A First Full-Year Guide at Roughly 70% Growth, Paid For With 350 Basis Points of Gross Margin and a $27 Billion Cash-Flow Swing

Results vs. Consensus

Year-over-Year Comparison

All income-statement figures above, including the prior-year comparatives, are taken from the condensed consolidated statements of income and the GAAP-to-non-GAAP reconciliation filed with the quarter, not from figures quoted verbally on the call. Note that beginning in Q1 FY27, NVIDIA's non-GAAP measures no longer exclude stock-based compensation, and the historical non-GAAP figures shown here have been recast on that basis by the company. Prior-year non-GAAP EPS is therefore not comparable to the $1.08 originally reported for that quarter under the old definition.

Sequential Comparison

Revenue Assessment

Adding $14,606M of quarterly revenue in thirteen weeks, on a base that was already $81,615M, is the single most arresting number in the release. The sequential increase alone exceeds the entire annual revenue of many companies in the S&P 500 semiconductor complex. Growth accelerated for a fourth consecutive quarter, which at this scale should not be arithmetically possible and is the strongest available evidence that the demand surface is still expanding faster than the base.

The composition matters more than the headline. Data Center at $89,023M is 92.5% of the company, and within it the two disclosed submarkets diverged: Hyperscale grew 13% sequentially while the AI Clouds, Industrial and Enterprise submarket grew 25%. Management now describes the non-hyperscaler leg as heading toward half the Data Center business. That framing was aspirational when the disclosure framework was introduced last quarter; at 45.3% of Data Center this quarter, it is close to fact. One caution on the submarket comparison. NVIDIA reclassified a customer from the non-hyperscaler bucket into Hyperscale during the quarter, citing a change in that customer's business model, and recast the prior period. The effect is not cosmetic: Q1 FY27 Hyperscale is now stated at $43,050M and the non-hyperscaler submarket at $32,196M, against the roughly $38B and $37B originally published for that quarter. The sequential growth rates in our tables are computed on the recast basis and are internally consistent, but anyone comparing them to figures published three months ago will find the two submarkets roughly $10B further apart than they appeared then: the reclassification moved about $5B, which widens the gap by twice that. The crossover narrative is intact directionally and further out in time than the original disclosure implied.

Half-year revenue of $177,837M is up 95.8% against the prior-year first half. Layering the $108.0B Q3 guide onto the two reported quarters gives $285,836M through nine months, against $147,811M in the equivalent FY26 period. On a Q4 that grows a conservative 13% sequentially, FY27 lands near $408B against FY26's $215,938M, a fiscal year in which NVIDIA nearly doubles at a scale where doubling was supposed to have stopped being available.

Margins Assessment

Reported margins were excellent and forward margins were cut, in the same twenty minutes. Gross margin of 75.0% held flat sequentially, which management attributed to Blackwell remaining the vast majority of revenue, so mix did not shift enough to move the ratio either way. Operating margin of 66.5% on a non-GAAP basis expanded 60 basis points sequentially and 540 basis points year over year, because operating expenses grew 11% against revenue's 18%.

The forward path is where the quarter changed. Management guided Q3 gross margin to 74.0%, then volunteered a Q4 trough of 71% to 72% and an FY28 settling range of 72% to 73%. Against the 75.0% just delivered, the trough is 300 to 400 basis points lower, roughly 350 at its midpoint. At the Q3 revenue guide, each 100 basis points of gross margin is roughly $1.08B of gross profit, so the guided step down in Q3 alone costs about a billion dollars against a flat-margin counterfactual.

Two things make this less alarming than the raw arithmetic. First, management said the price increases that restore margin have already been executed and take effect in Q1 of FY28, which is why the trough is a trough rather than a plateau. Second, the cost pressure is memory, and memory is scarce because the AI buildout that drives NVIDIA's revenue is consuming it. The margin hit and the revenue beat have the same root cause. That is a materially better problem than margin compression from competitive pricing, which is the version of this that would actually threaten the thesis. What it does mean is that the "mid-70s" gross-margin framework management carried for the last several quarters no longer exists, and any model still running 75% through FY28 is overstating earnings by roughly 250 basis points of revenue.

EPS Assessment

Non-GAAP EPS of $2.22 beat the $2.09 bar by 6.2% and grew 119.8% year over year. The operational contribution is nearly all of it: the beat traces to the revenue upside flowing through at a 66.5% operating margin, with the effective tax rate at 16.0% on a non-GAAP basis sitting inside the guided 16% to 18% band rather than below it. There is no tax-rate rescue in this print, which is worth saying because there was a modest one last quarter.

The GAAP line requires care. GAAP EPS of $2.46 exceeds the non-GAAP figure, which is the reverse of the usual relationship, and it does so because $7,771M of net gains on equity securities sit in GAAP other income and are excluded from non-GAAP. Sequentially GAAP EPS rose only 2.9% while non-GAAP rose 18.7%, entirely because those gains fell from $15,936M to $7,771M. As NVIDIA's equity portfolio grows (marketable equity securities of $42,783M plus non-marketable securities of $51,157M, $93,940M combined, against $35,137M at the prior fiscal year end), GAAP earnings will become progressively noisier and progressively less useful. Investors should anchor on the non-GAAP line and treat the equity gains as what they are: unrealised marks on stakes in privately held AI companies, taken through income.

On our updated numbers, FY27 non-GAAP EPS tracks roughly $9.20 and FY28 roughly $15.90, the latter already carrying the 72% to 73% gross-margin assumption management just gave. At the post-print close of $227.98 that is 24.8x FY27 and 14.3x FY28.

Segment Performance

NVIDIA reports on two cuts that do not reconcile line by line: two reportable segments, and two market platforms with the Data Center platform split into two submarkets. Both cuts sum to the same $96,221M total. The reportable-segment view is the accounting structure; the market-platform view is how management runs the business and how the Street should model it.

Reportable Segments

Market Platforms

Data Center: Hyperscale

Hyperscale revenue of $48,710M grew 13% sequentially and 102% year over year, adding $5,660M of quarterly revenue. Management tied the growth to Blackwell Ultra and made a point of noting that the customers absorbing this capacity are themselves reporting accelerating revenue and expanding margins, which is the return-on-investment argument the whole sector rests on. The forward framing was specific about timing: sequential growth in Q3 comes primarily from the non-hyperscaler leg, and hyperscale reaccelerates in Q4 as Vera Rubin supply builds.

