Summary

NVIDIA sells complete AI factories. The unit of sale is no longer a chip or even a server but a rack-scale system — 72 GPUs joined by NVLink into one machine — surrounded by CPUs, three or more networking fabrics, and a software stack that spans data preparation, pre-training, post-training and agentic inference. That integration is the company's actual moat: customers can mix and match components, and most lack the engineering depth to want to. It is also a pricing mechanism. Management now quantifies its own share of a customer's build in dollars per gigawatt of data-center capacity: about $18B per gigawatt in the Hopper generation, $25B with Grace Blackwell, and $40B with Vera Rubin, which began production shipments this month.

The customer problem is no longer whether AI compute pays for itself. On the Q2 call management argued that compute is fully utilized across every cloud it serves and that the constraint has moved to supply. The evidence supports the direction if not the certainty: hyperscale revenue grew 102% year over year, and the non-hyperscale bucket the company calls ACIE — AI clouds, industrial and enterprise, including neoclouds and sovereign buyers — grew 138%, reaching $40.3B and roughly 45% of Data Center revenue. Management expects that bucket to settle at about half the business.

That composition shift is the decisive uncertainty, because the two halves have very different balance sheets. Hyperscalers fund themselves. The AI labs and neoclouds do not, and the 10-Q says so in the company's own words: they are growing faster than their balance sheets and credit profiles can support, and lack the long-term contracts and investment-grade financing capacity to secure infrastructure independently. NVIDIA's response has been to supply the missing credit itself — $105B of lease guarantees behind an OpenAI affiliate signed in August 2026, $3.5B behind other AI clouds, nearly $50B invested in frontier labs, and take-or-pay revenue floors that let third-party lenders underwrite neocloud data centers.

The revenue is not fictitious; the compute is being used and the buyers are paying. What changed on August 26 is that the risk moved from a question about end demand to a measurable, dated, contractual exposure sitting on NVIDIA's own disclosures. The company that six months ago wrote that it had entered into no financing arrangements now has a guarantee table.

Why This Moment Is Different

The fiscal 2026 annual report, filed February 25, 2026, addressed customer financing in a single sentence under counterparty risk. It said the company had been asked to offer financing arrangements to support customers' data-center buildouts, and then stated plainly that it had not entered into any.

The 10-Q filed August 26, 2026 replaces that sentence with a table. Maximum gross exposure under guarantees is $108.5B: $3.5B of land, power and shell guarantees for AI cloud partners' lease obligations, and $105B of credit support entered into in August 2026 with affiliates of SB Energy Corp. on behalf of an affiliate of OpenAI Group PBC, covering roughly 4.25 gigawatts of IT load at the PORTS Technology Campus in Pike County, Ohio.

The structure is worth reading precisely, because it is more contained than the headline number suggests and more revealing than a summary allows. The guarantees are capped in aggregate and limited to defined portions of lease and power payments rather than the full cost of the site or all of the tenant's obligations. They become effective only as each of nine construction phases is completed, the first expected in fiscal 2029, and they decline over each phase's 20-year lease term as payments are made. In exchange, the campus exclusively hosts NVIDIA compute. NVIDIA holds a unilateral option to extend credit support to roughly 3.8 additional gigawatts.

One clause carries more information than the rest. The guarantees terminate on certain events including OpenAI achieving a satisfactory credit rating. The disclosure therefore states, as a contractual condition, that the counterparty does not have one today. NVIDIA is lending its balance sheet precisely because the tenant cannot yet rent 4.25 gigawatts on its own.

Three other commitments moved in the same quarter and belong in the same frame:

• Supply and capacity commitments rose from $119B to $279B in a single quarter, which management attributed primarily to the procurement of memory. Of the $279B, $92B falls due in the remainder of fiscal 2027 and $87B in fiscal 2028.
• Equity investments reached $99B held plus $25B committed. Purchases of equity securities were $42.4B in the six months ended July 26, 2026, against $1.2B in the comparable prior period.
• A new arrangement with select AI clouds totals $36B of NVIDIA commitments, typically six years. Under it, the AI clouds buy NVIDIA infrastructure and NVIDIA commits to buy cloud services back. The 10-Q notes those clouds may unilaterally stop providing the service to NVIDIA and sell the capacity to third parties at more advantageous rates.

