NVIDIA 2Q FY27 Earnings:Supply Sets the 70% FY28 Guide; Demand Would Support Close to 100%

NVIDIA designs accelerated computing platforms for AI. It reports two market platforms: Data Center, which at US$89.0bn was 93% of 2Q FY27 revenue and splits between hyperscalers and a second line covering AI clouds, industrial and enterprise customers; and Edge Computing, which covers workstations, consumer graphics, robotics and automotive. The current platform is Blackwell Ultra, with Vera Rubin entering production in August 2026. NVIDIA does not own fabs; TSMC manufactures its leading-edge silicon and memory comes from the three major DRAM suppliers.

NVIDIA reported 2Q FY27 revenue of US$96.2bn, up 106% YoY and 18% QoQ, against a consensus of US$92.3bn. Non-GAAP EPS of US$2.22 came in 6.0% above the US$2.09 consensus. Gross margin held at 75.0% on both a GAAP and non-GAAP basis, and operating income rose 124% YoY to US$63.7bn. The US$14.6bn added sequentially is the largest single-quarter increase the company has recorded.

The more consequential disclosure was forward-looking. NVIDIA guided 3Q FY27 revenue to US$108.0bn plus or minus 2%, about 3.6% above the US$104.2bn consensus, and then guided a full fiscal year ahead for the first time, putting FY28 revenue growth at approximately 70%. Management framed that number as a function of secured supply rather than expected demand.

Two features of the quarter deserve weight before the guide. Growth accelerated for a fourth consecutive quarter, from 62% in 3Q FY26 to 106% here, which is not the shape a business produces when it is running into a ceiling. And the acceleration came with a mix shift toward customers outside the hyperscalers, where NVIDIA sells complete systems rather than components.

A company capped by supply does not usually post four quarters of accelerating growth. The reconciliation is that NVIDIA has been adding capacity throughout, so each quarter delivered more than the last while still delivering less than customers asked for. The constraint shows up not in the reported growth rate but in the gap between the FY28 number management guided and the number it said demand would support.

Data Center revenue of US$89.0bn rose 117% YoY and 18% QoQ. Within it, Hyperscale revenue of US$48.7bn grew 102% YoY but only 13% QoQ, while AI Clouds, Industrial and Enterprise, the line covering neoclouds, sovereigns and enterprises, grew 138% YoY and 25% QoQ to US$40.3bn. Edge Computing added US$7.2bn, up 27% YoY.

The sequential split matters more than the annual one. Hyperscale decelerated to 13% QoQ while the non-hyperscaler line ran at 25%, so ACIE supplied US$8.1bn of the US$14.6bn added in the quarter. Management expects 3Q sequential growth to be driven primarily by ACIE again, with hyperscale reaccelerating in 4Q as Vera Rubin supply builds.

Two mechanical points sit under the mix. Neocloud installed capacity is expected to reach 8GW by the end of 2026 from about 3GW at the end of 2025, and these customers buy complete AI factories rather than accelerators, which is why NVIDIA captures more revenue per unit of capacity from them. Sovereign business, sold mostly through regional neoclouds, grew 35% QoQ and more than tripled YoY.

The composition of the growth reduces one risk and raises another. A customer base spread across sovereigns, enterprises and neoclouds is less exposed to any single hyperscaler cutting capex than the Hopper-era base was. It is also less creditworthy on average, which is the trade the balance sheet now carries and section 8 returns to.

Gross margin was 75.0% on both bases, up 2.6 percentage points YoY and flat sequentially. The annual improvement came from Blackwell Ultra mix; the sequential flatness reflects Blackwell remaining the large majority of revenue, so there was no mix change to move the number. Operating margin reached 66.2% as revenue grew 106% against a 55% increase in operating expenses.

The 1Q FY26 dip to 60.5% is not a trend point. It carries a US$4.5bn H20 charge taken when export restrictions stranded China-bound inventory, and margin recovered to 72.4% the following quarter. Read without that quarter, gross margin has climbed steadily from 72.4% to 75.0% across five quarters on Blackwell mix.

Operating leverage is the quieter part of the quarter. Operating expenses grew 55% YoY against 106% revenue growth, adding 5.4 percentage points to operating margin. Management guided full-year operating expense growth to the low 50s percent and attributed part of the restraint to internal use of AI tools, so the leverage should persist even as gross margin compresses.

Commitments to buy supply and capacity rose from US$119bn at the end of 1Q to US$279bn, an increase of US$160bn in one quarter that management attributed primarily to memory procurement. Cloud service agreements add US$29bn, data center leases not yet commenced US$25bn, equity investments US$25bn and capital expenditure US$8bn, taking total future commitments to US$366bn.

