Export controls meant to slow China’s AI progress just handed a Shanghai chipmaker a 2,000% revenue jump.
That number belongs to Biren Technology, a company most people outside the semiconductor world have never heard of. In the first half of 2026, its revenue grew nearly 2,000% year over year. Not 20%. Not 200%. Two thousand.
If you’re here because you want to understand AI agents rather than chip fabrication, stick with me. This story matters to you more than it looks, and I’ll explain why by the end.
What Biren actually sells
Biren makes AI accelerators. These are the specialized chips that do the heavy math behind every AI system you interact with, including the agents that book your appointments, summarize your inbox, or write code alongside you. Regular computer processors can technically do this work, but they’re slow at it. Accelerators are built specifically for the kind of repetitive number-crunching that neural networks demand.
For years, one company dominated this category almost entirely: Nvidia. AMD held a distant second place. If you were building AI anywhere in the world, you bought from them.
Biren debuted on the Hong Kong stock exchange in January 2026 and estimated revenue for the first six months of the year between 1.15 billion yuan and a higher ceiling, which represents up to a 22-fold surge. Analysts covering the company project operating revenue of $4.09 billion in 2026, $13.39 billion in 2027, and $27.38 billion in 2028, with growth rates of 165%, 228%, and 104% respectively.
Those are projections, not results. Analyst forecasts get revised constantly, and a company growing this fast from a small base can miss badly. But the direction is hard to argue with.
The part that’s genuinely counterintuitive
Export controls were designed to restrict China’s access to the most capable AI chips. The logic seemed sound: if you can’t buy the best hardware, you can’t train the best models.
What happened instead is that Nvidia and AMD exited the market, and the demand didn’t disappear with them. Chinese companies still needed accelerators. Chinese data centers still needed to be filled. So the money that used to flow to American chipmakers started flowing to domestic ones instead.
DIGITIMES estimates Chinese vendors will ship 2.123 million high-end cloud AI accelerators in 2026, a 136% year-on-year increase. Huawei is a major part of that, using its strength in system integration to package chips into complete systems. In March 2026, Huawei unveiled the Atlas 350, a single chip card delivering 1.56 quadrillion calculations per second, roughly three times the performance of Nvidia’s best China-legal chip.
Read that again. The best chip Nvidia was permitted to sell into China ended up slower than what a Chinese company built on its own.
Why this lands on your desk
Here’s my honest read as someone who spends her time explaining AI agents to people who don’t write code.
Most conversations about AI agents happen at the top of the stack. Which assistant is smartest. Which tool connects to your calendar. Which one hallucinates less. Almost nobody talks about the silicon underneath, because it feels like plumbing.
But the plumbing determines the price. Every agent that runs on your behalf, every request it makes, every step it takes to complete a task, burns compute. Compute costs money. And compute prices are set by how many companies are competing to sell you accelerators.
For years, that answer was basically one company. If a second and third serious supplier emerge, even in a market you don’t directly buy from, the pricing pressure eventually travels. Not immediately, and not evenly, but it travels.
There’s a second effect that’s less comfortable. A separate Chinese hardware ecosystem means separate software built to run on it. Over time, that can mean AI agents that behave differently, are trained differently, and answer to different rules depending on which side of the divide they were built on. Two AI worlds instead of one shared one. For anyone who cares about interoperability, that’s a real cost.
The lesson underneath the number
Restricting access to a technology creates urgency to build it locally. That’s not a new observation, but 2,000% is an unusually loud way to demonstrate it.
Biren is still small. A huge percentage gain on a modest starting figure is not the same as market leadership, and I’d be skeptical of anyone framing this as Nvidia’s obituary. Nvidia remains the dominant supplier in most of the world.
What changed is the number of credible answers to the question “who can build a competitive AI accelerator.” That number used to be very close to one. It isn’t anymore. Whatever the policy intent, that outcome is now permanent, and it will shape the cost and character of the AI tools sitting on your laptop for years.
🕒 Published: