Initiating Coverage at Hold: A Record Print With Data Center +154% and Blackwell Sampling, But the Mask-Change Slip and Full Valuation Keep Us Patient

Rating Action: Initiating Coverage

This is our initiation report on NVIDIA. We begin coverage at Hold — a constructive but deliberately patient stance on what has been the most important compounder in technology over the past 24 months and is now arguably the single most important equity in global markets. Our Hold rating is not a critique of operational execution, which we view as exemplary; it is a statement about price-to-thesis. The market has already extrapolated a multi-year continuation of the current Data Center compounding into the equity, and the upside-vs-downside asymmetry from current levels strikes us as roughly balanced rather than clearly skewed toward Outperform.

Three specific reservations underpin the Hold rather than Outperform:

1. Blackwell timing has slipped at the margin. A quarter ago the mass-production framing was "ramping in fiscal H2"; today it is "shipping out in Q4" and "into fiscal year '26." A one-quarter slip on a once-in-a-generation product cycle is not a thesis-breaker, but it does push a portion of the cycle peak into a window where comp dynamics are tougher and sell-side estimates have been built around an earlier ramp. This is a small information loss, not a large one.
2. Customer concentration is more concentrated than it looks. CSPs are ~45% of Data Center revenue, but the CSP bucket itself is dominated by a small handful of hyperscalers. The "consumer internet and enterprise" 50%+ slice contains additional concentration in a narrow set of frontier-model labs. The aggregate exposure to perhaps 8-10 ultimate buyers is, in our model, materially higher than headline disclosure suggests. We do not view this as immediate-print risk; we view it as a tail-risk to be priced.
3. Valuation discounts continuity. At the current setup, the equity is pricing the Data Center compounding to continue for several more years at growth rates that have rarely if ever been sustained at this absolute revenue scale. The bull-case math works; it just doesn't leave a comfortable cushion for the inevitable digestion phase that follows every multi-year capex super-cycle in this industry.

What gets us to Outperform: (a) the Q3 print delivers the Blackwell ramp on schedule with no further mask-driven slips, (b) the Q4 guide signals Blackwell at materially-above-management's-prior-framing for fiscal H1 FY26, (c) sovereign AI engagement count and revenue contribution scale beyond the "low-double-digit billions" framed today, or (d) a meaningful drawdown (15-20%) without thesis impairment that compresses the discounted-continuity premium. What gets us to Underperform: (a) sustained signals from hyperscalers of capex pause or pull-back, (b) a competitive datapoint — ASIC, AMD MI355/MI400, or hyperscaler internal silicon — that meaningfully reframes the merchant-share narrative, (c) further Blackwell execution slippage beyond the current single-quarter mask change.

Results vs. Consensus

Q2 FY2025 was a clean and material beat across every reported line. The print materially exceeded both management's prior outlook and Street consensus, and the Q3 guide did the same. The only soft spot in the print was the sequential gross-margin step-down — 75.7% non-GAAP from ~78.4% in Q1 — and that compression was substantively explained by Blackwell-related yield reserves and the H200 mix shift, not by competitive pricing pressure.

Quality of Beat

• Revenue: Fully organic and driven by the segment that matters. Data Center alone exceeded Street by ~$1.3B; Gaming, ProViz, and Automotive added another ~$0.1B at the margin. The beat is not a function of one customer over-pulling — per Colette, "compute revenue grew more than 2.5 times" and "networking revenue grew more than 2 times" YoY, suggesting broad strength across the stack.
• Gross margin: Non-GAAP 75.7% was at the high end of the prior guidance range. The QoQ compression of ~330bps is the only blemish on the print and is fully attributable to two non-secular factors: (i) inventory provisions for "low-yielding Blackwell material" associated with the mask-change yield improvement campaign, and (ii) richer mix of new Hopper products (H200) early in their cost curves. Neither is competitive pricing pressure, which is the only QoQ-margin driver that would have been thesis-relevant.
• Operating leverage: Non-GAAP opex grew ~12% QoQ on compensation expense and continued investment in the Blackwell platform. Even at that growth rate, opex as a share of revenue is dropping fast given the revenue compounder — the operating margin expansion is structural and persistent.
• Cash flow: $14.5B of operating cash flow in a single quarter is the standout. NVIDIA returned $7.4B to shareholders in Q2 (buybacks + dividend) and the board authorized a fresh $50B repurchase on top of the $7.5B remaining — a clear signal that management views the equity as a continued return-of-capital vehicle even as it compounds revenue at triple-digit rates.