The maturity profile is what makes the number legible. US$92bn falls in the remainder of FY27, US$87bn in FY28 and US$88bn in FY29, so 96% of the commitment lands inside the window the growth guide covers. This is cost contracted ahead of revenue that is forecast rather than ordered, and it is the clearest evidence that management believes the demand it described.

NVIDIA does not disclose units, so revenue per gigawatt is the closest available measure of content per unit of capacity. It has risen from about US$18bn on Hopper to US$25bn on Grace Blackwell and US$40bn on Vera Rubin, as the platform absorbed the CPU, three types of networking and now an inference accelerator. Revenue can therefore grow substantially even in a year where the gigawatts NVIDIA can equip are capped.

The downstream side is newer. NVIDIA has invested nearly US$50bn in frontier AI labs, guarantees up to US$105bn on the SB Energy PORTS-Pike campus that will host OpenAI under 20-year leases, and has arranged financing platforms with six infrastructure capital providers to raise over US$500bn of third-party capital. Management expects roughly a quarter of next year's business to come from labs relying on that support.

Management addressed the circularity objection directly, arguing the compute is fungible and redeployable and that the exposure is limited. The redeployability claim is credible: a Vera Rubin rack has many possible buyers in a market where demand exceeds supply. The claim holds only while that condition does, and the guarantees run 20 years.

The first half delivered US$177.8bn and the 3Q guide adds US$108.0bn, so nine months of FY27 are effectively booked at US$285.8bn. Reaching the US$395.1bn consensus requires a 4Q of US$109.3bn, which would be only 1.2% above the guided 3Q. The quarterly consensus stack implies US$117.7bn for 4Q, which is 9% sequential growth and consistent with Vera Rubin volume building.

That spread is why the FY27 base for the FY28 growth calculation is a range rather than a point. At the low end FY27 lands near US$395bn and 70% growth implies US$672bn; at the high end FY27 reaches US$404bn and 70% implies US$686bn. Either way the FY28 figure sits about US$100bn above where consensus currently is.

Consensus carries FY28 revenue of US$570.3bn from 41 analysts, which is 44.3% growth on the US$395.1bn FY27 consensus. The company guided approximately 70%. On the same base that implies about US$672bn, and on the higher FY27 base implied by the quarterly consensus stack it implies about US$686bn. The gap to consensus is therefore roughly US$100bn, or 18-20%.

Management also said demand would support growth close to 100%, which on the same base would be roughly US$800bn. The 30 percentage point spread between the guided and unconstrained figures is the quantity worth tracking, because closing any part of it requires no new demand, only more memory and more packaging capacity.

The margin guide is the price of the supply guide. Management reset expectations to 74.0% for 3Q, a 71-72% trough in 4Q, and 72-73% for FY28 once executed price increases take effect from 1Q. That is roughly 3.5 percentage points from peak to trough, of which about one point comes back. On the guided FY28 revenue range, each percentage point of gross margin is worth roughly US$6.8bn.

The FY28 margin guide should be treated as the less certain half of the outlook. Management said memory prices exceeded prior expectations and are heading higher into next year, then guided a margin that assumes price increases offset most of the remaining pressure. If memory inflation runs ahead of what has been contracted, the 72-73% band is where the guide breaks first, not the revenue number.

Four themes from the call carry beyond the quarter. Note that NVIDIA had not published its own transcript at the time of writing, so the following is summarised from a third-party transcript rather than quoted.

First, the guide is a supply figure. Asked what separates the guided number from the demand number, management said the entire supply chain is running flat out, that supply supports about 70% growth while demand is materially higher, and that the company would keep working to raise the figure. Management also said it guided a year ahead specifically so customers, suppliers and shareholders would plan against the same number.

Second, memory is what holds supply back, and management framed the scarcity as a symptom of the same AI buildout driving NVIDIA's own growth rather than an unrelated cost shock. The company said it is working with all three major memory suppliers to expand the capacity its roadmap requires, and expects supply to stay short at least through the end of FY28.

Third, on the financing structure, management acknowledged the circular-financing characterisation and rejected it, arguing the labs are once-in-a-generation companies whose growth is limited by compute rather than by demand, and that NVIDIA's exposure is bounded because the platform is fungible and redeployable. It also disclosed a revenue-sharing structure with neoclouds under which it provides a minimum revenue guarantee and shares in rental revenue above that floor.

