Anthropic has finalized a substantial compute expansion deal with Google and Broadcom, securing 3.5 gigawatts of processing capacity slated for activation in 2027. This commitment, the AI lab's largest to date, coincides with a tripling of its annual revenue. The agreement underscores a trend towards an AI oligopoly, where control over compute infrastructure, rather than model innovation, will dictate market dominance.
The deal substantially increases Anthropic's utilization of Google Cloud's tensor processing units (TPUs), building upon a prior agreement. Industry sources indicate the 3.5 gigawatt capacity is equivalent to the power consumption of hundreds of thousands of homes, with the majority housed in the United States.
This infrastructure surge is fueled by rapid commercial growth, with Anthropic's annualized revenue reportedly reaching $30 billion, a significant jump from $9 billion just over a year ago. The company boasts over 1,000 enterprise clients, each spending more than $1 million annually on its Claude models.
The consolidation dynamic is driven by compute access, which has become the primary gatekeeping mechanism in the AI sector. Anthropic's long-term, large-scale compute commitments represent a strategic land grab, effectively excluding competitors unable to secure similar capacity. This creates deepening dependencies and reinforces the positions of major players like Google Cloud and chip designer Broadcom, forming a vertically integrated, closed system.
Startups entering the market now face a landscape where prime compute resources are already claimed, making competition physically impractical. The immediate beneficiaries are Anthropic, Google Cloud, and Broadcom. Mid-tier labs and startups without comparable infrastructure deals will find it increasingly difficult to develop competitive frontier models. Geopolitically, this compute capacity is becoming a strategic national resource, granting leverage to companies that control it.
The AI industry is entering a phase defined by control over physical resources. Unlike ephemeral benchmark leadership, a 3.5-gigawatt infrastructure commitment signifies durable competitive advantage. This mirrors historical patterns in cloud computing and search, with AI now following a similar trajectory at an accelerated pace and with greater physical infrastructure constraints.