The money goes in circles
Nearly half the quantified AI demand on this map has seller money on the other side of it. Thirty deals among fourteen companies.
Hiding dims vendor capital and the commitments paired with it. Those dollars are still committed; the dimming marks where seller money sits on the other side.
Select any flow or company for the deal behind it.
Nvidia invests in, backstops, or leases from its own customers. Those customers buy Nvidia. AMD, Microsoft, and Google run the same loop with their own instruments.
The map holds the deals announced between spring 2025 and July 2026: equity stakes, warrants priced at a penny a share, purchase backstops, lease-backs, and one customer paid in its vendor's stock. Capital flows from sellers to buyers. Purchase commitments flow back. The question is how much of the announced demand stands on its own.
The instruments migrated
At the Morgan Stanley technology conference on March 4, 2026, Jensen Huang said the OpenAI and Anthropic stakes would be Nvidia's last direct equity bets in either company. The next chapter was already in production. On July 1, 2026, Nvidia formalized a backstop program that guarantees revenue floors for GPU clouds: $4.88B for SharonAI, an undisclosed sum for Firmus, 210,000 GB300s across the first two partners. Equity on the balance sheet became contingent liabilities beside it. The loop stayed the same, but the instrument got quieter.
The numbers did not get quieter with it. SemiAnalysis projects roughly $7.1T in AI debt outstanding by 2029, about $11.1T in cumulative AI capex from 2024 to 2029, and annual capex well north of $2T by 2028. The backstop program is how the loop continues at that scale without showing up as equity.
The last time sellers financed the buyers
Lucent committed $8.1B in vendor financing in fiscal 2000, 24% of that year's $33.6B revenue. Revenue then fell 69%, from $37.92B in calendar 1999 to $11.80B in 2002. Lucent took $2.2B in bad debt provisions in 2001 and another $1.3B in 2002. Nortel extended $3.1B and still had $1.4B outstanding; Cisco had promised $2.4B. Forty-seven CLECs went bankrupt between 2000 and 2003, and about $2T in telecom market capitalization was gone by September 2002.
Tomasz Tunguz ran the comparison in October 2025 and put Nvidia's direct investments at 67% of its $165B trailing revenue, against Lucent's 24%. That figure is stale, and stale in Nvidia's favor. The $110B numerator was the $100B OpenAI letter of intent plus the $10B Anthropic stake, and OpenAI's investment was ultimately signed at $30B. On the confirmed numbers, Tunguz's own method puts Nvidia at 24%: the same ratio as Lucent, not triple it.
The broader measure climbs back. Nvidia's equity, backstops, and lease-backs come to $58.3B, and the GPU-backed debt they enable adds $24.1B: 50% of the same revenue base, roughly double Lucent. Which figure is right depends on whether a take-or-pay floor and a lease-back count as vendor financing. They do, because the seller is on the hook either way.
Vendor financing as share of annual revenue
What the sellers lost last time
All three Nvidia bars use the $165B trailing revenue base from Tomasz Tunguz's October 2025 comparison, so they compare to Lucent, and to each other, on the same terms. As published, his 67% rested on a $110B numerator: the $100B OpenAI letter of intent plus the $10B Anthropic stake. OpenAI's investment was signed at $30B. Nvidia's revenue has also grown since late 2025, which pulls every Nvidia bar down from where it sits here. Sources: Tomasz Tunguz for the Lucent, Nortel, and CLEC figures; Paul Starr, The Great Telecom Implosion, for telecom market capitalization.
The case that this is fine
The strongest counterargument comes from J.P. Morgan Asset Management: the AI buildout is largely funded from hyperscaler free cash flow, which makes a loop between cash-rich companies closer to an accounting curiosity than a warning. It's a serious argument, and for the biggest buyers it describes the present accurately.
Some of the evidence here cuts the same way. Hide the vendor money and more than half the quantified commitments are still standing. The largest of them, OpenAI's reported $300B to Oracle and Oracle's reported $40B order to Nvidia, form a circle with no vendor equity anywhere in it. Circularity and vendor financing are not the same thing, and only the second one is drawn here.
Two numbers press against the free cash flow case. Bain estimates AI needs $2T in annual revenue by 2030 to fund the compute being built for it, and that the current trajectory leaves the industry roughly $800B short. Morgan Stanley expects AI-related global debt issuance to more than double in 2026, to roughly $570B. Free cash flow covers the buildout at today's scale, not at those numbers. The debt markets and the vendors themselves fund the rest.
Every deal on this map has a rationale on its own terms. Read together, they describe an industry vouching for its demand with its own balance sheet. Vendor financing defers the question of whether that demand exists.
Method and sources
Vendor capital means money or equity value moving from a seller of AI hardware or compute to a company that buys from it: equity, warrants, purchase backstops, lease-backs, and stock granted to a customer. Purchase commitments are the obligations flowing back. Deals were announced between spring 2025 and July 2026, and every deal card links its primary source. The readout counts only quantified deals; dashed flows have no disclosed amount and add nothing. Figures marked reported rest on press coverage without an SEC-grade filing.
Two exclusions are deliberate. Nvidia's $5B stake in Intel is off the map because Intel is a supplier building Nvidia-custom CPUs, not a customer buying Nvidia silicon, so no loop closes. Microsoft's reported $14B Nscale contract is drawn but unpaired: it is third-party demand, not a return flow to Nvidia, which holds the Nscale equity.
Where sources disagree, the map takes the filing and says so. Nvidia's CoreWeave stake is reported by some outlets as $2B for 9%; the Schedule 13G/A discloses 47.2M Class A shares, about 11.5% and $3.66B, and the map uses the filing. CoreWeave's total debt runs $21B, $24.9B, or $35B under different definitions; the map draws only the two GPU-backed term loans ($8.5B and $3.1B). The Oracle figures, roughly $300B from OpenAI and $40B to Nvidia, are reported and never SEC-confirmed. Morgan Stanley's 2026 debt-issuance forecast circulates as both $570B and more than $500B; the piece uses $570B. Cisco's roughly $900M vendor-financing writeoff appears in no source used here, so only its $2.4B commitment is stated. No reliable aggregate exists for total circular financing in AI; the readout sums only the deals on this map and claims nothing beyond them.
Sources: SemiAnalysis on the backstop program and debt projections, TechTimes on the 210,000 GPU program total, Futuriom on undisclosed program terms, TechCrunch on Huang's March 2026 remarks, Bain Global Technology Report (September 2025), J.P. Morgan Asset Management, and Tomasz Tunguz on the Lucent comparison.