August 26, 2026 14 min read

NVIDIA (NVDA) — Q2 FY2027 Earnings Briefing

Source: transcripts/NVDA/Q2_2027.txt — NVIDIA Q2 FY2027 earnings call, August 26, 2026.

Executive Summary

NVIDIA delivered another record, accelerating quarter: revenue more than doubled YoY to $96B, with data center demand broad enough to lift both hyperscale and non-hyperscaler segments to records.
The demand signal remains far above what NVIDIA can ship: customer forecasts exceed a supply-capped plan of ~70% growth in fiscal 2028.
The cost picture worsened: extreme memory pricing resets the gross-margin path lower through the first half of fiscal 2028.
The verdict: demand is not in question; execution is — capacity conversion, margin defense, and the financing structures underwriting frontier-lab growth.

Highlights and Lowlights

Highlights

Data center revenue rose 18% QoQ to $89B on sustained Blackwell strength, with hyperscale customers reporting accelerating growth and expanding margins.
The non-hyperscaler ACIE segment grew 138% YoY, led by NeoCloud, enterprise, and sovereign demand.
Vera Rubin entered production shipments with purchase orders from every major hyperscaler, AI cloud, and system OEM, and is expected to reach ~20% of data center revenue in Q3.
AWS expanded its partnership to deploy an additional 2 million GPUs through Q2 FY2029, adding Vera CPUs and NVIDIA’s physical-AI stack.
NVIDIA assembled financing platforms targeting over $500B of third-party capital with six global infrastructure partners to fund frontier-lab buildouts.
Cloud-industry backlog passed $2T, the demand evidence NVIDIA credits for keeping hyperscaler CapEx rising.
NVIDIA returned a record $26B to shareholders in Q2, split between repurchases and dividends.

Lowlights

Gross-margin expectations were reset lower on extreme memory pricing, with guidance for a Q4 bottom of 71–72% before only a partial recovery.
Customer forecasts point to demand doubling next fiscal year, and management conceded supply will bottleneck the business at least through fiscal 2028.
China data center compute fell to under 1% of Q2 data center revenue, and no China data center revenue is assumed in the forward outlook.
Roughly a quarter of next year’s business is expected to come from AI labs whose infrastructure NVIDIA helps finance — the “circular financing” critique management addressed directly.
Working capital stretched into the ramp: inventory reached $32B while days sales outstanding reached 60 days as large investment-grade customers received extended payment terms.
Analysts pressed on custom-silicon competition — an OpenAI in-house accelerator program and claims it outperforms Blackwell — while management dismissed such chips as single-service inference parts and reaffirmed the labs as long-term customers.

1. The Demand Backdrop: Acceleration Against a Supply Ceiling

The quarter extended an unusual streak: growth accelerated for a fourth consecutive quarter, with the surge driven by a global AI infrastructure buildout spanning hyperscalers, AI labs, AI natives, enterprises, and sovereign customers.

Demand evidence. CapEx by the top five hyperscalers is expected to approach $800B in 2026 and $1.3T in 2027, funded against a cloud-industry backlog that keeps climbing (see Highlights). Management pointed to compute’s visible payoff — hyperscaler revenue growth and margin expansion as GPU capacity comes online — as the reason the spending cycle keeps reinforcing itself.
Agentic workloads. Jensen Huang argued agents are the demand multiplier: the compute needed for an AI agent versus a human user runs 15–100x depending on the task, models keep getting larger and smarter, and reasoning agents push through multiple turns of tool use. He said most AI usage crossed into agentic in the last month.
Why guide a year out. Jensen said NVIDIA had never guided a full year ahead before, but land, power, and shell lead times now run 2–3 years, so NVIDIA works further upstream (memory suppliers, power generators, land/power/shell partners) and downstream to align planning; giving everyone “the same set of information” is the point.
The supply reality. Asked to rank constraints, Jensen declined to single one out: the entire supply chain is running flat out, more capacity comes online continuously, and yields keep improving. He said secured supply supports the guided trajectory while demand sits well above it, and that the unconstrained number would be “a lot higher” — the company is currently running at about 100% YoY growth on its own reckoning.

