Anthropic has opened the latest front in AI’s escalating power struggle: the chips beneath the models. The Claude maker is assembling a custom-silicon team to co-design hardware and software, saying the pairing should make its systems run “faster and more efficiently.”¹

The move follows a familiar calculation. As demand for Claude grows, Anthropic has secured computing arrangements with AWS, Google, Nvidia and AMD—but access to outside capacity is not the same as owning the roadmap. OpenAI introduced its Broadcom-built Jalapeño inference chip in June; Google DeepMind relies on Alphabet’s TPUs; and Meta has been developing MTIA accelerators. Anthropic is now joining the club rather than remaining a customer of it.¹

The engineering logic also has boosters inside the industry. OpenAI co-founder Greg Brockman called “ai for chip design” underrated, amplifying discussion of using AI itself to speed hardware-development loops.Greg Brockman (gdb)²

Then the contest shifts from the data center to the consumer. During a hectic week of releases—Anthropic’s Opus 5.5 followed 90 minutes later by OpenAI GPT-6 updates—Meta’s personal agent Muse reportedly seized attention with early growth that outpaced ChatGPT’s. Meta’s bet is that AI becomes habitual through smart glasses and a small Tamagotchi-style device, not merely another chatbot tab.³

Finally, the rivalry is heading into diplomacy. Trump and Xi are expected to discuss Silicon Valley’s allegations that Chinese companies, including Alibaba, use “distillation” to copy advanced U.S. AI technology—an accusation Beijing’s firms face as they build competing systems. Trump said AI, which he calls “super intelligence,” would be among the summit’s pressing issues, while signaling no imminent change in domestic regulation despite calls for tougher guardrails.⁴

The through line is control: Anthropic wants more control over compute, Meta over the consumer interface, and Washington over the technological boundary with China.