Meta Is Building a Cloud Business It Never Had — and the Timing Couldn't Be Worse

When Meta launched "Meta Compute" on July 1, the market read it as ambition — shares jumped 10% on the news that the company would sell raw GPU compute and cloud access to its proprietary AI models to outside customers ^[1]. The reading was wrong. Meta Compute is not a diversification play. It is a survival tactic, and the distinction matters because it reveals something structural about the AI buildout that no earnings call has stated plainly: Meta is the only company in the Magnificent Seven pouring hyperscaler-scale capital into AI infrastructure without a hyperscaler's existing revenue base to land it on. The numbers make the divide concrete. Amazon is spending $200 billion on AI data centers, but AWS grew 28% year over year to $37.6 billion in revenue and now contributes 59% of Amazon's operating profit ^[2][3]. Microsoft's Azure is growing 40%, and its Copilot products generate $37 billion in annual revenue — the company produced $77.4 billion in free cash flow last year ^[4][5]. Google Cloud grew 63% to $20 billion with a $460 billion backlog ^[2]. Each of these companies built its cloud franchise over a decade or more, and each can now feed new AI capacity into an existing customer base with predictable enterprise contracts. The hardware spend has somewhere to go. Meta has advertising. The ad business is formidable — $240 billion in forecast 2026 revenue, growing 22.3% on the back of AI-powered targeting — but it is not a cloud franchise ^[6]. It does not generate the kind of multi-year enterprise commitments that absorb server capacity predictably. And in the first quarter of 2026, Meta reported its first-ever decline in total daily active users ^[6]. The core business that funds the hardware buildout is showing signs of plateauing at the exact moment the buildout is accelerating. Zuckerberg has been unusually direct about the problem. On a shareholder call in early June, he framed selling excess computing capacity as a way to compensate for the company's lack of a dedicated cloud business ^[7]. CFO Susan Li went further, acknowledging the overbuild risk explicitly.

If we end up not needing as much as we anticipate, we can choose to bring it online more slowly or reduce our spending in future years. — Susan Li

This is not the language of a company executing a planned diversification. It is the language of a company hedging a bet it knows might not pay off. The launch of Muse Spark 1.1 on July 9 confirmed the pattern ^[8]. Meta's AI chief, Alexandr Wang, made the relationship between software and hardware explicit.

This is going to be served on top of the computer infrastructure that we've built. — Alexandr Wang

The software product exists to fill the hardware base, not the reverse. And the pricing tells the same story: Meta is charging roughly $1.25 per million input tokens and $4.25 per million output tokens, about 25% of what OpenAI and Anthropic charge ^[8]. That is a strategy best explained if the priority is keeping servers utilized rather than maximizing software revenue — though it could also serve market-entry or competitive-disruption goals. Either way, it sacrifices the premium margin a proprietary model could command. The market Meta Compute is entering makes the timing brutal. OpenAI, the sector's leader, lost $20.9 billion on $13.1 billion in revenue in 2025 — revenue grew 253% but losses grew faster. The company pays Microsoft $17.2 billion annually for cloud and R&D, and its IPO filing revealed the structural gap between AI software revenue and AI infrastructure cost. Meanwhile, a price war between OpenAI and Anthropic is compressing token prices just as enterprise customers are discovering they cannot afford the technology ^[9]. Sam Altman has given the phenomenon a name.

Today, 6.5 years later, that is about the per capita average in the world — Sam Altman

Companies including Microsoft, Uber, and Salesforce are now implementing token caps and rationing access ^[10]. Meta itself is not immune. In early June, Meta's own CTO, Andrew Bosworth, issued a strikingly candid internal warning ^[10].

All motion is not progress, and token usage alone is not a measure of impact of any kind. — Andrew Bosworth

The admission is remarkable: the company spending $125 to $145 billion a year on AI infrastructure is simultaneously cautioning its own employees that token usage proves nothing about value. The financial engineering required to sustain this buildout underscores how far Meta is stretching. The company has borrowed $55 billion, including $27 billion for the Hyperion data center, halted share buybacks, and is considering a new equity sale of tens of billions of dollars ^[7]. It has committed $600 billion to U.S. AI infrastructure over three years ^[11]. And it is simultaneously buying tens of millions of AWS Graviton cores — a company building its own compute empire while buying tens of millions of cores from a rival's silicon ^[12]. The stock has dropped more than 17% as investors watch free cash flow deplete without clear evidence of high-margin AI revenue to replace it ^[11]. The contrast with peers is not subtle. Artisan Value Fund explicitly dumped Meta and added Amazon in the first quarter, and its reasoning named the structural gap directly ^[13].

Additionally, we believe expenses are likely to grow faster than revenues in the near term as Meta invests heavily in AI. — Artisan Value Fund

The fund chose Amazon precisely because AWS contributes roughly 60% of operating income — the cloud revenue cushion Meta lacks and is now trying to build retroactively ^[13]. Even Meta's relationship with its own compute suppliers reveals the tension. CoreWeave, the neocloud provider, signed a $21 billion expanded deal with Meta this spring, bringing Meta's total contracted spend to $35 billion through 2032 ^[14]. Yet CoreWeave projects profitability only by 2028 and carries high debt — and Meta Compute now directly competes with the company Meta is paying billions to ^[14]. When Meta Compute launched, CoreWeave's stock fell as much as 17% ^[1]. The broader market is starting to price this structural divide. The Magnificent Seven lost $2.3 trillion in market value in June 2026 as AI capital expenditure across the sector exceeded $700 billion annually, with investors concerned that costs are depleting cash generation without clear evidence of monetization ^[15]. But the pain is not evenly distributed. Microsoft fell 20% despite having Azure and Copilot ^[15][4]. If the hyperscaler with the strongest software monetization story faces that level of skepticism, Meta — lacking both a cloud franchise and a proven AI software revenue stream — faces the same skepticism from a structurally weaker position. There is a counterargument worth taking seriously: the physical infrastructure market may still be supply-constrained, not oversupplied. Jefferies reported a 12-gigawatt global data center deficit in 2025, with cloud backlogs approaching $2 trillion ^[16]. If demand for compute continues to outstrip supply, Meta's capacity will find buyers regardless of the software layer's economics. But that argument assumes Meta can compete with established cloud providers on reliability, enterprise sales, and the ecosystem of services that make AWS, Azure, and GCP sticky — advantages built over decades, not months. What Meta Compute really reveals is the asymmetry at the heart of the AI buildout. The companies that built cloud franchises when the economics were favorable — before the current capex cycle, before the price war, before enterprises started rationing tokens — can now absorb AI infrastructure as an extension of an existing business. Meta is trying to build the cloud franchise and the AI infrastructure simultaneously, in a market where the software layer's economics are deteriorating by the quarter. The question Meta Compute poses is not whether Meta can compete with AWS. It is whether any company can build a cloud business from scratch in 2026 and make the numbers work.