2. Segment Performance: Hyperscale and ACIE

Segment	Q2 FY2027 revenue	QoQ	YoY
Total revenue	$96B	—	more than doubled
Data Center	$89B	+18%	—
— Hyperscale	$49B	+13%	—
— ACIE (AI clouds, industrial & enterprise)	$40B	+25%	+138%

Hyperscale. Sustained Blackwell strength drove the segment, reinforcing NVIDIA’s argument that more compute translates directly into more customer revenue as new GPU capacity comes online; hyperscaler customers reported accelerating revenue growth and expanding margins of their own.
ACIE is now half the franchise. The non-hyperscaler segment — sovereigns, regional NeoClouds, enterprise edge, and air-gap data centers — is expected to represent roughly half of the data center business, a share management says is the market’s invisible half because these customers buy full AI factory platforms rather than chips.
NeoCloud capacity is scaling fast. Partners using NVIDIA’s DSX reference designs are expected to exit the year with 8 gigawatts of installed capacity, up from approximately 3 gigawatts at the end of 2025, bringing capacity online faster and at lower token cost.
Sovereign AI is the sharpest edge. Sovereign revenue (largely via regional NeoClouds) grew 35% sequentially and more than tripled YoY — countries and regions can allocate land and power to neutral regional partners in ways they will not offer foreign hyperscalers. Examples ranged from Firebird in Armenia and Cassava across Africa to GMI Cloud in Taiwan, Yotta and Neysa in India, Firmus in Australia, YTL in Malaysia, Noetra in Japan, and LG and Hyundai in South Korea, plus a record slate of new supercomputers in Europe.
Enterprise and AI-native breadth. On a trailing 12-month basis, on-prem automotive revenue reached $8B while financial services, manufacturing, and healthcare combined for $7B; Hudson River Trading and Jane Street run AI factories for quantitative trading, Samsung uses cuLitho for computational lithography gains, and Bristol Myers Squibb is building on Vera Rubin. Global VC funding in AI reached over $400B in the first half of 2026 versus $265B in all of 2025, and roughly 20 companies now exceed $1B in annualized run-rate revenue, up from 13 in Q4.

3. Guidance and the Margin Reset

Metric	Guidance
Q3 FY2027 revenue	$108B ± 2%
Q3 FY2027 gross margin (GAAP & non-GAAP)	74% ± 50bp
Q4 FY2027 gross margin	bottom at 71–72%
FY2028 gross margin	settle at 72–73%
Q3 FY2027 OpEx	$9.2B GAAP / $9.0B non-GAAP
FY2027 OpEx growth	low 50s % YoY
FY2027 tax rate	16–18%, excluding discrete items
FY2028 revenue growth	~70% YoY, supply-constrained
Q3 data center mix	Vera Rubin about 20% of data center revenue

The guidance table is the full picture; the story behind it is the margin reset. Both GAAP and non-GAAP gross margins held at 75% in Q2, and Q2 OpEx rose 10% GAAP / 11% non-GAAP sequentially on high compute infrastructure costs plus compensation. But management said “extreme pricing conditions” in memory exceeded prior expectations and are headed higher into next year — and chose to reset expectations immediately to avoid letting the issue “linger as an open question.”

NVIDIA’s framing is that memory scarcity is largely a symptom of the same AI buildout driving its own growth, not an unrelated cost shock, and it is working with all three major memory suppliers to expand capacity. Executed price increases take effect in Q1; that is when margins are expected to settle at the fiscal 2028 level rather than recover to prior peaks.
Q3’s sequential revenue growth is expected to come primarily from ACIE within data center; hyperscale is expected to reaccelerate in Q4 and into fiscal 2028 as Vera Rubin supply builds.
Fiscal 2027 OpEx growth of low 50s percent reflects a broadening product portfolio and greater use of AI tools, which management says is already enhancing engineering productivity. The tax rate range is unchanged.

