On the morning of September 11, a chip company that had never turned a profit opened for trading in Shanghai. By the close, its shares were worth nearly three times the price fixed the night before. The celebratory number on the screen obscured a less comfortable fact: most of the company's sales come from a single customer — the same technology giant that backed it.
That company is Shanghai Enflame Technology (燧原科技), and its listing says as much about China's chip strategy as about one company's stock.
Key takeaways
- Enflame raised RMB 6.12 billion (≈ US$912M / HK$7.1B) in its Shanghai STAR Market IPO, selling 43.04 million shares at RMB 142.18 each.
- The stock closed at RMB 397 on debut, up 179%, valuing the company at about RMB 171 billion (≈ US$25.5B / HK$199B). One analyst, Shen Meng of Chanson & Co., cautioned that the pop reflected new-share demand more than fundamentals.
- Tencent was both the largest shareholder after the offering, at 17.95%, and the largest customer, accounting for 83.79% of 2025 revenue.
- Unlike peers focused on general-purpose GPUs, Enflame is betting on specialised, lower-power chips for inference — running trained models — and on its own software platform to reduce dependence on Nvidia's CUDA.
What the debut showed
Enflame sold 43.04 million new shares, equal to 10% of its enlarged capital. The online portion was subscribed 6,109 times, according to an exchange filing cited by Reuters, with more than seven million investor accounts applying. The offer price valued the company at roughly RMB 61.2 billion (≈ US$9.1B / HK$71B); the closing price put it near RMB 171 billion (≈ US$25.5B / HK$199B).
That is a remarkable first day, but it is not a verdict on the business. "The rise is driven by short-term incentives for new share subscriptions rather than investors' focus on the long-term fundamental development of the sector," said Shen Meng, director at Beijing investment bank Chanson & Co., speaking to Reuters. Enflame was one of a group of Chinese AI chip startups to reach public markets, following peers whose own debuts were also sharp.
The company's accounts explain why investors are interested and why they should be careful. Revenue rose 37% in 2025 to RMB 990.2 million (≈ US$148M / HK$1.15B), while the net loss narrowed to RMB 1.16 billion (≈ US$173M / HK$1.35B) from RMB 1.51 billion a year earlier. For the first nine months of 2026, Enflame forecast revenue of RMB 2.3–3 billion (≈ US$343–448M / HK$2.7–3.5B) and a net loss of RMB 700–860 million. It expects to break even or turn profitable in 2027 or 2028, depending on growth and margins.
Why inference matters
Enflame was founded in 2018 by Zhao Lidong and Zhang Yalin, both former AMD engineers. Its strategy differs from the general-purpose GPU path pursued by many rivals. The company is betting on specialised, more power-efficient processors aimed at inference — the stage where a trained large model (大模型) actually answers users, rather than the far more compute-hungry job of training it.
Inference is the faster-growing part of the market. It does not demand the same raw computing power as training, and it rewards efficiency, cost per query and ease of deployment. That plays to the strengths of smaller chip designers who cannot match Nvidia at the frontier. Enflame plans to spend most of the IPO proceeds developing its fifth- and sixth-generation AI chips, related software and large-scale computing systems.
The software is the other half of the strategy. Nvidia's CUDA ecosystem has become the industry standard, locking in developers through years of tools and familiarity. Enflame is building its own programming platform to compete, and other Chinese designers are offering ways to migrate CUDA code onto domestic hardware. This is less glamorous than a chip launch and more important: without a software path, a fast chip is hard to use.
The broader backdrop is China's effort to reduce dependence on foreign accelerators. Nvidia's share of China's AI accelerator server market has fallen from near-monopoly levels to an estimated 55%, according to data cited by Reuters, while domestic suppliers have gained ground. Export controls on advanced chips and manufacturing equipment have made that shift both more urgent and more expensive.
The customer question
The most striking number in Enflame's prospectus is not the IPO size. It is the 83.79% of 2025 revenue tied to Tencent, which held 17.95% of the company after the listing. A customer that is also a major shareholder is not unusual in China's tech sector, but it concentrates risk: a change in Tencent's orders would hit Enflame far harder than a diversified supplier.
For global readers, that is the lens worth keeping. The Chinese chip story is often told as a contest with Nvidia. The company-level story is more mundane and more revealing: fast revenue growth, persistent losses, a single dominant customer and a bet that inference plus a home-grown software stack can carve out a durable niche. Whether the IPO money buys that niche will be visible in the next few product cycles, not on the first day of trading.
Honest limitations
This article relies on Reuters reporting of the IPO and debut, a Reuters Breakingviews column, and a Forbes summary that cites the same filings. I have not read the prospectus in full or verified the figures independently. The oversubscription multiple, financials and customer concentration all come from company filings as reported.
Breakingviews also cites a very high first-quarter sales growth rate measured against a small base; the figure does not appear in the debut report, so I have treated it as context rather than a headline number. Forecast revenue and the break-even timeline are company projections, not results. This is not investment advice, and a first-day price move says little about long-term value.
I have not examined the technical merits of Enflame's chips against Nvidia's or compared its software stack with other domestic platforms. The "four little GPU dragons" grouping is a market label, not a technical category. Currency conversions use a November 2026 reference rate.
What readers can do now
- Read past the first-day pop. For any chip IPO, find revenue, customer concentration and path to profitability before reading the share price. Those three tell you more than the opening bell.
- Watch inference, not just training. The cheapest, most efficient chip for running a model may matter more to ordinary users than the fastest chip for building one.
- If you build on AI hardware, test the software path first. A domestic chip is only useful if your code can reach it. Migration tools are the quiet gatekeeper of this market.
Enflame's debut was a financial event. Its inference bet is a technological one. Only the second will still matter a year from now.
