Who Gets the TPU?

Programming note: we will return on Friday and take a look at Huawei's chip design strategy.

Google's Compute

Google should be the last company in the world to run short of AI compute.

It has its own cloud, its own data centers, and its own custom AI chips, known as Tensor Processing Units (TPUs). For a decade, Google had been building what every AI company now wishes it had: a private compute stack. The TPU was a vertically integrated hedge against Nvidia dependence, cloud scarcity, and the unpleasant experience of begging someone else for server time.

And yet, during Alphabet's earnings call last month, CEO Sundar Pichai had to explain to investors why Google Cloud was leaving billions of dollars on the table:

We see strong demand for infrastructure in Google Cloud. As I said earlier, in select cases, we are seeing demand for TPU hardware and others, data centers as well. We are modeling these out and working to allocate across these areas.

Obviously, we are compute constrained in the near term, and as an example, our cloud revenue would have been higher if we were able to meet the demand.

It is a remarkable admission. Google Cloud's backlog—the measure of contracted-but-unrecognized work—nearly doubled sequentially to a staggering $462 billion. There is a literal queue of enterprise customers standing outside the door, waving suitcases of cash, waiting for chips that do not exist yet.

And Wall Street has priced this queue directly into the stock. A year ago, Google traded at a relatively modest 19x P/E multiple, haunted by the narrative that generative AI would destroy its search monopoly. Today, it is a $4 trillion company trading near 29x P/E.

A significant part of that massive valuation re-rating is pure optimism over Google's TPU—the belief that vertical integration gives Google a durable advantage in a capacity-constrained market.

But there is a catch. The problem with building an incredibly successful alternative to Nvidia's chip monopoly is that once you prove it works, everyone wants to use it.

Every TPU Has Multiple Suitors

Historically, the TPU was Google's internal advantage. It was the unglamorous plumbing that powered Google Search, ran YouTube’s recommendation engine, and allowed Google DeepMind’s researchers to pursue ambitious, long-term scientific projects.

But today, the TPU is hopelessly oversubscribed, both internally and externally. The line of claimants is long, and every single one of them has a rational, high-stakes business case for why they deserve the next block of silicon:

• Gemini wants them because model quality is the foundation of Google's AI strategy.
• Search wants them because AI Overviews are defending the core search monopoly.
• Google Cloud wants them because paying enterprise customers are waiting, and that $462 billion backlog needs to be recognized as revenue.
• DeepMind wants them because long-term research is how Google invented the transformer in the first place.

On top of that, external partners like Anthropic have signed deals committing to use up to one million TPUs.

As veteran AI researcher Oren Etzioni told Bloomberg: "Inside Google, every TPU has three suitors. If you find yourself in the uncomfortable position where you have a pie-in-the-sky project and you are competing with a revenue-yielding customer, that's a tough position to be in."

The most striking evidence of this internal turf war comes from Pichai himself. He disclosed in a recent interview that he personally spends a dedicated hour each week reviewing compute allocation at the project and team level—looking at which teams are consuming which compute units. The CEO of Alphabet is not managing people or products in that hour. He is managing TPUs.

The Research Cost

This zero-sum math is quietly reshaping Google's legendary research culture. Historically, Google was the place researchers went to access the world’s best infrastructure with academic freedom. But in a constrained environment, speculative research has to compete directly against paying cloud customers.

According to Bloomberg, Google actually had to pause several internal research projects for an entire quarter in 2024 to free up compute for a major training run.

This explains the recent wave of high-profile researcher departures. Andrew Dai left after nearly 14 years at Google to found Elorian, focused on visual reasoning. David Silver—the star researcher behind AlphaGo and AlphaZero—left in late 2025 to found Ineffable Intelligence, which has since raised more than $1 billion in seed funding.

These researchers aren't leaving Google because startups have more chips; they are leaving because they want control over the chips they do have, without having to petition ten layers of corporate bureaucracy for server time.

Financializing the Silicon

To solve this bottleneck, Google has to build more capacity. But building data centers is an incredibly capital-intensive construction project that requires land, power, cooling, and massive upfront cash.

Alphabet's capital expenditures are already projected to reach a historic $180 billion to $190 billion this year, with CFO Anat Ashkenazi signalling that 2027 outlays will be "significantly" higher. But there is a limit to how much capex a public company can pile onto its balance sheet before Wall Street starts to panic about margins.

This explains the strategic logic behind Google's recent joint venture with private equity giant Blackstone. Google and Blackstone are partnering to launch a new, privately funded AI cloud company built entirely around Google's TPUs. Blackstone is committing $5 billion in initial equity, with plans to bring 500 megawatts of capacity online.

Why let a private equity firm buy your prized, highly proprietary chips and run them in a separate cloud? Because it turns the TPU from an internal capex burden into a financial asset.

By offloading the massive cost of building the actual physical data centers to Blackstone's balance sheet, Google can scale TPU access, satisfy its waiting cloud customers, and avoid scaring its own shareholders with an infinite capex ramp.

And Google is not alone in discovering this playbook. In October 2025, Meta signed a $27 billion financing deal with alternative asset manager Blue Owl Capital to fund Hyperion, its largest-ever data centre project in Louisiana, targeting more than two gigawatts of compute capacity. The structure was nearly identical: the tech company contributed the strategic asset, private capital provided the infrastructure balance sheet. AI compute is becoming an asset class.

The TPU began its life over a decade ago as an obscure internal hardware project designed to save Google money. Today, it is a global bottleneck, a source of internal political warfare, and the foundation of a multi-billion-dollar private equity joint venture. That trajectory—from internal cost-saving measure to global bottleneck to private equity asset class—is not what anyone planned.

It is the defining lesson of the AI boom: vertical integration does not eliminate scarcity. It just moves the scarcity inside the house.

Signal Stack

The operating reality beneath the headlines.

• Frontier Labs Don't Use Most AI Compute (Epoch AI) – OpenAI, Anthropic, and xAI together controlled under 30% of the world's operational AI compute at end of 2025, meaning the majority of the buildout is being consumed by players nobody is talking about.
• SpaceX-xAI Seeks More Compute Customers (Data Center Dynamics) – Elon Musk's AI lab is quietly pivoting from model developer to compute landlord, with orbital data centers the stated endgame.

📊 Data > Narrative

We pull key data points to show you the mathematical reality of what's happening in tech.

The Data: SoftBank shares jumped 49% since May 20, pushing the company's market cap above ¥45 trillion. The catalyst was a report that OpenAI is preparing to file for an IPO. The move was enough to push SoftBank past Toyota in market capitalization—excluding treasury shares—for the first time since 2000. SoftBank has committed close to $65 billion to OpenAI in total, targeting a roughly 13% stake by October.

The Takeaway: Toyota built the most valuable industrial company in Japanese history across a century of manufacturing. SoftBank surpassed it on the news that a company it owns a slice of might soon file paperwork. That inversion is the market's clearest statement yet about where it believes value creation is migrating.

You received this message because you are subscribed to ARPU newsletter. If a friend forwarded you this message, sign up here to get it in your inbox.