Google Cloud Selling $35B/GW TPU Systems, Eyes Mid-100s Growth

Full analysis

Your draft

Google Cloud is booking full TPU systems as gross revenue to outside shell companies that then rent the chips to Anthropic and others, and SemiAnalysis says that mechanic pushes GCP toward triple-digit growth in 2027 while Gemini gets starved of compute internally. What's being decided, for anyone building on top of these labs, is whether frontier compute is quietly consolidating under a landlord whose incentives are rent, not research. This is a Type 1 read: the SPV structures and multi-year TPU commitments named in the quote are already signed, and they set the queue you'll be waiting in through 2027.

The Skeptic

SemiAnalysis is good, but a $3-EPS, mid-100s-growth call is a buy-the-stock thesis that happens to come with a research methodology. Gross revenue booked to special-purpose vehicles is the oldest trick for making growth look clean. You ship the chip, book the whole system price, and the utilization risk sits in an entity you half-control. Whether that $35B/GW is a sale or vendor financing dressed as one determines if any of this survives an audit. And notice the plot: Google is selling the compute that trains Anthropic, the model that competes with Gemini. Either Kurian is very clever, or Google's own internal ROI on Gemini is worse than anyone will say out loud. For the PM: Google may be making more money renting shovels than digging, which tells you what they think of their own gold.

The Safety Lens

Anthropic built its whole brand on the Responsible Scaling Policy: compute thresholds it can pause at, conditions it controls. Read the quote again. More than 20% of TPU shipments through 4Q27 go to Anthropic, on top of hundreds of thousands of chips already rented, through SPVs whose reason to exist is GCP revenue. A pause condition you can't enforce because your landlord reallocated the pod is not a pause condition. If a constrained quarter arrives and Google moves TPUs to Meta or back to Gemini, Anthropic's safety levers bend to a commercial counterparty's spreadsheet. For the PM: the safety-first lab now rents its safety brakes from the company it's racing.

The Compute Pragmatist

The $35B/GW number is doing something. Equivalent NVIDIA H100/H200 builds have tracked closer to $10-15B per gigawatt at scale. So either TPU v6 systems deliver two-to-three times the performance-per-dollar, or SPV buyers are paying a fat premium for supply certainty and ecosystem lock-in. I'd bet mostly the latter, because when you've committed to four-plus gigawatts of pre-sold capacity, the scarce thing isn't flops, it's a guaranteed delivery slot. That backlog cuts both ways. It de-risks Google's capex if the chips ship on schedule. It becomes the 2027 story if delivery windows slip and SPVs start renegotiating. For the PM: paying $35B a gigawatt buys you a place in line, not a discount.

The Builder

If you call Anthropic's API, you now sit two layers down from a Google capacity decision nobody disclosed to you. Your latency SLA, your rate limits, your ability to get burst inference on launch day: all contingent on how much TPU the relevant SPV holds versus what Google rents back that quarter. When Anthropic ships a new model and everyone hammers it, the queue you're stuck in was set by a contract between Google Cloud and an entity you'll never see a logo for. You can't instrument that. Your only real hedge is a second provider wired up and eval'd before you need it. For the PM: your uptime dashboard is green while the plumbing underneath is somebody else's rental agreement.

The Researcher

The "financialization of GCP" framing is a clean story, and clean stories about messy intra-firm politics deserve suspicion. The genuinely consequential claim buried in here isn't the EPS math. It's that Thomas Kurian won the compute-allocation fight against DeepMind. If that's true, Google decided that renting compute at high margin beats spending it on its own frontier model. That's an organizational incentive shaping frontier AI timelines, which is more interesting than any revenue-recognition question. But I'd want more than one SemiAnalysis post before I treat "Gemini is starved" as fact rather than narrative. For the PM: the interesting bit is Google choosing to be the landlord over being the tenant.

Where they part ways

Three real disagreements. The Compute Pragmatist reads the $150B+ backlog as a de-risking event; the Skeptic reads gross-booked SPV revenue as exactly the thing that looks firm until it isn't. Both are looking at the same number and seeing opposite balance sheets. The Researcher thinks the Kurian-over-DeepMind story is the signal; the Skeptic thinks a Gemini starved of compute is the tell that Gemini's ROI never justified the spend. Same fact, one calls it strategy, the other calls it surrender. And the Safety Lens and the Builder agree on the mechanism but split on who should care: the Builder wants a failover provider, the Safety Lens says a failover doesn't fix a lab whose scaling policy runs on rented silicon.

What it hinges on

Two things. First, is that $35B/GW a genuine arm's-length sale or vendor financing that inflates GCP's top line, because that decides whether the mid-100s growth is real or accounting. Second, does Google honor Anthropic's delivery slots under stress, because that decides whether "compute access as a safety lever" means anything. The council leans skeptical on the revenue framing and genuinely worried on the dependency. Before you take the SemiAnalysis growth number as gospel, watch whether Alphabet breaks out TPU-system sales as a line item and whether an auditor blesses the gross treatment. Before you bet your product on Anthropic uptime, wire a second model behind a router and eval it now, not on launch day.

Prediction: In Alphabet's Q4 2026 earnings report (early Feb 2027), management will NOT disclose a discrete "TPU systems sold to third-party SPVs" revenue line, keeping the gross-booking mechanic folded inside Google Cloud revenue.

Confidence: Medium. Hyperscalers consistently bury structural revenue detail, and disclosure invites the auditor scrutiny they have every reason to avoid.

Why: The whole appeal of the SPV structure is that it makes GCP growth look clean while keeping the utilization and financing risk off the visible balance sheet, and breaking it out as a line item would hand analysts exactly the tool they'd need to test whether it's a real sale or vendor financing. Google has a long track record of reporting Cloud as a single blended number and resisting segment-level granularity even when investors ask. The opposite outcome, voluntary disclosure of a high-margin new stream, only happens if management wants to hard-anchor the mid-100s growth story publicly, and they gain more by letting SemiAnalysis carry that narrative than by putting their own name on a number an auditor has to defend.

Revisit by 2027-02-15: We're right if Alphabet's Q4 2026 filing and call keep TPU-system SPV sales inside blended Google Cloud revenue with no discrete breakout. We're wrong if they report a separate line item or give a specific dollar figure for third-party TPU system sales.