--- title: "Iluvatar CoreX took its general-purpose GPU public — can \"train and infer on one chip\" work?" date: 2026-10-09 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-iluvatar-corex-gpu language: en --- # Iluvatar CoreX took its general-purpose GPU public — can "train and infer on one chip" work? > Shanghai's Iluvatar CoreX (天数智芯) became one of the first Chinese general-purpose GPU makers to list publicly, betting on chips that handle both AI training and inference. A look at its products, financials, and the hard road from demo to ecosystem. Most AI chips pick a lane: train the model, or serve it. A Shanghai company bet that one piece of silicon could do both — and after years of heavy losses, it just rang the bell in Hong Kong. Shanghai Iluvatar CoreX Semiconductor (天数智芯) listed on the main board of the Hong Kong Stock Exchange in January 2026 under code 9903.HK. It is one of the first Chinese designers of general-purpose GPUs (通用GPU, GPGPU) to reach public markets, and its story captures both the promise and the pain of China's push for compute independence. ## What "general-purpose GPU" actually means here Nvidia's lead is not just hardware; it is a software world (CUDA) where almost every AI model already knows how to run. A GPGPU is a programmable accelerator meant to do the same broad job — train and serve models, run scientific workloads — rather than a fixed-function chip built for one task. Iluvatar splits its line into two families: the Tianchu (天垓) series for training, and the Zhikai (智铠) series for inference, sold as cards, servers, and clustered systems. The company says it was the first in China to mass-produce both a training and an inference GPGPU, and the first Chinese designer to reach that milestone on a 7nm process. Those "firsts" matter less than the underlying claim: one architecture, used across the AI workflow, should be cheaper to adopt than stitching together separate chips. ## The numbers behind the listing The prospectus and post-listing disclosures give a clear, if unfinished, picture. Revenue rose from about 189 million yuan in 2022 to 290 million yuan in 2023 and 539 million yuan (≈ US$74.9 million) in 2024 — a compound annual growth rate near 69%. In the first half of 2025, revenue reached 324 million yuan (≈ US$45 million), up about 64% year on year. Shipments of general-purpose GPU products climbed from 7,800 units in 2022 to 16,800 in 2024, with 15,700 in the first half of 2025 alone. By mid-2025 the company said it served more than 290 customers, with over 900 deployments and compatibility across roughly 450 AI models. Research and development spending reached 773 million yuan (≈ US$107 million) in 2024 — more than revenue that year, the normal profile of a chip firm buying its way to the next node. ## Why it is still losing money Like most of its peers, Iluvatar is not yet profitable. Building a GPU and the software stack around it costs more than early sales return, and the company has said it expects to keep investing heavily. The plan is to put the majority of IPO proceeds into R&D — a bet that scale and a deeper ecosystem will eventually flip the math. The harder competitor is not another Chinese startup but Nvidia's installed base and software. Iluvatar's answer is a self-built software suite that mirrors mainstream frameworks (PyTorch, TensorFlow, PaddlePaddle) so customers can move work over with low friction. Whether "low friction" is low enough is the real test, and it is one no single benchmark settles. ## Where the chips actually land Iluvatar targets the pragmatic middle: financial services, healthcare, transportation, and research clusters where organizations want local, controllable compute and are less obsessed with topping a leaderboard. A single card is described as able to run models in the tens-of-billions-of-parameters range, with cluster deployments holding precision differences under 1% versus the source model — figures the company discloses, not yet independently audited. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the domestic GPGPU race will be won less by peak specs than by which vendor makes migration painless for enterprises that cannot afford to rewrite everything. ## What readers can do now - If you evaluate AI infrastructure, treat "GPGPU" and "NPU" as different tools: GPGPUs aim at flexible training plus inference; don't assume one replaces the other. - If you source compute in China, pilot a domestic GPGPU on a real workload before committing — software compatibility, not FLOPS, decides success. - If you follow the sector, watch recurring revenue and customer count more than launch-day valuation; those show whether the ecosystem is real. ## Honest limitations Financial figures come from the company's Hong Kong IPO documents and listing disclosures (primary sources). Performance claims such as per-card model size and cluster precision are company-disclosed and not independently audited. The competitive comparison to Nvidia is qualitative; no benchmark in this article is verified by a third party. This is analysis, not investment advice, and avoids political or ideological framing. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-iluvatar-corex-gpu Free to quote with attribution and a link to the original.