Anthropic Hits $30B Revenue, Signs Multi-Gigawatt TPU Deal with Google and Broadcom

In September 2025, Broadcom CEO Hock Tan dropped a bombshell during an earnings call: a single unnamed customer had placed a $10 billion order for custom TPU racks. The disclosure set off a guessing game across Wall Street and Silicon Valley. Three months later, Broadcom confirmed what many had suspected — the mystery buyer was Anthropic, and the relationship had already grown to $21 billion in total commitments.

What makes this arc remarkable is how quickly it has escalated. In barely six months, Anthropic has gone from an anonymous line item in Broadcom's financials to the company's defining customer relationship. Mizuho analyst Vijay Rakesh now projects $21 billion in Broadcom AI revenue from Anthropic in 2026 alone, doubling to $42 billion in 2027. Bloomberg Intelligence frames the broader opportunity at $40-$50 billion. For a company whose entire AI revenue trajectory was questioned by skeptics just a year ago, Anthropic has become Broadcom's proof point — and its biggest concentration risk simultaneously.

The most striking detail in this announcement may not be the dollar figures but the unit of measurement. The deal is denominated in gigawatts — 3.5 GW of TPU compute capacity starting in 2027, expanding from 1 GW currently being supplied. To put this in perspective, 3.5 gigawatts is roughly the output of three large nuclear power plants, or enough electricity to power a city of over two million people.

This signals a fundamental shift in how AI infrastructure deals are structured and understood. We have moved from counting GPUs and chips to counting power consumption as the binding constraint. The implication is that AI compute scaling is no longer limited primarily by chip supply or capital — it is increasingly constrained by energy infrastructure. Anthropic's willingness to commit to this scale also reveals something about its internal demand forecasts: the company is betting that customer growth will continue at a pace that justifies power-plant-scale compute buildouts years in advance.

Anthropic's $30 billion revenue run rate now exceeds OpenAI's approximately $24 billion, a reversal that would have seemed implausible even six months ago. The speed of the shift is as notable as the milestone itself: Anthropic was at $9 billion at year-end 2025, $14 billion in February, and $30 billion by early April. Its $1M+ enterprise customer base doubled from 500 to over 1,000 in under two months.

However, run rate is a snapshot metric — it annualizes the most recent period's revenue and projects it forward. It does not account for churn, seasonality, or whether growth will sustain. Anthropic is still a private company and does not disclose profitability, margins, or actual trailing revenue. The $30 billion figure tells us that demand for Claude is extraordinary right now, but it says less about whether this pace is durable.

Perhaps the most strategically significant dimension of this news is what it reveals about the evolving chip landscape. Anthropic trains and serves Claude across three distinct hardware platforms: Google TPUs (via Broadcom), Amazon Trainium (via AWS's Project Rainier with approximately 500,000 Trainium 2 chips), and Nvidia GPUs. This makes Anthropic the only frontier AI model provider operating across all three major silicon architectures.

This multi-vendor strategy serves multiple purposes. It reduces dependence on any single supplier — particularly Nvidia, whose GPU dominance has given it extraordinary pricing power. It also creates competitive leverage, as each cloud provider has incentive to offer favorable terms to keep Anthropic's workloads on their platform. The scale of the TPU commitment — 3.5 GW — suggests Anthropic sees custom silicon as increasingly competitive with Nvidia's general-purpose GPUs for its specific workloads.

Buried in Broadcom's announcement was a notable caveat: "The consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic's continued commercial success." This is not standard boilerplate. It is an explicit acknowledgment that the deal's value is contingent, not guaranteed — making it more of a capacity reservation than a firm purchase order.

This caveat matters because it exposes the reflexive nature of the AI infrastructure boom. Broadcom's stock rose 3-4% on the news, and analysts are projecting tens of billions in AI revenue through 2031. But those projections rest on the assumption that Anthropic's revenue growth will not only sustain but accelerate enough to justify consuming 3.5 GW of compute. If enterprise AI spending cools, if a competitor leapfrogs Claude, or if Anthropic's margins prove unsustainable at scale, the capacity commitment could shrink. Bank of America analyst Vivek Arya focused on the positive — the deal removes the overhang of Google insourcing fears — but the demand-dependent structure means investors are ultimately making a leveraged bet on Anthropic's future, not just Broadcom's.