Segment Performance

Data Center: The Whole Equity Story

Data Center revenue of $26.3B (+154% YoY, +16% QoQ) is now ~88% of the company. Every other segment is rounding error in the equity case. Within Data Center, compute revenue grew more than 2.5x YoY and networking more than 2x YoY — the platform breadth is widening, not narrowing. Per Colette, CSPs were ~45% of Data Center revenue and "more than 50% stemmed from the consumer, Internet, and enterprise companies." That is a meaningfully more diversified mix than the bear-case framing of a hyperscaler-only buyer base, although the CSP bucket itself remains concentrated.

"Hopper demand is strong and Blackwell is widely sampling. We executed a change to the Blackwell GPU mask to improve production yields. Blackwell production ramp is scheduled to begin in the fourth quarter and continue into fiscal year '26. In Q4, we expect to ship several billion dollars in Blackwell revenue." — Colette Kress, CFO — the single most important sentence on the call

The H200 ramp began in Q2 with shipments to large CSPs and consumer-internet and enterprise customers. The H200 offers ~40% more memory bandwidth than the H100 and is the "bridge" SKU that absorbs Hopper-class demand into Q3 while the Blackwell yield-improvement work is completed. Hopper supply availability has improved, which is what allows the Q3 guide to step up to $32.5B even with Blackwell still pre-ramp.

The Spectrum-X (Ethernet for AI) story is the second-order positive that the print under-emphasized in our view. Networking revenue increased 16% sequentially, and Ethernet for AI revenue specifically "doubled sequentially." Spectrum-X has X-AI as an early anchor customer "to connect the largest GPU compute cluster in the world." Colette guided Spectrum-X to "a multi-billion dollar product line within a year" — that is a real second leg of the Data Center story beyond the GPU itself, and it is being priced more cautiously by the Street than we think it should be.

Sovereign AI: Colette upsized the FY25 sovereign-AI revenue framing to "low-double-digit billions this year" — a meaningfully larger number than the prior framing. Specific deployments referenced include Japan's National Institute of Advanced Industrial Science and Technology building "AI Bridging Cloud Infrastructure 3.0" with NVIDIA. Sovereign demand is structurally additive because it is dual-vendor by political mandate (national autonomy considerations) but it is a leading indicator for FY26 revenue more than a near-term Q3/Q4 lever.

China: "Our data center revenue in China grew sequentially in Q2 and is a significant contributor to our data center revenue. As a percentage of total data center revenue, it remains below levels seen prior to the imposition of export controls." This is the most that NVIDIA management is willing to disclose on China, and it is the structural overhang we view as the largest single tail-risk in the forecast. The export-control regime can tighten further, and the H20 / China-specific SKU revenue contribution — which is what we believe drives the China sequential growth — is not separately broken out.

Inference share: Colette framed inference as "more than 40% of our data center revenue" over the trailing four quarters. This is the most important disclosure on the call from a long-term thesis perspective: the ASIC competitive pressure is concentrated in inference workloads (because they are more volume-elastic and TCO-sensitive than training). NVIDIA holding 40%+ of Data Center revenue from inference and citing recent MLPerf inference leadership is the strongest counter-narrative to the "ASICs will erode merchant GPU share at the inference end" bear thesis. We weight this disclosure heavily.

Assessment: Data Center is firing on every cylinder we can identify. Compute and networking both compounding north of 2x YoY, H200 ramping at the largest CSPs, Spectrum-X on track to multi-billion run-rate, sovereign AI scaling into low-double-digit billions, and inference market share holding via MLPerf leadership. The only soft spot is the Blackwell mask-change slip, which is real but bounded.

