NVDA's $165 Billion Question: What NVIDIA's 10-K Reveals

NVIDIA builds the computing infrastructure for artificial intelligence — GPUs, networking, and software platforms that power everything from ChatGPT to autonomous vehicles. Last year, it generated $216 billion in revenue with 75% gross margins, returning $41 billion to shareholders while spending only $6 billion on capex.

By every traditional measure, this is the most dominant operating performance of any technology company in history. Operating cash flow of $102.7 billion. Return on invested capital of 87.3%. A fabless business model where TSMC handles manufacturing and NVIDIA captures the economics. If you stopped reading at the earnings release, you'd see a flawless growth story.

But the FY2026 10-K, filed February 25, 2026, tells a more nuanced story. Scattered across five separate footnotes — supply commitments, cloud agreements, investment obligations, operating leases, and product warranties — NVIDIA has accumulated $165.4 billion in total commitments. Every category is growing 2 to 5 times faster than revenue. A $17 billion deal with Groq created $14.4 billion in goodwill on zero revenue. And a $40 billion equity investment portfolio financially entangles NVIDIA with its own customer ecosystem. These aren't reasons to sell. They're the specific conditions that determine whether 25x forward earnings is a reasonable price or an aggressive one.

NVIDIA's AI Computing Platform

NVIDIA's dominance begins with CUDA, the software ecosystem that locks AI researchers and engineers into NVIDIA hardware. Over a decade of library development, framework integration, and developer tooling has created switching costs that no competitor has matched. On top of CUDA, NVIDIA's NVLink and NVSwitch proprietary interconnect technology — which grew networking revenue 142% to $31.4 billion in FY2026 — makes multi-GPU clusters increasingly dependent on NVIDIA's full-stack architecture.

The result is a platform company, not a chip company. Data Center accounts for 89.7% of revenue ($193.7 billion), with Gaming (5.1%), Professional Visualization (2.5%), and Automotive (2.7%) as side businesses. Within Data Center, networking is now the fastest-growing segment — at $31.4 billion, it grew 2.4 times faster than compute ($162.4 billion, +59%). This shift matters because it deepens lock-in: once a hyperscaler deploys NVLink fabric across a cluster, switching GPUs means replacing the interconnect too.

This platform dominance explains the operating metrics — and also explains why the balance sheet commitments are so large. When you own the standard, you commit to maintaining it.

The Revenue Machine

The FY2026 headline numbers are extraordinary by any measure, but the most compelling signal in the filing isn't the revenue growth everyone already knows — it's the $9 billion that customers deposited before NVIDIA even shipped the product.

Customer advances reached $9.0 billion in FY2026, up 2.5 times from $3.6 billion in FY2025. This is cash that hyperscalers and enterprise buyers paid upfront to secure future GPU shipments. At 4.2% of revenue, it's not transformative — but as a directional signal, prepayments growing faster than revenue means demand is intensifying, not plateauing.

The ROIC decline deserves attention. Even as revenue grew 65%, ROIC fell from 100.6% to 87.3% because capital deployed grew faster — driven by the Groq deal ($17 billion), expanding equity investments ($22.3 billion in non-marketable alone), and inventory that more than doubled. NVIDIA is still earning exceptional returns, but the incremental dollar of capital is working less hard than the last one.

The demand picture has one important nuance: hyperscaler revenue accounts for "slightly over 50%" of Data Center revenue according to CFO commentary, but growth was "led by the rest" of Data Center customers. Revenue is diversifying even as concentration metrics worsen — a paradox explained by the fact that NVIDIA's largest customers are growing their absolute spending while smaller customers are growing their share faster.

NVIDIA's customers deposited $9 billion in advance payments in FY2026 — 2.5 times the prior year — the strongest demand signal in the filing and a direct counterpoint to concerns about AI spending sustainability. If customer advances decline below $6 billion in a future quarter, that signal weakens.

The $165 Billion Balance Sheet

This is where the 10-K diverges from the earnings narrative. The FY2026 filing reveals $165.4 billion in total commitments — not in a single table, but scattered across five separate footnotes that nobody aggregates. We built a combined scorecard.

