● dense/○ explainer
aileena·machina
markets · 2026 · may

NVIDIA is buying its own demand.

An $18B public book that reads like a map of the AI stack, plus $100B+ of private commitments routed back into GPU orders. Here’s the flywheel — and where it breaks.

— tldr —

equity out, revenue back, stake up, repeat. $18.4B public, $40B+ private in 2026, plus up to $100B into OpenAI. cleanest vendor-financing critique you’ll ever see.

NVIDIA sells the shovels. It also now owns equity in the miners, the toll roads, and the company that makes the shovel’s steel. The thing it invests in, over and over, is its own demand.

NVIDIA writes a check into a company; the company uses it to buy GPUs; the order lands on NVIDIA’s income statement; the company’s valuation rises; NVIDIA’s stake appreciates. Then it does it again, with more money.

the loopequity out → revenue back → stake up → repeat.

It comes in two books — the public one in a 13F, and the private one in strategic commitments. They are very different sizes.

↓ book one ↓

book 1 — the public 13F

NVIDIAthe axleequity + chipshyperscalersOpenAI · MSFTAnthropic · xAIneocloudscapex spenddatacenters · power7 month buildGPU orders100k chipsback to Jensen$30B OpenAI$5B→$25B+ Intel$19.4B Nebius~100k chipsrevenue backthe loop$ $ $equity → revenue
1
2
3
4
5

the loop that keeps spinning until it doesn’t

① NVIDIApublic 13F · $18.4B
Q1 2026 book
② hyperscalersOpenAI up to $100B
~10 GW systems
④ capex$19.4B Nebius pact
~100k GB300 chips
③ gpu ordersCoreWeave
stake +110%
⑤ private book$40B+ committed
in 2026 alone
↻ spins faster
where does the money REALLY come from?

As of March 31, 2026, NVIDIA’s disclosed equity book was worth about $18.4 billion, up from $13.1B a quarter earlier. The holdings aren’t random tech names — they are the layers of the AI stack.

13F in plain wordsthe public filing where big funds disclose what stocks they own.

Intel at ~51.6% — the foundry + x86 hedge. The position went from $5B in to $25B+ out.

foundrythe factory that makes chips. owning intel = a manufacturing hedge.

CoreWeave sits at #2, around $3.66B. Share count grew 95% to 47.2M; the position appreciated +110%.

neoclouda new-style GPU cloud that rents compute. coreweave is the biggest.

Synopsys ~10.4% (EDA — the software you design every chip with). Coherent #4 ~$1.86B (optics/photonics, 7.79M shares, new position). Nokia ~7.3% (optical transport + networking).

EDAelectronic design automation — chip-design software. synopsys is one of two.

Read that list as a thesis, not a portfolio. Compute (CoreWeave), manufacturing (Intel), design tools (Synopsys), optics (Coherent), and networking (Nokia). NVIDIA is buying a toll booth on every road its chips travel.

↓ book two — the bigger one ↓

book 2 — the private commitments

The 13F is the tip. The body is the strategic book — over $40 billion committed in 2026 alone, on top of ~$17.5B into private companies and infra funds the prior fiscal year.

scalepublic $18.4B. private $40B+ this year. very different sizes.

The anchor: OpenAI — up to $100B, ~10GW of systems. OpenAI’s own CFO said the quiet part out loud: “most of the money will go back to Nvidia.”

stagedconditional, slow to deploy — but the intent is unambiguous.

Nebius — $2B in pre-funded warrants, earmarked to deploy 5+ gigawatts by 2030. CoreWeave — ~$6.3B backstop to buy unsold compute, plus ~$860M data-center lease guarantee. Lambda — ~$1.5B, plus seats in xAI and Anthropic mega-rounds.

backstopNVIDIA promises to buy compute the neocloud can’t sell. floor under demand.

why it’s rational

It turns a chip sale into a compounder. Sell a GPU once, book the margin once. Own equity in the buyer and you also capture the value the GPU creates. Intel: $5B → $25B+. CoreWeave: more than doubled.

compoundersame dollar, two payoffs — margin on the sale + appreciation on the stake.

It allocates scarce supply to real builders. When GPUs are rationed, an equity check steers chips toward customers who will deploy at scale — and locks in CUDA + NVLink as the default.

CUDA / NVLinkNVIDIA’s software + interconnect. the moat that keeps customers stuck.

It also hedges the whole stack — foundry, EDA, optics, transport — and seeds its own ecosystem, so every funded neocloud is a distribution channel that isn’t a hyperscaler trying to build its own chips.

the risk — round-tripping

The same dollar can travel in a circle — NVIDIA → OpenAI → NVIDIA — and on each lap everyone’s numbers go up, even though no new outside money entered the loop. That’s vendor financing, and at $100B scale it’s unprecedented.

vendor financingyou lend the buyer the money they use to buy from you. lucent & nortel did this.

The equity book and the core business are the same bet — in an AI-capex downturn, revenue and the portfolio fall together. The reflexive loop deflates on the way down, faster.

reflexivityprices that go up because prices went up. works both ways.

Critics put it bluntly: the math only works if the returns arrive before the money runs out. A large slice of the market is now a leveraged bet that AI scaling continues — with NVIDIA’s balance sheet as the axle.

the bottom line

NVIDIA became the central bank of the AI economy — it prints capacity, lends it to its colonies, and takes equity in return. The flywheel is the bull case and the bear case wearing the same coat. The only real question is how long the music plays.

← dispatch