Gaming: Healthy and Underdiscussed

Gaming revenue of $2.88B (+16% YoY, +9% QoQ) was a clean beat against a low-bar Street setup. Console, notebook, and desktop all grew sequentially and channel inventory is "healthy." The RTX install base is now 100M+ devices, with 600+ AI-powered applications and games — this is the underappreciated AI-PC tailwind that the equity narrative does not currently price. NVIDIA ACE (a suite of generative-AI technologies for RTX AI PCs) shipped its first commercial use case (Mecha BREAK using the Minitron-4B small language model on-device) in the quarter.

Assessment: Gaming is doing its job — growing, generating cash, and seeding the on-device AI inference ecosystem. It is not the equity story but it is no longer a drag.

Professional Visualization: AI Use Cases Driving Acceleration

ProViz revenue of $454M (+20% YoY, +6% QoQ) accelerated on AI fine-tuning and Omniverse workloads. Foxconn is using Omniverse to power digital twins of the physical plants that produce NVIDIA Blackwell systems — a self-referential proof point that has both narrative and operational value. Mercedes-Benz signed a multi-year Omniverse Cloud contract for industrial digital twins. WPP is using USD NIM microservices for content creation pipelines for customers including Coca-Cola.

Assessment: ProViz is now a pure AI play. Industrial-AI digital twins are a genuine emerging vector that benefits the GPU-plus-Omniverse-software stack uniquely.

Automotive & Robotics: Multi-Billion-Dollar Multi-Year Build

Automotive revenue of $346M (+37% YoY, +5% QoQ) growth was driven by new self-driving customer ramps and AI cockpit solutions. Colette framed automotive as "multi-billion dollars in revenue across on-prem and cloud consumption." NVIDIA Isaac Robotics platform customers now include Boston Dynamics, BYD Electronics, Figure, Siemens, and Teradyne Robotics — the customer count is widening across humanoids, robot arms, and mobile robots.

Assessment: Automotive and robotics are 2026/2027 revenue ramps with strong directional positives but no near-term print-mover impact.

Key Topics & Management Commentary

Overall Management Tone: Confident, expansive, and notably more emphatic on the "two simultaneous platform transitions" framing than on prior calls. Jensen led with the structural narrative — general-purpose computing transitioning to accelerated computing, and human-engineered software transitioning to generative-AI software — rather than the quarter's specific datapoints. That tone choice is consistent with a management team that is increasingly answering ROI-skeptic questions with a long-cycle, civilizational-shift narrative rather than quarter-to-quarter execution metrics. The framing is not new, but the conviction with which it is delivered has stepped up.

The Blackwell Mask Change: One Quarter, Likely Not Two

The single piece of new information on the call — the Blackwell mask change — was disclosed clinically and accompanied by sufficient color to constrain the downside. Jensen stated the change is complete, with no functional changes; the company is now sampling functional Grace Blackwell systems across roughly 100 different system configurations. Production starts in fiscal Q4 (the November-January quarter) with several billion dollars of Blackwell revenue expected to ship in Q4 itself.

"The change to the mask is complete. There were no functional changes necessary. And so we're sampling functional samples of Blackwell — Grace Blackwell in a variety of system configurations as we speak." — Jensen Huang, CEO

The "several billion dollars" framing for Q4 Blackwell revenue is meaningful because it is directional rather than precise. A few interpretations: (a) management is intentionally setting a low bar to over-deliver, (b) production yields are still being ramped and the precise output is itself uncertain, (c) systems-integration timing at the customer level (the GB200 NVL72 rack is a 3,000-pound liquid-cooled system that requires close-to-data-center final assembly) is the gating factor for revenue recognition rather than chip output. Our base case is a combination of (a) and (c).

The H200 / Hopper bridge is the right structural answer to the mask-change slip. Hopper supply has improved, demand "remains strong" per Jensen's closing remarks, and customers are not waiting for Blackwell — they are ordering Hopper now while Blackwell capacity stands up. This is the operational pattern we want to see during a product-cycle transition: the in-market product carries the revenue while the next product samples and ramps in parallel, rather than customers pausing orders.

