Compute Equals Commitments
NVIDIA and CoreWeave both beat revenue. Both fell. The 10-Ks confirmed what the series predicted: the risk was never demand. It was the commitments.
NVIDIA reported a $68.1 billion quarter on Tuesday evening — up 73% year-over-year, beating consensus by $2 billion, with Q1 guidance of $78 billion that crushed every estimate on the Street.[1] CoreWeave reported a $1.57 billion quarter the next evening — more than doubling year-over-year, beating by $20 million, with backlog swelling to $66.8 billion.[2] Both stocks fell — NVIDIA 5.5% on Wednesday, its worst session since April, and CoreWeave as much as 13% in after-hours trading.[3]
Two revenue beats. Two sell-offs. Same week, same sector, same message from the market: we believe in the demand. We’re worried about the commitments.
Over the past three weeks, I published a three-part series examining the structural architecture of AI infrastructure risk: NVIDIA’s four channels of neocloud control (”Jensen’s COMECON“), CoreWeave’s debt mechanics and Magnetar’s hedge (”What a GPU Debt Crisis Would Look Like“), and the dissonance between credit ratings and CDS pricing on GPU-backed debt (”Two Markets, One Asset“). This week’s filings are the first earnings test of each framework. Here’s what was confirmed, what changed, and what sharpened.
Jensen Huang repeated his axiom throughout the NVIDIA call: “Compute equals revenues.”[4] He’s right — in the generative AI economy, compute is the production input. Compute also equals commitments: non-cancellable purchase orders, long-term debt, and forward obligations placed before demand is fully known. That corollary is what the series mapped and what this week’s 10-Ks made visible.
NVIDIA: The COMECON deepens
In “Jensen’s COMECON,” I mapped NVIDIA’s quadruple role — simultaneously supplier, investor, guarantor, and customer of its own neocloud ecosystem — and documented how that architecture creates a reflexive financing loop in which money circulates through entities that each record it as a distinct financial claim on the same underlying asset. The Q4 10-K reveals how far that architecture has extended.
The number: purchase obligations of $95.2 billion, up from $16.1 billion a year ago — a nearly six-fold increase in twelve months.[5] Total supply obligations reached $117 billion, approaching NVIDIA’s $131 billion in annual operating cash flow.[6] The COMECON framework predicted that NVIDIA’s commitment architecture would deepen as the bilateral deal system scaled. The magnitude exceeded what the framework implied. No semiconductor company has ever carried purchase obligations at this scale relative to revenue. The filing warns these obligations are “expected to continue to grow and become a greater portion of supply.”[7]
The cause is structural. TSMC now demands longer-term contracts and advance payments to build custom fabrication and advanced packaging capacity.[8] Each generation requires more custom CoWoS packaging, more HBM per unit, and longer lead times. NVIDIA’s CFO was candid: inventory and purchase commitments are “further out in time than usual.”[9] The COMECON framework explains why: NVIDIA isn’t just buying chips for itself. It’s securing supply for an entire ecosystem of satellite companies whose ability to service their debt depends on NVIDIA delivering the hardware on schedule. The commitment architecture serves the bilateral deal system.
Michael Burry flagged the same $95.2 billion figure within hours of the filing, drawing a comparison to Cisco.[10] The parallel was a version of what I’d already mapped — a supplier extending commitments, anticipating perpetual growth. Bernstein’s Stacy Rasgon offered the bull case: the commitments ensure NVIDIA becomes “the most dependable supplier” in a market he estimates could reach $1.4 trillion, noting demand shows “absolutely zero signs of slowing.”[11] Both readings acknowledge the same fact: NVIDIA’s operating model has structurally changed. This is no longer a company that manufactures to demand. It’s a company that commits to supply chains years in advance, at scale, and trusts that demand will arrive to fill the capacity.
NVIDIA can absorb a demand pause: the company generated $97 billion in free cash flow in FY2026 and carries non-GAAP gross margins of 75%.[12] Yet absorption has a price. Last May, export controls stranded H20 inventory and purchase commitments overnight, producing a $4.5 billion charge.[13] The commitment ratchet only turns in one direction. At $95.2 billion, the ratchet has never been tighter. And in the COMECON architecture, NVIDIA’s commitments flow downstream: if NVIDIA is locked into $95 billion of supply, the neoclouds deploying that hardware are locked into the debt that funds it.
CoreWeave: The interest-income squeeze arrives
In “What a GPU Debt Crisis Would Look Like,” I flagged CoreWeave’s Q3 interest coverage ratio of 0.17x: for every dollar of interest owed, operating income covered seventeen cents. Q4 is worse. Q1 guidance is worse still.