We define a Commitment Coverage Ratio (CCR) as total obligations divided by annual operating cash flow. NVIDIA's CCR is 1.61x — meaning the company would need to dedicate 1.61 years of current cash flow to cover all commitments, assuming no growth. Cash plus one year of OCF ($62.6B + $102.7B = $165.3B) barely covers total obligations with less than $1 billion to spare.

The 1.61x ratio is manageable at current growth rates. The problem is the trajectory.

Every obligation category grew faster than revenue. Supply commitments nearly tripled — adding $44.9 billion in the final quarter alone. Cloud commitments more than doubled. Non-marketable equity investments grew 556%, driven by AI ecosystem bets. Future data center leases — which don't appear on the balance sheet until they commence — tripled to $22.7 billion.

The filing's own language reveals the flexibility, and its limits:

"We enter into agreements with our supply vendors that allow them to procure inventory based upon our defined criteria, and in certain instances, these agreements are cancellable, able to be rescheduled, or adjustable for our business needs prior to placing firm orders. Changes to these agreements may result in additional costs."

"Cancellable" sounds reassuring until you read "may result in additional costs." The cost of exiting $95.2 billion in supply commitments is nowhere disclosed — and substantially all of it is due through FY2027, concentrating the risk in a single year:

"As of January 25, 2026, these commitments were $95.2 billion, of which substantially all will be paid through fiscal year 2027."

Then there's the off-balance-sheet infrastructure build. NVIDIA is signing 20-year data center leases for R&D facilities — a structural shift from the fabless, asset-light model that generated 87% ROIC:

"Between fiscal years 2027 and 2030, we expect to commence leases with future obligations of $22.7 billion, primarily data center leases to support our research and development efforts, with lease terms of 1.8 to 20 years."

Whether these commitments matter depends entirely on growth. At 30%+ revenue growth, cash flow absorbs the commitments easily. Below that, the commitment-to-revenue ratio worsens and write-downs become probable. NVIDIA's FY2026 10-K reveals $165.4 billion in total commitments across five separate footnotes — supply, cloud, investment, lease, and warranty obligations that equal 1.61 times annual operating cash flow and are growing 2 to 5 times faster than revenue.

The Inference Bet

If the $165 billion in commitments reveal how NVIDIA is financing its growth, the Groq deal reveals where it thinks growth is going. In December 2025, NVIDIA paid $17 billion for a technology that competes with its own GPUs.

"In December 2025, we entered into a non‑exclusive license agreement with Groq, Inc., or Groq, for its language processing unit technology and hired certain Groq employees. No customer contracts, existing products, or equity interests were purchased."

This is not a traditional acquisition. NVIDIA didn't buy Groq's business — it licensed the language processing unit (LPU) technology and hired Groq's engineers. No customer contracts, no products, no equity stake. The deal produced $14.4 billion in goodwill and a $2.5 billion developed technology intangible asset with a five-year useful life — meaning $500 million in annual amortization starting FY2027.

"We recorded $14.4 billion of goodwill and a $2.5 billion developed technology intangible asset, valued using a cost‑to‑recreate methodology with a five‑year useful life. Goodwill, primarily attributable to the workforce and future development of the licensed technology, was recorded in the Compute & Networking reporting unit."

The accounting tells the story. NVIDIA's goodwill is backed by workforce and R&D potential — not by an operating business with revenue, customers, or products. Compare this to how peers have deployed goodwill:

AMD paid $49 billion for Xilinx and got an operating FPGA business with $3.7 billion in revenue and an established customer base. Broadcom paid $61 billion for VMware and got $13 billion in software revenue. Only NVIDIA paid a premium for a technology license with no revenue stream. This is either visionary — securing inference technology at a $13.4 billion effective cost after the tax deduction on goodwill — or a $14.4 billion impairment risk if LPU technology doesn't appear in NVIDIA products within 3-5 years.

The Groq deal connects to a broader pattern: NVIDIA's equity investment portfolio reached $40 billion in FY2026. Non-marketable investments surged from $3.4 billion to $22.3 billion in a single year, with another $17.7 billion in publicly-held equity. Some of these investees are NVIDIA's direct GPU customers — creating a circular capital flow where NVIDIA finances the companies that buy its products.

NVIDIA paid $17 billion for a non-exclusive Groq technology license that created $14.4 billion in goodwill on zero revenue — the only major tech acquisition in recent years where the goodwill is backed by workforce and R&D potential rather than an operating business.

