Xia Lixue's team does not build chips. That is the point.
His company, Shanghai-based Infinigence AI, sells the layer between silicon and models — software that lets a single training job run across accelerators from Huawei, Moore Threads, AMD and Nvidia at once. On October 12 it closed a RMB 2.3 billion Series C, about $344 million, the largest single round yet recorded in China's AI-native infrastructure sector.
Cumulative funding is now RMB 4.3 billion — and a confidential Hong Kong filing follows.
It owns almost none of the hardware it sells.
A Tsinghua bet on the middle layer
Infinigence was spun out of Tsinghua University's electronics department in 2023. It was initiated by professor Wang Yu and co-founded by his former doctoral student, now chief executive, Xia Lixue.
The strategy is called "Agentic Infra": a compute distribution center that pools hardware, a token factory that serves models as a platform, and an industry-solutions arm that pushes tokens into real applications.
The funding cadence has been quick. A near-RMB 500 million Series A in September 2024 gave way to an A+ of similar size in November 2025 and more than RMB 700 million in May 2026, before this round.
What mixed-chip training actually means
The claim that matters is technical. Infinigence says its thousand-card platform — built to scale to 10,000 cards — has trained large models across four domestic accelerator families alongside AMD and Nvidia hardware, at up to 97.6% compute utilization.
Mixed-chip training is hard for a plain reason. Parallelism strategies and toolchains differ vendor by vendor. Pooling rival architectures usually means rewriting the training stack from scratch.
With China Telecom, the company says it ran China's first large-scale cross-region mixed-training test over more than 4,000 kilometres, linking clusters in Hami, Xinjiang, with Shenzhen to train a 30-billion-parameter model.
On inference, its model-service platform's daily token usage grew 40-fold from December 2025 to July 2026, the company says. Its cross-cluster architecture cut time-to-first-token by 51.5% and cost per token by 37.5%.
Those are the company's own numbers. Treat them as company figures, not a benchmark.
The platform reaches 37,000 petaflops across 16 mainstream chip types. Infinigence frames the problem as M×N: M models against N chips, glued by scheduling software rather than a single vendor's stack.
The pattern behind the round
Infinigence is not alone in betting that the scarce asset is software. On October 8, Nvidia-backed Upscale AI launched Token Fabric, a platform for connecting chips from different suppliers without rebuilding the network around each one.
As we wrote in October, the fight has moved from training throughput to inference economics — the bet behind Positron AI's $875M round.
Export controls sharpen the stakes. With access to top Nvidia silicon restricted, China's answer is a scheduler that treats many chips as one machine.
The IPO in the background
Bloomberg reported that Infinigence confidentially filed for a Hong Kong IPO, seeking several hundred million dollars at a pre-offer valuation of RMB 14.3 billion. A listing could arrive as early as the first half of 2027.
Backers include Tencent, Baidu and Lenovo, alongside Sequoia China, Qiming and Shunwei. Earlier investors include Lenovo Capital, Xiaomi and the AI lab Z.AI.
The round is closed. The filing is in. What stays open is whether a neutral layer can remain neutral when every chip vendor wants the stack to be its own.