AI · Semiconductors

Why AI Labs Are Suddenly Building Their Own Chips

DeepSeek, Zhipu, OpenAI and Anthropic have all moved into custom silicon this year. The shift from renting GPUs to designing dedicated inference chips is redrawing the map of who controls the economics of artificial intelligence.

For most of the AI boom, the recipe looked simple: train a big model, then rent thousands of GPUs from Nvidia to serve it. That arrangement is now cracking. As inference — actually running models for users — starts to eat 30% to 50% of AI revenue, the companies behind the models have decided the chip underneath them is too important to leave to someone else.

From software to "full stack"

The key shift: algorithms are now defining hardware

OpenAI has already shipped its first self-designed inference chip. Anthropic has been reported in talks with Samsung to co-develop an ASIC purpose-built for its Claude models. In China, DeepSeek is developing its own AI silicon aimed at inference, while Zhipu has opened technical discussions with domestic chip designers to build a dedicated processor for its industry models. The common thread: instead of fitting an algorithm to whatever chip is available, these labs want a chip shaped around their own model's math.

Three things worth knowing

Inference, not training, is the new battleground: training happens once; serving a model to millions of users happens forever. When inference becomes the dominant cost, a chip tuned to that exact workload can cut both the bill and the wait.
Different industries "make chips" for different reasons: phone makers build on-device SoCs with NPUs to balance battery and experience; carmakers build safety-grade driving chips; AI labs build cloud inference ASICs to cut cost and lock in supply. Same word, very different strategies.
Capital is the gatekeeper: DeepSeek's reported first funding round exceeded 50 billion yuan, among the largest single rounds in Chinese AI history. Building fabs-adjacent custom silicon is only possible once a lab has that kind of war chest to fund both the model and the hardware.

Why it matters beyond one company

The move is a bet that AI's center of gravity is shifting from "who has the biggest model" to "who controls the cheapest, most reliable way to run it." Nvidia's GPUs remain the default, but every lab designing its own inference chip is quietly building a second option — one that reduces dependence on a single supplier and tailors the hardware to its own software.

If the trend holds, the next decade of AI competition may be decided less by benchmark scores and more by who owns the silicon their intelligence runs on. The labs that treat chips as a core competence, not a commodity purchase, are positioning themselves for exactly that world.