What 25x Forward Earnings Requires

NVIDIA trades at approximately 25 times consensus FY2027 earnings, implying roughly 25% annual EPS growth sustained for three or more years. The question isn't whether NVIDIA deserves a premium — it does, by every operating metric. The question is what specific filing-level mechanisms could slow the implied growth rate.

The filing reveals three headwinds. First, the effective tax rate is rising: from 12.0% in FY2024 to 13.3% in FY2025 to 15.1% in FY2026, with management guiding 17-19% for FY2027. At the 18% midpoint, the 2.9 percentage point ETR increase would add approximately $4.1 billion in annual tax expense — a direct headwind to earnings growth. Second, Groq intangible amortization adds $500 million annually to operating expenses. Third, the $165 billion in commitments must be absorbed without margin compression — a manageable task at 30%+ growth, but increasingly difficult below that threshold.

Relative to AI semiconductor peers, NVIDIA is not expensive:

At 34.2x EV/EBITDA, NVIDIA is the cheapest AI semiconductor leader relative to its profitability. AMD trades at 77.5x despite margins less than a fifth of NVIDIA's. Broadcom trades at 50.9x with half the growth rate. The risk isn't relative valuation — it's absolute: what has to go right for 25x forward earnings to hold.

We identified five filing-derived monitoring thresholds that make the conditional nature of the valuation explicit:

At approximately 25x forward earnings, the market is pricing NVIDIA for continued 25%+ EPS growth. The FY2026 10-K supports that trajectory — $102.7 billion in operating cash flow, $9 billion in customer prepayments, and 75% gross margins leave room for execution. But it also documents the mechanisms that could slow it: a rising effective tax rate adding $4 billion in annual drag, $165 billion in commitments that must be absorbed without margin compression, and $14.4 billion in goodwill backed by technology potential rather than revenue. If three or more of the five monitoring thresholds reach amber, the implied growth rate faces compression risk. Watch the Q1 FY2027 10-Q filing (expected May 2026) for the first quarterly update on commitment levels, customer concentration, and Groq integration progress.

NVIDIA trades at 25 times forward earnings, implying 25% annual EPS growth for three or more years — a target complicated by a rising effective tax rate (guided to 17-19%, up from 15.1%), $165 billion in commitment absorption, and $500 million in annual Groq intangible amortization.

Frequently Asked Questions

What does NVIDIA do?

NVIDIA designs graphics processing units (GPUs) and full-stack data center platforms for AI training, inference, and high-performance computing. Its CUDA software ecosystem and NVLink/NVSwitch interconnect technology create switching costs that make its platform the industry standard for AI workloads. In FY2026, 89.7% of revenue ($193.7 billion) came from Data Center, with Gaming, Professional Visualization, and Automotive comprising the remainder.

What is the $165 billion in total obligations?

The FY2026 10-K discloses obligations across five separate footnotes: supply and manufacturing commitments ($95.2 billion), cloud service agreements ($27.0 billion), investment commitments ($11.4 billion), operating leases and future leases ($25.6 billion), product warranty ($2.8 billion), and other commitments ($3.4 billion). Totaled, these reach $165.4 billion — equivalent to 1.61 times NVIDIA's $102.7 billion in annual operating cash flow. No single table in the 10-K presents this aggregate view.

Are NVIDIA's supply commitments cancellable?

Partially. The 10-K states they are "cancellable, able to be rescheduled, or adjustable for our business needs prior to placing firm orders." However, the filing also notes that "changes to these agreements may result in additional costs." The cost of exiting is not disclosed, and substantially all of the $95.2 billion is due through FY2027, concentrating the exposure.

Who are NVIDIA's largest customers?

The 10-K discloses two customers above 10% of revenue: one at 22% (approximately $47.5 billion) and another at 14% (approximately $30.2 billion). They are not named. Accounts receivable concentration is worse: three customers represent 56% of $38.5 billion in receivables, with a $4 million allowance for doubtful accounts — essentially zero provisioning against top-customer default risk.

What is the Groq deal and why does the goodwill matter?

