Compute Stopped Being a Moat. It Is Becoming a Rental Market.

Meta is reportedly in talks to lease Anthropic up to $10 billion of AI data center capacity over two years. Those two companies are racing to build the same thing. One of them is offering to rent the other the ground to build it on.

For three years the industry has treated compute as the moat. Buy the GPUs, sign the power contracts, lock the capacity, and competitors simply cannot follow. That story is quietly coming apart, and the deals from the past week show how.

**The supply side got crowded**

SpaceX is negotiating to sell dedicated AI data center capacity to the Pentagon in a multi-billion dollar deal. A rocket company is now a compute vendor. Reflection AI committed to spend over $1 billion with Nebius for Nvidia-powered capacity through 2029, buying rather than building. And Mitsubishi Heavy Industries is working with Nvidia on modular cooling and power infrastructure for large-scale AI data centers.

Read those together and a pattern shows up. The people selling compute are no longer just the three hyperscalers. They are neoclouds, industrial conglomerates, aerospace firms, and now frontier labs with capacity they overbuilt.

**Overbuilding has a cost, and leasing is how you pay for it**

The unglamorous explanation for the Meta deal is the correct one. Capacity built ahead of demand is enormously expensive to hold idle. Depreciation runs whether or not the racks are busy. Leasing converts a stranded asset into revenue, and if the buyer happens to be a competitor, the finance team does not especially care.

That is a rational move. It is also an admission. You do not rent out your moat.

**What this reprices**

If compute is becoming a utility, the returns migrate to whatever stays scarce. Three candidates are visible in the deal flow.

Power and thermals. Every large deal in the past month has an energy or cooling component attached. The Mitsubishi work is infrastructure, not intelligence, and it is where the physical constraint actually bites.

Distribution. If capability is rentable and, as Moonshot's open-weight Kimi K3 suggests, increasingly cheap, then owning the customer relationship matters more than owning the model.

Specialised silicon. Etched is reportedly raising at roughly a $20 billion valuation. General-purpose compute commoditises fastest. Purpose-built silicon is a bet that the next margin sits below the model layer.

**The honest caveat**

One lease does not make a market, and the Meta talks are early. Compute is still genuinely scarce at the frontier, and access to the newest chips still separates labs. The claim here is narrower: the direction of travel is from strategic asset toward priced commodity, and 2026 is the year that started showing up in the deal terms rather than in the commentary.

**What to do with this**

If you are evaluating an AI company, stop treating announced compute commitments as a competitive advantage and start reading them as a cost structure. A $1 billion capacity commitment is a fixed obligation, not a moat, and it should be underwritten like one.

Watch for the second lease. One frontier lab renting to another is an anomaly. Two is a market. If a second such deal lands this quarter, the repricing is real, and compute-heavy valuations built on scarcity assumptions deserve a much harder look.

You can follow the compute deals as they land in our AI deal tracker.