China just built a major AI model without Nvidia chips. Now OpenAI has found ways to run on far fewer of them, cutting inference costs by more than half. Even so, Nvidia stock rose.

That is the puzzle. OpenAI is one of Nvidia’s (NVDA) biggest customers. Yet the shares climbed even as it moved to need fewer chips.

OpenAI Cuts Inference Costs on Two Fronts

The first front is software. The Information reported that OpenAI engineers cut inference costs by more than half with new optimization methods. OpenAI has not published the technical details.

The savings reduce the number of Nvidia chips needed to handle some ChatGPT traffic. They could also let OpenAI lower prices or raise usage limits.

The second front is hardware. On June 24, OpenAI and Broadcom (AVGO) unveiled Jalapeño, its first custom chip. OpenAI said early tests point to far better performance per watt than today’s leading chips, with a nine-month design.

The first chips will deploy at a gigawatt scale by the end of 2026, with Microsoft as the lead partner. Nvidia still runs most of OpenAI’s inference, even as OpenAI funds its Broadcom chip partnership.

Big Tech Races to Build Its Own Chips

OpenAI is not alone. Google has built tensor processing units since 2016, and Amazon followed with its own. Research firm TrendForce projects ASIC-based systems will reach 27.8% of AI server shipments in 2026, the highest since 2023.

By TrendForce’s count, custom chips are set to grow faster than Nvidia’s GPUs for the first time. Suppliers like Broadcom and Marvell have become key custom chip makers in the build-out.

Sanctions are pushing the same trend in China. Meituan recently trained its 1.6 trillion parameter LongCat-2.0 model on China’s domestic chips, without any Nvidia hardware.

Why Nvidia Stock Keeps Rising

The threat is real, but the numbers explain the calm. Nvidia stock rose nearly 2% on June 30, near a $4.8 trillion value. Nvidia’s latest results showed data-center revenue up 75% to a record $62.3 billion in a single quarter.

Most of the pressure sits at inference, not training. Nvidia still dominates model training, where its CUDA software has locked in developers since 2006. Custom chips rarely match that flexibility.

Nvidia is also defending the inference layer it is accused of losing. At GTC, Nvidia said its upcoming Rubin platform cuts inference costs per token by up to 10 times compared to Blackwell. Cheaper inference also tends to lift usage and total compute with it.

Not everyone is convinced. Some investors have rotated into rival chip stocks, betting the inference shift compounds. Yet Nvidia guided to this quarter without counting any China sales, and still sees record demand.

Nvidia still sells every chip it can make. The real test is whether its biggest customers can cut it out faster than the market grows.

Source: BeInCrypto