Assessment: The mask change is contained. The Q3 guide of $32.5B with Blackwell still pre-ramp is the validation. We will look closely at the Q3 print for any further timing adjustments to Blackwell output, and at the Q4 guide for any indication that the "several billion dollars" framing is being upsized or downsized.

The CapEx-ROI Debate: Jensen's First-Principles Answer

The single most consequential analytical exchange on the call was Jensen's response to Toshiya Hari (Goldman Sachs) and Matt Ramsay (TD Cowen) on the customer ROI question. The framework Jensen offered is the cleanest articulation we have yet heard from management of why hyperscaler capex on NVIDIA infrastructure is not a bubble.

"The people who are investing in NVIDIA infrastructure are getting returns on it right away. It's the best ROI infrastructure — computing infrastructure investment you can make today." — Jensen Huang, CEO — in response to Matt Ramsay, TD Cowen

The argument has three layers, in his own framing: (1) Cost displacement — accelerated computing replaces general-purpose computing for data processing, recommender systems, and search at lower TCO and lower energy consumption ("not unusual to see someone save 90% of their computing cost"). (2) Rental absorption — CSPs build out NVIDIA capacity and immediately rent it to AI startups and frontier-model labs at high utilization, generating ROI for the CSP without dependence on internal use cases. (3) Self-use leverage — hyperscalers' own ad-targeting, recommender, and search systems are migrating to generative-AI architectures, generating internal ROI from the same capacity.

The argument is well-rehearsed and largely valid in our view. The weakness is that it is unfalsifiable at the present quarter — it will be falsified or confirmed by hyperscaler capex commentary on their own earnings calls over the next 2-3 quarters. If hyperscalers begin to signal capex digestion, the argument starts to crack regardless of the sophistication of the framework. We treat the framework as plausible but not yet decisive.

Assessment: Jensen's ROI argument is the strongest version of the bull thesis we have heard articulated. It is internally consistent and supported by the cost-displacement and rental-absorption mechanics. It is not yet validated by hyperscaler-side disclosure. Watch hyperscaler 2H 2024 earnings reports for confirmation or contradiction.

The Two Platform Transitions Framing

Jensen returned multiple times to the dual-transition framing: general-purpose compute to accelerated compute, and human-engineered algorithms to generative-AI software. This is the structural argument for why the current capex cycle is different from prior semiconductor cycles — it is not a single product cycle but two simultaneous architectural shifts both of which require GPU compute.

"The world is moving from general purpose computing to accelerated computing. And the world builds about $1 trillion dollars' worth of data centers... in a few years will be all accelerated computing." — Jensen Huang, CEO

The $1 trillion installed-data-center-base-to-modernize TAM framing is now standard NVIDIA management vocabulary. It has been rebutted by skeptics on the grounds that not all of that installed base will be GPU-accelerated and the migration timeline is multi-decade. We think the skeptic is right on the qualifier and management is right on the directional vector. The relevant analytical question is not "is the TAM real" but "what share of that TAM does NVIDIA capture against ASIC, AMD, and hyperscaler custom silicon competition." The answer to that is not in this print; it is in the next 6-8 prints.

Enterprise AI: Real but Not Yet a Print-Mover

Colette devoted significant prepared-remarks time to enterprise AI — the NIM microservices, NIM Agent Blueprints, and NVIDIA AI Enterprise software platform. Customers cited included Amdocs (30% customer-service cost reduction with the smart agent), AT&T (70% cost savings and 8x latency reduction on call transcription), ServiceNow (Now Assist), SAP, Cohesity, Snowflake, Lowe's, and Aramco. Software/SaaS run-rate was guided to "approach a $2 billion annual run rate exiting this year."

Assessment: Enterprise AI is real. The customer roster is broad and the per-GPU-per-year pricing ($4,500) is straightforward. But $2B run-rate against $30B per quarter of revenue is <2% — this is a leading-indicator vector for FY26-27 software margins, not a Q3/Q4 print-mover. The directional positive is the breadth of customer adoption rather than the dollar contribution.

