A Shanghai chip designer that few outside the industry had heard of sold shares to the public in Hong Kong.
Its pitch: a general-purpose GPU built on a 7-nanometer process, made for China's AI training needs.
The listing landed as Beijing pushes hard for chips it can call its own — and investors showed up.
Key takeaways
- Iluvatar CoreX (天数智芯) listed on the Hong Kong Stock Exchange on 8 January 2026 under ticker 9903.HK, pricing at HK$144.60 and raising about HK$3.7 billion (≈ US$474M).
- Its IPO market capitalization was about HK$35.4 billion (≈ US$4.5B).
- Before the listing it closed two 2025 private rounds: a D round above RMB 1.4 billion (≈ US$197M / HK$1.5B) and a D+ round of RMB 2.05 billion (≈ US$289M / HK$2.26B) at a ~RMB 12 billion (≈ US$1.69B / HK$13.2B) pre-money valuation.
- Per its prospectus (citing Frost & Sullivan), it is the first Chinese firm to mass-produce both training- and inference-class general-purpose GPUs on a 7nm process.
The listing, in numbers
Iluvatar CoreX (上海天数智芯半导体) was founded in Shanghai in 2015 and builds general-purpose GPUs for AI training, inference and high-performance computing. Its product lines — the Tiangai (天垓) training series and Zhikai (智铠) inference series — are designed as domestically controllable alternatives to foreign GPUs.
The verified IPO figures:
- Listing date: 8 January 2026 on HKEX, ticker 9903.HK.
- Issue price: HK$144.60 per H share.
- Shares issued: about 25.43 million H shares.
- Amount raised: about HK$3.7 billion (≈ US$474M).
- IPO market cap: about HK$35.4 billion (≈ US$4.5B).
- Cornerstone investors: 18 parties including ZTE (Hong Kong), UBS AM Singapore, Fourth Paradigm (第四范式), ChinaAMC (Hong Kong) and Hua Sheng Tian Cheng, taking roughly 43% of the offer.
The funding blitz that preceded it
The IPO capped a busy 2025 on the private market. According to the prospectus and listing documents:
- D round: more than RMB 1.4 billion (≈ US$197M / HK$1.5B), with backers including Sequoia China and Yuanhe Fund.
- D+ round: RMB 2.05 billion (≈ US$289M / HK$2.26B), at a pre-money valuation of about RMB 12 billion (≈ US$1.69B / HK$13.2B), with investors such as Xingnahe Capital and Quzhou Zhizao.
- Cumulative private funding: above RMB 5.52 billion (≈ US$777M / HK$6.1B) per Caixin, with a peak valuation of RMB 12 billion.
That trajectory — a large late-stage round, then a public offer months later — is the standard script for a Chinese chip firm racing to fund fabs, tape-outs and software stacks before competitors do.
Why a GPU matters here
A general-purpose GPU is the workhorse behind every large model (大模型). Training and serving models demands parallel compute (算力) at a scale few other chips provide. US export controls have restricted China's access to the most advanced foreign GPUs, which turned domestic alternatives from a nice-to-have into a strategic priority.
Iluvatar's claim to distinction is process maturity. Per its prospectus (citing Frost & Sullivan), it is the first Chinese company to achieve mass production of both inference- and training-class general-purpose GPUs using an advanced 7nm process. Shipments rose from 7,800 units in 2022 to 16,800 in 2024, with 15,700 in the first half of 2025 alone.
The financial reality check
The growth is real but the losses are still substantial. Revenue climbed from RMB 1.89 billion (2022) to RMB 2.89 billion (2023) to RMB 5.40 billion (2024), then RMB 3.24 billion in H1 2025 (up 64% year on year), with gross margin improving to about 50%. Yet net losses were heavy — RMB 554 million, RMB 817 million and RMB 892 million across 2022–2024, and RMB 609 million in H1 2025 — driven by relentless R&D.
In other words, the market is pricing a future where China's AI compute (算力) is built on local silicon. The company is not yet profitable, and may not be for years.
What readers can do now
- If you follow semiconductors: track Iluvatar's shipment growth and gross margin — the path to profit runs through volume, not just design wins.
- If you invest in HK listings: read the prospectus's loss trajectory before treating the 7nm story as a moat; peers Moore Threads and Biren listed around the same window.
- If you use Chinese AI services: domestic GPUs are the reason many local models kept training through export curbs — a quieter form of supply security.
Honest limitations
Key figures — HKEX listing 8 Jan 2026, ticker 9903.HK, issue price HK$144.60, ~25.43M H shares, ~HK$3.7B raised, ~HK$35.4B market cap, 18 cornerstones (~43% of offer) including ZTE HK/UBS AM Singapore/Fourth Paradigm/ChinaAMC HK/Hua Sheng Tian Cheng; 2025 D round >RMB 1.4B (≈ US$197M / HK$1.5B) with Sequoia China/Yuanhe, D+ round RMB 2.05B (≈ US$289M / HK$2.26B) at ~RMB 12B pre-money with Xingnahe/Quzhou Zhizao; cumulative private funding >RMB 5.52B per Caixin; revenue RMB 1.89B/2.89B/5.40B (2022–24) and RMB 3.24B H1 2025 (+64% YoY), gross margin ~50%, net losses RMB 554M/817M/892M and RMB 609M H1 2025; shipments 7.8k/12.7k/16.8k (2022–24) and 15.7k H1 2025; "first China 7nm mass-produced GPU" per prospectus citing Frost & Sullivan — come from the company's listing documents / investor-relations site (ir.iluvatar.com), Eastmoney, Caixin and China Tech News. HK-dollar figures use ~HK$7.8 = US$1 for the USD equivalent; RMB figures use fixed rates RMB 7.1 = US$1, RMB 1 = HK$1.1 (e.g., 1.4/7.1 ≈ 197M; 2.05/7.1 ≈ 289M; 12/7.1 ≈ 1.69B). The "first 7nm" claim is the company's, sourced to Frost & Sullivan via the prospectus — not independently re-audited here. Not investment advice. Current to 9 November 2026.