"We expect sequential growth to be driven primarily by ACIE with data center, while growth in hyperscale is expected to reaccelerate in Q4 and into fiscal year '28, as supply of Vera Rubin grows over time."
— Colette Kress, EVP and CFO

The industry backdrop management cited: cloud backlog above $2 trillion, and top-five hyperscaler capital expenditure reaching nearly $800B in 2026 and $1.3 trillion in 2027. Alongside that, an expanded AWS agreement covering an additional two million GPUs deploying from this quarter through Q2 of FY29, plus Vera CPUs both standalone and integrated with Rubin.

"AWS is deploying an additional 2 million GPUs starting this quarter through the second quarter of fiscal '29 along with Vera, CPUs, some integrated with Rubin, other standalone."
— Colette Kress, EVP and CFO

Assessment: A 13% sequential quarter at a $48.7B run rate is a good result that reads as a deceleration only because the other submarket did 25%. The AWS commitment is the more important disclosure, because it is a multi-year unit commitment rather than a spend forecast, and it extends visibility into FY29. The honest read is that hyperscale is now the ballast leg: large, supply-gated, and growing at a rate set by NVIDIA's own allocation decisions rather than by customer appetite.

Data Center: AI Clouds, Industrial & Enterprise

The non-hyperscaler submarket reached $40,313M, up 25% sequentially and 138% year over year, adding $8,117M of quarterly revenue against hyperscale's $5,660M. It is now 45.3% of Data Center and, if both submarkets held this quarter's growth rates, would pass hyperscale in two quarters. Management guided hyperscale to reaccelerate in Q4, which pushes that crossover out. Management attributed growth to AI-cloud capacity additions serving enterprises, AI startups and sovereign customers, plus hyperscalers themselves buying capacity from AI clouds to supplement their own builds. That last point deserves attention: it means some revenue booked here is ultimately hyperscaler demand routed through a third party.

The supporting datapoints management supplied were unusually specific. AI-cloud partners are expected to exit the year with 8 gigawatts of installed capacity against roughly 3 gigawatts at the end of 2025. Sovereign business grew 35% sequentially and more than tripled year over year. Venture funding provides the demand pipeline behind the AI-native customers.

"Global VC funding in AI, roughly 70% of which is spent on compute exceeded $400 billion in the first half of 2026, surpassing the $265 billion raised in all of 2025."
— Colette Kress, EVP and CFO

On the enterprise side, trailing-twelve-month on-premises revenue reached $8B in automotive, with financial services, manufacturing and healthcare contributing $7B combined. Named deployments spanned quantitative trading, computational lithography and pharmaceutical research.

Assessment: This is the segment that justifies the multiple, and it is doing what we argued last quarter it would do. Two caveats keep us from treating it as pure diversification. First, the hyperscaler-buying-from-AI-clouds dynamic means the customer-concentration reduction is smaller than the segment split implies. Second, this is the customer base that NVIDIA is financing, guaranteeing and taking equity in, so the receivables build and the growth rate are not independent facts. The growth is real; the credit quality behind it is now a variable we have to track rather than assume.

Edge Computing

Edge Computing revenue of $7,198M grew 13% sequentially and 27% year over year. Management described a split market: strong Blackwell workstation sales offset by slower consumer PC sales, with the consumer weakness attributed to elevated memory and system prices. The same memory scarcity compressing gross margin is therefore also suppressing unit demand at the low end of this platform. Product news skewed toward the physical-AI roadmap, including a Windows PC partnership, an expanded robotaxi platform ecosystem, an open humanoid-robot reference design and a full-stack robotics safety system.

Assessment: At 7.5% of revenue, Edge is not going to move the model, and the consumer softness is a cost pass-through problem rather than a demand problem. Its strategic value is as the distribution surface for the CUDA ecosystem into robotics and automotive, where the revenue is years out. We model it as a low-double-digit grower and would not change the thesis on a quarter either way.

Key Operating and Balance-Sheet Indicators

Key Topics & Management Commentary

Overall Management Tone: Management was confident on demand and unusually pre-emptive on the two things it knew would be attacked, leading with the gross-margin reset and addressing the circular-financing charge by name rather than waiting to be asked. The full-year framing is itself new ground: management noted it has "never forecasted or never guided to a year in advance." The weakest ground was quantification: management was precise about gigawatts, purchase orders and commitment tables, and consistently non-specific about the cash timing and the credit terms behind them.

1. The First Full-Year Guide, and What "Supply-Constrained" Is Carrying

NVIDIA has never guided a fiscal year in advance. This quarter it did, and the number was roughly 70% revenue growth for FY28 against a Street carrying roughly 44%. The framing was as important as the figure: management presented 70% as what supply permits, not what demand supports, and said customer forecasts point to a doubling.

"We expect to grow revenue by approximately 70% in fiscal 2028. This is a supply-constrained outlook."
— Colette Kress, EVP and CFO

"Customers forecasts point to our growth, doubling next year. However, as I mentioned earlier, we expect to grow approximately 70% as we are supply constrained."
— Colette Kress, EVP and CFO

The stated rationale for breaking the no-annual-guide convention was coordination rather than promotion. Management wanted customers, suppliers and shareholders working from one set of numbers, because all three are committing capital against it.

"It is the case that we've never forecasted or never guided to a year in advance. And even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%."
— Jensen Huang, President and CEO

Assessment: The guide is credible precisely because it is a supply statement backed by the $279B of purchase commitments disclosed in the same document. A demand forecast at this horizon would be worthless; a supply commitment is verifiable and expensive to be wrong about. The risk this creates is asymmetric in a way worth naming: by putting a number on FY28, management has converted an open-ended growth story into a target the company can now visibly miss. If supply improves and NVIDIA delivers 85%, the guide looks conservative; if a single link in the chain fails and it delivers 55%, the credibility cost is far larger than the revenue shortfall.

2. The Gross-Margin Reset

Management retired the "mid-70s" gross-margin framework it had carried for several quarters, and did so unprompted in prepared remarks. The new path is 74.0% in Q3, a 71% to 72% trough in Q4, and 72% to 73% in FY28.

"As you are already aware, we are experiencing extreme pricing conditions in memory. The magnitude of the price increase has exceeded our prior expectations and are headed even higher into next year. As a result, we are resetting expectations today."
— Colette Kress, EVP and CFO

"We expect margins to bottom in Q4 in the 71% to 72% range before settling at 72% to 73% in fiscal year '28, as executed price increases take effect in Q1."
— Colette Kress, EVP and CFO

Management framed the disclosure choice explicitly: "We want to be direct about this rather than let it linger as an open question." They also argued the cost pressure is not the ordinary kind, because the same demand driving NVIDIA's revenue is what makes memory scarce.