Guidance & Outlook

The Q3 guide of $32.5B is the constructive surprise. Above the ~$31.7B Street consensus by ~$800M, the guide implies +8% sequential growth and +79% YoY — on a base that is itself growing at +122% YoY. That magnitude of compounding at $30B+ revenue per quarter is unprecedented in semiconductor history. The implicit decomposition: continued growth of Hopper architecture (H200 ramping at large CSPs, ongoing H100 demand), sampling of Blackwell products through the quarter (no material revenue), and modest contribution from Gaming/ProViz/Auto.

The 75% non-GAAP gross margin guide is the bridge metric to watch. Down ~70bps QoQ on Hopper-mix-shift to new SKUs (the H200 specifically) and continued Blackwell-related provisions. Full-year FY25 GM guided to "mid-70%" range — consistent with 75% framing, leaving Q4 GM at potentially the low-to-mid 70%s as Blackwell production yields ramp through their early-life cost curve. This is normal cycle dynamics for a major new product introduction; the trajectory matters more than the level for FY26 modeling.

FY25 implied: H1 FY25 revenue of $56.1B (Q1 $26.0B + Q2 $30.0B) plus Q3 guide of $32.5B places 9-month FY25 revenue at approximately $88.6B. If Q4 lands in the $36-39B range (consistent with low-to-mid-single-digit sequential growth despite Blackwell ramp providing positive mix), full-year FY25 revenue tracks toward $124-128B — roughly +120% YoY. That places the FY25 exit run-rate well above $140B annualized.

Capital return: $7.4B returned in Q2 (buybacks plus dividend); $50B fresh repurchase authorization on top of $7.5B remaining; dividend per share increased earlier in the year. The capital-return cadence has stepped up materially, and the message from the board is that even at current valuation the equity is viewed by management as a continued return-of-capital vehicle. This is signal, although it should be weighted against the fact that NVIDIA's capital-return as a percentage of free cash flow remains modest by mature-tech-platform standards.

Analyst Q&A Highlights

Blackwell Production Ramp and Mask Change

• Vivek Arya, Bank of America Securities: Asked about the Blackwell mask change — whether there were back-end packaging changes and how the Q4 ship target is preserved despite the design adjustment. Jensen confirmed the mask change is complete with no functional changes; sampling continues across ~100 system configurations.
  Assessment: The clinical disclosure on the mask issue was the right management posture. We treat it as a contained, one-quarter timing variance.
• Joe Moore, Morgan Stanley: Asked about Hopper-vs-Blackwell coexistence and whether Blackwell deployment will be in new clusters or intermixed. Jensen pointed to broad CSP deployment of both, with H200 and Hopper remaining state-of-the-art for current workloads while Blackwell stands up for next-plateau compute.
  Assessment: The "intermix" answer matters — it implies that customers are not pausing Hopper orders ahead of Blackwell, which is what we want to see during a product-cycle transition.
• Aaron Rakers, Wells Fargo: Asked about Rack-Scale system mix in the Blackwell cycle, specifically GB200 NVL72 and the leveraging of NVLink. Jensen clarified that NVIDIA designs the rack but sells in disaggregated components — ODM/OEM partners (Quanta, Foxconn, HP, Dell, Lenovo, Supermicro, Gigabyte, ASUS) integrate and assemble close to the data center.
  Assessment: Important read — NVIDIA is explicitly not a vertical integrator. The supply chain disaggregation is a deliberate architectural choice to scale to the world's data center footprint. This addresses the "should NVIDIA do more vertical integration" question that has been actively debated by the Street.

Customer ROI and CapEx Sustainability

• Toshiya Hari, Goldman Sachs: Asked Jensen what NVIDIA watches internally to gauge customer ROI and capex sustainability. Jensen offered the three-layer framework: cost displacement, rental absorption, and self-use leverage. On the related sovereign AI question, Colette upsized FY25 sovereign revenue to "low-double-digit billions."
  Assessment: The single most consequential answer of the call. The framework is rehearsed and well-articulated; its validity is hyperscaler-disclosure-dependent and will be tested over the next 2-3 prints.
• Matt Ramsay, TD Cowen: Asked about the bifurcation between customers spending to push the AGI frontier (where ROI questions don't gate spend) versus customers focused tightly on near-term ROI. Jensen's response leaned hard into the "best ROI infrastructure investment you can make today" framing, citing data processing TCO savings as the floor.
  Assessment: Jensen explicitly does not concede the bifurcation framing. His implicit view is that there is no meaningful customer cohort that is pausing on ROI grounds — the floor case (data processing displacement) generates ROI even before the AI use cases.