4. The Platform Roadmap: Vera Rubin, Vera CPU, Groq, Networking

Vera Rubin began production shipments in August with purchase orders from every major hyperscaler, AI cloud, and system OEM; management expects it to be the fastest ramp in company history and the mix target is in Highlights. It delivers 30x higher throughput per megawatt and 35x lower token costs versus Grace Blackwell Ultra, the payoff of extreme co-design across GPU, CPU, NVLink, scale-up and scale-out networking, systems, algorithms, and software. The AWS expansion gives it a marquee deployment: AWS adds Vera CPUs — some integrated with Rubin, some standalone — and will serve NVIDIA’s Nemotron family on Bedrock and SageMaker while adopting Omniverse, Cosmos, Isaac, and Jetson for warehouse robotics (deployment scale and schedule in Highlights).
Vera CPU is in full production as a standalone product that expands the addressable market: it completes agentic tasks 1.8x faster on the spec benchmark and offers 5x the bandwidth per watt of any other data center CPU. Deployments are expected from every major hyperscaler, NeoCloud, AI lab, and system OEM, with shipments underway to lead partners OCI, SpaceX AI, and AWS. Demand stands at approximately $20B for total server CPUs, and CPU revenue is preliminarily expected to more than double in fiscal 2028. Grace, introduced in 2021, has already exceeded $5B in trailing-twelve-month revenue.
Groq 3 LPX, the first rack-scale LPU system, is in full production and posts nearly 4x the tokens per second of the next-best alternative on the Artificial Analysis benchmark. Volume shipments begin later this quarter, with Nebius the first adopter; the product targets services that need maximum interactivity at high ASPs, while most data centers will remain Vera Rubin plus NVLink 72.
Networking posted another record quarter, with revenue up 18% sequentially; Spectrum-X Ethernet grew 2.6x YoY, and management now describes NVIDIA as on track to be the largest and fastest-growing network company. NVLink 72 rack-scale systems are in their third generation, and the platform now spans multiple networking types — scale-up, scale-out, scale-across, security, and multicampus.
Model universality. Management’s pitch is that NVIDIA is the only platform running every frontier model, closed or open, and that open and closed model adoption are both skyrocketing: nearly all open models run on NVIDIA, open models let every enterprise and country build proprietary AI, and the CUDA footprint from PCs and workstations to robots and on-prem data centers keeps the platform fungible across the full AI life cycle.

5. The Gigawatt Monetization Ladder

Era	NVIDIA revenue opportunity per gigawatt
General-purpose computing	~$3–5B
Hopper	~$18B
Grace Blackwell	~$25B
Vera Rubin	~$40B
Beyond Vera Rubin	Higher still — the goal is to maximize compute per site

Jensen’s framing: the goal is to put as much compute as possible on each plot of land and gigawatt — the “perfect answer” he described is effectively infinity per gigawatt — and every generation so far has moved in that direction.
The driver is platform breadth: each generation absorbs more of the AI factory (CPU, GPU, NVLink, networking, and now the LPU), while a single fungible system spans data preparation, pre-training, post-training, and agentic inference — preserving the customer’s investment across the AI life cycle and keeping NVIDIA’s share of the data center TAM rising.
Customers race to each new generation because the productivity X-factor compounds: Jensen cited return on invested capital of under a year on $50B data-center projects, which is why rental economics remain “insanely profitable” even at today’s scale.

6. Financing the Buildout: Frontier Labs and the NeoCloud Model

The problem being solved. Frontier AI labs have extraordinary training and inference demand, but they are growing faster than their balance sheets and credit profiles can support; their growth is limited by compute, not technology or customer demand.
NVIDIA’s direct commitment. NVIDIA has invested nearly $50B in the frontier labs — a meaningful commitment that management says is a small fraction of expected free cash flow over the same period.
The financing platforms. Six leading infrastructure capital providers — Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR — are partnering with NVIDIA on platforms that will raise third-party institutional capital for lab infrastructure at attractive rates (scale in Highlights).
Anchor sites and commitments. NVIDIA secured land, power, and shell capacity with SoftBank Energy at its Portsmouth campus, where an initial 4.25 gigawatts is expected to host OpenAI workloads; each generation of systems there could represent approximately 1.5 million GPUs, and the site can support multiple upgrade cycles over 20 years — the essential economics being a long-lived site with repeatedly upgradable compute. OpenAI’s existing and planned commitments represent approximately 12 gigawatts of NVIDIA compute; for another frontier lab, NVIDIA will provide selective credit enhancement for nearly 2 gigawatts, complementing that lab’s independent capacity.
Answering “circular financing.” Management expects the criticism and rejects it: these are once-in-a-generation companies whose technology leadership, customer traction, and usage are proven; NVIDIA’s risk is limited because its compute is fungible, durable, and redeployable; and the compute shipped to financed labs is expected to be consumed by investment-grade customers or those backed by one. Still, management disclosed the share of next year’s business expected to depend on these labs (see Lowlights).
The NeoCloud revenue-share model. Instead of requiring a single long-term offtake that lenders can underwrite, NVIDIA offers a take-or-pay commitment on a portion of a facility’s capacity — a minimum revenue guarantee that gives lenders confidence — and shares in revenue above that floor. Independent capital still underwrites every deal; as management put it, “We’re not making loans.” In this structure NVIDIA gets paid twice: once on the hardware sale and again through a recurring share of rental revenue, a model management says could drive billions of revenue over the medium to long term.