Management addressed the obvious characterization directly on the call, acknowledging that some observers will call the arrangement circular financing and saying the company sees it differently — the labs are once-in-a-generation businesses, the compute is fungible and redeployable, and the risk is therefore limited. The most load-bearing disclosure in that answer was quantitative rather than rhetorical: the CFO said demand from the AI labs against which NVIDIA expects to leverage its balance sheet should contribute roughly a quarter of next year's business.

How the Business Makes Money

Products and go-to-market. Two reportable segments, which sum to the $96.221B quarter: Compute & Networking segment revenue of $88.299B, up 114% year over year, and Graphics segment revenue of $7.922B, up 46%. The company also reports by market platform, and reorganized that view this quarter into just two: Data Center ($89.023B) and Edge Computing ($7.198B). Data Center splits into Hyperscale ($48.710B) and ACIE ($40.313B). One company was reclassified from ACIE to Hyperscale during the quarter with prior periods recast, so the sub-segment growth rates are not strictly comparable to previously published figures.

Stated growth vectors, in the company's own framing. Management describes three engines: an architecture that runs every model, closed and open; a full-stack AI factory that raises revenue captured per gigawatt of customer capacity; and a CUDA ecosystem reaching customers with no interest in designing custom silicon. Concrete commitments behind that: Vera Rubin entered production with purchase orders from every major hyperscaler, AI cloud and system OEM and is expected to be about 20% of Data Center revenue in Q3; the Vera CPU is in full production with CPU revenue expected to more than double in fiscal 2028 against a stated server-CPU opportunity around $20B; and Groq 3 LPX, the first rack-scale product from the Groq license, is in full production with Nebius as first customer. NVIDIA paid $2.944B under the Groq arrangement in the six months and carries $986M of accrued purchase consideration.

Customer concentration is the number that moved. For Q2 fiscal 2027, one direct customer was 16% of total revenue, against two at 23% and 16% a year earlier — a genuine improvement in revenue concentration. Receivables went the other way. Five direct customers accounted for 22%, 14%, 13%, 11% and 10% of the accounts receivable balance as of July 26, 2026 — 70% among five names, against 56% among three at January 25, 2026, on a receivable balance that itself grew from $38.5B to $63.1B. Indirect concentration is disclosed only qualitatively: the company estimates one AI research and deployment company contributed a meaningful amount of revenue by purchasing cloud services from NVIDIA's customers.

Manufacturing. Fabless. Wafers from TSMC and Samsung, memory from SK hynix, Micron and Samsung, CoWoS packaging, assembly and test through Hon Hai, Wistron and Fabrinet. The supply chain is concentrated in Asia with expansion into the U.S. and Latin America underway. The memory dependency is now the margin story rather than a footnote.

Top filed risks, from the annual report. Supply-demand mismatch from long lead times and non-cancellable orders; dependence on third-party foundries and packaging; customer and partner inability to secure power, land and capital; counterparty risk on commercial arrangements; export controls; and investment-portfolio sector concentration. The annual report also discloses roughly 6,000 employees in Israel supporting networking R&D, a securities class action relating to 2017–2018 channel inventory statements, and a September 15, 2025 preliminary finding by China's antitrust regulators that NVIDIA's export-control compliance violated conditions of the Mellanox approval.

China is effectively zero and is guided as zero. The annual report states the company was effectively foreclosed from China's data-center compute market as of the end of fiscal 2026. In the first half of fiscal 2027 NVIDIA took a $0.4B charge on H200 excess inventory and purchase obligations; shipments made under the U.S. licences were less than 1% of Data Center revenue; those units carry a 25% tariff on U.S. importation that the company says it has been unable to pass to customers; and management called current Hopper shipments dilutive to corporate gross margin. There is no China data-center compute revenue in the outlook.