Geographic Mix and China

• Ben Reitzes, Melius Research: Asked about the 10-Q geographic disclosure showing US revenue down sequentially while several Asian geographies up. Colette explained that the disclosure reflects "who we invoice to" rather than where the product ultimately deploys — the geographic shift reflects ODM/SI integration locations, not end-customer geography. China revenue does include gaming, data center, and automotive in aggregate.
  Assessment: Useful clarification. The 10-Q geographic breakdown is materially less informative about end-customer geography than headline reading suggests. We discount geographic shifts in the 10-Q accordingly.

Gross Margin Trajectory and Blackwell Math

• Stacy Rasgon, Bernstein Research: Asked whether Q4 Blackwell revenue is additive to Hopper Q4 strength, and whether the FY25 mid-70% GM framing implies a Q4 GM step-down to ~71-72%. Colette confirmed Hopper grows in 2H separate from Blackwell additionality but pushed back on the 71-72% Q4 GM math, suggesting that figure is too low.
  Assessment: The most analytically sharp question of the call. Colette's pushback on the implied Q4 GM is mildly constructive — it suggests the "mid-70%" framing is closer to 73-74% than to 71-72%, which keeps the FY25 EPS algorithm meaningfully above where the implied Q4 math would land it.

Liquid Cooling, Power, and the FY26 Setup

• Timothy Arcuri, UBS: Asked about liquid cooling adoption rates as a potential gating factor on Blackwell ramp shape, and about FY26 risk factors (power supply, model size dynamics). Jensen confirmed strong CSP adoption of both air-cooled (Blackwell HGX) and liquid-cooled (Grace Blackwell) configurations, and reiterated the "next year is going to be a great year" framing on Data Center FY26.
  Assessment: The liquid-cooling concern that has circulated on the buyside is overstated, per management. NVIDIA offers both form factors and CSPs are deploying both. FY26 framing is constructive but unspecific — we will look for incremental color in subsequent prints.

Vertical Integration and Margin Profile

• C.J. Muse, Cantor Fitzgerald: Asked about whether NVIDIA's accelerated annual product cadence raises the case for greater vertical integration, and the implications for margin profile. Jensen pushed back hard: NVIDIA explicitly does not want to vertically integrate beyond the architectural design and component supply, because the ODM/OEM ecosystem (Quanta, Foxconn, HP, Dell, Lenovo, Supermicro, etc.) provides the supply-chain reach and customer-bespoke integration that NVIDIA could not replicate.
  Assessment: Strategic clarity. The ODM/OEM ecosystem is a moat and a scalability mechanism, not a margin-leak. We update our model to reflect that vertical-integration upside is not on management's roadmap.

What They're NOT Saying

1. Customer concentration within the CSP bucket: Colette disclosed CSP revenue mix at ~45% of Data Center but did not break out concentration within that bucket. Our prior is that 4-5 hyperscalers (Microsoft, Meta, Google, Amazon, plus a smaller cluster of next-tier CSPs) drive nearly all of that revenue. Management has consistently declined to disclose this and we read the silence as conscious — the concentration is more concentrated than buyside conventional wisdom assumes.
2. Specific China data-center revenue dollar figure: "Significant contributor" but "below pre-export-control levels" is the disclosure. We have repeatedly looked for a dollar number or a percentage and management has consistently declined. Our prior is that China data center revenue is meaningful but an asymmetric tail-risk on further export-control tightening.
3. Blackwell pricing and per-system economics: The $30 billion Q2 print includes essentially zero Blackwell revenue; the Q4 "several billion dollars" framing is the first datapoint, but ASP, per-GB200-rack pricing, and gross margin contribution are not disclosed. The Street is forced to model Blackwell from third-party teardowns and inferred ASPs — this is a material modeling uncertainty that the company is choosing not to resolve.
4. The MI300X and MI355 competitive response: Colette's "very competitive going-forward" framing on China is the closest she came to addressing AMD's MI300X / MI325X / MI355 roadmap. Jensen's MLPerf citation is the implicit response on inference benchmarks, but there is no direct engagement with AMD's specific accelerator competitive positioning. We read this as conscious — engaging would acknowledge competitive pressure that NVIDIA prefers to leave un-named.
5. Hyperscaler internal silicon (TPU, MTIA, Trainium, Maia): Not mentioned once in prepared remarks or Q&A. The threat from hyperscaler custom silicon for inference workloads is the largest medium-term competitive risk in our framework, and management's silence on the topic is conspicuous. The implicit posture is "our customers will continue to buy our merchant GPU because TCO and ecosystem leadership," but we would be more comfortable if management addressed the topic directly.