"Tighter memory supply is a symptom of the same demand surge that's driving our own growth."
— Colette Kress, EVP and CFO

Assessment: Two facts sit uneasily together. The argument that memory scarcity is endogenous to the AI boom is analytically correct and does distinguish this from competitive margin erosion. But management said prices are "headed even higher into next year," which means the 71% to 72% trough is a forecast of a cost input they have just told us they previously mis-forecast. The word "executed" doing work in the pricing sentence is the mitigant: the offsetting price increases are contracted, not planned. We take the trough at face value and treat a deeper one as the single most likely negative revision over the next two quarters.

3. Vera Rubin Enters Full Production

Vera Rubin began production shipments in August, with purchase orders already in hand from every major hyperscaler, AI cloud and system original-equipment manufacturer. Management expects it to be the fastest product ramp in the company's history and put a number on the Q3 contribution.

"We commenced production shipments of Vera Rubin earlier this month, having already received purchase orders from every major hyperscaler, AI cloud and system OEM, we expect Vera Rubin to mark the fastest product ramp in NVIDIA's history."
— Colette Kress, EVP and CFO

The performance claim is a generational step rather than an increment: "Vera Rubin exemplifies this, delivering 30x higher throughput per megawatt and 35x lower token costs relative to Grace Blackwell Ultra." Management guided Vera Rubin to about 20% of Data Center revenue in Q3, which on Data Center revenue near $100B at Q2 mix implies roughly $20B of Vera Rubin revenue in its first quarter of shipment.

Assessment: Reaching a fifth of Data Center revenue in the first quarter of shipment is a steep ramp on any comparison, and it is the operational proof point that last quarter's Q3-production commitment was met. It also explains the inventory build and part of the margin path, since new-platform ramps carry start-up costs before yields mature. The claim to watch is not the ramp rate but the throughput-per-megawatt figure: if 30x holds in customer deployments, the power constraint that gates the whole industry loosens by more than any supply-chain fix could deliver.

4. Revenue Per Gigawatt: The Ladder That Frames the Whole Model

The most useful modelling disclosure on the call was the progression of NVIDIA's revenue opportunity per gigawatt of data-centre capacity, which has risen as the company has taken over more of the rack.

"Since Hopper, our revenue opportunity has grown from roughly $18 billion per gigawatt to $25 billion with Blackwell, to $40 billion with Vera Rubin, which now spans Vera CPU, Rubin GPU, NVLink, InfiniBand or Ethernet and Groq LPU announced earlier this week."
— Colette Kress, EVP and CFO

That is a 39% step from Hopper to Blackwell and a further 60% step to Vera Rubin, achieved by attaching CPU, multiple networking layers and now an inference accelerator to what used to be a GPU sale. Management was explicit that the ladder continues, and anchored it against the general-purpose computing era.

Assessment: This reframes NVIDIA's growth as partly independent of gigawatts deployed. Even in a scenario where global AI power capacity growth slows, NVIDIA's revenue can keep compounding by capturing more dollars per gigawatt. It also quietly reprices the competitive threat: a rival selling an accelerator competes for a shrinking share of the $40B, not for the whole thing. The constraint on this argument is customer tolerance. Each step up in NVIDIA content per gigawatt is a step down in the customer's ability to source alternatives, and at some point that becomes a procurement objective rather than an accepted cost.

5. The Circular-Financing Charge, Answered by Name

Management raised the circular-financing critique itself, in prepared remarks, before any analyst asked. The disclosure sequence was: roughly $50B already invested in frontier AI labs, financing platforms with six institutional capital providers targeting over $500B of third-party capital, and then the concession and the rebuttal.

"We recognize the scale of this support, and we know some will call this circular financing. We see it differently."
— Colette Kress, EVP and CFO

The defence rested on three claims: that the labs' constraint is compute rather than technology or customers, that NVIDIA's investment is small relative to its own cash generation, and that the hardware is redeployable if a customer fails.

"First, we've invested nearly $50 billion in the Frontier AI labs. This was a meaningful commitment, but it represented a small fraction of our expected free cash flow over the same period."
— Colette Kress, EVP and CFO

Then came the number that matters most, and the one that will define the bear case for the next year.

"For context, we expect demand from the AI labs for which we expect to leverage our balance sheet to contribute toward roughly a quarter of our business next year."
— Colette Kress, EVP and CFO

Assessment: Volunteering that a quarter of next year's revenue depends on customers whose purchases NVIDIA is helping to finance is a genuine act of disclosure, and management deserves credit for making the bears' strongest argument for them. It does not dispose of the argument. The fungibility defence is the load-bearing claim: if a lab fails, the compute is redeployable only if there is unmet demand at the moment of failure, which will be true in a supply-constrained market and false in the scenario where a lab fails because demand disappointed. The defence is therefore strongest exactly when it is least needed. We do not think this breaks the thesis at current growth rates, but it converts a former equity-only risk into a credit risk, and it should be underwritten as one.

6. PORTS-Pike: A $105 Billion Guarantee Behind 4.25 Gigawatts

Subsequent to quarter end, NVIDIA entered guarantees providing credit support for the land, power and shell buildout of roughly 4.25 gigawatts at SB Energy's PORTS-Pike campus in Ohio, which will exclusively host NVIDIA infrastructure under twenty-year leases to OpenAI.

"Our guarantee obligations are capped at a total of $105 billion and become effective in phases as certain conditions are met, including data centers becoming ready for service, with the first expected in fiscal 2029."
— NVIDIA CFO Commentary, Q2 FY2027

The exposure declines as OpenAI fulfils lease payments, and NVIDIA holds an option to provide credit support for roughly 3.8 additional gigawatts as the site scales. The offsetting revenue case is sized directly.

"Each generation of NVIDIA infrastructure deployed at PORTS-Pike could represent approximately 1.5 million NVIDIA GPUs, or approximately $150 billion to $200 billion in NVIDIA revenue."
— NVIDIA CFO Commentary, Q2 FY2027

Management separately disclosed that OpenAI's existing and planned commitments represent approximately 12 gigawatts of NVIDIA compute, and that a second frontier lab will receive selective credit enhancement for nearly 2 gigawatts.