CoreWeave’s revenue beat masked three numbers the market couldn’t ignore. Q1 guidance of $1.9–2.0 billion missed the $2.29 billion consensus by 15%.[14] Full-year 2026 capex guidance came in at $30–35 billion on projected revenue of $12–13 billion — a capex-to-revenue ratio between 2.3x and 2.9x.[15] Q1 interest expense guidance hit $510–590 million, consuming roughly a quarter to a third of Q1 revenue, against adjusted operating income of $0–40 million.[16]
The interest trajectory I tracked in the GPU Debt Crisis piece has continued on the same vector: quarterly interest expense has nearly quadrupled over the past 15 months, from $149 million in Q4 2024 to a guided midpoint of $550 million in Q1 2026, while adjusted operating income has compressed toward zero.[17] The company carries $21.4 billion in total debt.[18]
The defense rests on contracted demand. CEO Intrator cited $66.8 billion in signed agreements, up $11.2 billion sequentially, with average contract length extending to 5 years and no single customer accounting for more than 35% of backlog.[19] Management claims substantially all capex is tied to signed contracts. The weighted-average interest rate fell by 300 basis points in 2025, and no significant debt matures until 2029.[20] Over the lifetime of the backlog, the math may work: $66.8 billion in contracted revenue against $21.4 billion in current debt is a ratio the credit markets can underwrite.
However, a maturity wall isn’t the only failure mode. Q4 revealed a mechanism the GPU Debt Crisis piece anticipated but couldn’t yet quantify: the interest-income squeeze is not a phase the company passes through on the way to profitability. It is a structural feature of the growth itself. Every new contract requires capex. Every dollar of capex requires debt. Every dollar of debt generates interest. And interest is now growing faster than operating income, not because demand is weak, but because the capital structure of serving demand costs more than the demand currently pays. The bull case requires revenue to scale for several consecutive quarters while interest compounds. The bear case requires it to miss once.
The circular chain I mapped in “Jensen’s COMECON” — NVIDIA sells GPUs, invests the proceeds, backstops the capacity, then buys the cloud services — is now visible in a single quarterly filing. NVIDIA’s quadruple role deepened further in January: a 13% equity stake, a $2 billion fresh investment, a $6.3 billion capacity guarantee through 2032, and $27 billion in cloud service agreements.[21]
Two markets, same verdict
In “Two Markets, One Asset,” I documented the dissonance between the CDS market — pricing CoreWeave at 640–700 basis points, implying a 40–55% five-year default probability — and rating agencies assigning AAA ratings to data center debt backed by the same underlying asset class. Since publication, that dissonance has sharpened. By mid-December, Bloomberg data showed CoreWeave’s five-year CDS at 773 basis points, implying roughly 47.5% default probability.[22] No publicly available CDS reading exists for February, but the bond market offers a proxy: the 9.25% 2030 secured notes traded at a Z-spread of 646 basis points as recently as February 23.[23] Spreads had not materially tightened even before earnings, despite a patron that just doubled down. A 13% after-hours equity drop and a Q1 guidance miss are not the conditions under which bond spreads compress.
Revenue doubled. The backlog grew by $11 billion. The interest-to-revenue ratio is heading toward 30%. The equity market heard the revenue beat and sold anyway. The credit market doesn’t look inclined to tighten spreads. When both markets reach the same conclusion from different vantage points, the signal is structural, not episodic. The CDS market was pricing structural risk, not a revenue miss. Q4 confirms it.
The parallel
Both companies delivered growth that would have been unthinkable two years ago. Both are locked into forward commitments that assume AI demand will not only continue to grow but also accelerate. The simplest explanation is that both stocks sold off because guidance didn’t clear the sky-high bar. NVIDIA’s $78 billion wasn’t enough above the $72 billion consensus. CoreWeave’s Q1 missed entirely. The structural explanation — that the market shifted from pricing demand to pricing commitments — only becomes falsifiable over the next two to three quarters. But the structural explanation is what connects two sell-offs at different layers of the same stack in the same 48-hour window.
The answer depends on who bears the downside. NVIDIA’s $95.2 billion in purchase obligations sits atop $97 billion in annual free cash flow and 75% gross margins. A demand pause means write-downs and margin compression: painful but survivable, exactly as the H20 charge demonstrated. CoreWeave’s $21.4 billion in debt sits atop near-zero adjusted operating income and capex running at 2.3x-2.9x revenue. A demand pause means an interest-to-revenue spiral with no free cash flow cushion. Same structural exposure, radically different survivability.