Market Reaction

• After-hours move (Aug 28 evening): NVDA traded down approximately 6-7% in the after-hours session immediately following the press release and conference call. The decline was driven by a combination of the gross-margin step-down (75.7% non-GAAP, ~330bps QoQ), the Blackwell mask-change disclosure (perceived as a timing slip), and a "bar was high" dynamic in which the +4.5% revenue beat and above-Street Q3 guide were not large enough to clear elevated expectations.
• Next-day intraday context: Pre-market and early Thursday Aug 29 trading saw the stock recover a portion of the after-hours decline as the gross-margin and Blackwell-timing concerns were re-contextualized through sell-side analysis emphasizing the structural execution and the Q3 guide quality.
• Pre-print context: NVDA had appreciated ~150% over the trailing twelve months entering the print; expectations were elevated and a "buy-the-rumor, sell-the-news" dynamic was reasonable on a stock that has run that far. The 6-7% drawdown is roughly the magnitude that AI-momentum names with similar setups have absorbed on similar prints.
• Volume: Materially elevated vs. trailing 30-day average, consistent with index repositioning around what is now the world's most heavily-weighted technology equity in major benchmarks.

The market reaction is rational, in our view, and roughly proportionate to the print's actual information content. The headline beat was clean but not sufficient to warrant a "buy the print" rally given the elevated entry; the gross-margin step-down is bounded but real; the Blackwell mask change is contained but slips visible revenue into Q4 from where bull-case modeling had it. None of this is thesis-impairing, and we expect the stock to digest the print over 1-2 weeks rather than to break trend. We do not view the reaction as an entry-level opportunity given the bias of the multiple to discount continuity.

Street Perspective

Debate: Is the gross-margin step-down a one-quarter blip or the start of a multi-quarter normalization?

Bull view: The 75.7% non-GAAP GM is exactly where management guided, and the QoQ compression is fully explained by Blackwell yield reserves and Hopper SKU-mix dynamics, both of which are non-secular. The Q3 guide of 75% (down ~70bps further) is consistent with the new-product-cycle cost curve and will reverse as Blackwell yields ramp through fiscal H1 FY26. Structural GM should re-expand to the high-70%s through FY26.

Bear view: Gross margin in the mid-70%s is unsustainable on a multi-year basis as ASIC competition and hyperscaler internal silicon force ASP normalization at the inference end of the workload spectrum. The Q3/Q4 GM step-downs are the early signal of a structural mean-reversion to the high-60%s/low-70%s, not a one-cycle Blackwell artifact.

Our take: Bull view is closer to right for FY25-26; bear view captures a real long-run risk for FY27+. Near-term, the GM trajectory is yield-curve and mix-curve driven, not competitive-pricing-driven. But we agree with the bear that a sustained 75%+ GM at $150B+ of annual revenue is a remarkable outlier in semiconductor history and that mean-reversion pressure builds over time. We model 73-75% non-GAAP GM through FY26 and a gradual step-down toward 70-72% over FY27-28.

Debate: Does the Blackwell mask change matter to the FY26 algorithm?

Bull view: The mask change is one-quarter timing and is operationally complete. Several billion dollars of Blackwell revenue in Q4 plus a full ramp through FY26 leaves the FY26 algorithm essentially unchanged from where the Street had it pre-print. The H200 / Hopper bridge keeps Q3 guidance moving up, and customers are not pausing Hopper orders. Net: no FY26 estimate cuts warranted.