Assessment: The structure is more defensible than the headline number suggests. The guarantee is capped, phased, amortising against lease performance, secured against a physical site rather than a company, and first effective in fiscal 2029, which is more than a year of OpenAI operating history away. Against it sits $150B to $200B of revenue per infrastructure generation across a site expected to support multiple upgrade cycles over twenty years. That said, $105B is equivalent to roughly 46% of NVIDIA's $228,984M of shareholders' equity at quarter end, contingent on the lease performance of a single private counterparty, and its accounting treatment as leases commence is not yet visible to investors. This is the disclosure we will be reading the FY27 10-K most carefully for.

7. The NeoCloud Revenue Share: Getting Paid Twice

Management detailed a commercial structure that is new in kind, not just in size. Rather than requiring an AI-cloud operator to lock its entire capacity to one long-term offtake agreement to satisfy lenders, NVIDIA provides a floor and shares in the upside.

"NVIDIA provides a take-or-pay commitment on a portion of the facility's capacity, a minimum revenue guarantee that gives lenders the confidence to underwrite the project, and in exchange, we share in a portion of the NeoCloud's revenue earned above that floor. Independent capital still underwrites every deal on its own merits. We're not making loans."
— Colette Kress, EVP and CFO

"In this model, we get paid twice, once on the hardware sale and again, through the share of rental revenue, a highly reoccurring stream layered on top of a one-time equipment purchase."
— Colette Kress, EVP and CFO

Management sized the medium-to-long-term opportunity as "billions in revenue," which is deliberately imprecise, and disclosed $36B of AI cloud agreements plus $20B of data-centre leases intended for reassignment to third parties.

Assessment: Strategically this is the most interesting disclosure on the call. A recurring, usage-linked revenue stream attached to installed hardware is exactly the mix shift that would justify a higher multiple, because it converts a cyclical equipment sale into an annuity. The catch is what the take-or-pay commitment is: NVIDIA is short a minimum-revenue guarantee on capacity it does not operate. The economics work while utilisation is high and rental rates hold. In a downturn the revenue share goes to zero at exactly the moment the take-or-pay obligation bites. The structure is well-designed for the environment we are in and poorly designed for the one the bears are forecasting, which is a reasonable trade to make but should be understood as a directional bet, not a hedge.

8. Working Capital: Sixty Days to Pay

Days sales outstanding rose to 60 from 45, and management gave the reason without being asked.

"Days of sales outstanding increased to 60 days, reflecting extended payment terms for large purchases by certain investment-grade customers to be shipped over multiple quarters."
— Colette Kress, EVP and CFO

The consequence flowed straight through the cash flow statement. Receivables rose $22,346M, inventory $5,784M, prepaid and other assets $5,497M, for a $30.7B net working-capital use. Operating cash flow fell to $24,077M from $50,344M, and free cash flow to $21,341M from $48,554M. NVIDIA issued $25.0B of senior unsecured notes in the same quarter, taking long-term debt from $7,469M at the prior year end to $32,366M, and still returned $25,779M to shareholders.

Assessment: Three separate things are being conflated in the commentary around this number and they deserve separating. The inventory build is unambiguously good: it funds the Vera Rubin ramp management just told us is 20% of Q3 Data Center revenue. The receivables build is defensible on its stated terms, since investment-grade counterparties on multi-quarter shipment schedules genuinely do warrant longer terms. The debt issuance is the item that changes the picture. A company generating this much operating cash does not need $25B of notes for general corporate purposes unless it intends to fund something the operating cash flow will not cover, and the commitment tables in the same document show what that is. This is the metric we will track most closely: if DSO holds near 60 and inventory normalises, the Q2 cash conversion was a timing event. If DSO drifts past 65 while receivables keep outgrowing revenue, NVIDIA is financing demand and the reported growth rate is partly a credit decision.

9. Supply Commitments More Than Double to $279 Billion

The largest single change in the disclosures is on the obligation side of the ledger.

"Our commitments increased from $119 billion last quarter to $279 billion, primarily related to the procurement of memory."
— NVIDIA CFO Commentary, Q2 FY2027

Total future commitments reach $366B across supply and capacity ($279B), cloud service agreements ($29B), data-centre leases not yet commenced ($25B), equity investments ($25B) and capital expenditure ($8B). A further $56B sits in the additional-commitments table for AI cloud agreements and third-party leases. Of the $279B, $92B falls in the remainder of FY27 and $87B in FY28.

Assessment: This is the concrete backing for the 70% FY28 guide, and it is why we treat that guide as a supply statement rather than a forecast. It is also, read the other way, $279B of largely non-cancellable purchase obligation concentrated in the exact input whose price management just told us it mis-forecast. Locking volume in a scarce market is correct; locking it at prices set during the scarcity is how gross margin gets a multi-year rather than a two-quarter problem. The FY28 settling range of 72% to 73% embeds an assumption about the average price inside those commitments that investors cannot independently verify.

10. Vera CPU and Groq 3 LPX: Two New Silicon Legs

Two products moved from announcement to production this quarter. The Vera CPU is now in full production, with shipments underway to lead partners and a stated performance advantage on agentic workloads. Management sized the existing business and the trajectory.

"Our Grace CPU introduced in 2021 has been a great success, with revenue on a trailing 12-month basis, exceeding $5 billion."
— Colette Kress, EVP and CFO

"our preliminary expectation is for CPU revenue to more than double in fiscal '28, positioning us as one of the world's leading server CPU suppliers."
— Colette Kress, EVP and CFO

Separately, Groq 3 LPX, the company's first rack-scale inference accelerator following last year's Groq partnership, entered full production with volume shipments expected later this quarter. Management claimed nearly four times the tokens per second of the next best alternative on the Artificial Analysis benchmark.

Assessment: The CPU line is now a real business rather than an attach-rate story, and doubling from a $5B trailing base takes it past $10B against the roughly $20B of server-CPU demand management says it currently sees. That visibility is largely captive, since these CPUs are bought alongside racks NVIDIA has already sold, which makes the doubling more credible than a share-gain forecast against incumbent x86 vendors would be. Groq 3 LPX is the more interesting strategic move: it fills the high-interactivity, high-cost-per-token niche where a general-purpose GPU is the wrong tool, which is precisely the gap that inference-specific competitors have been aiming at.

11. The Demand Argument, Restated

Asked whether recursive self-improvement and progress toward general intelligence would inflect demand further, management declined the framing and substituted a narrower, more falsifiable one.

"I think the most important thing that matters for the industry is that; one, AI is now doing productive and useful work; two, AI is generating profitable tokens; and three, if we had more compute, we could generate more profitable tokens, which results in more profit for all of the services. This is the exact phase where we're at, which is the reason why everybody is leaning in."
— Jensen Huang, President and CEO

The quantitative version was the compute multiplier from agentic workloads: "The amount of compute necessary for an agent versus a human using it is probably 15 to 100 times depending on the type of problem you're trying to solve."

