Anthropic is giving Google what every new chip platform needs: a customer large enough to force the rest of the business into existence.
Before Anthropic, Google's TPUs were principally an internal advantage and a service rented through Google Cloud. Anthropic's commitment—up to one million TPUs, more than one gigawatt in 2026 and another 3.5 GW of Broadcom-supported capacity starting in 2027—is helping turn them into something closer to Nvidia's offer: merchant systems, software, installation, support and financing wrapped around the silicon.
That does not put Google at Nvidia's revenue scale, customer breadth or installed base. It does pull Google toward Nvidia's business model. Alphabet now has $43.8 billion of current credit backstops tied to data centres and an agreement to provide an estimated $24.1 billion of additional backstops for data-centre and energy infrastructure. That is $67.9 billion of current and future exposure before the separate $7.6 billion of financial guarantees. Nvidia has disclosed $108.5 billion of gross guarantees behind AI infrastructure.
Google's disclosed backstop capacity is therefore about 63% of Nvidia's guarantee book. The filings do not identify Anthropic as the counterparty behind every Google backstop, so that ratio is a company-level comparison, not an allocation to one customer. But Anthropic is the named customer showing why Google needs the machinery at all.
Anthropic turns the TPU into a merchant product
Anthropic's own announcements supply the hard edge of the programme.
- In October 2025 it said it would gain access to up to one million Google TPUs, worth tens of billions of dollars, bringing well over one gigawatt online in 2026.
- In April 2026 it announced multiple gigawatts of additional next-generation TPU capacity starting in 2027. Broadcom's filing put approximately 3.5 GW of that access through Broadcom and made Anthropic's consumption conditional on its continued commercial success.
- Anthropic said it still trains and runs Claude across AWS Trainium, Google TPUs and Nvidia GPUs. Amazon remains its primary cloud and training partner.
Alphabet's June 10-Q makes the strategic change explicit. Google Cloud now includes product sales, primarily TPU systems. It began recognising revenue from a limited number of TPU supply agreements in the second quarter, with the significant majority expected in 2027.
That is the line between an internal accelerator and a merchant product. Google is no longer only using TPUs to lower the cost of Gemini or renting them inside its own cloud. It is selling complete systems—hardware, software, installation, support and warranty—to customers that want specialised, high-scale infrastructure on their own premises.
Anthropic is the reference customer and the forcing function. A commitment this large requires Google to make the TPU consumable outside its own estate, coordinate Broadcom and data-centre partners, support customer deployments and help make the resulting capacity financeable. One buyer does not make a broad merchant market, but it can make Google build the product and operating system needed to serve the next one.
The $150 billion is a programme, not a delivery
The Financial Times reported this month that Google had assembled roughly $200 billion of interlocking contracts to put more than $150 billion of TPUs into Anthropic's hands. The programme is real. The verb travelling with it is not: Google has not already supplied $150 billion of chips. The FT figure is a reconstruction of a multi-year financing and supply network, much of it for capacity that starts in 2027, rather than a shipment total or one announced invoice.
Anthropic's disclosures establish the chips, the scale and the timetable. They do not establish $150 billion of completed deliveries. The FT's reported >$150 billion chip total is about three quarters of the wider $200 billion network it reconstructed; Reuters separately reported, citing The Information, a $200 billion five-year Anthropic commitment to Google Cloud and chips. That would average $40 billion a year, but neither company has published that contract or split it between hardware, cloud consumption and services.
That distinction matters because Broadcom itself describes the next 3.5 GW as contingent. A capacity commitment whose use depends on the customer's continued commercial success is not the same accounting fact as hardware accepted and revenue recognised.
Anthropic is pulling Google toward Nvidia's business model
Broadcom is the development and supply partner. Private credit buys the equipment, leases turn an unrated lab's payments into financeable paper, and backstops give lenders another balance sheet to look through to. Anthropic's demand is pulling those pieces together. The package starts to resemble what Nvidia sells: not a chip, but a financed route from silicon to an operating cluster.
That is the useful meaning of Google rising toward Nvidia's level. It is not market-share parity, and it is not proof that TPUs match Nvidia across every workload. It is Google learning to sell and finance a second full stack rather than keeping its silicon advantage inside the company.
The balance sheet is now part of the product
Alphabet's filing says the current credit derivatives have terms of up to 15 years. Their notional rose from $16.9 billion at December 2025 to $43.8 billion in June, a 159% increase in six months. The recognised liability was $815 million, 1.9% of the notional—not a forecast of the ultimate loss, but the accounting value carried for the risk at quarter-end.
The maximum is not the expected bill. Exposure declines as the underlying parties pay, and Google can assume or sublease the facilities after a default. The future $24.1 billion is also subject to final terms. The honest description is therefore $67.9 billion of current and future backstop capacity, not $67.9 billion owed.
