Google's Direct TPU Sales Are a Margin Story, Not a Re-Rate Story

Google announced in April that it would start selling its Tensor Processing Units—the chips it spent a decade building to run its own AI—to a select group of third-party data-center operators, instead of renting them out only through its cloud.Google would sell its TPUs to a select group of third-party operators The market has read this as the opening of a slow-motion hunt at NVIDIA. Alphabet trades at roughly 17x trailing earnings17x TTM P/E on a $4.1T market cap because the market prices it as an advertising business, not an AI-infrastructure compounder, so the bet writes itself: make Google a merchant AI-chip vendor like NVIDIA, and the stock finally collects the AI-infrastructure premium its front-page competitor enjoys.

That fantasy assumes the re-rate comes from selling silicon. Google's own numbers say the value lives somewhere else.

The chip is where Google gives value away, not where it earns it—and the Anthropic deal is the proof.

Google's biggest external TPU customer, Anthropic, is the cleanest illustration of the two business models sitting side by side. The October 2025 expansion gave Anthropic access to up to a million TPUs and well over a gigawatt of capacity, a deal worth tens of billions.access to up to one million TPUs and well over a gigawatt Independent estimates split that exposure in a revealing way: roughly 400,000 Ironwood chips went to Anthropic as a direct hardware sale, around $10 billion of finished racks, with Broadcom as the merchant vendor; the other 600,000 are rented through Google Cloud under about $42 billion of contracted capacity.400k TPUv7 Ironwood units (~$10 billion) sold by Broadcom directly to Anthropic; 600k rented through GCP, ~$42 billion in RPO Same customer, same chip, two completely different economics.

The rented half is the whole point. Google keeps ownership of the fleet, keeps the software stack top-to-bottom, keeps the recurring revenue, and keeps its operating margin on every token. The sold half is a one-time, capital-hungry transfer in which Google ships the silicon, pockets a merchant chipmaker's margin, and hands the buyer the integration work that produced the advantage in the first place. A merchant silicon business is how NVIDIA makes its $60.4 billion data-center compute quarter and its roughly 80% shareroughly 80% of the AI accelerator market and $60.4B in data center compute revenue; building one from a standing start means undercutting that incumbent while accepting thinner margins and heavier capex than Google's existing cloud line.

Look at where the TPU's economics already show up on the income statement. In the second quarter, Google Cloud revenue rose 82% year over year to $24.8 billion, and its operating income more than tripled to $8.8 billion—a roughly 35% operating marginCloud revenue surging 82% to $24.8 billion and operating income more than tripling to $8.8 billion on a hyper-scale cloud, with a $514 billion contracted backlog behind it$514 billion backlog of contracted work. That margin expansion is the TPU working. Because Google runs its frontier models on its own silicon, it can price Gemini roughly 58% below Anthropic's Claude or 60% below OpenAI's flagship on a per-token basis while still clearing a margin58% lower than Anthropic's Claude Opus 4.7 and 60% lower than OpenAI's GPT-5.5, an edge competitors cannot simply buy. None of that shows up as "TPU revenue." All of it shows up as cloud margin.

The merchant-silicon pivot does not fund itself on those terms. Alphabet booked negative free cash flow in the quarter—$39.1 billion of operating cash flow against $44.9 billion of capex$39.1 billion of operating cash flow and $44.9 billion on capital expenditures—raised its 2026 capex guidance to roughly $205 billionraised its 2026 capex guidance to $205 billion, and, for the first time since its 2004 IPO, sold $80 billion of new equity$80B equity raise to pay for the buildout. So the story running behind the chip sales is that shareholders are being diluted to fund the very fleet whose silicon Google is now considering shipping to third parties at a lower margin than it could rent. That is a strange capital allocation if the goal were to improve per-share economics. It makes sense only if the goal is share-and-scale in a market big enough to matter at Alphabet's size—a $4.1 trillion company needs a very large hardware line before it moves the per-share needle.

The harder question is whether the TPU's advertised edge even survives the trip out the door. The 4x cost-performance claims, the 44% lower all-in total cost of ownership versus NVIDIA's GB200, the 52% lower cost per effective petaflop Anthropic gets against the GB300—these are all real, but they are Google-stack artifacts.all-in TCO per TPUv7 Ironwood chip approximately 44% lower than GB200; Anthropic ~52% lower TCO per effective PFLOP versus GB300 They depend on Google's power-efficient design, its self-owned software, its networking, and the tight fit between its models and its silicon. Hand a bare chip to a third-party operator running its own fleet, with its own software and utilization, and the margin of superiority compresses toward the general-purpose average. NVIDIA's 80% share is not falling because TPUs conquered on merit alone. It is falling, slowly, because customers such as Anthropic are deliberately hedging a single-vendor dependence—Anthropic runs TPUs, Amazon's Trainium, and NVIDIA GPUs in paralleldiversified compute strategy utilizing Google's TPUs, Amazon's Trainium, and NVIDIA's GPUs—and because NVIDIA's own Rubin ramp has slipped. Attribute some of the apparent challenger success to customer hedging and a rival's stumble, not to a demonstrated product rout.

That is why the real moat battle is software, not silicon. Google and Meta are quietly pushing TorchTPU, an effort to run PyTorch natively on TPUs, which is an attack on the CUDA switching costsTorchTPU aimed at running PyTorch on TPUs, targeting switching costs locked into CUDA that have held NVIDIA's ecosystem together for two decades. If that works, the TCO advantage transfers to outsiders and a true merchant business becomes plausible. If it does not, Google sells fast bare chips into a world whose engineering habits are still wired to CUDA.

The re-rate, if it comes, will be visible in the margin line, not in chip unit shipments.

The silicon sale is the headline; the cost engine is the story. The TPU has already demonstrated its value to shareholders—not as a product to be parceled out to third parties, but as the reason Google's own AI is cheaper to serve than anyone else's and its cloud is growing profitably at 82%. The market is already pricing much of that through Cloud's margin expansion. Selling the chip outright hands that moat to others, earns a thinner margin for doing so, and costs real shareholder capital upfront. It is a reasonable side business and a strange reason to re-rate an entire company. Watch Google Cloud's margin and whether TorchTPU makes the edge portable; the stock's multiple will follow those, not the count of racks sold to data-center operators.