Meta Is Entering a Business That Can't Turn a Profit

This month CoreWeave signed contracts to rent A100 chips — Nvidia's workhorse, launched in 2020 — through 2029. Six-year-old chips, booked for another three years, at prices the company says are at or above where they were years ago ^[1]. CoreWeave's chief executive, Mike Intrator, put the demand picture plainly.

pricing for prior generation SKUs is at or above where it was years ago. — Mike Intrator

Everything that should make a compute business profitable is in place. CoreWeave holds a $104 billion backlog that is effectively sold out, backed by 4.2 gigawatts of contracted power ^[2]. Pricing holds flat. The hardware holds value. And the company still won't turn a profit until fiscal 2028, carrying 6.7 times net debt to EBITDA to get there ^[2]. The reason isn't the GPUs. It's everything around them — power, cooling, land, networking — the capital intensity of the surrounding infrastructure that eats the margin no matter how full the order book is. That is the business Meta just walked into. Meta's chatbots hold a sliver of the market against OpenAI and Alphabet, and analysts read its $130 billion to $145 billion in AI capital spending as money it can't find enough internal use for ^[3]. So this summer it launched Meta Compute, selling the excess capacity to outside customers ^[4]. The market cheered the pivot — Meta's stock jumped 10% — and punished the incumbent, wiping 50% off CoreWeave's shares and $49 billion off its value ^[4][3][5]. The read was that a giant entering compute-as-a-service would commoditize the neoclouds' margins. The market had the direction wrong. Meta is entering the one business where the pure-play specialist, with every advantage, still can't make money. And it is arriving late, into a market that is already crowding. In July a secondary exchange opened for used Nvidia GPUs, trading H100s and A100s at 40% to 60% below list price with real-time price indexes — the same market structure that exists for oil and grain ^[6]. SpaceX and Bitcoin miners are piling in, TeraWulf signing a $19 billion deal with Anthropic ^[5]. The hyperscalers Meta means to disrupt hold $460 billion backlogs and their own custom silicon ^[7]. QumulusAI is undercutting on marginal cost ^[8]. Meta is even, at the same time, one of Microsoft Azure's largest AI customers and a direct competitor to it, while negotiating a $10 billion, two-year lease to Anthropic — the same circular pattern that already runs through the chip layer, now one floor down ^[9][10]. The proof was on the table before Meta moved. The company that does exactly what Meta is now attempting, with demand sold out and pricing flat and hardware that holds value for nine years, still can't turn a profit. The chips will be fine. The business won't. Moving down the stack doesn't close the gap; it just moves with you.