Too Much AI Computing, Too Few Paying Customers

Meta has a problem no other hyperscaler has faced. Amazon and Google built AI infrastructure on top of existing cloud businesses that could sell computing power to outside customers. Meta built AI infrastructure for its own advertising engine, and spent $115–145 billion in 2026 capex doing it ^[1]. The ad business can't absorb all that capacity. So Meta is launching "Meta Compute" to sell the surplus externally, with Zuckerberg framing it plainly:

We haven't done that yet, because we think that we have a use for the compute. But obviously, if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out. — Pavel Durov

^[1] The market's reaction told you what the announcement meant. Meta's shares rose 10% on the cloud news. CoreWeave dropped 10.8% and Nebius dropped 12.4% the same day ^[1]. Investors didn't read Meta's entry as proof that AI computing demand was growing. They read it as a new competitor arriving with excess inventory, threatening the specialized AI cloud providers who already exist. Meta is adding surplus capacity to an external market where enterprise buyers are pulling back ^[2][3]. And they are pulling back with reason. Ninety-five percent of organizations report no measurable return on AI investments ^[4]. AI computing costs now exceed human labor costs at multiple firms — Nvidia's own Bryan Catanzaro says for his team, compute costs dwarf employee costs ^[5]. An unnamed enterprise received a $500 million monthly bill for unrestricted Claude access. Uber exhausted its 2026 AI budget by April. Microsoft canceled internal Claude Code licenses at $2,000 per engineer per month ^[2]. Meta's own CTO reversed internal AI policy, telling employees not to use AI tools just for the sake of using them ^[3]. PNC's CEO put the demand-side math in terms any bank customer would understand:

Any impact that AI can have on the productivity of a bank, that productivity can be taken away by the cost of tokens. — Bill Demchak

^[6] To be clear, AI is producing returns in some deployments — just not the ones Meta needs to buy its excess capacity. Meta's own ad-targeting AI generated $240 billion in forecast revenue this year, and Baidu's AI revenue surged 49%, now more than half its general business ^[7][8]. But those are internal platform deployments where the company that built the AI is also the company consuming it. The external enterprise layer — where Meta now needs paying buyers — is where returns are absent. Companies that buy AI computing from outside vendors are the ones exhausting budgets without results, not the platforms running AI for their own products. The contradiction is precise. The hyperscaler without a cloud business built more AI capacity than it can consume and needs external buyers. Those external buyers are scaling back because the computing costs more than the work it replaces. The supply-side overbuild meets the demand-side retreat at the midpoint of the value chain — the commercial cloud layer where AI was supposed to graduate from internal experiment to sold product. The early evidence says that external market was smaller than the $1 trillion-plus in combined hyperscaler capex assumed across 2025–2026 ^[9]. It may still grow. Meta's ad-targeting returns and Baidu's revenue surge show AI can work when the builder is also the user ^[7][8]. But the BIS warned in June that five hyperscalers' capex is outpacing earnings and free cash flow, comparing the pattern to canal mania and the dot-com boom ^[9]. The June 18–22 sell-off wiped $500 billion-plus from hyperscalers in days, driven not by financing costs but by the absence of returns ^[10]. As Tom Essaye of Sevens Report Research put the dot-com parallel:

While people connected to the internet, their connection wasn't nearly as profitable as quickly as everyone assumed. — Tom Essaye

^[11] The crack is not in the technology. It is in the arithmetic between what was built and what it earns — and it is widest exactly where Meta is now trying to sell.