Assessment: Declining to trade on an artificial-general-intelligence narrative and reframing on unit economics is the right instinct and a notable tonal change from a management team that has previously leaned into the bigger story. The three-part formulation is also the correct bear checklist. Each clause is independently falsifiable, and the one that would break first is the second: if token generation stops being profitable for NVIDIA's customers, the buildout stops regardless of how much compute would theoretically help.

12. China Remains Zeroed Out

China contributed effectively nothing and is assumed to contribute nothing going forward.

"In Q2, we shipped less than 1% of our total data center revenue in Hopper 200 products to customers based in China in accordance with the U.S. government licenses. Current Hopper shipments are dilutive to corporate gross margins. And given ongoing geopolitical uncertainty, there is no China data center compute revenue in our forward outlook."
— Colette Kress, EVP and CFO

Assessment: The disclosure that current China shipments are dilutive to gross margin is new and cuts against the reflex assumption that a China reopening would be straightforwardly positive. On these terms it would add revenue and subtract margin, which given the margin narrative around this print is not the trade the market would want right now. We continue to carry China at zero and treat any resumption as unmodelled optionality, with the caveat that its margin effect is negative at current product economics.

Guidance & Outlook

The Q3 guide of $108.0B beats the roughly $104.0B the Street carried by 3.8% and implies another $11.8B of net new quarterly revenue. Management was specific about where it comes from: primarily the AI Clouds, Industrial and Enterprise submarket, with hyperscale reaccelerating in Q4 as Vera Rubin supply builds. Vera Rubin is guided at about 20% of Data Center revenue in Q3, which on Data Center revenue near $100B at Q2 mix is roughly $20B in the platform's first quarter of shipment.

Implied quarter-over-quarter ramp. Nine-month revenue on the guide reaches $285,836M. For FY27 to land near $408B, Q4 needs roughly $122B, or about 13% sequential growth. That is in line with the 12.1% step FY25 delivered and well below FY26's 19.5%, so it is not a demanding requirement. The FY28 guide of roughly 70% growth off that base implies approximately $690B, or quarterly revenue averaging around $173B.

Street position. Consensus entering the print carried roughly $92.1B for Q2, roughly $104.0B for Q3 and roughly 44% growth for FY28. All three were beaten. The FY28 gap is the material one: applied to our $408B FY27 base, a 26-point difference in the growth rate is roughly $106B of FY28 revenue, and estimate revisions of that magnitude do not complete in a single session. Offsetting it, gross-margin lines across published models will move down 200 to 300 basis points for FY28, so the EPS revision will be smaller than the revenue revision and will take longer to settle.

Guidance style. Newly transparent, and structurally more exposed as a result. Management beat its own Q2 revenue guide by 5.7%, came in under its own operating-expense guide, missed its own informal days-sales-outstanding forecast, and abandoned its own gross-margin framework, all in one quarter. The pattern is a company that guides revenue conservatively and has been guiding costs optimistically. Applying that pattern forward: treat $108.0B as a floor and 71% to 72% as the optimistic end of the Q4 margin range.

Analyst Q&A Highlights

What Actually Backs the First Full-Year Guide

The opening question went straight at the two novel features of the guide: why break a long-standing practice of not forecasting a year out, and what separates the 70% supply number from the roughly 100% demand number. Management's answer was about visibility earned upstream and downstream rather than about demand confidence, describing multi-year lead times on land, power and shell that force the company to plan further into the pipeline than a chip supplier normally would. The stated motive for publishing the number was to put customers, suppliers and shareholders on identical information because all three are committing capital against it.

Q: "I wonder if you could give us color on the 70%, and what gives you the confidence to guide a full year out if you haven't been doing that? And then what's the gap between that amount of growth in the 100% demand growth. What is the kind of key constraint that separates those numbers? And could you close those gaps over time?"
— Joseph Moore, Morgan Stanley

A: "It is the case that we've never forecasted or never guided to a year in advance. And even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%. And we're going to continue to work with our supply chain to increase on that. But what we wanted to do is to be consistent with everybody from our customers, our shareholders, our supply chain, everybody sees the same view."
— Jensen Huang, President and CEO

Assessment: The answer was responsive on motive and thin on mechanism. What makes the guide credible is not the reasoning offered on the call but the $279B commitment table filed alongside it, which is the auditable version of the same claim. Management should be held to the framing it chose: this is a supply forecast, so a miss would be a supply-chain execution failure, not a demand disappointment, and should be judged that way.

Composition of the Step-Up, and the Unconstrained Number

A follow-up pressed for the building blocks of the FY28 increase across products, how much of it is price rather than volume, and what the number would be without the supply ceiling. Management engaged the last part directly and the first two only in general terms, restating the ladder of revenue per gigawatt and the split between hyperscale and everything else rather than attributing the uplift to specific product lines. The pricing component of the guide was not quantified at all.

Q: "I was just wondering if you could talk us through the contributors, does that increase across the different products, Vera and CPUs and Groq and everything else? And also you talked about your price increase that takes effect in Q1. So I assume some of this is pricing. And I guess I'm also curious, maybe I'm squeezing too many questions in here, but I'm also curious just it's a constrained number. What would it be if it wasn't constrained?"
— Stacy Rasgon, Bernstein Research

A: "The unconstrained would be a lot higher. It's -- we grew 100% year-over-year this year. The unconstrained is significant. And so we're just going to have to work hard to get more capacity. And we have a large supply chain. We have a really gigantic supply chain. And so we have incredible partners and we've secured a lot of supply, but we just need a lot more."
— Jensen Huang, President and CEO

Assessment: The unanswered half is the more useful one. Management has told us prices rise in Q1 of FY28 and that FY28 revenue grows roughly 70%, but not how much of the 70% is price. That distinction determines whether the FY28 gross-margin recovery to 72% to 73% is a margin story or a mix story, and it is the single most valuable disclosure still outstanding. Expect it to be pressed again at the November call.

Sizing the Ecosystem Exposure, and Financing Customers Who Build Rival Silicon

The most pointed exchange added the disclosed commitments to roughly $500B and asked three things: whether that is the full extent of the ecosystem exposure, what the cash component is in FY28, and how the company reconciles funding frontier labs that are simultaneously designing their own inference chips. Management answered the third question at length and the first two barely. The competitive-silicon response drew a platform-versus-chip distinction, arguing that customer-designed accelerators are inference-specific parts for one cloud while NVIDIA sells a full-lifecycle factory that runs anywhere. On the investment logic, the answer was unhedged conviction.

