One in Seven AI Deals Is Now a Compute Contract. One in Thirty Is an Acquisition.

Nine months ago the AI industry bought companies four times more often than it bought computing capacity. That ratio has inverted, and the inversion is larger than almost anything that happened at the model layer in the same period.

These are figures from our own deal tracker, covering 543 recorded deals dated October 2025 through August 4, 2026.

| Period | Acquisitions | Compute contracts | |---|---|---| | Q4 2025 | 16.5% | 3.9% | | Q1 2026 | 10.6% | 5.3% | | Q2 2026 | 7.7% | 18.4% | | Q3 to date | 3.1% | 14.6% |

Acquisitions fell to roughly a fifth of their old share. Compute contracts rose to nearly four times theirs. The crossover happened in Q2.

## What the mix is telling you

An acquisition is a bet on a team. A compute contract is a bet on a bottleneck. When the mix moves this hard in three quarters, it means buyers have concluded that the scarce input is megawatts and accelerators rather than talent or product, and that a contract gets you there faster than a cap table does.

## The buy side changed as well

Of the 48 compute deals we recorded since May 1, eight have a state or public entity on one side of them. Japan's government, with Nvidia, backing Noetra. The European Commission and EuroHPC in the Tarragona gigafactory. Spain's SETT. The UK government. Saudi Arabia's SDAIA with Nvidia, and separately its Public Investment Fund. The US State Department wiring capacity into the Philippines. The Department of Defense contracting with SpaceX.

That is one in six. Sovereign compute has stopped being a white paper and become a line item.

The second shift is quieter. Labs are now each other's suppliers. Meta contracted compute with Anthropic in July. Nvidia and OpenAI did a compute deal the same month. Anthropic signed with Volta and Bitdeer in August. And the money behind several of these is private credit, not venture: Blackstone and Coatue in Firmus, Apollo and Blackstone on an Anthropic compute facility. Capacity is being financed the way toll roads are financed, on contracted cash flows, while the application layer still raises equity.

## The honest caveats

Our tracker is a sample of publicly reported deals, not the whole market, and it skews toward deals that got English-language coverage. "Q3 to date" is five weeks, not a quarter, because the newest deal in our corpus is dated August 4. Small denominators move percentages around.

What survives those caveats is the direction. Four consecutive quarters, both series moving the same way, with the compute share more than tripling and the acquisition share falling by three quarters. That is hard to read as noise.

## What to do with this

**If you are raising, pitch against the bottleneck rather than the category.** The capital that used to buy companies is now buying capacity. Anyone selling power, cooling, networking or silicon is pitching into a slice of deal flow that quadrupled in nine months. Anyone selling an application is competing for a slice that is shrinking, and needs a sharper answer on why an incumbent would buy rather than build.

**Watch the acquisition rate as your bubble gauge, not valuations.** Compute contracts are pre-committed spend and cannot deflate quickly, which makes them a lagging signal. Acquisitions are discretionary and move first. If the acquisition share is still near 3% when Q4 numbers land, the industry has stopped believing it can buy its way into position, and that tells you more about the top of this market than any headline valuation will.