Moxin raises nearly $140 million for sparse inference chips
Chinese chipmaker Moxin has raised nearly $140 million as sparse inference moves from model technique to a commercial hardware strategy.
The financing
Chinese AI-chip company Moxin Artificial Intelligence has completed a nearly 1 billion yuan C+ round, according to QbitAI. The financing reportedly arrived only five months after a similarly sized previous round, while the company said its third-quarter revenue grew several-fold.
Moxin is building a full-stack sparse-inference platform spanning algorithms, compilers and accelerator hardware. Its approach combines weight sparsity with activation sparsity, aiming to avoid spending compute on parameters or token paths that contribute little to a particular output. That design is especially relevant to mixture-of-experts models, whose conditional execution has helped make sparse computation a central theme in the post-DeepSeek efficiency race.
From benchmark to deployment
The company says its hardware has won in multiple categories of MLPerf Inference for three consecutive editions. It also reports a 2.88-times throughput improvement and 7.64-times model compression in remote-sensing workloads. Those figures are company claims rather than independently audited evidence, but they point to a more important shift: sparse acceleration is being sold as an end-to-end operating model, not merely as a software optimization.
Moxin is also involved in a Shanghai computing project and is preparing a new SP200 card and SECA software stack. The immediate question is whether its performance can survive outside carefully selected workloads and translate into repeatable customer deployments.
Why it matters
The round is significant because it gives a Chinese accelerator company substantial capital while the market is rethinking how much inference infrastructure dense models require. If Moxin can turn sparsity into predictable economics across models and customers, it could strengthen China’s domestic alternative to imported accelerators. If the gains remain workload-specific, the financing will look more like a bet on architecture than proof of a new hardware category.
Uncle Cat take: Moxin’s nearly 1 billion yuan follow-on round is meaningful, but its real valuation will depend on SP200 deployments beyond remote sensing—not another MLPerf headline.