Why every AI giant is suddenly building its own chips

In the past eighteen months, eight of the biggest names in AI have all started doing something that used to be unthinkable for a software company: designing their own chips. OpenAI, Anthropic, Google, Amazon, Microsoft, and IBM on one side of the Pacific; DeepSeek and Zhipu on the other. Not one or two of them, not a slow migration — a simultaneous rush. The question worth asking isn't whether they'll pull it off. It's why all of them moved at once. The demand side is a wall, not a preference. A single AI query draws roughly ten times the electricity of a plain web search ^[1], and flagship Nvidia servers pull up to 1,200 watts under load ^[2]. A Meta engineer put the problem in terms no amount of clever software can dodge.

It's a fundamental physical problem. — Meta

That is the constraint driving the rush: inference cost and power are physical problems, and no optimization of the model itself makes a watt cheaper. The inventory is the point. OpenAI's Jalapeno, co-designed with Broadcom, is an inference-only chip that runs at 700 watts and cuts inference cost roughly in half ^[3]. Amazon's Project Rainier is an $11 billion cluster of half a million proprietary Trainium chips, claiming 30 to 40 percent better price performance than Nvidia ^[4]. Microsoft's Maia 200, its second-generation accelerator, runs GPT-5.2 at 30 percent better performance per dollar ^[5]. Google, which trained Gemini 3 entirely on its own TPUs, is now selling them outward — a multibillion-dollar deal to supply Meta and up to a million chips to Anthropic ^[6]. Anthropic is building an in-house silicon team to co-design hardware and software around Claude, while keeping AWS, Google, Nvidia, and AMD as suppliers ^[7]. IBM hardcoded the Arm instruction set into its mainframe processor so enterprises can run AI inference on the same silicon as their core workloads ^[8]. And in China, DeepSeek and Zhipu are building their own chips for a different reason — export controls cut them off from Nvidia — but landing in the same place ^[9][10]. None of this would be happening at once if the barrier to entry were still what it was. Chip design used to be a multi-year, multi-billion-dollar moat — the reason software companies bought chips instead of making them. That moat is collapsing, and the thing collapsing it is the same AI that needs the chips. TSMC's AI design tools now solve tasks that took human engineers two days in five minutes ^[2]. Samsung cut chip verification from over a month to two days using Anthropic's Claude Code ^[11]. Cadence's AI agent delivers up to 10x productivity, and its chief says the company will move from licensing tools to renting virtual engineers.

ChipStack represents a major leap in our design-for-AI and AI-for-design strategy, applying agentic AI directly to our customers’ front-end flows to tackle the growing complexity and scale of modern chips. — Anirudh Devgan

OpenAI designed Jalapeno in nine months using its own models to accelerate the work ^[12]. The loop is the story. The same AI that created the demand for custom silicon is now being used to design it. The competition has shifted from whose model is smartest to whose silicon runs it cheapest — and AI is the engine on both sides of that shift. None of this means Nvidia is dying. Its revenue grew 85 percent to $81.6 billion, CUDA still holds 900-plus libraries, and custom silicon is additive rather than a replacement — AI ASIC shipments are growing 44.6 percent a year against 16.1 percent for GPUs, but the total capex pie, around $725 billion in 2026, is expanding fast enough for both ^[13]. The tell is subtler: Nvidia is now financing its own customers, including a $105 billion lease guarantee for OpenAI's Ohio data center ^[14]. Pricing power propped up by financial engineering is a different thing than pricing power earned by product. The barrier that kept software companies out of hardware was time, money, and expertise. AI is eroding all three at once. The loop is tightening, not widening — and the next chip war will be fought by companies that, a year ago, had never designed one.