By ByteBot
China’s AI Chip IPO Wave Reaches Peak FOMO
Biren Technology is launching a $300 million Hong Kong IPO this month, becoming the third Chinese AI chip company to go public in two weeks. The timing tells the story: Moore Threads raised $1.13 billion and surged 425% on December 5. MetaX followed today with a 569% first-day pop. And Biren just received regulatory approval on December 15. This isn’t careful industrial policy—it’s a feeding frenzy.
Investors are treating domestic GPU makers like lottery tickets. Moore Threads’ IPO was oversubscribed 4,000 times, with retail investors getting a 0.036% allocation rate. You had better odds finding a 4090 during the crypto boom. MetaX’s oversubscription was even higher. And now Biren, valued at $2 billion pre-IPO, wants its piece of the mania.
But scratch beneath the hype and the fundamentals look questionable.
The Numbers Don’t Add Up
Moore Threads raised over a billion dollars, popped 425% on day one, then announced it would park 90% of the IPO proceeds in bank products instead of R&D. That’s not confidence in your technology—that’s a cash grab dressed as tech sovereignty.
The company is unprofitable. In the first half of 2025, Moore Threads posted 701 million yuan in revenue against 271 million yuan in losses. Now it’s sitting on 8 billion yuan in fresh capital that will mostly earn interest instead of funding chip development. When your burn rate suggests you don’t actually need the money for operations, what were investors buying?
Biren’s $2 billion pre-IPO valuation sounds impressive until you remember Nvidia’s market cap is $3 trillion. That makes Biren worth 0.067% of the incumbent it’s supposedly challenging. The David versus Goliath narrative works better when David has, you know, a working slingshot.
Biren’s BR100: Show Us the Benchmarks
Biren claimed in 2022 that its BR100 GPU matches Nvidia’s H100 performance. Great—where’s the proof? Three years later, we still don’t have independent benchmarks, peer-reviewed tests, or real-world deployment data. Just press releases and investor decks.
The claimed specs raise eyebrows. Biren says the BR100 delivers 256 TFLOPS of FP32 performance compared to the H100’s 60-67 TFLOPS. That’s a 4x advantage on paper. But the H100 has double the memory bandwidth (3.35 TB/s versus 1.6 TB/s), uses newer HBM3 memory, and actually ships to customers who can verify the numbers.
Hardware specs are marketing. The question developers actually care about: Can you train a GPT-4 scale model on it? Can you run inference at competitive speeds? Where’s the MLPerf score? Without public benchmarks, Biren’s performance claims have the credibility of a startup calling itself “the Uber of X” without showing revenue.
Export Controls Backfired
Washington’s strategy was straightforward: cut off China’s access to Nvidia’s H100 and A100 GPUs to slow their AI development. What actually happened? A $10 billion domestic investment wave, government backing for four competing GPU startups, and three IPOs in two weeks.
Before the U.S. tightened export controls in April 2025, ByteDance, Alibaba, and Tencent stockpiled over $5 billion worth of Nvidia’s H20 chips—a China-specific variant that itself was designed to comply with earlier restrictions. ByteDance alone ordered roughly one million H20 GPUs. When that supply line got cut, the same companies pivoted to backing domestic alternatives. ByteDance invested in Moore Threads. Tencent backed Enflame Technology. Alibaba started building custom chips that employees say match the H20’s performance.
The U.S. loses a $10 billion annual market. China gets inferior chips but maintains domestic AI development. Both sides lose, just differently. Export controls succeeded in fragmenting global AI infrastructure—whether that serves American strategic interests is debatable.
Software Is the Real Moat
Even if the BR100 matched H100 hardware specs—and there’s no evidence it does—Biren faces a bigger problem: CUDA. Nvidia spent 18 years building the software ecosystem that makes its GPUs actually useful. PyTorch, TensorFlow, and every major deep learning framework assume CUDA is available. Performance comes from software optimization as much as silicon.
Chinese GPU makers are building that ecosystem from scratch. Biren likely has proprietary frameworks with sparse documentation and tiny developer communities. AMD’s ROCm, an open-source CUDA alternative, has been in development for eight years and still trails Nvidia in real-world performance. Huawei’s Ascend platform uses the CANN framework, which barely exists outside China.
For developers, this means porting CUDA code to Chinese GPUs requires significant work with uncertain performance outcomes. And while Biren tries to catch up to 2022’s H100, Nvidia ships Blackwell and plans Rubin. The target keeps moving.
Forced Migration Coming for Chinese Developers
Here’s the practical impact: If you’re deploying AI on Alibaba Cloud or Tencent Cloud in 2026, you’ll be coding for Biren or Moore Threads GPUs whether you want to or not. It’s not a technical choice—it’s geopolitical reality.
ByteDance invested in Moore Threads and stockpiled billions in Nvidia chips before the ban. Tencent backed Enflame and purchased 2 billion yuan worth of GPUs in Q1 2025 for its ChatGPT-like app Yuanbao. Alibaba’s been deploying custom chips since early 2025. Chinese authorities have also instructed these companies to pause H20 purchases pending a “national security review” and are requiring data centers to prioritize domestic chips.
When the platforms deploy domestic GPUs at scale in 2026-2027, Chinese developers have no choice. The migration is coming whether the technology is ready or not. If you’re building for that market, start planning now. Test workloads on ROCm to understand porting effort. Avoid CUDA-specific optimizations that won’t port to Chinese frameworks. Budget for 20-40% performance degradation. Or face emergency rewrites when your cloud provider swaps the hardware underneath you.
The Big Picture: Nobody Wins
The IPO wave proves investor desperation, not technical validation. Biren might successfully raise $300 million and become one of three viable domestic GPU makers alongside Moore Threads and MetaX. The chips will probably work well enough for inference workloads in Chinese clouds. They’ll struggle with cutting-edge model training. That’s adequate for a captive market created by government mandates.
The U.S. retains the technological lead but loses access to a massive market. China gets chips that work domestically but can’t compete globally. Nvidia’s monopoly cracks in one geography but doesn’t break overall. And developers get to maintain two separate codebases for geographic fragmentation.
Progress looks different when nobody actually wins. But at least the IPO investors got their 4,000x oversubscription stories.
