As clean as an earnings print can get
Summary
NVIDIA delivered another beat-and-raise quarter, with revenue of $96.2bn (+106% YoY, +18% QoQ) and non-GAAP EPS of $2.22, ahead of consensus on both lines and extending a beat streak that now runs to fourteen consecutive quarters. The headline event, however, was not the quarter, it was the guidance. For the first time in its history, NVIDIA issued a full-year-forward outlook, guiding FY2028 revenue growth of approximately 70%, against a Street modelling roughly 44%. That implies revenue of close to $700bn versus consensus near $570bn, and management was explicit that the figure is supply-constrained, not demand-constrained, with underlying customer demand tracking closer to 100%. Beneath the headline, the Data Center mix is rebalancing in a way that should ease long-standing customer-concentration concerns: non-hyperscale revenue grew 25% sequentially and 138% year-on-year versus hyperscale at 13%, and is on track to become the larger half of the segment. The one blemish is a guided gross margin step-down to 74% in Q3 and a 71–72% trough in Q4 on acute memory cost inflation but we regard this as a repeat of the Hopper-to-Blackwell transition, when margins compressed through the ramp and were fully reclaimed at volume. Finally, supply commitments more than doubled in a single quarter to $279bn, which we read as both a powerful demand signal and a material escalation in execution risk at a scale that now leaves NVIDIA very little room for error.
Key Takeaways
1. Another beat on a beat — and the sequential compounding is the real story
NVIDIA beat on every line that matters, and did so against a consensus that had already positioned above the company's own guidance.
Revenue of $96.2bn came in 4.5% above the $92.07bn consensus and 5.7% above the $91.0bn guidance midpoint. Non-GAAP EPS of $2.22 beat the $2.09 consensus by 6.3%. Data Center revenue of $89.0bn (+117% YoY) landed near the top of the estimate range and roughly 4% above the $85.7bn midpoint. Gross margin held at 75.0% on both a GAAP and non-GAAP basis, and operating margin expanded 540bps YoY to 66.2%.
The company had beaten consensus in each of the thirteen preceding quarters. This makes fourteen. But the more instructive metric is the unbroken sequential progression, which is what distinguishes a structural buildout from a cyclical order surge:
$44.1bn → $46.7bn → $57.0bn → $68.1bn → $81.6bn → $96.2bn → $108.0bn (guided)
That is six consecutive quarters of sequential growth culminating in a Q3 guide of $108.0bn, +12% QoQ and roughly +90% YoY, comfortably ahead of the $105bn consensus. Growth has now accelerated for four straight quarters on a base that has more than doubled. Doubling revenue at $46bn is impressive; doing it at $96bn while simultaneously guiding to a further 12% sequential step is a demand signal of a different order.
Nothing in this print suggests demand is abating.
2. The first-ever full-year guide: 70% growth versus a Street at 44%
The genuine surprise arrived not in the release but on the call, when CFO Colette Kress told analysts NVIDIA expects FY2028 revenue growth of approximately 70%.
To frame the magnitude: consensus had modelled roughly 44% growth, implying FY2028 revenue of about $570bn. NVIDIA's guide implies $690–700bn, more than $100bn above where the Street sat, and a figure that would make NVIDIA the second-largest US technology company by revenue, ahead of both Apple and Alphabet and behind only Amazon.
Three features of the guide deserve emphasis:
It is a supply number, not a demand number. Kress was explicit that this is a supply-constrained outlook, that customer forecasts point to growth roughly doubling, and that supply is expected to remain the bottleneck at least through the end of FY2028. Huang reinforced it directly on the call: demand is much greater than 70%, but supply allows them to confidently deliver 70%. If the supply chain expands faster than planned, the number goes up.
It is unprecedented, and deliberately so. Huang confirmed NVIDIA has never guided a year in advance. His stated rationale was that customers, shareholders and suppliers are all committing enormous resources and should be working from the same set of information. We take this at face value as it is also, in effect, a commitment device that raises the cost of a miss considerably.
It is backed by contracted inputs, not aspiration (see takeaway 5).
For a market that has spent much of 2026 debating whether AI infrastructure spending is approaching an air pocket, a company at this scale voluntarily underwriting 70% growth twelve months out while stating that true demand is nearer 100%, is about as direct a rebuttal as the data permits. It does not settle the debate about eventual returns on AI capex. It does, in our view, put the near-term "blip" thesis to bed.
3. Data Center concentration risk is quietly resolving itself
One of the most durable bear arguments on NVIDIA has been customer concentration: a business where a handful of hyperscalers accounted for the overwhelming majority of Data Center revenue carries obvious fragility, since a single budget review at a single customer could reset the growth rate.
That picture is changing, and this quarter NVIDIA disclosed the split for the first time:
| Data Center subsegment | Q2 FY27 | QoQ | YoY |
|---|---|---|---|
| Hyperscale | $49bn | +13% | — |
| ACI&E (NeoClouds, sovereigns, enterprise, industrial) | $40bn | +25% | +138% |
Non-hyperscale is now 45% of Data Center revenue and growing at roughly twice the hyperscale rate. Management guided Q3 sequential growth to be driven primarily by ACI&E, with hyperscale re-accelerating in Q4, and stated plainly that the non-hyperscale business will represent roughly half of Data Center going forward. On current trajectories it becomes the larger half.