Fourth, on competition from customers designing their own inference silicon, management drew the distinction between a single-purpose chip for one cloud and a platform spanning the full AI lifecycle across every cloud. The argument is coherent for training and mixed workloads. It is weaker for steady-state inference at a single large buyer, which is precisely where the custom parts are aimed.

Operating cash flow of US$24.1bn was 40% of GAAP net income of US$59.7bn. The reconciliation is working capital and non-cash gains: receivables consumed US$22.3bn, inventories US$5.8bn and prepaid items US$5.5bn, while US$7.8bn of gains on equity securities sit in net income without generating cash. Days sales outstanding rose to 60 from 45 on extended payment terms for large multi-quarter agreements.

NVIDIA funded the gap rather than slowing down. It issued US$25.0bn of senior unsecured notes, taking long-term debt from US$7.5bn to US$32.4bn, and still returned a record US$26.0bn to shareholders through US$19.7bn of buybacks and US$6.0bn of dividends. That is 60% of free cash flow year to date against a policy of 50% or more.

Extending terms to customers who are themselves capital-constrained is new this quarter. A 15-day move in days sales outstanding on a US$96bn quarterly revenue base is roughly US$16bn of additional financing NVIDIA is extending, and it lands on the same customer set the company is already supporting with guarantees and equity. Watching days sales outstanding is now as informative as watching the revenue line.

GAAP earnings should be read with care for the same reason. Gains on equity securities contributed US$7.8bn in the quarter and US$23.7bn across the first half, which is why GAAP net income rose only 2% sequentially while non-GAAP net income rose 18%. As NVIDIA's stakes in private AI companies grow, the GAAP line will increasingly track the marks on those holdings rather than the operating business.

NVIDIA trades at 25.3x forward earnings, 14.1x forward EV/Sales and 27.5x forward EV/EBITDA against peer medians of 60.3x, 15.2x and 31.6x. The forward P/E median is not a usable benchmark. It is set by Intel at 60.5x and Arm at 112.5x, both on forward earnings bases that are small relative to their market values, with Qualcomm at 15.6x at the other end. A median drawn from that spread describes dispersion rather than a fair multiple.

Broadcom is the closest comparison on scale and AI exposure, at 31.7x forward earnings and 31.6x forward EV/EBITDA against NVIDIA's 25.3x and 27.5x. The gap reflects a forward earnings base that is both large and growing fast enough to compress the multiple. Whether it makes the shares attractive depends on the FY28 outcome, which is the open question rather than a settled input.

Our stance is structurally positive over the long term and cautious over the next several quarters. The long-term case rests on the platform position, content per gigawatt rising from US$18bn to US$40bn across three generations, and demand that exceeds what NVIDIA can supply. The near-term caution rests on the margin trough, the working-capital build and the counterparty exposure set out in sections 6 and 8.

Zero One Investment Research does not carry a rating or target price on NVIDIA, and this note does not take a view on whether the current share price is cheap or expensive. The company sits in broad coverage rather than focused coverage, so the multiples above are drawn from market data and consensus estimates rather than from a Zero One forecast model, and no forecast financial statements accompany this note.

Counterparty credit is the first-order risk and it is new this cycle. NVIDIA carries US$108.5bn of guarantees, nearly US$50bn of frontier-lab equity, and expects roughly a quarter of FY28 revenue from customers management describes as unable to finance their own infrastructure. A funding failure at one large lab would hit revenue, the guarantee, and the equity stake at the same time.

Memory cost inflation is the second. The FY28 gross margin guide of 72-73% assumes executed price increases offset most of the remaining pressure, on the same call where management said memory prices are heading higher than expected. A trough below 71% or an FY28 band below 72% would be the first sign that pass-through is incomplete.

Commitment overhang is the third. US$279bn of supply commitments, 96% of it falling within three fiscal years, is cost contracted against forecast rather than ordered revenue. If growth comes in materially below the guide, that obligation does not scale down with it.

Two persistent risks remain unchanged. China Data Center compute is assumed at zero in the outlook, which removes downside but means any reopening is upside the company is not modelling. And the largest customers continue to design their own inference silicon while relying on NVIDIA capital, a tension that has no resolution visible in this release.

What would change the view in either direction is the memory contract price series. Continued deceleration through the fourth quarter turns part of the 30 percentage point gap between guided and unconstrained growth into revenue that requires no new demand. Re-acceleration puts the FY28 margin band at risk before it puts the revenue guide at risk.

The tables below carry reported figures only. NVIDIA sits in broad coverage, so no Zero One forecast columns accompany them.