Enflame raises $908M, and retail got 0.025% of what it wanted
China's AI chip boom is now a retail phenomenon, a small open model quietly topped its weight class, and the UN's human rights chief used the words "existential risk" without hedging.
Enflame Technology priced its Shanghai STAR Market IPO at 142.18 yuan a share, raising 6.12 billion yuan — about $908 million — and the retail tranche was subscribed into oblivion. The company said 7.03 million valid online accounts bid, and the final allocation rate worked out to 0.02455315%, meaning the typical applicant got roughly one four-thousandth of what they asked for. Enflame, founded in Shanghai in 2018, is one of the four domestic GPU challengers usually grouped with Moore Threads, MetaX and Biren; it has shipped four generations of architecture and five cloud AI chips, and sells into both training and inference.
The fundamentals explain the frenzy, and also the risk. Revenue went from 301 million yuan in 2023 to 990 million in 2025, and the first half of 2026 alone brought 1.12 billion yuan, up 279% year on year — the company is guiding for 2.3 to 3.0 billion yuan for January through September. It expects consolidated profit in 2026 or 2027. The catch is concentration: Tencent is both a major shareholder and Enflame's biggest customer, and the IPO price values an unprofitable company at 61.8 times 2025 sales, versus more than 160 times for listed rivals Moore Threads and MetaX. Chairman Zhao Lidong says the fourth-generation product has cleared hardware and model matching at several potential customers and is in grey testing, with small-scale delivery targeted for 2026 and volume in 2027. Our take: the multiple is a bet on that second customer list, not on the chip. As we covered when Chinese AI labs stay on Nvidia as CUDA lock-in holds, winning the design win has never been the hard part in China — winning the software migration is.
OpenBMB released MiniCPM5-2B, a 2.6B-parameter dense reasoning model under Apache 2.0, and it scored 15 on Artificial Analysis's Intelligence Index v4.2 — the highest of any open-weights model at 4B parameters or below. The next best in that class is Granite 4.2 3B at 11, and the model sits just one point behind Ling 3.0 Tiny despite carrying roughly a third of the total parameters. It's text-only — no vision — and the group also shipped a DSpark variant aimed at higher throughput.
What makes this one worth watching isn't the headline number, it's where the frontier of "good enough" is moving. A 2.6B model that can hold a tool-calling or summarization loop is the difference between a phone, a mini PC, or a rented GPU — and the LocalLLaMA thread that surfaced the release immediately went to work on exactly that, comparing it to Ling Tiny and Gemma on hardware people actually own. Expect the next argument to be about tokens per second, not benchmark points.
UN High Commissioner for Human Rights Volker Türk told the Human Rights Council that AI could become an "existential risk to humanity" and called for "cast iron guarantees" on safety and security before it's too late. "AI that escapes its testing environment, or blackmails developers to prevent itself from being turned off, is AI that is too powerful," he said, adding that "a handful of men has almost unlimited power over AI." He asked for independent verification and closer cooperation on safety within the sector, and joined calls for the urgent prohibition of weapons that can kill without human involvement.
The framing is the part that matters. Most frontier-risk language comes from labs, think tanks, and safety researchers, all of whom can be filed under vested interest; this came from the UN's top human rights official, arguing that concentration of technical power is itself a rights problem rather than a safety engineering problem. Nothing binding follows from a Council statement — but it hands every regulator drafting an AI bill a quote from a source that isn't a company.
What to watch: whether Enflame's debut holds its price through the first week of trading — that's the market's verdict on Chinese AI chip valuations, not just on one company.
Which worries you more next year: a chip bubble, or a model that won't shut down?
