absorb a real demand downshift — and the Q4 print does not yet provide evidence of resilience to that scenario.

Our take: The bull is closer to right. The three-scaling-laws framework is internally consistent, externally validated by the operational data (multiple scaling axes simultaneously consuming compute), and supported by the breadth of the Q4 demand composition (CSP, consumer internet, enterprise, automotive, sovereign all growing). The bear's hyperscaler-capex-deceleration concern is a 2026/2027 question rather than a Q4 FY25 question; if it emerges, it will show up with lead time. We treat the bull case as the central case post-Q4.

Debate: Is the gross-margin recovery to mid-70s by year-end credible?

Bull view: Yes. The trough is a deliberate, well-flagged consequence of Blackwell ramp cost intensity (multiple configurations, air- and liquid-cooled, several networking options, expedited manufacturing). As yields improve and configuration mix normalizes, the structural margin profile of the franchise (mid-70s+) reasserts. Kress has been disciplined on margin guidance through the cycle and her commitment carries weight.

Bear view: Partially. The 71% to mid-70s ramp implies ~200bps per quarter expansion in 2H FY26 — a steep recovery that depends on Blackwell yield curves not previously executed at this scale. Tariff exposure is an unquantified additional risk. If yields lag or tariffs materialize, the year-end framing is at risk and the margin reset could be more durable than the framing implies.

Our take: Bull is right, but the back-half ramp pace is the single most important execution data point to monitor. We model gross margin in the 71–72% band through 1H FY26 with a meaningful step in 2H to mid-74–75% by Q4 FY26. Tariff exposure is a watch-item but not currently a base-case headwind.

Model Implications

Valuation impact: Our Q3-update fair-value framework was anchored on FY26 EPS of ~$5.85 at a 30–35x forward multiple, producing a $175–205 fair-value range that anchored the Hold. Updating for the higher FY26 EPS midpoint of ~$6.85 and holding the multiple range constant, fair value moves to $205–240. Given the structural framework reset (three scaling laws, Blackwell execution proof, sovereign-AI second leg, automotive at ~$5B FY26), we are willing to apply a modestly higher multiple range of 32–38x to reflect the multi-year cycle de-risking, putting our Q4-update fair-value range at $220–260. This range is meaningfully above current trading levels and supports the upgrade to Outperform.

Thesis Scorecard Post-Earnings

Overall: Ten bull points (eight existing strongly confirmed plus two new positive vectors), with one bull point committed-but-execution-pending (margin recovery). On the bear side: one mitigating, one converting from variable to known constant, two watch-items, one transient weakness (Gaming), and the valuation-pricing-perfection point partially mitigated by multiple compression and operational de-risking. The framework reset is unambiguously positive, and the upgrade to Outperform is grounded in operating fundamentals rather than price-action timing.

Action: Maintaining Outperform. The Q4 FY2025 print is the cleanest framework-reset quarter NVIDIA has delivered. Blackwell executed at $11B in its first full quarter as the fastest product ramp in company history; the three-scaling-laws reframe answers the post-DeepSeek bear case at the structural level; FY26 framework (Q1 $43B guide, Automotive ~$5B, OpEx mid-thirties growth, mid-70s margin exit rate) confirms the multi-year cycle; sovereign AI and enterprise breadth are real second-leg demand vectors. Every operating thesis point we underwrote at the Q3 upgrade has confirmed or improved. Our fair-value framework moves to $220–260, extending the constructive case rather than creating a new one.