Q: "When I just add up everything that's in the CFO commentary, I get to a number of about $500 billion or so, obviously, over the next several years. So I had a few questions related to that. First is, is that the takeaway that the sum of all your ecosystem investments over the next several years is in that ballpark, or are there other equity or other investments that could still be ahead?"
— Vivek Arya, Bank of America Securities

A: "investing in these companies are once in a generation opportunity. I think the only regret that I have is that I didn't invest more and sooner. And both of the -- 2 of the companies will likely go public soon, and others will follow. And these will be some of the most consequential technology companies in history."
— Jensen Huang, President and CEO

Assessment: This is the exchange to reread. Two direct, quantitative questions were asked and neither was answered: no confirmation that roughly $500B is the ceiling, and no FY28 cash figure. The follow-on from the finance side addressed only the supply-commitment portion, which is the least controversial component and the one already tabulated in the filing. "I didn't invest more and sooner" is a revealing answer to a question about risk limits, because it is an argument for less restraint offered in response to a request for a boundary. The absence of a stated cap is now itself a disclosed fact, and until one exists the exposure should be modelled as open-ended.

Where the Supply Chain Actually Binds

A question sought a rank ordering of the binding constraints across data-centre power and shell availability, memory, and foundry capacity, on the theory that knowing which one binds hardest tells you where the 70% could become 85%. Management declined to rank them, describing the entire chain as running at capacity and framing relief as continuous rather than discrete, arriving through daily yield improvement rather than a single unlock.

Q: "or rank order some of the most acute constraints, whether that be things like data center, power and shell availability, DRAM, wafer foundry availability, et cetera? If you could maybe help us understand which are the biggest among those, that would be very helpful."
— James Schneider, Goldman Sachs

A: "I think the answer is our entire supply chain is challenged. And it's everybody is really running flat out. And more capacity is coming online all the time, which is one of the advantages of what's going to happen this year. It's not going to come online in just an instance in time, but it's going to come online every day."
— Jensen Huang, President and CEO

Assessment: A refusal to rank is a legitimate answer when every input genuinely binds, and the $160B increase in supply commitments concentrated in memory does suggest memory is first among equals. But a supply chain described as uniformly stretched is also an answer that cannot be checked. The practical consequence for modelling is that there is no identifiable catalyst that converts the 70% into a higher number, which means the guide should be treated as a ceiling for FY28 rather than the conservative base case the tone implied.

Inference Share, Agentic Workloads, and Where the Inference Accelerator Fits

A question on inference market share, agentic workload evolution and the role of the newly productionised inference accelerator drew the clearest product-strategy answer of the call. Management described the AI lifecycle in four phases and argued that the rack-scale architecture's value is fungibility across all of them, so a single deployment serves data preparation, pre-training, post-training and agentic inference. The inference accelerator was positioned as a bolt-on for a specific niche rather than a replacement, and management was unusually explicit in bounding its role.

Q: "There's tremendous investor focus on your inference market share. Can you speak to the evolving workloads you're seeing with the agentic AI and how you see your share evolving here over time, particularly when you reflect on the growing value of the TAM you're seeing with each new full stack generation, your expectation for greater growth from ACIE, and then also including Groq 3 LPX?"
— Christopher Muse, Cantor Fitzgerald

A: "And then for services that would like to have super high interactivity, super high-speed token generation done, the throughput is going to be a lot lower. The cost per token will be higher, but you could associate it with high ASP services. And so for those companies, you could bolt on one of our Groq accelerators. I'm super excited about that. But the vast majority of the world's data centers will just be Vera Rubin and NVLink 72."
— Jensen Huang, President and CEO

Assessment: Management deliberately capped expectations for the inference accelerator in the same breath as launching it, which is unusual and reads as credible rather than promotional. The strategic content is the admission embedded in the answer: there exists a workload class where the general-purpose rack is the wrong economic tool. That is exactly the wedge inference-specific competitors have been driving at, and NVIDIA has now productionised a dedicated part to close it rather than argue the gap does not exist.

Scaling Revenue Per Gigawatt Past Vera Rubin

The closing question asked whether the revenue-per-gigawatt ladder continues past Vera Rubin, and whether capacity deployment scales linearly or unlocks in steps. Management answered the first part affirmatively and put the whole progression in historical context, dating it back to the general-purpose computing era, then argued the ladder is good for customers as long as productivity, durability and fungibility rise alongside the price.

Q: "So as we think about the path even beyond Vera Rubin, we think about Vera Rubin Ultra and so on and so forth. Should we really conceptualize like 40 billion goes to 60 billion, 80 billion. And then I guess, underneath of that question is, how do we kind of think about your ability to scale the capacity deployments -- is it a linear function?"
— Aaron Rakers, Wells Fargo

A: "We started in the world of general purpose computing during Moore's Law. We were probably pick your favorite number, but I'm going to go with something like $5 billion, $3 billion per gigawatt of compute with general-purpose computing. And then eventually, with Hopper, it was 18. Now Grace Blackwell is 25. Next, Vera Rubin is 40. And after that, it's going to be higher."
— Jensen Huang, President and CEO

Assessment: "After that, it's going to be higher" is a commitment to keep raising content per gigawatt, and it is the single most important growth lever that does not depend on the world building more data centres. The supporting claim offered in the same answer, that return on invested capital for customers is now under a year on $50 billion facilities, was attributed to something heard rather than measured and should not be carried into a model. If that payback period is real, the ladder can keep climbing; if it is anecdote, the customer tolerance for $60B and $80B per gigawatt is the constraint nobody has tested.

Whether Open Models Help or Hurt

A question probed a genuine tension in the thesis: much of NVIDIA's demand comes from frontier closed-model labs, so open models reaching frontier capability could damage the very customers driving the buildout. Management rejected the zero-sum framing, arguing both categories are growing, that essentially all open models run on NVIDIA because of installed-base breadth, and that open models are what allow enterprises and countries to build proprietary systems at all.

