Azarudeen Mohamed Ali @Azarudeen17: "📈 Anthropic has a demand..."

United Arab Emirates

📈 Anthropic has a demand problem, there's too much: Anthropic just announced it will access multiple gigawatts of next-generation tensor processing unit (TPU) capacity starting in 2027, through a new long-term deal with Google and Broadcom. The company's run-rate revenue has crossed $30 billion, up from $9 billion at the end of 2025, fueled by enterprise adoption of Claude Code and the seep of Claude Cowork into workflows. Breaking down the buzz: Broadcom filed a regulatory disclosure today confirming a long-term agreement to develop and supply custom TPUs for future generations of Google's chips, and separately that Anthropic will access approximately 3.5 gigawatts of next-generation TPU-based compute through Broadcom starting in 2027. This follows an earlier $21 billion order Anthropic placed with Broadcom across two waves in fiscal 2025, covering around one million Google TPU Ironwood racks for delivery by late 2026. The infrastructure story is now in two acts: near-term capacity already locked in, and the next generation already committed before it exists. Let's understand how a TPU differs from Nvidia graphics processing unita (GPUs). Google's TPUs are application-specific integrated circuits (ASICs), purpose-built for the dense matrix operations that dominate AI training runs and large-scale inference. Analysts at SemiAnalysis estimate the total cost of ownership for Ironwood TPUs deployed internally at Google runs roughly 44% lower than an equivalent Nvidia system, and that even at external customer pricing with Google's margin on top, TPUs can still deliver around 30% lower cost than Nvidia's current flagship graphics processing unit GPU platform. Nvidia GPUs, by contrast, offer broad programmability across workload types, which is why Anthropic keeps them in the stack; not every workload tolerates the architectural rigidity of an ASIC. Anthropic spreads workloads across Google TPUs, Amazon's custom Trainium chips and Nvidia GPUs, matching each to the platform best suited for the task. Why this matters: The demand driving this infrastructure build came from a culture war. When the Pentagon labeled Anthropic a supply-chain risk after the company refused to sign a contract without explicit prohibitions on autonomous weapons and mass domestic surveillance, Claude surged to the top of the Apple App Store as users switched in protest. Altman acknowledged in an internal memo that OpenAI's decision to announce its own Pentagon deal hours after Anthropic was blacklisted "looked opportunistic and sloppy," and the company subsequently renegotiated its contract to add clearer surveillance restrictions. The episode handed Anthropic a reputational tailwind that boosted revenue, as more than 1,000 business customers now spend over $1 million on Claude annually, 2x the February figure. Layer Claude Cowork's seep into enterprise workflows on top of that, and the demand signal became impossible to plan around conservatively. Anthropic is routing its highest-volume, most cost-sensitive workloads toward TPUs, which is precisely the slice of spend where Nvidia's pricing power has historically been strongest. As power constraints rather than chip supply emerge as the key bottleneck for AI infrastructure, the cost-per-watt efficiency of custom silicon becomes a structural argument rather than a procurement preference. Total compute demand is growing fast enough that Nvidia likely keeps selling everything it makes through the near term Copy my portfolio for steady and consistent growth, $NVDA (NVIDIA Corporation)

Not investment advice. The author may have financial interests in the mentioned instruments.

NVDA

NVIDIA Corporation

229.85

-5.89 (-2.50%)

1 Mentioned

NVDA

NVIDIA Corporation

Strictly Necessary Cookies

Performance Cookies

Functional Cookies

Targeting Cookies