Meanwhile, Q4 13F filings show the smart money repositioning. Magnetar sold 13.8 million shares during the quarter — a 16.8% reduction — while maintaining put protection through June at an $82.50 strike.[24] Coatue exited its entire position. The stock closed at $97.63 before earnings and hit $84.64 at its after-hours low — sixteen cents above Magnetar’s put strike. The hedge that looked distant three weeks ago is now at the door.
Verdict
The market stopped asking whether AI demand is real and started asking what happens to $95 billion in purchase orders and $21 billion in debt if it even pauses.
The three-part series mapped the architecture: NVIDIA’s quadruple role binding the ecosystem into a single correlated bet, the debt mechanics that make growth and fragility two sides of the same quarterly filing, and the credit markets already pricing the structural risk that rating agencies refuse to acknowledge. This week’s filings didn’t change the architecture. They confirmed it — with bigger numbers, at every layer of the stack, in a 48-hour window.
NVIDIA’s purchase obligation spike is not a crisis. It is a permanent change in how the most valuable semiconductor company on Earth operates — committing at a historic scale, years before demand is known, because the COMECON architecture demands it. CoreWeave’s debt stack is not yet a crisis either — the maturity schedule is sensible, and the backlog is real. However, both balance sheets now carry forward commitments of a size and duration without precedent in the semiconductor or cloud infrastructure industries. The revenue case may prove them right. The filings show what it costs if it doesn’t.
One question the filings don’t answer: what happens when a hyperscaler spending $200 billion a year on infrastructure decides it can’t keep going. The commitment chain runs NVIDIA → neoclouds → hyperscalers. The cash flows in the other direction.
Notes
[1] NVIDIA Q4 FY2026 earnings release, February 25, 2026. Revenue: $68.13 billion vs. $66.21 billion consensus (LSEG). Data center revenue: $62.3 billion, up 75% YoY. Q1 FY2027 guidance: $78 billion ±2%, vs. consensus of approximately $72 billion (LSEG). Guidance explicitly excludes data center compute revenue from China.
[2] CoreWeave Q4 FY2025 earnings release, February 26, 2026. Revenue: $1.57 billion vs. $1.53 billion consensus (FactSet). Revenue more than doubled YoY. Backlog: $66.8 billion, up from $55.6 billion in Q3 2025.
[3] NVIDIA (NVDA) closed at $184.89 on February 26, down approximately 5.5% — the worst single-session decline since April 16, 2025. Market data via Yahoo Finance; session context via CNBC, February 26, 2026. CoreWeave (CRWV) closed at $97.63 on February 26; traded as low as $84.64 in after-hours, down approximately 13.3% from the regular-session close. Note: NVIDIA figure is regular session; CoreWeave figure is after-hours low.
[4] NVIDIA Q4 FY2026 earnings call transcript, February 25, 2026. Jensen Huang used variants of the “compute equals revenues” formulation multiple times across prepared remarks and Q&A. Transcripts via Shacknews, PYMNTS.
[5] NVIDIA Form 10-K, FY2026 (filed February 2026). Purchase obligations: $95.2 billion, up from $16.1 billion in FY2025 (5.9x increase). The filing states purchase obligations are “expected to continue to grow and become a greater portion of supply.”
[6] Total supply obligations of $117 billion per Sherwood News analysis of NVIDIA 10-K (February 27, 2026), inclusive of purchase obligations and other contractual commitments. B-tier: journalist calculation from A-tier filing. NVIDIA FY2026 operating cash flow: approximately $131 billion; free cash flow: $97 billion. Note: operating cash flow is the pre-capex figure; free cash flow deducts capital expenditure.
[7] NVIDIA Form 10-K, FY2026.
[8] NVIDIA 10-K FY2026: TSMC required longer-term contracts and advance cash payments for custom semiconductor fabrication and advanced packaging capacity. CoWoS packaging capacity constrained per TSMC CEO C.C. Wei (TSMC earnings call, January 2026).
[9] NVIDIA Q4 FY2026 earnings call. CFO Colette Kress: “Inventory grew 8% quarter over quarter, while purchase commitments also increased significantly, and we have strategically secured inventory and capacity to meet beyond the next several quarters, further out in time than usual.”
[10] Michael Burry, “Short Thought: Nvidia Ratchets Up the Risk,” Cassandra Unchained (Substack), February 26, 2026. Burry’s core claim: the $95.2 billion in purchase obligations is “structural to new trajectory, not temporary,” drawing a Cisco 2000–2001 parallel — supply commitments built anticipating 50% annual growth, subsequently written down by approximately 40% when demand collapsed. The structural architecture of NVIDIA’s commitment ratchet was mapped in “Jensen’s COMECON,” The AI Realist, February 11, 2026 — fifteen days before Burry’s post.
