Two AI-chip startups in Shanghai share the same office park, the same foundry queue and the same handful of potential customers. Within a year, one may simply absorb the other. The joke inside the industry is that China has more AI-accelerator startups than it has hyperscalers to buy from them — and the math is starting to force a reckoning.
The consolidation of China's AI-chip sector is no longer a forecast. It is happening, in both formal mergers and quiet ecosystem tie-ups.
A real merger, not a rumor
- In May 2025, Hygon (海光信息) and Sugon (中科曙光) announced a strategic reorganization — the first absorption merger under China's revised restructuring rules issued 16 May 2025. The combined entity was estimated at over 400 billion yuan (≈ US$56 billion).
- The stated logic was explicit: move "from lone soldiers to a corps" (从单兵作战转向兵团作战) — chip design plus systems integration under one roof, so a customer can buy a working system instead of a component and a promise.
The crowded field behind it
China's domestic accelerator market is crowded with well-funded names, often called the "four little dragons" (四小龙) of domestic GPUs: Moore Threads (摩尔线程), Muxi (沐曦), Enflame (燧原) and Biren (壁仞). Add Cambricon (寒武纪), Iluvatar CoreX (天数智芯) and a tail of smaller designers, and the field has more entrants than the domestic market can sustain at scale.
Each one must fund its own tape-outs — hundreds of millions of yuan per chip — its own software stack, and its own fight for the same customers: a few hyperscalers and state-backed compute platforms. That is an expensive way to stay fragmented.
The IPO wave that exposed the divide
The past year turned the private race into a public one:
- Moore Threads listed on the STAR Market on 5 December 2025, raising about 8 billion yuan (≈ US$1.1 billion) and drawing an outsized first-day move that reflected scarce supply more than earnings.
- Muxi (沐曦股份) listed on the STAR Market on 17 December 2025 at 104.66 yuan (≈ US$14.70) a share, raising roughly 4.2 billion yuan (≈ US$590 million); the National AI Industry Investment Fund took a strategic stake.
- Biren (壁仞科技) shifted from a planned A-share listing to Hong Kong, debuted on the HKEX on 2 January 2026, and closed its first day up 76%, raising about HK$5.583 billion (≈ US$715 million).
- Enflame (燧原科技) restarted IPO tutoring in November 2025 targeting the STAR Market; its shareholders include Tencent, the National Big Fund and Meitu.
Why consolidation is the logical endgame
Fragmentation is the enemy of software. Every independent chip needs its own compiler, drivers and model ports — and developers will not learn ten toolchains. The company that wins is the one whose ecosystem becomes the default, not the one with the highest peak spec on a slide.
Cambricon (寒武纪) shows the alternative path: it swung to profit in 2024 and its market value climbed to roughly 561.6 billion yuan (≈ US$79 billion) at its 2025 peak, making it the bellwether the rest are measured against. Profitability, not a listing, is the real moat — and only a consolidated field can produce more than one such winner.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the domestic accelerator market will consolidate around a handful of ecosystem owners, because software compatibility — not raw chip specs — decides who developers actually adopt.
Huawei's ecosystem as the quiet consolidator
Beyond formal M&A, consolidation is happening through partnerships. Huawei's Ascend (昇腾) stack pulls many smaller players into one software and hardware camp, reducing the number of truly independent ecosystems even without a merger filing. A startup that ports to Ascend is, in effect, joining a coalition — a softer form of consolidation that may matter more than any single deal.
The strategic tie-ups worth watching
The pattern points to three kinds of combination:
- Design-plus-systems (the Hygon–Sugon template): chip maker joins system integrator to sell finished compute.
- Foundry-and-stack alignment: designers converging on shared software layers to shrink the developer burden.
- State-backed aggregation: funds like the National AI Industry Investment Fund taking strategic stakes that nudge rivals toward one ecosystem.
None of this requires every startup to vanish. It requires the field to shrink from dozens of solitary efforts to a few durable camps — exactly what happened in GPUs elsewhere.
Honest limitations
- The Hygon–Sugon deal and the IPO timelines are drawn from established financial-media and exchange reporting; we did not independently confirm post-merger financials or audit the listing figures.
- First-day share moves (for example Biren's 76% gain) reflect scarce float and sentiment, not fundamentals; we cite them as market signals, not recommendations, and express no view on any security.
- The "four little dragons" framing is industry shorthand, not an official classification; the full list of domestic AI-chip entrants is longer and shifts as firms list or restructure.
- Not investment advice. Cambricon's profitability and valuation are reported figures and have been volatile; we did not verify them beyond cited reporting.
What readers can do now
- Track the Hygon–Sugon integration as the template: watch whether combined R&D and customer wins beat the sum of the parts — that is the real test of "corps versus soldiers."
- For investors, judge domestic chip firms by ecosystem adoption (developers, model ports) and a path to profit, not by listing-day hype.
- For engineers, the consolidating market raises the value of portable skills — CUDA-to-Cambricon or Ascend porting, verification, driver work — that survive any single company's fate.