Q: "There's a lot of talk about that these models could gain share for workload in the U.S. You're obviously well positioned with Nemotron, but on the other hand, a lot of the end demand is being driven by these big frontier model companies. So there's a lot of investors that equate open models as being negative for the growth of those companies. So how do you sort of put and take that? Do you see the rise of open models as being good for NVIDIA or ultimately negative?"
— Timothy Arcuri, UBS

A: "The world will need both closed models and open models. And both closed models and open models are skyrocketing in use. Near most, I would say, nearly all open models run on NVIDIA, and the reason for that is because NVIDIA's footprint around the world is the highest, and our architecture is the most fungible."
— Jensen Huang, President and CEO

Assessment: The architecture-neutrality argument is sound and is genuinely a structural advantage: NVIDIA is indifferent to which model wins, which is a better position than any model developer holds. The answer sidesteps the sharper version of the question, which is not about NVIDIA's share of open-model inference but about whether commoditised frontier capability compresses the economics of the labs whose capital expenditure NVIDIA is now partly guaranteeing. Given the disclosure that roughly a quarter of next year's business comes from balance-sheet-supported labs, those two questions are no longer separable.

What They're NOT Saying

1. No cap on ecosystem exposure. Asked directly whether roughly $500B represents the total of ecosystem investments and guarantees or whether more is coming, management did not answer. No ceiling was offered, and the follow-on addressed only the supply-commitment portion. Until a limit exists, the exposure is open-ended by disclosure.
2. No FY28 cash requirement. The same question asked what portion of the commitments is cash in FY28 and got no figure. The commitment table shows $120B falling in the remainder of FY27 and $100B in FY28, but which of that is cash out the door versus contingent is not broken out.
3. No split of the FY28 guide between price and volume. Management confirmed price increases take effect in Q1 of FY28 and guided roughly 70% growth, but never sized the pricing component. This determines whether the margin recovery to 72% to 73% is durable.
4. No ranking of supply constraints. The rank-ordering question was declined in favour of "our entire supply chain is challenged." Without knowing which input binds hardest, there is no way to identify what would convert 70% into a higher number.
5. No FY28 gross-margin sensitivity to memory. The 72% to 73% FY28 range embeds an assumed average memory price inside $279B of commitments. No sensitivity was given, and management said prices are heading higher into next year.
6. No forward DSO target. Having missed the informal mid-50s expectation set last quarter with a 60-day result, management offered no forecast for where receivables settle. This was the most conspicuous omission on the cash-flow side.
7. No accounting treatment for the PORTS-Pike guarantees. The $105B is disclosed as maximum gross exposure, phased and amortising, but how it will be recognised as leases commence in fiscal 2029 was not addressed.
8. No identification of the second frontier lab. Selective credit enhancement for nearly 2 gigawatts was disclosed for an unnamed lab. Concentration risk cannot be assessed without knowing whether the counterparty overlaps with existing exposures.
9. No customer-level detail behind the receivables build. "Certain investment-grade customers" is the only characterisation given for a $22.3B increase. Whether this is two counterparties or twenty is materially different for risk.
10. No revenue quantification for the NeoCloud revenue share. "Billions in revenue over the medium to long term" is the only sizing offered for a structure management presented as strategically transformative, and no disclosure of the take-or-pay obligation's magnitude accompanied it.
11. No comment on the $25B debt issuance rationale beyond "general corporate purposes." A company producing this much operating cash issuing $25B of notes invites a specific explanation, and the prepared remarks did not offer one.
12. No explanation of the $2,944M "Groq, Inc." line. A cash outflow of that size appears for the first time in financing activities, labelled only with the counterparty name. Financing classification rules out a straightforward asset purchase, and neither the call nor the commentary addressed what it represents.

Market Reaction

• Pre-print setup: NVDA closed the August 26 session at $209.66, entering the print up 12.4% year to date against the S&P 500's 12.1%, up 15.3% over trailing twelve months and up 6.7% over the trailing thirty days. The 52-week closing range into the print was $165.17 to $235.74, placing the stock roughly 11% below its 52-week closing high.
• After-hours move: The initial reaction to the release was negative, then reversed higher during extended trading as the call progressed. The sequence tracks the order of disclosure: the gross-margin reset landed first, the FY28 guide of roughly 70% came in the prepared remarks and overrode it.
• Reaction session: The stock closed August 27 at $227.98, up 8.7% or $18.32, after opening at $222.86 (a 6.3% gap) and trading a $220.90 to $230.47 range. The close finished near the upper half of the day's range, so the gain held into the bell rather than fading.
• Volume: 298.9M shares against a 118.1M thirty-day average, or 2.5 times normal.
• Index and peer context: The S&P 500 rose 0.7% on the session, so roughly eight points of the move were stock-specific. The semiconductor complex followed: a large networking and custom-silicon peer added 4.5%, a memory supplier 2%, and a rival processor vendor 4%.

The price action resolved a genuine two-sided print in favour of growth, and it did so on the second-order disclosure rather than the quarter itself. A 4.5% revenue beat and a 3.8% forward-quarter guidance beat were, by NVIDIA's recent standards, ordinary; the stock did not move 8.7% on those. It moved on a growth rate for a fiscal year the Street had been modelling twenty-six points lower, delivered with a supply-commitment table behind it.

The reversal from a negative initial print reaction to a strongly positive close is itself the useful signal. It says the market read the margin reset first, priced it, then repriced when the FY28 number arrived, and concluded that 26 points of revenue growth is worth more than 250 to 350 basis points of gross margin. That arithmetic is correct on our numbers: the FY28 revenue upgrade adds far more to earnings than the margin cut removes. What the tape did not price, in our reading, is the cash-conversion change. Free cash flow falling $27.2B sequentially while revenue rose $14.6B is the kind of divergence that usually attracts attention, and it did not appear to weigh on the session at all.

Positioning going in was moderate rather than stretched. The stock had underperformed its own earnings growth badly over the trailing twelve months, gaining 15.3% while quarterly revenue doubled, which means the multiple had been compressing for a year into the print. That compression is the reason an 8.7% move was available on a beat of this size, and it is also why we think the re-rating has further to run rather than having been spent in a single session.

Street Perspective

Debate: Is the Margin Reset a Two-Quarter Problem or a Structural Repricing?

Bull view: The bull case holds that the trough is shallow, dated and already solved. Management has executed price increases that take effect in Q1 of FY28, the margin recovers to 72% to 73% within four quarters, and the cost pressure comes from memory scarcity caused by the same AI buildout driving revenue, so it is a symptom of strength rather than weakness. On this reading, a 250-basis-point structural haircut against roughly 70% revenue growth is a trade any investor would take.