What the Street and Public Conversation Say

The news layer around this print is thin and should not be leaned on. A search pull completed August 27, 2026 at 00:34 UTC returned five items for NVDA, of which two are usable: two CNBC pieces dated August 26 covering the results and the memory-price question, both reporting a beat with shares slightly lower. A third item is NVIDIA's own release for the first quarter of fiscal 2027 carrying an inferred publication date — a stale document that would misstate the quarter if read as current. A fourth is a price-target promotion piece with an inferred date and no analytical content. The fifth is an unrelated Indian brokerage listing page. This is not coverage; it is a thin, partly mis-dated sample, and the analysis above rests on the filings instead.

The sell-side questions on the call are more informative for what they did not ask. Five analysts asked about demand: the durability of the 70% guide, inference share, open models, recursive self-improvement, and the ranking of supply constraints. One — Vivek Arya of Bank of America — summed the commitment tables to about $500B and asked whether that was the total and how it squares with OpenAI and Anthropic designing their own silicon. Stacy Rasgon of Bernstein worked out that the fiscal 2028 guide implies roughly $200B above the prior three-year framing and asked what the unconstrained number would be; management said it would be a lot higher. Nobody asked about the receivables build, the 70% concentration in the receivable balance, or what happens to the $105B guarantee if the first phase does not reach service in fiscal 2029. The balance-sheet question is being left to the filings.

One claim from the call should not be carried forward as fact. Management said return on invested capital for AI data centers is now under a year, prefaced by "I heard the other day." That is hearsay from the CEO, not a company metric, and no filing supports it.

The Conviction Chain

Claim 1 — Demand exceeds supply, and the constraint is real rather than rhetorical. Evidence: revenue accelerated for a fourth consecutive quarter to +106% year over year; supply and capacity commitments more than doubled to $279B with $92B due within the year; management guided fiscal 2028 to roughly 70% while stating customer forecasts point to demand doubling. Confirms it: Q3 delivering near the $108B guide; Vera Rubin reaching the stated ~20% of Data Center revenue; supply commitments continuing to convert into shipments. Breaks it: a quarter where revenue lands at the low end while inventory and commitments keep climbing — the signature of commitments made against a forecast rather than an order book.

Claim 2 — Gross margin bottoms at 71–72% and recovers, because memory pricing is a pass-through the company can re-price. Evidence: management guided Q3 to 74.0%, a Q4 trough of 71–72%, and 72–73% for fiscal 2028 as executed price increases take effect in Q1, attributing the reset to memory costs that exceeded prior expectations and are headed higher. Confirms it: the Q4 trough landing in range and Q1 fiscal 2028 showing the price increases. Breaks it: a Q4 print below 71%, or a fiscal 2028 margin guide that slips again — which would mean NVIDIA is absorbing memory inflation rather than passing it on, and that its pricing power over customers it is simultaneously financing is weaker than assumed.

Claim 3 — The financed half of demand pays. This is the weakest link and the reason conviction is medium. Evidence for: guarantees are capped, phased, limited to defined portions of lease and power payments, and terminate on the counterparty reaching a satisfactory credit rating; the compute itself is fungible and redeployable; independent capital underwrites each project and NVIDIA states it is not making loans; roughly a quarter of next year's business is the stated exposure, not the majority. Evidence against: days sales outstanding moved from 45.4 in Q1 to 59.6 in Q2 on the company's own arithmetic, which management attributed to extended payment terms for large multi-quarter purchases by investment-grade customers; receivable concentration rose to 70% across five names; the entire structure is six months old and has never been tested through a funding drought; and the equity stakes NVIDIA holds in the ecosystem are marked to market through earnings, so a repricing of the counterparties hits revenue quality and reported profit at the same time. Breaks it: a receivable write-down, a guarantee phase triggering, an ACIE growth stall, or a neocloud failing to reach financial close on a project NVIDIA has floored.

Claim 4 — Reported profit quality is deteriorating at the margin even as operations improve. Evidence: GAAP net income rose 2% sequentially while non-GAAP net income rose 18%, because the GAAP figure was flattered by $12.8B more equity gain in Q1 than in Q2. Gains from equity securities were $7.771B in Q2, 10.9% of pre-tax income, against 7.2% a year earlier. Six-month operating cash flow was $74.4B against $118.0B of net income — a 63% conversion rate — with the gap accounted for by $23.7B of non-cash equity gains, a $24.6B receivable build and a $10.2B inventory build. Confirms it: continued divergence between GAAP and non-GAAP EPS growth. Breaks it: cash conversion recovering toward historical levels as the multi-quarter shipments behind the extended terms are collected.

