Biren Technology grew revenue by nearly 2,000% year-over-year in the first half of 2026. Biren Technology also expects that half-year revenue to land somewhere starting at 1.15 billion yuan — a figure that would barely register as a rounding error in Nvidia’s quarterly data center line. Both facts are true at once, and holding them together is the only honest way to read what is happening inside China’s accelerator market.
I spend most of my time on agent architectures: inference graphs, tool-calling loops, KV-cache behavior under long-horizon planning. So my interest in Biren is not really about who wins a market share chart. It is about which hardware the next generation of agent runtimes gets written against, because that decision tends to calcify for a decade.
Percentages are a function of where you started
A 2,000% jump sounds like a phase change. Mechanically, it is what happens when a small company with a small base suddenly becomes the default option in a large market. Nvidia and AMD have exited the high-end China market. Export controls removed the incumbent. Demand did not go anywhere. Someone had to fill the socket.
The volume picture backs that up. DIGITIMES estimates Chinese vendors will ship 2.123 million high-end cloud AI accelerators in 2026, up 136% year-on-year. That is a real market with real racks behind it, not a subsidy mirage. Biren, which listed on the Hong Kong exchange in January, is riding a demand curve it did not have to create.
Analyst projections stretch further out: operating revenue of $4.09 billion, $13.39 billion, and $27.38 billion across 2026 through 2028, with growth rates of 165%, 228%, and 104%. I would treat those numbers with the skepticism any researcher applies to a three-year hockey stick, especially given the distance between a 1.15 billion yuan half-year and a four-billion-dollar annual figure. Sell-side models for newly listed hardware firms are directional at best. The direction, though, is not ambiguous.
Huawei set the performance floor, which changes the argument
In March 2026, Huawei introduced the Atlas 350, a single chip card rated at 1.56 quadrillion calculations per second — close to three times the performance of Nvidia’s best China-legal part. That comparison matters less as a benchmark than as an end to a specific conversation. The old assumption was that domestic silicon was a compromise you accepted for supply reasons. Against the export-compliant Nvidia SKUs specifically, that framing no longer holds.
Which pushes the entire competition into software. And software is where agent workloads get interesting.
Agent inference stresses different silicon than training does
Training is a FLOPS and interconnect problem. Agentic serving is not. An agent doing multi-step reasoning with tool calls generates a long, latency-sensitive sequence of short forward passes, each dragging a growing KV cache behind it. The performance ceiling is set by memory bandwidth, cache paging strategy, scheduler behavior under bursty concurrency, and how well the runtime handles speculative decoding and prefix reuse across sibling agent branches.
Almost none of that lives in the chip spec sheet. It lives in kernels. Paged attention implementations, fused attention variants, quantized matmul paths, graph capture to kill per-step launch overhead, continuous batching schedulers — every one of these has been tuned for years against one vendor’s architecture by thousands of engineers who never had to think about portability.
That is the actual moat, and it is why a 2,000% revenue figure does not automatically translate into a competitive agent stack. A domestic accelerator can match peak throughput on paper and still lose badly on a ReAct loop with 32k context because the attention kernel falls off a bandwidth cliff at a sequence length nobody profiled.
What I would watch instead of revenue
If you want to know whether this shift is durable, ignore the growth rate and track these:
- Whether major open-source inference servers merge first-class backends for domestic accelerators, maintained by people who are paid to keep them working.
- Time-to-support for new attention variants. When a new sparse or linear attention scheme lands in a paper, how many weeks until it runs efficiently on non-Nvidia hardware in China?
- Multi-node behavior for agent fleets, where thousands of concurrent sessions with heterogeneous context lengths hammer the scheduler and the interconnect together.
- Tail latency under load, not median throughput. Agent chains multiply p99 delays across every step.
Export controls handed Biren and its peers something money cannot buy: a captive market with no incumbent and enormous urgency. That gets you the revenue chart. What it does not get you is a mature kernel ecosystem, and the second problem is harder and slower than the first.
The interesting outcome is not that China builds competitive accelerators. That question is largely settled. It is that a second, independent agent-serving stack now has both the demand and the funding to exist. Two divergent software ecosystems, each optimized for different silicon, is a more consequential development for anyone building agents than any single quarter’s growth number.
🕒 Published: