China just built a major AI prototype without Nvidia chips. Now OpenAI has found ways to run on much fewer of them, cutting inference costs by more than half. However, Nvidia stock rose.
This is the puzzle. OpenAI is one of Nvidia’s (NVDA) largest customers. However, shares rose even as the need for fewer chips moved.
OpenAI reduces inference costs on two fronts
The first front is software. Information I mentioned OpenAI engineers cut inference costs by more than half using new optimization methods. OpenAI has not published technical details.
The savings reduce the number of Nvidia chips needed to handle some of the ChatGPT traffic. They could also allow OpenAI to lower prices or raise usage limits.
The second front is hardware. On June 24, OpenAI and Broadcom (AVGO) unveil Jalapeno, her The first custom chip. OpenAI said early tests indicate much better performance per watt than today’s leading chips, with a nine-month design lead.
The first chips will be deployed on a gigawatt scale by the end of 2026, and Microsoft will be the lead partner. Nvidia still runs most of OpenAI’s inferences, even as OpenAI funds it Broadcom chipset partnership.
Big tech companies are racing to build their own chips
OpenAI is not alone. Google has been building tensor processing units since 2016, and Amazon has followed suit with its own unit. Research company TrendForce Projects ASIC-based systems will reach 27.8% of AI server shipments in 2026, the highest percentage since 2023.
According to a TrendForce count, custom chips are set to grow faster than Nvidia GPUs for the first time. Suppliers have become Broadcom and Marvell Major custom chip makers In construction.
Sanctions are pushing the same trend in China. Meituan recently trained a LongCat-2.0 model containing 1.6 trillion parameters Local china chipswithout 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 value of $4.8 trillion. The latest from Nvidia results Data center revenue showed a 75% increase to a record high of $62.3 billion in one quarter.
Most of the pressure is in inference, not training. Nvidia still dominates model training, with its CUDA software securing developers since 2006. Custom chips rarely match this flexibility.
Nvidia also defends the inference layer it is accused of missing. At GTC, Nvidia said its upcoming Rubin platform reduces per-code inference costs by up to 10 times compared to Blackwell. Cheaper heuristics also tend to increase usage and overall compute with it.
Not everyone is convinced. Some investors have turned to… Competitor chip stocks,Betting on transformational inference compounds. However, Nvidia headed into the quarter without counting any sales in China, still seeing record demand.
Nvidia is still selling every chip it can make. The real test is whether its largest customers are able to divest from it faster than the market grows.
this post After China, OpenAI dumps Nvidia: So why is NVDA stock up? appeared first on BeInCrypto.