Financial Proof

Margins and growth rates are computed from the underlying statement values, not transcribed. Two year-over-year cells are marked n/a because the fiscal 2025 Q3 and Q4 revenue lines are absent from the vendor statement file; they are not zero and were not estimated. The FY2027 Q2 row is taken from the primary filings because the vendor file still ends at Q1. Q2 operating cash flow is the six-month figure less Q1, both from the filings.

Balance-sheet movement over the six months to July 26, 2026: receivables $38.5B to $63.1B; inventories $21.4B to $31.6B (119.3 days on Q2 cost of revenue); marketable equity securities $12.9B to $42.8B; non-marketable securities $22.3B to $51.2B; long-term debt $7.5B to $32.4B against $24.9B of debt issued; total assets $206.8B to $320.3B. Uses of cash in the half — $39.0B of buybacks, $6.3B of dividends, $4.4B of capital expenditure and $42.4B of equity purchases — totalled $92.2B against $74.4B of operating cash flow, a $17.8B shortfall funded by the debt raise. A $25B commercial paper programme is in place with nothing drawn.

Valuation. Under the company's own guidance — Q3 at $108B, roughly 70% fiscal 2028 growth, 72–73% gross margin and a 16–18% tax rate — and a desk assumption of 35% operating-expense growth off an approximate $34B fiscal 2027 base, fiscal 2028 EPS models to $15.27–$15.88. At $209.66 that is 13.2x–13.7x. The operating-expense assumption is the softest input and is the desk's, not the company's; doubling it still leaves the multiple in the mid-teens. The market is not paying a growth multiple for guided 70% growth.

Positioning. Short interest was 292.7M shares at the July 31, 2026 settlement, down from 324.1M at July 15, at 2.3 days to cover. There is no squeeze setup and no crowded short.

Stock Setup

The stock closed at $209.66 on August 26, down 1.59% into the print, on 1.54x average volume. RSI is 46.3 on a Wilder basis — mid-range, no extreme in either direction. Price sits below the 10- and 20-day averages and above the 50- and 200-day, with a golden cross intact and an 18.6-week base. It is 11.4% below the 52-week high.

The one-year record is the most interesting line in the table and does not appear on the chart: NVDA is up 15.5% over twelve months with negative one-year alpha, meaning it lagged the market over a period in which quarterly revenue went from $46.7B to $96.2B. Multiple compression has absorbed essentially all of the fundamental improvement. That is the setup — a business accelerating into a de-rating — and it is why the entry is defined against the support cluster rather than chased at the pivot.

Entry zone $190–$208, where the 50-day, the anchored VWAP, the shelf and the 200-day cluster. Stop $182, below the point of control. Target $280, about 18x modelled fiscal 2028 EPS and roughly 18% above the prior high.

Catalysts

• Q3 fiscal 2027 results, November 17, 2026. The Q4 gross-margin trough of 71–72% gets confirmed or denied one quarter ahead of the fact, and the first Vera Rubin revenue quarter is measured against the stated ~20% of Data Center revenue.
• Q1 fiscal 2028 pricing. Management said the memory pass-through arrives as executed price increases in Q1. That is the direct test of pricing power.
• Definitive agreements on the $500B financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. The 10-Q describes memorandums of understanding that may not lead to definitive agreements.
• First PORTS-Pike construction phase, expected fiscal 2029 — the point at which the first tranche of the $105B guarantee becomes effective.
• Any frontier-lab IPO. Management said two of the invested labs will likely go public soon, which would move a portion of the $51.2B non-marketable book to marketable and put a public price on the ecosystem stakes.
• Jensen Huang at the Goldman Sachs technology conference September 10 and the GTC Berlin keynote October 21.

Risks

• Counterparty. $108.5B of guarantees, $99B of equity stakes and $36B of cloud commitments against customers the company itself describes as outgrowing their credit profiles.
• Memory cost. The margin reset was driven by memory pricing that exceeded management's own prior expectations and is guided higher into next year. NVIDIA buys from three suppliers in a market its own demand is tightening.
• Concentration. 70% of a $63.1B receivable balance across five names.
• Custom silicon. Both OpenAI and Anthropic are designing their own inference chips; an analyst cited OpenAI's public claims for a part benchmarked against Blackwell. Management's answer was platform breadth, not a technical rebuttal.
• Export control. China data-center compute is zero and guided zero. The unresolved Mellanox antitrust finding is a live tail risk to the networking business specifically.
• Reflexivity in reported earnings. Equity gains are 10.9% of pre-tax income and the holdings are concentrated in the same ecosystem that generates the revenue, so a downturn compounds through both lines at once.

Kill Tests

The thesis is wrong, and the position should be exited rather than trimmed, if any of the following occurs:

1. A receivable provision or write-down against any named ACIE customer.
2. Q4 fiscal 2027 gross margin prints below 71%, or fiscal 2028 margin guidance is cut again.
3. ACIE revenue declines sequentially in any quarter while supply commitments remain above $200B.
4. A guarantee phase is triggered, or the SB Energy arrangement is restructured before the first phase reaches service.
5. Fiscal 2028 revenue guidance is revised below roughly 50% growth.

Cross-References

• The August 22, 2026 tracked-company profile framed the same question before the print — that NVIDIA was building the financing layer beneath its own demand — and asked who ultimately funds the capacity. This dive is the post-print answer, with the guarantee table the profile could not yet cite.
• NVDA sits in 30 watchlists including the focus list, and is a multi-year outperformer currently in pullback on the August 26 monster read: up 853% over five years but 15.5% over one. Over that same twelve months MU returned 706%, WDC 488% and STX 412%. The memory and storage suppliers captured the re-rate; the company selling the systems did not.
• The memory story runs both ways. NVIDIA's supply commitment jump from $119B to $279B is attributed primarily to memory procurement, and it announced a multiyear technology partnership with SK hynix in the quarter — direct corroboration for the memory-supercycle track, from the buyer's side of the table.
• Vera Rubin racks are named as running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius; Nebius is also the first Groq 3 LPX customer. Two of those are tracked neocloud names, and the neocloud cohort closed uniformly lower on August 26 — the counterparties NVIDIA is underwriting are the ones the market repriced the same day.
• The $500B financing platform partners overlap the AI-power-delivery track's capital layer, and the 4.25-gigawatt PORTS-Pike campus is a power-and-shell asset before it is a compute asset.

Sources

Primary filings and company documents:

• NVIDIA Form 10-Q, quarter ended July 26, 2026 (filed 2026-08-26)
• NVIDIA Form 10-K, fiscal year ended January 25, 2026 (filed 2026-02-25)
• NVIDIA Form 8-K (filed 2026-08-26)
• Press release — Q2 fiscal 2027 results, 8-K EX-99.1
• CFO commentary on Q2 fiscal 2027 results, 8-K EX-99.2
• NVIDIA Q2 FY2027 earnings call transcript, August 26, 2026 (Yahoo Finance, powered by Quartr) — attributed management and analyst statements only; the filings above own the actuals.

News characterized in this dive:

• Nvidia earnings live updates: Q2 results, memory prices and AI outlook — CNBC, 2026-08-26
• Nvidia shares slide slightly despite Q2 beat, top tech analysts weigh in — CNBC, 2026-08-26. The pull returned only the quote-page URL (https://www.cnbc.com/quotes/NVDA) for this item, not an article permalink; cited as characterized, not as a resolvable source.

Repo receipts:

• income-statements, cash-flow-statements.json, balance-sheets.json, short-interest.json
• 10k-business, 10k-risk_factors.json
• company-facts, key-financials.json, 10Q-2026-08-26.html
• latest (pull completed 2026-08-27T00:34:11Z, 5 articles, 2 usable)
• focus (price, RSI, moving averages, base and pivot; August 26, 2026 close)
• 2026-08-22-NVDA
• 2026-08-26-monster-scan