Nvidia's version is larger and less subtle. Its July-quarter disclosure carried $3.5 billion of land, power and shell guarantees for AI-cloud partners plus $105 billion tied to a 4.25 GW Ohio campus expected to host Nvidia infrastructure under 20-year OpenAI leases. On its call, Nvidia also described take-or-pay support for part of a neocloud's capacity in exchange for a share of rental revenue above the floor.
Anthropic is helping Google enter the same market Nvidia has already decided it must underwrite. Nvidia still has the larger guarantee book and a far broader customer base.
Google still sells Nvidia—and does not disclose what it buys
There is no public dollar figure for Alphabet's purchases from Nvidia. Alphabet does not break servers into TPU, GPU, CPU, memory or networking spend, and Nvidia does not name Google revenue.
What Google does say is enough to reject the simple replacement story. At Cloud Next in April it called Nvidia GPUs a core part of its accelerator portfolio and said it would be among the first clouds to offer Vera Rubin NVL72 systems. Its new Virgo network is designed to connect either TPU 8 superpods or Nvidia Vera Rubin racks. Google Cloud's customer promise is choice even while Google's own silicon becomes a product.
Installed-base estimates circulate, but an installed base mixes generations, purchases and time. It is not a dollar total readers can reconcile to either company's accounts. This is where a neat point estimate would be invented. We are not going to.
The contradiction is only apparent. Google can reduce its dependence on Nvidia for internal and Anthropic workloads while continuing to sell Nvidia instances to CUDA-native customers. It can be Nvidia's cloud partner and its most credible custom-silicon rival at the same time.
What it does to the two models
Our Alphabet model already carries TPU system sales inside Google Cloud rather than as a new business. The reported $200 billion cannot be added again: Alphabet does not say whether it sits inside the $514 billion Cloud backlog, how much is hardware, or when the consumption portion becomes revenue.
The backstops do not change the model either. Charge the entire $67.9 billion against equity, assume every current and future backstop is final, fully drawn and recovered at zero, and the ceiling is $5.52 per Alphabet share. That is 1.8% of the model's $311.82 base fair value. By comparison, lowering the model's opening quarterly Cloud growth from 10% to 8% costs $8.93 a share; taking one turn off the exit revenue multiple costs $43.23. The ordinary operating assumptions matter more than an intentionally punitive treatment of the backstops.
Our Nvidia model already names custom ASICs taking inference workloads as part of its bear case: $67.73 a share against $252.93 in the base case. The Google-Anthropic programme belongs in that direction, but it supplies no Nvidia revenue displaced, comparable rack count or price concession. Moving Nvidia's assumed quarterly rack-price drift from −1% to −2% costs $27.17 a share. Nothing in the Google disclosures tells us to make that move.
Neither model changes. Anthropic is helping prove that Google's alternative stack can be sold and financed. It does not yet tell us how much Nvidia revenue that stack replaces.
What to watch
- TPU system revenue in 2027. Alphabet says the significant majority of revenue from its limited supply agreements lands then. A product-sales split would show whether Google has a merchant chip business or one very large customer.
- The credit-derivative notional and liability. $43.8 billion and $815 million are the two numbers to roll forward. The first shows gross exposure; the second shows how the accounting view of risk changes.
- Anthropic's 3.5 GW condition. Broadcom made consumption dependent on continued commercial success. Deployment, not announced capacity, settles how firm the programme is.
- A second external TPU buyer at comparable scale. Google says it has a limited number of system agreements but names no concentration. One more disclosed customer would show whether Anthropic created a platform or merely one enormous programme.
- Google's Nvidia disclosure. Vera Rubin availability proves the partnership continues. A GPU count, purchase commitment or hardware mix would finally put a number on it.
Anthropic is helping Google rise toward Nvidia's level as a financed systems vendor. It has not put Google at Nvidia's commercial scale. The next proof is another external customer, not another Anthropic tranche.
Alphabet's current credit-backstop notional, recognised derivative liability, additional future-backstop agreement, remedies, TPU-system revenue policy and 2027 timing come from its June 2026 Form 10-Q. Anthropic disclosed the one-million-TPU, >1 GW and “tens of billions” terms in October 2025, and the multi-GW expansion and multi-platform hardware mix in April 2026; Broadcom's 8-K supplies the 3.5 GW and commercial-success condition. The >$150B chip and ~$200B network figures are Financial Times reporting, not company disclosures; the five-year $200B commitment is Reuters reporting citing The Information. Google's continuing Vera Rubin offer is from its Cloud Next announcement. Nvidia's guarantees are disclosed in its Q2 FY2027 commentary. All ratios, annualisations, per-share bounds and sensitivities are R40 arithmetic; model values are assumptions, not guidance.
