The AI Investment War

Money in AI no longer flows in one direction. Chipmakers invest in the labs that buy their chips. Clouds invest in the labs that rent their servers. Labs pay their own shareholders hundreds of billions for compute — and are sometimes paid back in their supplier’s stock. The result is a web of dependencies unlike anything in tech history, and one property runs through all of it: the same capital keeps changing hands.

That circularity is the through-line of this piece. Follow any large arrow far enough and it tends to loop back to where it started. Whether that loop is a flywheel or a trap depends on one number — real end-user demand — that no contract can manufacture.

The map below reconstructs the web as of July 2026. It follows the structure of the industry itself: four horizontal layers, from consumer applications at the top down to the silicon everything runs on.

The four layers

The industry stacks into four tiers, and almost every deal on the map is a wire between two of them.

• Application — where users meet AI: ChatGPT, Claude, Gemini, Llama-powered products, Grok.
• Model — the frontier labs training foundation models. This is where capital pools.
• Compute — hyperscaler clouds and dedicated data-center projects: Azure, AWS, Google Cloud, Oracle, Stargate, Colossus.
• Silicon — the chips underneath: Nvidia and AMD GPUs, Broadcom-built custom accelerators, Google’s TPU, Amazon’s Trainium.

The map

How to read it. Purple arrows are equity: the arrow points from the shareholder to the company it owns a piece of. Green arrows are commercial contracts: the arrow points from the customer to the supplier it pays. In both cases, arrows follow the money. Click a company to isolate every deal it touches; click an arrow for the terms; use the toggles to show equity or contracts alone.

The model layer: two capital vortexes

Nearly every dollar on the map is ultimately pulled toward one of two labs. They raised the most, and they committed the most — often to the very investors who funded them.

OpenAI converted its Microsoft partnership into a formal 27% stake during its October 2025 restructuring, then went on a spending spree that dwarfs the money it raised: $300B to Oracle, $250B to Azure, 10 GW of custom chips from Broadcom, and 6 GW of GPUs from AMD. Nvidia committed up to $100B in return for 10 GW of its own systems. The most unusual deal runs backwards: AMD granted OpenAI warrants for roughly 10% of AMD itself, vesting as OpenAI deploys AMD hardware — a supplier paying its customer in equity to win the order.

Anthropic built the opposite structure: instead of one anchor partner, it took money from three rival hyperscalers. Amazon invested $13B with $20B more pledged; Google committed up to $40B; and in January 2026, Microsoft and Nvidia added roughly $15B at a $350B valuation. Each investment came bundled with a spending commitment flowing back — over $100B to AWS, $200B to Google Cloud, and about $30B to Azure. By mid-2026 Anthropic had raised again at a $965B valuation, with memory-chip makers Samsung, Micron, and SK Hynix joining the cap table.

The shapes differ, but the mechanic is the same: money arrives as investment and leaves as a compute bill, often to the same counterparty. That is the pattern worth decoding before reading the rest of the map.

The deal structures, decoded

The headline numbers hide four repeating financial structures. Naming them makes the whole map legible.

Why do deals take these shapes rather than plain cash-for-goods? Because compute is scarce and trust is expensive. A lab that commits $200B of future spend gets priority access to capacity that would otherwise be rationed. A cloud that pre-sells that capacity can borrow against the backlog to build the data centers. A chipmaker that takes equity in its customers locks in demand and shares the upside if the bet pays off. Every structure trades cash flexibility today for a binding claim on the AI build-out tomorrow — which is exactly why they interlock into loops.

The compute layer: everyone builds, everyone rents

The compute layer is where the biggest absolute numbers live. Stargate — the $500B venture owned by SoftBank, OpenAI, Oracle, and MGX — plans roughly 7 GW of capacity across the US, UAE, Norway, and Argentina. Oracle, a founding equity partner, is also its lead builder, and OpenAI’s $300B contract transformed Oracle’s cloud business overnight — while forcing Oracle to borrow heavily to pour the concrete.

The strangest compute story is Colossus. After SpaceX absorbed xAI in February 2026, it began renting out its Memphis supercomputer like real estate: the entire 300 MW, 220,000-GPU Colossus 1 site went to Anthropic — a direct competitor of xAI’s Grok — for $1.25B a month, while Google signed up for $920M per month of capacity through 2029. Even the fiercest rivals in the model layer are customers of each other one layer down. Compute is fungible; brand loyalty stops at the rack.

The silicon layer: Nvidia’s money boomerang

Nvidia sits at the bottom of the map and touches nearly every arrow above it. Its investment portfolio — up to $100B in OpenAI, ~$10B in Anthropic, $2B in xAI, $5B in Intel — reads like a list of its own largest customers. Money leaves Nvidia as equity and returns as GPU purchase orders. That round trip is the single most-discussed loop on the map.

The counterweight is Broadcom, the quiet winner of the war. It co-designs Google’s TPUs (a long-term agreement running through 2031), builds Meta’s custom accelerators through 2029, and is producing 10 GW of custom chips for OpenAI. Its overall AI revenue is projected near $46B for 2026. Every hyperscaler’s plan to reduce Nvidia dependence runs through Broadcom — which is why Broadcom wins no matter who wins the model layer.

The circularity problem

Follow the purple and green arrows in a loop and the concern becomes concrete. Here is the canonical example, one step at a time:

Revenue is real at every step, and each company can book it honestly. But a meaningful share of the industry’s growth is the same capital circulating between a handful of balance sheets. Cloud providers report record AI revenue partly funded by their own investments; investors mark up stakes in labs whose spending commitments they underwrite. The loop is not fraud — it is leverage. It amplifies whatever is happening underneath it.

Bull and bear: two readings of the same map

The same web supports opposite conclusions. What separates them is not the deals — everyone sees the same arrows — but a bet on demand and on how binding the commitments really are.

What would tell them apart? Each reading makes falsifiable predictions worth watching:

• Watch the top of the map. If application-layer revenue (subscriptions, API usage, ads on AI products) keeps compounding, the bull wins by default. If it stalls while compute commitments keep growing, the gap is the bubble.
• Watch whether commitments convert to cash. These are multi-year, milestone-contingent numbers. If reported cloud revenue tracks the headline contracts, the deals are real; if the contracts quietly shrink or slip, they were optionality dressed as demand.
• Watch the private marks. A down-round at OpenAI or Anthropic would ripple through every investor that holds them — the clearest single signal that the loop has started running in reverse.
• Watch Nvidia’s concentration. The more of Nvidia’s growth traces back to companies Nvidia funded, the more the boomerang, not new buyers, is driving the number.

Key numbers

Figures reflect public reporting as of July 2026. Committed amounts are multi-year and often milestone-contingent — treat them as order-of-magnitude, not booked revenue. Private valuations are last-round marks, not cash-flow-based. The map is a snapshot; deals of this size change monthly.