The underlying drivers are broad and geographically diversified: NeoCloud partners are on track to exit the year at 8 gigawatts of installed capacity, up from roughly 3GW at end-2025; sovereign AI grew 35% sequentially and more than tripled YoY; and the named partner list now spans Armenia, Africa, Taiwan, India, Australia, Malaysia, Japan and Korea alongside the established NeoClouds. Enterprise on-premise revenue reached $8bn in automotive and $7bn across financial services, manufacturing and healthcare on a trailing-twelve-month basis.
This matters for two reasons. First, it materially dilutes single-customer dependency. Second and more importantly it is evidence that this is a genuine computing platform shift rather than a capex cycle at four companies. Sovereigns and enterprises buying full-stack AI factories is demand that simply does not exist in a narrative where a small number of hyperscalers are overbuilding.
We would note the counterweight for balance: the faster-growing half is also the segment with a more varied credit profile than the hyperscalers, and NVIDIA is increasingly involved in the financing architecture that underpins parts of it. That is a trade-off worth monitoring, not a reason to discount the diversification benefit.
4. Gross margins: the report's only blemish, and we don't think it's one
The single negative in the release was the margin path. Management reset the entire trajectory in one motion rather than drip-feeding it:
75.0% (Q2 actual) → 74.0% ±50bps (Q3E) → 71–72% (Q4E trough) → 72–73% (FY2028)
The cause is unambiguous and was named directly: extreme memory pricing, with Kress conceding the magnitude has exceeded prior expectations and is heading higher into next year.
We do not read this as a structural impairment, for two reasons.
First, we have seen this exact pattern before. During the Hopper-to-Blackwell transition, gross margin compressed through the ramp — 72.5% in Q2 FY26 on a non-GAAP basis, with full-year FY26 GAAP margin at 71.1% before recovering to 75.0% GAAP and 75.2% non-GAAP by Q4 FY26 as Blackwell shipped at volume. Ramp economics on a new architecture are front-loaded: yields improve, fixed costs amortise across larger volumes, and mix normalises. With Vera Rubin now in full production and guided to roughly 20% of Data Center revenue in Q3, scaling from next quarter onward, we would expect a directionally similar recovery.
Second, NVIDIA is passing cost through. Management confirmed executed price increases take effect in Q1 FY2028, which is what drives the recovery from the 71–72% trough to 72–73%. Pricing power of that order, at this scale, in a market where customers are supply-rationed, is itself a significant competitive datapoint.
Where we would temper the optimism: input cost inflation is real and persistent, and memory is not a component NVIDIA can design around in the near term. Kress made the fair point that memory scarcity is itself a symptom of the same AI demand surge driving NVIDIA's growth but that does not make it costless. Price increases may not fully offset, and we think it is reasonable to expect margins to settle slightly below 75% rather than fully reclaiming it. At roughly $700bn of revenue, a 72–74% gross margin remains an extraordinary outcome. This is a quibble about the last two points of margin on a business compounding at 70%, and it should be treated as such.
5. Supply commitments more than doubled — a demand signal and an execution warning
Supply and capacity commitments rose from $119bn to $279bn in a single quarter, primarily related to memory procurement for the Vera Rubin platform. Total future commitments now stand at approximately $366bn.
We read this in two ways.
First, as confirmation of demand. A company does not contract $279bn of supply, more than doubling in ninety days unless it has high conviction in the order book behind it. Kress framed it precisely this way on the call: the commitments are essential for the Vera Rubin ramp this year and next, the largest portion falls within the first three years, and the scale of contracted supply is what underpins confidence in the revenue guide. This is the mechanism that converts the FY2028 number from a forecast into something closer to a schedule. It is also the strongest available answer to anyone arguing the 70% guide is promotional.
Second, and more soberly, as a step-change in execution risk. At this scale there is effectively no margin for error. A $279bn supply book locked in against a revenue plan leaves very little room for a design slip, a yield miss, a packaging bottleneck or a schedule delay and any of those now translates into materially larger absolute losses than would have been the case even two quarters ago. The commitments are largely non-cancellable; the demand behind them, however strong, is not contracted with the same rigidity.
This is not a hypothetical concern. NVIDIA has encountered execution issues before, and the most recent example is directly relevant to the platform now ramping. Rubin Ultra was originally unveiled at GTC 2026 as a four-die design, four near-reticle-sized compute dies with 16 HBM4E stacks and 1TB of memory in a single CoWoS-L package. Within roughly three months, industry reporting indicated the four-die configuration had been abandoned due to substrate warpage in CoWoS-L packaging, with thermal and structural stresses at that die count rendering yields uneconomical. The design reverts to a dual-die package, with NVIDIA reportedly attempting to recover system-level performance through a 2+2 board-level arrangement on the Kyber rack.
