When the Customer Is Also the Competitor
A report surfaced on June 29 that Google told Meta in March 2026 it could not deliver all the compute capacity Meta had asked for, throttling access to some of the Gemini-class infrastructure Meta was relying on for internal AI work. Both companies declined to comment to the Financial Times, but the detail matters enormously: the two firms had signed a cloud agreement worth more than 10 billion dollars over six years in August 2025. Meta was, in other words, one of Google Cloud's largest and most strategic customers, and it still got capped.
We have argued for months that the AI buildout has crossed from a money problem to a physics problem, and this is the cleanest proof yet. When a vendor tells a customer paying it billions that it simply cannot supply what was ordered, the constraint is no longer price. It is power, land, and silicon delivery schedules. For every CIO who assumed a signed cloud contract guaranteed capacity on demand, the Google-Meta episode is a cold splash of reality.
The Numbers Behind the Squeeze
Google has not been shy about the scale of the problem. VP Amin Vahdat said in November 2025 that the company must double its AI capacity every six months just to keep up with demand, a growth rate that no construction pipeline can sustain indefinitely. The company has responded by lifting its 2026 capital expenditure expectations to between 180 and 190 billion dollars and signaling that 2027 will bring significant further increases. Those figures are eye-watering, and yet they are still not enough to satisfy every customer in full.
The funding side tells the same story. Google has reportedly sought roughly 84.75 billion dollars in equity to bankroll AI infrastructure and committed to a monthly lease worth around 920 million dollars for SpaceX and xAI data center capacity. When a company with Alphabet's balance sheet is raising tens of billions and renting capacity from a rocket company's AI affiliate, the message is unmistakable: even the hyperscalers cannot build fast enough, so they are buying, borrowing, and rationing all at once.
Why Meta Was Buying From a Rival in the First Place
It is worth pausing on the strangeness of the relationship. Meta runs one of the largest private GPU fleets on earth and has committed to hundreds of billions in its own data center buildout. That it was still leaning on Google Cloud for compute shows how acute the shortage has become: when you cannot build your own capacity fast enough, you rent from whoever has spare megawatts, even a direct competitor in advertising, social, and AI models. The 10 billion dollar deal was a pragmatic admission that no single company can self-supply during the crunch.
That pragmatism cuts both ways. By selling Meta compute, Google booked revenue and filled capacity, but it also took on the awkward position of rationing a competitor's growth. When the squeeze came, Google protected its own workloads and other customers first, leaving some of Meta's internal projects short. Any enterprise that depends on a vendor which also competes with it should study this carefully, because priority during a shortage is decided in a room you are not in.
Capacity Is the New Competitive Moat
There is a strategic dimension here that goes beyond one bilateral deal. For years the cloud advantage was measured in services, regions, and managed tooling. In the AI era it is increasingly measured in raw, deliverable accelerator capacity, the one thing that cannot be conjured by a software update. Google protecting its own workloads when supply tightened was not pettiness, it was the rational behavior of a company that knows compute is now its scarcest and most valuable asset. The provider with the most physical capacity wins, and everyone else negotiates from weakness.
That reframing should shape how executives evaluate their cloud partners. The relevant due-diligence question is no longer just uptime and pricing, it is how much power and silicon a vendor has actually secured and energized for the years you plan to grow into. A hyperscaler sitting on contracted gigawatts and a full GPU delivery pipeline is a fundamentally safer bet than one with a slicker console but a thinner supply position. The Google-Meta episode hands buyers a new lens, and it is pointed straight at the substation.
The Lesson for Enterprise Buyers
If Meta can be capped, so can you. The practical takeaway for CTOs and CIOs is that committed-capacity language in cloud contracts is now the most important clause to negotiate, more important than headline pricing or discount tiers. A contract that promises a rate but not a guaranteed quantity of accelerators is, in a shortage, a promise to sell you something that may not exist when you need it. Reserved capacity, firm delivery dates, and penalty clauses for under-delivery are no longer luxuries.
The deeper lesson is diversification. Relying on a single hyperscaler for AI compute concentrates your fate in that vendor's allocation decisions. We expect more enterprises to spread workloads across multiple clouds and neoclouds, not for resilience against outages, but for resilience against rationing. The Google-Meta story is a preview of the next two years: budgets will keep growing, and the binding constraint will keep being whether anyone can actually deliver the machines.