In December 2025, NVIDIA entered a non-exclusive license agreement with Groq for its language processing unit technology for $17.0 billion ($13 billion at closing plus $4 billion payable). NVIDIA recorded $14.4 billion in goodwill and a $2.5 billion intangible asset. No customer contracts, products, or equity interests were purchased. Pro forma revenue was "not material." This is the only major tech deal in recent years where goodwill is backed entirely by workforce and R&D potential rather than an operating business.

How does NVIDIA's ROIC compare to AMD and Broadcom?

NVIDIA's FY2026 ROIC is 87.3%, compared to AMD's 6.0% and Broadcom's 48.4%. Despite the wide gap, NVIDIA's ROIC declined 13.3 percentage points year-over-year even as revenue grew 65%. This reflects a growing capital base from the Groq deal, expanding equity investments, and rising inventory — incremental capital is earning lower returns as the company scales.

Is NVIDIA investing in its own customers?

Yes. NVIDIA's equity investment portfolio reached $40 billion in FY2026 — $22.3 billion in non-marketable equity (up from $3.4 billion) and $17.7 billion in publicly-held equity, including Intel shares with $6.6 billion in unrealized gains. Some investees are NVIDIA's direct GPU customers, creating a circular capital flow where NVIDIA finances the companies that buy its products.

What should investors watch in NVIDIA's next quarterly filing?

Five metrics derived from the FY2026 10-K: (1) supply commitment level — did it exceed $100 billion, indicating accelerating obligations? (2) A/R concentration — did the top customer exceed 25% of receivables? (3) customer advances — did prepayments sustain above $9 billion? (4) effective tax rate — is it tracking toward the 17-19% guidance? (5) Groq-related disclosures — any product announcements or goodwill impairment language in the notes.

Is NVIDIA's operating cash flow quality good?

Yes, with caveats. Operating cash flow of $102.7 billion represents 47.6% of revenue, which is strong by any standard. Cash conversion of 85.6% is healthy. However, accounts receivable grew 67% versus revenue growth of 65%, and Q4 days sales outstanding reached 51 days according to CFO commentary — early signs of working capital stretch that bear monitoring.

How does NVIDIA's valuation compare to peers?

NVIDIA trades at 34.2x EV/EBITDA versus AMD's 77.5x and Broadcom's 50.9x. On a forward earnings basis, NVIDIA is the cheapest AI semiconductor leader relative to its growth rate. The risk is absolute rather than relative: 25x forward earnings requires sustained 25%+ EPS growth, and the filing documents specific mechanisms — tax rate increases, commitment absorption, and Groq amortization — that could slow it.

Methodology

Data Sources

This analysis is based on NVIDIA's FY2026 10-K filed February 25, 2026 (view the full filing), the accompanying 8-K earnings release and CFO commentary, and NVIDIA's FY2025 10-K for year-over-year comparisons. Financial metrics were calculated using MetricDuck's automated SEC filing analysis pipeline, which extracts XBRL-tagged data from over 5,000 US public companies. Peer data for AMD and Broadcom was sourced from their respective FY2026 and FY2025 10-K filings through the same pipeline.

Limitations

• Total obligations are an estimate. The $165.4 billion figure aggregates five footnotes that the company does not combine in a single disclosure. Some obligations may overlap or have netting arrangements not disclosed at this level of granularity.
• Cancellation costs are unknown. The filing states commitments are semi-cancellable but does not quantify exit costs. The true locked-in exposure could be materially lower than $165.4 billion.
• Customer identity is inferred. The 10-K does not name customers representing 22% and 14% of revenue. Our characterization of these as likely hyperscalers is based on cross-referencing prior disclosures and industry context.
• Forward earnings multiples use consensus estimates. The ~25x forward PE uses analyst consensus for FY2027 EPS, which may be revised. The valuation analysis is directional, not precise.
• Peer fiscal years differ. NVIDIA's FY2026 ends January 2026, AMD's ends December 2025, and Broadcom's ends November 2025. Comparisons use the most recent full fiscal year for each company, not identical calendar periods.
• Groq deal accounting is preliminary. The purchase price allocation is based on initial estimates and may be revised during the measurement period (up to one year from closing).

Disclaimer

This analysis is for informational purposes only and does not constitute investment advice. The author does not hold positions in NVDA, AMD, or AVGO. Past performance and current metrics do not guarantee future results. All data is derived from public SEC filings and may contain errors or omissions from the automated extraction process.