Meta Just Did the Neoclouds a Favour

The market read Tuesday's headline as new supply. It is actually new demand.

I am writing this from the Maldives, where I have spent the week doing very little on purpose. Being far from the screens has a way of making market reactions look clearer, not blurrier. And the reaction I watched from here this week deserves a closer look.

On Tuesday, Bloomberg reported that Meta plans to sell AI compute to outside customers. The reaction was instant. CoreWeave fell 14 percent. Nebius fell 17 percent. IREN fell 13 percent. Meta jumped 9 percent. And within hours, the word “overcapacity” was back in circulation, as if it had been waiting by the door.

The logic of the selloff is simple. A giant is entering the rental market. More supply means more competition means lower prices. Sell the landlords.

I think that logic has the trade exactly backwards. Let me walk through why.

Meta Is the Tenant

Start with what Meta has actually been doing with its money, not what a headline implies it might do.

Meta signed a deal with Nebius worth roughly 27 billion dollars. It signed contracts with CoreWeave worth well over 14 billion. Trackers who map this market site by site put Meta’s contracted third party capacity at more than 5 gigawatts across cloud and colocation deals in the first six months of this year alone. Since early 2024 the running total is close to 10 gigawatts. And the majority of Meta’s capacity additions are now coming through third parties, not through its own builds.

Sit with that for a second. The company the market just decided will flood the rental market with spare compute is the single hungriest tenant in that market. It is renting other people’s data centres at a record pace because it cannot build its own fast enough.

Nobody leases a second apartment because their own house has too many empty rooms.

So Why Sell Compute at All?

This is the fair question, and it has a good answer. Selling compute is not what a company does when it has too much. It is what a company does when compute has become absurdly profitable to own.

Elon Musk proved this a few months ago. SpaceX signed compute deals at rates estimated to run three to four times what ordinary providers charge per megawatt. The contracts are enormous, the margins are extraordinary, and they carry 90 day cancellation rights on both sides. At that kind of pricing, a couple of hundred megawatts pointed at one customer can generate something like 10 billion dollars of revenue a year. Musk did not find excess capacity. He invented a new market segment: premium compute, available now, priced accordingly.

Very few companies can play in that segment, because it requires having gigawatts already built and the balance sheet to carry them without a long term offtake. Meta is one of them.

And that is only one of Meta’s options. I count four. Keep training frontier models. Scale its ads recommendation systems, which appear able to profitably absorb ten times the current compute, because every extra gigawatt shows up directly in ad performance. Sell access to frontier models as a service, the way Amazon does with Bedrock, and there are credible indications Meta is already in talks with at least one frontier lab on exactly that. Or rent raw capacity at SpaceX style pricing to whoever needs it most urgently.

Four uses. All high margin. All fighting for the same compute.

This Is New Demand, Not New Supply

Here is the point I keep coming back to, because it is the whole thesis in one sentence.

Tuesday’s announcement created no new supply. Not one wafer. Not one megawatt. Not one transformer. The physical constraints that have defined this entire cycle, the packaging lines, the memory allocation, the multi year grid queues, are exactly as binding today as they were on Monday.

What the announcement did create is a reason for Meta to acquire even more compute. When every gigawatt you control can be pointed at four different high margin uses, the rational move is not to slow down your procurement. It is to accelerate it. If the superintelligence effort needs the capacity, it gets it. If it does not, the same capacity earns premium rent instead. There is no version of that decision tree where Meta buys less compute.

Which is why I expect Meta to remain one of the largest sources of contracted backlog growth for the neoclouds, including the very names the market sold on Tuesday. Satellite and permit tracking shows Meta’s two largest campuses alone have 2.5 gigawatts under construction. My view is that Meta’s 2027 capital spending will come in far above what the market currently expects.

So the sequence on Tuesday was this. A tenant with a near infinite appetite announced it had found a way to make compute even more profitable. The market responded by selling the tenant’s landlords.

What Would Change My Mind

I always try to name the thing that would break the argument, so here it is.

If Meta stops contracting third party capacity, that matters. If rental rates for GPU hours start falling in a sustained way, that matters even more, because rental rates are the one signal in this industry that cannot be faked. And if the neocloud backlogs stop growing while all this Meta capacity comes online, then the overcapacity crowd will have been early rather than wrong, and I will say so.

None of those things is happening. The backlogs are growing. The rates are firm. And the hungriest tenant in the market just told you it wants more.

The market sold the landlords because a tenant announced he would like to collect rent too. I have read a lot of market reactions over the years. This one will not age well.

Neel Khokhani, Founder and CEO Epochal Corporation, @neel_epochal