7. Capital Returns, Working Capital, and China

Capital returns hit a record in Q2 (see Highlights) — $20B through repurchases and $6B through dividends of $0.25 per share — equal to 60% of free cash flow year to date against a plan to return 50% or more. Management intends to increase returns of excess free cash flow net of strategic uses.
Working capital absorbed cash ahead of the Vera Rubin launch — inventory and days sales outstanding both stepped up (see Lowlights) as extended payment terms for large investment-grade customers taking delivery over multiple quarters lifted receivables.
China remains effectively closed: Q2 included only a token amount of China-licensed Hopper product revenue (see Lowlights), those shipments dilute gross margin, and no China data center compute revenue is assumed in the forward outlook given ongoing geopolitical uncertainty.

Conclusion

The quarter leaves NVIDIA with a demand problem only in the sense that it cannot serve all of it: the growth signal broadened beyond hyperscalers while supply remains the binding constraint for at least another year. The decision to guide a full year ahead is a statement of confidence in visibility, but it also accepts that upside stays theoretical until capacity converts. The margin reset is the more consequential development — cost inflation from memory is arriving just as mix shifts toward the most capable systems, and management is choosing to absorb it while defending platform economics. Meanwhile, the financing web around frontier labs keeps the structural question alive about how much demand is intermediated, rather than simply earned. On balance, backlog, platform breadth, and roadmap argue for durability; the burden of proof has moved from demand to execution — conversion, mix, and margin recovery.

Artifacts and Follow-Ups

Key Artifacts — highest priority

Gross-margin trajectory. The memory-driven reset is now the most contested item in the story. Baselines: the Q3 guide at 74% ±50bp, a Q4 bottom at 71–72%, and a fiscal 2028 settle at 72–73% as executed price increases take effect in Q1. Watch whether the trough holds, and whether memory pricing commentary worsens or stabilizes.
Supply conversion versus the fiscal 2028 plan. Baseline: ~70% YoY revenue growth, explicitly supply-constrained, with bottlenecks expected to persist at least through fiscal 2028’s end. Confirm or refute via management’s language on closing the supply-demand gap and any change to the multi-year capacity outlook.
Vera Rubin ramp. Baselines: about 20% of Q3 data center revenue, described as the fastest product ramp in company history, with orders from every major hyperscaler, AI cloud, and system OEM and AWS deployments starting now. Watch Q4 mix and ramp linearity.
Financing exposure to AI labs. Baselines: nearly $50B invested; over $500B of third-party platforms being raised; ~12 GW of OpenAI commitments; ~2 GW of credit enhancement; roughly a quarter of next year’s business from financed labs. Watch new commitments, cash deployment, and guarantee structures.
Demand indicators into year-end. Baselines: cloud backlog above $2T; top-five hyperscaler CapEx projected at $1.3T in 2027; NeoCloud capacity exiting the year at 8 GW; sovereign revenue more than tripled YoY. Watch Q3 bookings and whether hyperscale growth actually reaccelerates in Q4 as guided.

Secondary Tracking Items

Operating-expense growth and the sequential step-up in Q3 (see guidance table).
Days sales outstanding and inventory as Vera Rubin shipments scale.
Shareholder-return pace against the stated free-cash-flow plan.
Groq 3 LPX volume shipments to early adopters (Nebius first) later this quarter.
Networking growth and Spectrum-X attach rates; NVLink 72 third-generation ramp.
Gaming and other non-data-center segments, conspicuous by their absence from this call’s commentary.
Management appearances: Goldman Sachs Communacopia (September 10) and GTC Berlin (October 21); Q3 call scheduled for November 17.