First Amazon built Trainium. Then Google built TPUs. Microsoft went with custom silicon. OpenAI reportedly explored acquisitions. Now Anthropic is in the game—early-stage work on a custom AI server chip, with Samsung in the mix as a potential manufacturer.
This isn't ambition run amok. It's arithmetic. When you're spending billions annually on inference and training, the gap between "good enough" GPUs and "optimized for our models" silicon is measured in hundreds of millions of dollars a year. Nvidia takes a healthy margin on every H100. At Anthropic's scale, that margin justifies the NRE (non-recurring engineering) cost of going custom.
The Vertical Integration Play
What makes this interesting isn't the chip itself—it's what it signals about Anthropic's strategic posture. They just settled a fraught dispute with the US government over Fable 5 and Mythos 5 model releases, agreeing to proactively detect and address security risks. They account for 43% of all H1 2026 VC funding in AI (alongside OpenAI, $217B of a $510B total). They're no longer a startup scrambling for compute—they're an infrastructure company.
When you're spending billions annually on inference, the gap between "good enough" GPUs and custom silicon is measured in hundreds of millions per year.
The Samsung partnership talks are notable too. It suggests Anthropic is thinking beyond "we'll design it and TSMC will build it"—they're exploring alternative supply chains, which in 2026 is it own kind of strategic moat.
What This Means for the Field
If you're not building your own silicon at this scale, you're leaving money on the table. But more importantly, the AI infrastructure layer is consolidating. The winners won't just be the model developers—they'll be the ones who control the full stack from silicon to API.
OpenAI has the partnership with Microsoft. Google has TPUs. Amazon has Trainium and Inferentia. Anthropic is now joining the club. The moat isn't just the model—it's the hardware underneath it.
The question is whether custom silicon matters when model architecture is changing every 6-12 months. My read: it matters more for inference (where the workloads are more stable) than training. But at Anthropic's current trajectory, they'll need both.