[11] Bernstein analyst Stacy Rasgon, cited in Sherwood News, February 27, 2026. Rasgon characterized the $95 billion in supply commitments as ensuring NVIDIA becomes “the most dependable supplier” in a market he estimates could reach $1.4 trillion. Separately: “Demand is showing absolutely zero signs of slowing.” B-tier: analyst commentary, attributed.
[12] NVIDIA Q4 FY2026 earnings release. FY2026 free cash flow: $97 billion ($35 billion in Q4 alone). Q4 non-GAAP gross margin: 75.2%; GAAP gross margin: 75.0%. $41 billion returned to shareholders via buybacks and dividends.
[13] NVIDIA Q1 FY2026 earnings (reported May 2025). Charge of approximately $4.5 billion related to H20 inventory and purchase commitments, following April 9, 2025, export controls on H20 chips. Per Rev.com earnings call transcript.
[14] CoreWeave Q4 FY2025 earnings call, February 26, 2026. Q1 2026 revenue guidance: $1.9–2.0 billion vs. $2.29 billion consensus per FactSet.
[15] Ibid. Full-year 2026 guidance: revenue $12–13 billion, capital expenditure $30–35 billion. Capex-to-revenue ratio: low end $30B/$13B = 2.3x; high end $35B/$12B = 2.9x.
[16] Ibid. Q1 2026 interest expense guidance: $510–590 million. Adjusted (non-GAAP) operating income guidance: $0–40 million. Interest as percentage of Q1 revenue: $510M/$2.0B = 25.5% to $590M/$1.9B = 31.1%. Note: EBITDA, the metric lenders typically use for debt coverage, is substantially higher than adjusted operating income due to depreciation on CoreWeave’s large asset base.
[17] Q4 2024 interest expense: $149 million (CoreWeave 10-K FY2024). Q4 2025: $388 million (Q4 FY2025 earnings release). Q1 2026 guidance midpoint: $550 million. Increase: $550M / $149M = 3.69x over approximately 15 months.
[18] CoreWeave Q4 FY2025 earnings release. Total debt: $21.37 billion as of December 31, 2025.
[19] CoreWeave Q4 earnings call. Backlog: $66.8 billion, up $11.2 billion sequentially. Average contract length: approximately five years. No single customer exceeds 35% of backlog, improved from approximately 85% concentration at the beginning of 2025.
[20] Ibid. Weighted average interest rate declined approximately 300 basis points during 2025, generating roughly $700 million in annualized savings per CFO Nitin Agrawal. No significant debt maturities until 2029.
[21] NVIDIA press release, “Nvidia and CoreWeave Strengthen Collaboration to Accelerate Buildout of AI Factories,” January 26, 2026. $2 billion equity investment at $87.20/share; 13% ownership stake. CoreWeave Q4 earnings call: $6.3 billion capacity guarantee through 2032; $27 billion in cloud service agreements.
[22] TheValueist (@TheValueist), X post, December 12, 2025, citing Bloomberg CDS series CZ037664. CoreWeave 5-year CDS at 773.480 bps, up from 371.040 bps on September 26, 2025. Implied five-year default probability of approximately 47.5% under 40% recovery assumption. Most recent publicly available CDS reading. First documented in “Two Markets, One Asset,” The AI Realist, February 21, 2026.
[23] BondbloX, CoreWeave 9.25% 2030 secured notes, Z-spread of 646.2 bps, accessed February 23, 2026 (pre-earnings). Post-earnings bond spread data was not publicly available at time of publication. Note: Z-spread is not directly comparable to CDS spread but provides a directional proxy for credit market pricing on CoreWeave secured debt.
[24] Magnetar Capital Q4 2025 13F filing (positions as of December 31, 2025): sold 13.8 million shares (-16.8%), retaining approximately 68.2 million shares (~23% of company). Magnetar executed 1,204 sale transactions totaling $3.46 billion in proceeds over the six months ending December 31, with zero purchases. Put collar filed via Schedule 13D/A: 392,632 shares with $82.50 puts and $200 calls, expiring June 18, 2026. Earlier collar: $70 puts, $160/$175 calls, expiring March 20, 2026. Coatue Management exited entirely in Q4 (-6.7 million shares, -100% of position) per Q4 13F. Fidelity reduced by 18.9 million shares (-70.4%). After-hours low of $84.64 on February 26, 2026, per Yahoo Finance — $0.16 above the $82.50 June put strike.