Bear view: A one-quarter ramp slip on a generational product cycle compresses the cycle's revenue contribution into a narrower fiscal window and increases the probability that some portion of the demand timing will shift to later quarters or to incumbent SKUs that carry lower ASPs. The FY26 algorithm relies on Blackwell ramping at expected rates; any further slips compound the pressure.

Our take: Bull view is right on the central-tendency case; bear view captures a real second-derivative risk if a second Blackwell timing event were to occur. We model FY26 Data Center revenue of ~$200-220B (vs. consensus near $200B), with Blackwell as the dominant driver and Hopper bridging into fiscal H1. Sensitivity to additional Blackwell slippage: each one-quarter slip removes roughly $5-8B from FY26 revenue (Blackwell pull-forward dollars that get pushed into FY27).

Debate: Is the equity discounting too much continuity?

Bull view: The current valuation reflects a multi-year compounding of Data Center revenue at growth rates that, while decelerating from +154%, will still meaningfully exceed Street consensus through FY26. The base of $30B/quarter is too large to compound this fast without disrupting global semiconductor supply chains, but NVIDIA's vertical-integration of architecture and software (CUDA + ROCm-equivalent + NIMS) creates a wider moat than prior semis cycles permitted. The premium is justified.

Bear view: The equity is pricing a multi-year continuation of capex super-cycle that requires hyperscalers to maintain their current pace of NVIDIA spend through FY27. Any sign of capex digestion at any major buyer (Meta, Microsoft, Google, or Amazon) would compress the multiple meaningfully. The valuation has no margin for normal-course-of-business demand modulation.

Our take: The bear view is closer to right on the asymmetry of the setup. We do not view the current valuation as obviously wrong; we view it as offering a poor risk/reward for new capital being deployed today. The first hyperscaler signal of capex digestion will compress the multiple by 20-30% in the absence of an offsetting Blackwell-driven beat. We are comfortable being patient and waiting for either a meaningful drawdown or a cleaner entry post-Blackwell-ramp validation.

Model Implications

Valuation framing: At a 30-35x forward P/E on FY26 estimates (a multiple that has historically been awarded to growth tech franchises with similar compounding profiles but at materially smaller absolute revenue scale), the implied fair-value range lands roughly at-market to modestly below current trading levels. We do not see the equity as obviously expensive at our base-case FY26 numbers; we see it as fully priced — offering call-option upside if Blackwell exceeds and downside if hyperscaler capex normalizes. This is the textbook setup for a Hold rating — a high-quality franchise at a price that already discounts continued execution.

The asymmetric scenarios that would change our rating: (i) Blackwell ramp contributes $8B+ to fiscal Q4 with constructive H1 FY26 framing → Outperform pull; (ii) any major hyperscaler signals capex deceleration in their Q3 or Q4 print → Underperform pull; (iii) export-control regime tightens further on the China-specific SKU revenue stream → modest Underperform pressure but probably bounded by sovereign-AI offset.

Thesis Scorecard at Initiation

Overall: Five bull points are largely confirmed by this print, and five bear points remain live or neutral. The central tension is that operational execution is essentially flawless and the multi-year setup is constructive, but the equity has already priced that constructive view. The Hold rating is a price judgment on a high-quality franchise, not an execution judgment. NVIDIA is in our universe of names we want to own at the right price; the right price is below current levels.

Action: Initiating at Hold. Constructive operational bias. Upgrade triggers: (a) Q3 print delivers Blackwell ramp on schedule with no further timing slippage and Q4 guide above $36B; (b) sovereign AI engagement count and revenue contribution scale meaningfully above the current "low-double-digit billions" framing; (c) drawdown of 15-20% without thesis impairment that compresses the discounted-continuity premium. Downgrade triggers: (a) any major hyperscaler signals capex deceleration on their next earnings call; (b) further Blackwell execution slippage; (c) export-control tightening that meaningfully impairs the China data center contribution. We will revisit on the Q3 FY25 print.