China's $344M bet: Infinigence makes AI run on 16 chip types at once
Shanghai-based Infinigence AI just raised $344M to network 16 different chip types into a single supercomputer using software.
Funding · Source: TechNode
What happened
Shanghai-based Infinigence AI just closed a 344 million dollar Series C round. This brings their total funding to 643 million dollars since launching in 2023. The startup builds software that forces different AI chips to work together. They also confidentially filed for a Hong Kong IPO targeting the first half of 2027.
The platform connects 37,000 petaflops of computing power across 16 mainstream chip types. It supports mixed-chip training using both Chinese and international processors. This means developers can train and run models without relying on a single hardware vendor. Heavyweights like Tencent, Baidu, Lenovo, and Sequoia China backed the round.
The scale is accelerating fast. Infinigence reports daily token usage jumped 40-fold between December 2025 and July 2026. Their cross-cluster architecture cuts the time to generate the first token by 51.5 percent and drops the cost per token by 37.5 percent. Major Chinese model builders like Zhipu AI and Moonshot AI are already using the infrastructure.
Key facts
- $344.2 million — Series C funding raised by Infinigence AI
- 16 — Mainstream chip types supported by the platform
- 37,000 petaflops — Computing capacity connected across the infrastructure
- 51.5% — Reduction in time to generate the first token
- 40-fold — Increase in daily token usage from Dec 2025 to July 2026
Why it matters
The AI bottleneck is shifting from acquiring the best hardware to maximizing the hardware you actually have. By abstracting the compute layer, Infinigence allows developers to mix and match whatever chips they can get their hands on. You no longer need a uniform cluster of top-tier processors to train a frontier model. Software is bridging the gap left by hardware shortages.
This turns a massive supply chain constraint into a pure engineering problem. Export controls forced Chinese companies to get creative with heterogeneous compute. Now that forced innovation is creating a resilient infrastructure layer that threatens the lock-in power of dominant chipmakers. If AI workloads can run seamlessly across 16 different chip architectures, the premium on any single brand of silicon drops significantly.
For builders
Hardware abstraction lowers compute costs
You do not need to overpay for unified processor clusters anymore. Tooling that supports mixed-chip inference drops the cost per token and speeds up generation times. Startups building AI applications win by paying less for compute.
Infrastructure is the real AI moat
Everyone is fighting over foundational models but the real money is flowing into the plumbing. Building deployment and scheduling software for heterogeneous compute is a massive market. Founders who solve infrastructure bottlenecks will capture enterprise budgets.
My take
I always say constraints breed the best engineering. America tried to choke China's AI progress by cutting off the best chips, but they just forced them to build software that makes the hardware brand irrelevant. When you ban the hardware, you just accelerate the software.
Original reporting: TechNode. This is my rewrite and opinion.