NVIDIA Q2: After $96.2B of Revenue, the Risk Perimeter Reaches Power and Credit

For the quarter ended July 26, 2026, NVIDIA reported $96.221 billion of revenue, up 18% sequentially and 106% year over year. Data Center revenue was $89.023 billion, up 18% and 117%, respectively. GAAP and non-GAAP gross margin were both 75.0%, and GAAP operating income was $63.734 billion. Q3 revenue is guided to $108 billion, plus or minus 2%, with no China Data Center compute revenue assumed.

The mix inside Data Center is more informative than the headline. Hyperscale revenue was $48.710 billion, up 13% sequentially. AI Clouds, Industrial and Enterprise revenue was $40.313 billion, up 25% sequentially and 138% year over year. Incremental demand is broadening beyond a small set of hyperscalers into AI-native companies, enterprises, sovereigns and hyperscalers consuming infrastructure through AI clouds.

Growth quality cannot be read from the income statement alone. Operating cash flow fell to $24.1 billion from $50.3 billion in Q1. Accounts receivable reached $63.1 billion and DSO rose from 45 to 60 days because NVIDIA extended terms on large multi-quarter agreements with certain investment-grade customers. Inventory increased to $31.6 billion from $25.8 billion ahead of Vera Rubin. Revenue, receivables, inventory and cash flow are no longer moving together.

Growth did not slow, but the engine is becoming a full-platform expansion

The Blackwell Ultra ramp pushed Data Center to another record. Compute & Networking revenue was $88.299 billion, up 18% sequentially and 114% year over year. NVIDIA says Blackwell represents the vast majority of revenue, so the transition did not create a financial air pocket. Vera Rubin is in full production, with racks running at CoreWeave, Google Cloud, Microsoft Azure, Oracle and Nebius.

Customer mix matters more. ACIE grew almost twice as fast sequentially as hyperscale, supporting broader demand. NVIDIA did reclassify one customer from ACIE to hyperscale and recast prior periods, so the comparison must use the restated series rather than treating classification as a new customer event.

China Hopper shipments were less than 1% of Data Center revenue, and Q3 again assumes zero China Data Center compute revenue. The quarter was not driven by a China reopening; China is also absent as a buffer. Export policy and local substitution remain scenarios outside management's base case.

Margin remains exceptional while the constraint shifts into memory and system synchronization

A 75% gross margin shows that Blackwell Ultra mix still covered ramp cost. Q3 margin is guided to 74%, however, and management indicated on the call that the trough could be 71% to 72% in Q4 before stabilizing at 72% to 73% in fiscal 2028. That path combines Vera Rubin launch costs, memory inflation and system bill of materials; it is not a single demand signal.

The hardest supply evidence is contractual. Supply and capacity commitments jumped from $119 billion to $279 billion, primarily for memory. Roughly $92 billion falls in the remainder of FY2027, $87 billion in FY2028 and $88 billion in FY2029. NVIDIA is securing several years of HBM and related capacity, trading supply certainty for greater forecast risk.

If end demand continues higher, those commitments protect delivery. If product cadence, customer finance or campus activation slips, inventory and non-cancellable procurement can amplify margin pressure. The question has moved beyond HBM availability to whether each generation of memory, GPUs, networking and racks reaches customer acceptance in the same window.

NVIDIA is helping customers secure land, power and credit, not merely selling equipment

CFO commentary lists $366 billion of future commitments: $279 billion of supply and capacity, $29 billion of cloud services, $25 billion of uncommenced data-center leases, $25 billion of equity investments and $8 billion of capital expenditure. These are not one class of debt or an immediate cash outlay. Together, they show NVIDIA reserving positions across supply, R&D infrastructure, model ecosystems and data-center capacity.

A separate $56 billion schedule covers third-party AI-cloud agreements and uncommenced leases. The model can extend beyond upfront infrastructure sales: when criteria are met, NVIDIA can participate in revenue generated by AI clouds from third-party customers. Upside is recurring participation; downside introduces utilization, customer credit and lease-transfer execution.

Guarantees are the change that deserves explicit pricing. Existing land, power and shell guarantees for AI-cloud partners carry $3.5 billion of maximum exposure. The initial 4.25GW PORTS-Pike phase carries a guarantee cap of $105 billion, activated in phases as facilities become ready for service and declining as OpenAI makes lease payments. Total maximum guarantee exposure is $108.5 billion. It is neither a realized loss nor an immediate $108.5 billion payment, but it is a conditional credit ceiling.

1. 01
   Supply commitment

   NVIDIA reserves wafers, memory, packaging and systems, taking demand-forecast risk.

2. 02
   Land, power and shell

   Developers build campuses while NVIDIA credit support improves financeability.

3. 03
   System sale and lease

   Hardware acceptance creates revenue; some leases are expected to be reassigned.

4. 04
   Customer cash flow

   AI-cloud and model-company utilization ultimately must cover rent and compute payments.

The useful call signal is the boundary among supply, margin and finance

Management said customer forecasts imply demand growth could be closer to doubling, with supply limiting deliverable growth. That shifts the near-term bottleneck from end demand into HBM, packaging, networking, racks and campus activation. It supports the cycle while making revenue more dependent on synchronization across companies and engineering layers.

The margin path matters more than the current 75%. Rubin launch costs and memory prices are expected to push Q4 lower before price and efficiency actions stabilize fiscal 2028. If growth continues and margin follows that bridge, the decline is transition cost. If margin falls while receivables, inventory and guarantees keep rising, the issue becomes capital efficiency.

The right financing question is not whether every arrangement is circular. It is whether risk is priced and isolated. Watch guarantee activation, collateral and recourse, successful lease assignment, revenue share generated by independent third-party demand, and NVIDIA's cash obligation after customer default. The $108.5 billion ceiling should neither be booked as a loss nor dismissed as a footnote.

Five hard measures now define earnings quality

First, delivery against the $108 billion revenue guide. Second, a consistent bridge among Rubin revenue, inventory turns and margin. Third, whether DSO retreats from 60 days and operating cash flow reconnects with profit growth. Fourth, whether $279 billion of supply commitments expands again. Fifth, whether PORTS-Pike and other guarantees reach formal ready-for-service, financing and lease-transfer milestones.

The counter-case is also observable: memory and system cost rises faster than pricing, Rubin activation lags inventory, customer terms extend further, or campuses require more NVIDIA credit support. Any one can be manageable. In combination, they would require NVIDIA to be valued less like a pure high-margin platform and more like an infrastructure orchestrator with project-finance exposure.