The circular AI trade, mapped

That AI capital spending has gone circular is, by now, a consensus observation: chipmakers take equity in the startups that buy their chips, hyperscalers fund the labs that rent their clouds, and the same dollars seem to lap the track more than once. Rather than restate the thesis, this issue traces one node, NVIDIA, through named and sourced relationships, and lets the loop draw itself.

The investor is the supplier. Figure AI closed a more-than-$1B Series C at a $39B post-money valuation in September 2025, with NVIDIA, Microsoft, and Amazon all on the cap table Reported · 0.8. Figure's robots, in turn, run on NVIDIA: Figure 02 does all inference on dual onboard NVIDIA RTX GPUs and trains its models on NVIDIA H100s Reported · 0.9. So NVIDIA's investment dollars come back as hardware revenue. The investor is the supplier.

Figure is the purest expression of it. A $39B post-money valuation, set in September 2025, on a company still years from meaningful revenue, underwritten in significant part by the three firms whose chips and cloud it will consume. That is not a knock on Figure's technology; it is an observation about where the $39B came from. A startup priced like a mid-cap, funded by its own suppliers, is the loop compressed into a single line on a cap table.

The co-investors are the customers. Here is where the loop tightens. The other two names on Figure's cap table, Microsoft and Amazon, are also NVIDIA's two largest customers: NVIDIA's top two (unnamed) buyers were ~39% of Q2 FY2026 revenue, and its top four ~61% Reported · 0.8, with Microsoft widely reported as the largest (a ~200,000-unit GB300 order in November 2025). So the dollars NVIDIA books selling GPUs to Microsoft ($281.7B in revenue) and Amazon ($716.9B) Confirmed partly fund a startup that buys still more NVIDIA. Money moving in three named directions, one chipmaker at the center.

The customers are also trying to escape. The same hyperscalers funding the loop are quietly trying to leave it, by building their own accelerators: Amazon has Trainium, Microsoft has Maia Reported · 0.8. But the escape hatch runs straight back through NVIDIA's own supply chain, because Microsoft's Maia 200 uses SK hynix HBM3E Reported · 0.8, the very memory NVIDIA depends on. You can design around NVIDIA's GPU; you cannot easily design around NVIDIA's suppliers.

We have seen vendor-financed booms before. In the late-1990s telecom buildout, equipment makers like Lucent and Nortel lent their own customers the money to buy their gear; the revenue looked spectacular right up until the carriers couldn't pay, and the receivables unwound as fast as they had built. The mechanism today is gentler, equity stakes and cloud commitments rather than vendor loans, but the question it raises is identical: how much of the reported demand would survive if the funding stopped? That is not a call for a bust. It is a reason to separate, in your own model, the AI revenue that is arm's-length from the AI revenue the seller helped pay for.

The chokepoint underneath. Which is the concentration that actually matters. NVIDIA itself is sole-sourced on its two most critical inputs: TSMC is its only foundry, per NVIDIA's own 10-K Confirmed · 0.9, and SK hynix ($68.3B in revenue Confirmed) is the sole supplier of 12-layer HBM3E for the GB300 Reported · 0.8. And SK hynix feeds both NVIDIA and Microsoft's would-be NVIDIA-killer. The whole stack, incumbent and challenger alike, funnels through two companies in Taiwan and Korea.

The one outside the loop. Tesla. It designs its own AI silicon (AI5 and AI6, fabbed by TSMC and Samsung under a $16.5B Samsung deal Reported · 0.9), it is not on anyone's cap table, and it does not buy NVIDIA for Optimus's brain. Vertically integrated, not vertically financed, the counter-model to everything above.

What the graph shows. Plotted in the Tettares supply graph, this stops being anecdote and becomes structure: NVIDIA carries three confirmed buyers (Amazon, Microsoft, Figure); TSMC and SK hynix each flag “high” customer-concentration on the Herfindahl measure; and the ownership layer flags the NVIDIA, Microsoft, and Amazon stakes in Figure as strategic. That last pattern, a supplier holding equity in the company it sells to, is the lock-in that never appears on either party's income statement. It is exactly what our analytics are built to surface.

What we're watching upstream

• How much hyperscaler AI capex is vendor-financed (equity, cloud credits, strategic rounds) versus arm's-length demand.
• Maia and Trainium volumes, the real test of whether the customers can leave the loop, and how much SK hynix HBM they consume getting there.
• Any new NVIDIA strategic stake in a compute customer: the loop widening.
• TSMC and SK hynix capacity allocation, the bottleneck the whole circle depends on.

Map it and the picture is less a supply chain than a flywheel: NVIDIA sells to the firms that fund the startups that buy NVIDIA, while everyone, even the would-be escapees, draws from the same two upstream suppliers. None of that makes the demand fake. But it does mean “AI demand” and “AI concentration” are increasingly the same number, viewed from different angles. Read the cap table alongside the order book.