Bear view: The bear camp argues the reset is the first crack in the pricing power that justified the multiple. Management explicitly said memory prices exceeded prior expectations and are heading higher into next year, which means the 71% to 72% trough is a forecast built on the same inputs that were just wrong. Worse, $279B of supply commitments locks in volume at prices set during the scarcity, converting a cyclical cost spike into a multi-year gross-margin ceiling. Passing costs through requires customers to keep accepting price increases at $40B per gigawatt.

Our take: The bulls have the better of the direction and the bears have the better of the confidence interval. The endogeneity argument is genuinely distinguishing, and the word "executed" on the price increases is the strongest fact in the discussion, because contracted pricing is not a plan. But management has now missed on this input once, and the FY28 range depends on average prices inside a commitment book investors cannot see. We model 72.5% for FY28 and treat a Q4 print in the high 60s as the most likely negative surprise of the next two quarters, because it would signal the FY28 settling range is too high as well. Each 100 basis points of FY28 gross margin is roughly $6.9B of gross profit on our revenue estimate, or about $0.24 of EPS, so even a 250-basis-point shortfall against the guided range costs about $0.60. That is survivable in isolation. What it would not survive is the premise that management guides its cost line accurately, which is worth more to the multiple than sixty cents is to the earnings.

Debate: Is the Balance Sheet Financing Real Demand or Manufacturing It?

Bull view: The bull case is that this is vendor financing of a familiar and profitable kind, executed against exceptionally strong collateral. The customers are compute-constrained rather than demand-constrained, the hardware is fungible and redeployable, the guarantees are capped and amortising against lease performance, third-party capital underwrites each deal on its own merits, and the roughly $50B invested is small against NVIDIA's cash generation. The revenue-share structure converts a one-time equipment sale into a recurring stream, which should command a higher multiple, not a lower one.

Bear view: The bear camp contends that a supplier extending payment terms, taking equity in its buyers, guaranteeing their landlords and providing minimum-revenue floors on their capacity is no longer purely selling. The disclosed facts support the concern: DSO stretched to 60 days from 45, receivables consumed $22.3B in a quarter, $25B of debt was raised, guarantee exposure reached $108.5B, and management itself said roughly a quarter of next year's business comes from labs whose purchases lean on NVIDIA's balance sheet. If any of that demand is credit-created, the growth rate overstates the underlying market.

Our take: Both descriptions are accurate, and the difference between them is entirely a function of whether end demand holds. The fungibility defence is the crux, and it has a specific weakness: redeployment works only if there is unmet demand at the moment a customer fails, which is true in the current shortage and false in the scenario where a lab fails because demand disappointed. The defence is therefore strongest when it is least needed. We do not think this breaks the thesis, because the growth is visible in customers' own reported results and the shortage is real. But it changes what kind of security this is. NVIDIA now carries credit exposure to its customer base, and it should be underwritten with the metrics you would apply to a lender: receivables ageing, counterparty concentration and the trajectory of DSO. Those disclosures do not currently exist at the required granularity, and the absence is itself a reason to size positions with more care than a quarter ago.

Debate: Does 14x FY28 Earnings Mean the Stock Is Cheap or That Nobody Believes the Guide?

Bull view: The bull case is straightforward arithmetic. On the guided FY28 revenue growth and the guided FY28 margin, non-GAAP EPS reaches roughly $15.90 and the stock trades near 14 times that. For a company compounding revenue at roughly 70% with 66% operating margins, that is a multiple normally attached to businesses with no growth at all. The stock returned 15.3% over trailing twelve months while quarterly revenue doubled, so the de-rating has already happened and the re-rating has not started.

Bear view: The bear camp argues the low multiple is the market pricing the probability that FY28 does not happen as guided, and that this is rational rather than an inefficiency. A quarter of the revenue is balance-sheet-supported, the margin path has already been revised down once, $279B of commitments become stranded costs if demand slows, and a first-ever annual guide is a first-ever opportunity to miss one. Terminal-value scepticism about the durability of a buildout at this scale is also embedded in the multiple, and no single quarter resolves it.

Our take: The bears are right that the multiple is a probability weighting rather than a mispricing, and wrong about the weight. The FY28 guide is backed by $279B of purchase obligations, purchase orders from every major customer class for Vera Rubin, and a two-million-GPU commitment from a single hyperscaler running into FY29. That is a far more verifiable basis than a demand forecast, and it deserves a higher probability than 14 times implies. The genuine uncertainty is not whether FY28 revenue arrives but what it converts into, given the margin path and the cash-conversion change. We would rather own the growth at 14 times with a known margin ceiling than pay a full multiple for certainty that was never available in this name.

Model Update Needed

Valuation impact: Fair value range moves to $285 to $350, from $250 to $310. At the post-print close of $227.98 the range implies 25% to 54% upside, with a midpoint near $318 implying roughly 39%. The range is built on FY28 non-GAAP EPS of approximately $15.90 at 18 to 22 times, a deliberately conservative band that discounts for the margin-path risk and the newly material credit exposure. The stock currently trades at 24.8 times FY27 and 14.3 times FY28. The catalysts between now and the November 17 third-quarter call are the memory-price trajectory, evidence on whether DSO stabilises near 60 days, and any disclosure that puts a ceiling on ecosystem exposure.

Thesis Scorecard Post-Earnings

Overall: Thesis strengthened on growth, weakened on quality, and the two do not net to zero in either direction. Every demand pillar we carried into this quarter was confirmed, and the central one was upgraded from a cumulative multi-year framework to a specific full-year guide backed by purchase commitments, which is a materially better artifact. Against that, two new bear points enter the scorecard at meaningful status, and both are management-disclosed rather than analyst-inferred: a gross-margin path 250 to 350 basis points below the framework carried for the past several quarters, and a cash-conversion and counterparty-credit profile that did not exist as a risk category a quarter ago. The prior quarter's read that NVIDIA had reframed itself from a GPU monopoly into a diversified compute platform remains correct. What this quarter adds is that it is also becoming a capital provider to its own end market, and that is a business with a different risk profile and, eventually, a different multiple.

Action: Maintaining Outperform with lower conviction. Fair value range $285 to $350. The FY28 guide is worth more than the margin reset costs, and at 14.3 times a year management has underwritten with $279B of supply commitments, the risk/reward remains favourable. Hold existing positions and add on weakness toward the low $200s. Three things would change the rating rather than the estimates: a Q4 gross margin below 70%, DSO drifting past 65 days with receivables again outgrowing revenue, or any expansion of guarantee exposure without a disclosed ceiling. The November 17 third-quarter call is the next test, and the Vera Rubin ramp against the guided 20% share of Data Center revenue is the specific commitment to grade.