The point is not that the redesign is fatal — a board-level workaround may well preserve system performance, and TSMC's CoPoS packaging could resolve the underlying constraint, albeit not in volume before late 2028. The point is that a packaging decision cascaded through HBM specification, interconnect and networking architecture across an entire product generation, and it did so within a quarter of the product being publicly demonstrated. Against a $279bn supply book and a 70% growth commitment, that is precisely the category of risk that now carries outsized consequences.
This is the item we will be watching most closely. The demand picture is, on current evidence, extraordinary. The question for the next several quarters is execution.
Appendix — Figures
A. Q2 FY2027 delivered versus expectations
| Metric | Actual | Consensus | Company guidance | vs Consensus | vs Guidance |
|---|---|---|---|---|---|
| Revenue | $96.22bn | $92.07bn | $91.0bn ±2% | +4.5% | +5.7% |
| Non-GAAP diluted EPS | $2.22 | $2.09 | n/a | +6.3% | n/a |
| GAAP / non-GAAP gross margin | 75.0% / 75.0% | ~75.0% | ~75.0% | In line | In line |
| Data Center revenue | $89.0bn | $85.7bn | n/a | +3.9% | n/a |
| GAAP diluted EPS | $2.46 | n/a | n/a | n/a | n/a |
GAAP EPS exceeds non-GAAP because $7.77bn of net gains on equity securities sit in GAAP other income and are excluded from non-GAAP.
B. Guidance issued versus consensus
| Metric | NVIDIA guidance | Consensus | Delta |
|---|---|---|---|
| Q3 FY27 revenue | $108.0bn ±2% | $105bn | +2.9% |
| Q3 FY27 gross margin | 74.0% ±50bps | 74.8% | -80bps |
| Q3 FY27 non-GAAP opex | $9.0bn | n/a | n/a |
| Q3 FY27 non-GAAP EPS | Not guided | n/a | ~$2.45 implied on our estimates |
| FY28 revenue growth | ~70% | ~44% | +~26pts |
| FY28 revenue (implied) | ~$690–700bn | ~$570bn | +~21% |
| FY28 gross margin | 72–73% | Not separately published | — |
| Q4 FY27 gross margin | 71–72% (trough) | Not separately published | — |
Q3 FY27 guidance assumes zero Data Center compute revenue from China.
C. Headline P&L progression
| $m unless stated | Q2 FY26 | Q1 FY27 | Q2 FY27 | QoQ | YoY |
|---|---|---|---|---|---|
| Revenue | 46,743 | 81,615 | 96,221 | +18% | +106% |
| Data Center | 41,096 | 75,246 | 89,020 | +18% | +117% |
| Edge Computing | 5,647 | 6,369 | 7,201 | +13% | +27% |
| GAAP gross margin | 72.4% | 74.9% | 75.0% | +10bps | +260bps |
| Non-GAAP gross margin | 72.5% | 75.0% | 75.0% | — | +250bps |
| Operating income (GAAP) | 28,440 | 53,536 | 63,734 | +19% | +124% |
| Operating margin | 60.8% | 65.6% | 66.2% | +60bps | +540bps |
| Non-GAAP net income | 24,763 | 45,548 | 53,954 | +18% | +118% |
| Non-GAAP diluted EPS | $1.01 | $1.87 | $2.22 | +19% | +120% |
| GAAP diluted EPS | $1.08 | $2.39 | $2.46 | +3% | +128% |
D. Data Center composition
| Q2 FY27 | % of DC | QoQ | YoY | |
|---|---|---|---|---|
| Hyperscale | $49bn | 55% | +13% | — |
| ACI&E (NeoCloud / sovereign / enterprise / industrial) | $40bn | 45% | +25% | +138% |
| Total Data Center | $89.0bn | 100% | +18% | +117% |
E. Gross margin path — precedent and guidance
| Period | GAAP gross margin | Context |
|---|---|---|
| Q2 FY26 | 72.4% | Blackwell ramp |
| FY26 (full year) | 71.1% | Ramp year, plus H20 charges |
| Q4 FY26 | 75.0% | Blackwell at volume — margin reclaimed |
| Q1 FY27 | 74.9% | |
| Q2 FY27 | 75.0% | Reported |
| Q3 FY27E | 74.0% ±50bps | Vera Rubin ramp begins; memory inflation |
| Q4 FY27E | 71–72% | Guided trough |
| FY28E | 72–73% | Price increases effective Q1 FY28 |
F. Commitments and scale
| Item | Prior quarter | Q2 FY27 | Change |
|---|---|---|---|
| Supply and capacity commitments | $119bn | $279bn | +134% |
| Total future commitments | n/a | ~$366bn | — |
| Q2 capital returned to shareholders | — | ~$26.0bn | — |
| Remaining buyback authorisation | — | ~$99.0bn | — |
| Quarterly dividend per share | $0.01 (Q2 FY26) | $0.25 | 25x |
