China has just built a large AI model without using Nvidia chips, and now OpenAI has found a way to reduce the use of those chips by more than half. This cuts inference costs by more than half. Even so, Nvidia’s stock price is still rising
This is the puzzle: OpenAI is one of Nvidia’s biggest customers (NVDA), yet the stock price is rising even as demand for the chips is being reduced.
The first is software. The news outlet The Information reported that OpenAI engineers have reduced inference cost by more than half through new optimization approaches. OpenAI hasn’t disclosed technical details yet.
These savings help reduce the number of Nvidia chips needed to support some of ChatGPT’s usage, and could also allow OpenAI to lower prices or limit usage more efficiently.
The second is hardware. On June 24, OpenAI and Broadcom (AVGO) unveiled their first custom chip, Jalapeño. OpenAI said early testing shows better performance per watt than the leading chips currently available, with just nine months of design.
We’ve designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products. Chips are foundational to the AI… pic.twitter.com/mHU7DaMMTi
— OpenAI (@OpenAI) June 24, 2026
The first chips will be deployed at the gigawatt scale by the end of 2026. Microsoft is the key partner. Nvidia will continue to handle most of OpenAI’s inference, even though OpenAI is investing in chip-related collaboration with Broadcom.
Big Tech is accelerating the development of its own chips
OpenAI isn’t alone. Google has developed Tensor Processing Units since 2016, and Amazon has also developed its own. Research firm TrendForce estimates that systems using ASICs will account for 27.8% of AI server shipments in 2026—the highest share since 2023.
According to a TrendForce survey, custom chips are growing faster than Nvidia GPUs for the first time. Suppliers such as Broadcom and Marvell have become major custom chip manufacturers, helping drive this infrastructure shift.
Sanctions are also pushing the same trend in China. Recently, Meituan trained its LongCat-2.0 model—1.6 trillion parameters—on chips in China without using any Nvidia hardware at all.
The reason Nvidia’s stock keeps climbing
The threat is real, but the numbers offer some comfort. Nvidia shares rose nearly 2% on June 30, valuing the company at about $4.8 trillion. The latest earnings report from Nvidia shows that data-center revenue increased 75% to a new quarterly record of $62.3 billion.
Most of the pressure will be on usage (inference) rather than training models. Nvidia still dominates the model-training share, with CUDA software locking in developers since 2006. Most custom chips still can’t match this flexibility.
Jensen Huang is basically saying custom chips from $GOOGL, $AMZN, $AMZN, $MSFT & $META struggle to compete with $NVDA because Nvidia is singularly focused on AI acceleration at a scale no one else matches. CUDA is the default language developers build on, NVLink is still ahead… pic.twitter.com/KQYQT8CWYO
— Shay Boloor (@StockSavvyShay) February 1, 2026
Nvidia is also defending its position in inference processing, which it’s been accused of losing as well. At GTC, Nvidia said its upcoming Rubin platform could cut inference cost per token by as much as 10x compared with Blackwell. Lower inference costs often lead to more usage and higher total compute volume as a result.
But some people are still not convinced. Some investors have shifted to holding competing chip stocks, expecting that the inference transition trend will accelerate even more. However, Nvidia set this quarter’s target without counting sales in China, and it still sees unprecedented demand.
Nvidia still sells all the chips it can produce. The real test is whether the company’s biggest customers will decide to stop using Nvidia faster than the market’s growth rate.