NVIDIA has put its balance sheet behind 209 transactions since January 2024. Thirty of the companies on the receiving end were funded from two or more separate points in the same loop. This is the map.
Point-in-time note. Transaction data covers January 2024 through August 27, 2026 and is not updated after publication.
The question is not whether NVIDIA invests in its customers. It discloses that. The question is what the aggregate structure looks like when you draw every edge at once.
NVIDIA-attached round value: $314.9B, of which 95.4% points at compute consumers · 363 edges, 305 nodes
The map sorts every funding entity into one of four roles. Anchor is NVIDIA itself, split between direct balance-sheet activity and NVentures. Circular nodes are companies that take NVIDIA capital and spend a meaningful share of it on NVIDIA compute: OpenAI, Anthropic, CoreWeave, Nebius. Supply nodes sit upstream: Intel, Marvell, Synopsys, Coherent, Lumentum, Nokia. Peer platforms such as Databricks, Hugging Face and Anysphere sit alongside rather than inside the compute relationship.
The distinction between those roles is not cosmetic, and it runs in opposite directions. When NVIDIA takes a stake in Coherent or Synopsys, NVIDIA is the customer. It buys optical components and design tools from them. Capital goes out and purchase orders go out; nothing returns as NVIDIA revenue. That is supply security, not circularity. The circular question only arises on the demand side, where the recipient of capital turns around and buys compute.
The interesting property is not any single edge. It is that the circular nodes are themselves active allocators. OpenAI has made 35 investments of its own since January 2024. Anthropic has made 35. Capital that enters the loop at the top does not stop at the first hop.
Deal count has been roughly stable. Committed round value has not.
| Year | NVIDIA-attached deals | Round value |
|---|---|---|
| 2024 | 62 | $24.2B |
| 2025 | 88 | $113.6B |
| 2026 (through Aug) | 59 | $177.0B |
2026 has produced fewer transactions than 2025 and more than seven times the 2024 round value. That is concentration, not breadth. A small number of very large commitments carries the total.
The five largest NVIDIA-attached commitments account for the overwhelming majority of the $314.9B figure.
| Date | Counterparty | Round value | Type |
|---|---|---|---|
| 2026-02-27 | OpenAI | $122.0B | Late-stage round |
| 2025-10-15 | Aligned Data Centers | $40.0B | M&A, whole |
| 2025-11-18 | Anthropic | $30.0B | Series G |
| 2026-01-06 | X.AI | $20.0B | Series E |
| 2024-11-20 | X.AI | $6.0B | Series C |
| 2024-09-11 | OpenAI | $6.6B | Non-convertible debt |
| 2025-09-18 | Intel · supply side | $5.0B | Follow-on equity |
Round value is the full size of the round, not NVIDIA's allocation. Per-investor amounts are not disclosed in syndicated transactions. Every figure here therefore represents capital NVIDIA participated in, not capital NVIDIA committed. The distinction matters and the gap is unknowable from the outside.
The $314.9B aggregate is the least useful number in this report, because it sums three structurally different things. Separating them is the first thing any reader should do.
| Direction | Deals | Round value | Share |
|---|---|---|---|
| Demand side, recipient consumes NVIDIA compute | 202 | $300.4B | 95.4% |
| of which named platforms | 14 | $235.5B | 74.8% |
| of which venture portfolio | 188 | $64.9B | 20.6% |
| Supply side, NVIDIA is the buyer | 7 | $14.5B | 4.6% |
The supply-side stakes are the ones most often miscounted in the circular-financing discussion: Intel at $5B, Marvell, Coherent, Lumentum and Synopsys at $2B each, Nokia at $1.01B, Corning at $500M. Every one is disclosed as follow-on or preferred equity. None is a prepayment or a capacity commitment. Investing in a supplier does not generate revenue for the investor, and at 4.6 percent of the total it would not move the argument even if it did.
The venture portfolio is not a neutral third category. Figure AI, Anysphere, Thinking Machines, Reflection AI, Safe Superintelligence, Fireworks.ai, Skild AI, World Labs, Scale AI, Crusoe: these are compute consumers, structurally the same side of the ledger as the named platforms, just earlier and smaller. Grouping them with the platforms is the honest read.
So the shape is lopsided. Strip out the seven supplier stakes and 95 percent of the mapped capital points at companies that buy compute. That is the concentration observation, and it survives the objection rather than depending on it.
The amber rings in the map mark companies capitalized from two or more separate points in the loop. These are the structurally interesting cases, because they are where the network stops being a hub-and-spoke and starts being a mesh.
What it shows: the AI buildout is being financed, in part, by its largest supplier, and the recipients of that financing are themselves financing the next layer down. Capital and demand are travelling the same edges.
What it does not show: whether any of this is improper, unusual for the sector, or economically unsound. Strategic investment by a platform vendor into its ecosystem is ordinary corporate behavior. Intel, Google and Microsoft have all run comparable programs. The map is a description of structure, not a verdict on it.
The analytical value is in the shape. A hub-and-spoke network in which the hub is also the primary revenue beneficiary has different failure characteristics than a diversified capital base. If demand for compute slows, the same edges that carried capital outward carry the impairment back. That is a concentration observation, not a prediction.
Two hundred and nine transactions. Two hundred and ninety counterparties. Thirty companies funded from more than one point in the same loop.
The structure was always disclosed. It had just never been drawn.