---
title: "China's GPU Makers Are Learning the Moat Is Code"
date: 2026-11-16
category: Chips & Compute
site: NeuroAI
canonical: https://neuroai.site/a/na-chips-cuda-moat-software
language: en
---

# China's GPU Makers Are Learning the Moat Is Code

> Moore Threads and MetaX now sell GPUs into Chinese data centres. The hard part is not silicon but persuading developers to abandon twenty years of CUDA code.

A team spends a quarter porting a training job to a domestic accelerator. The model runs. It produces correct outputs. Then the engineers profile it and find that one attention kernel is still written for the other vendor's hardware, so it falls back to a slow path and the whole job runs at a fraction of the expected speed. Nothing is broken. The chip is fine. The code is the problem.

This is the wall that Chinese GPU companies hit once they clear the hardware bar, and it is the part of the story that gets the least attention because it cannot be expressed in a spec sheet.

## Two strategies, one problem

Moore Threads (摩尔线程) and MetaX (沐曦) took different routes to the same destination, and the contrast is instructive.

Moore Threads built MUSA, launched in 2022, around translation. Its MUSIFY tool converts CUDA C++ source into MUSA C++ source, which the MUSA compiler then turns into binaries for its own instruction set. The pitch is that existing code needs little or no rewriting. The company has open-sourced Torch-MUSA, OpenCV-MUSA, and vLLM-MUSA, and says its PyTorch backend covered more than 1,050 operators as of November 2024.

MetaX went the other way with MXMACA. It developed its own GPU IP, instruction set, and architecture, and offers a C-like programming model that is source-compatible with mainstream heterogeneous computing syntax but is not a CUDA transpiler. In practice, developers port rather than recompile — more work up front, but no dependency on another vendor's language definition.

Both approaches are legitimate engineering choices, and both run into the same economic fact: the value of a GPU is set by the software that already runs on it. Nvidia's advantage is not that its chips are fast. It is that a decade and a half of kernels, libraries, debugging tools, and half-forgotten Stack Overflow answers are written against its programming model.

## Porting is not one problem, it is four

In practice the work breaks into a stack of increasingly annoying problems rather than one migration.

- **Source compatibility.** Can the code compile? This is the easy layer, and the one vendors demo.

- **Library parity.** Does the vendor have equivalents for the linear algebra, FFT, collective communication, and inference libraries the model actually calls? A missing routine in a collective library can cost more than a missing instruction.

- **Numerical behaviour.** Do results match bit-for-bit, or close enough for training to converge? Different fused-multiply-add ordering or reduced-precision defaults can change loss curves in ways that are hard to debug and easy to blame on the chip.

- **Tooling.** When something is slow, can you find out why? Profilers, memory analysers, and kernel debuggers are the difference between a week of tuning and a month of guessing. This is the least glamorous layer and the one that decides whether a platform gets used twice.

There is a fifth, quieter problem: model architectures change every few months. A port that is validated against today's attention variant may need revalidation against the next one, and each revalidation consumes the same scarce engineers.

## The business case is being built on volume, not benchmarks

What has changed is that both companies now have public capital and shipping products.

- Moore Threads listed on the Shanghai STAR Market (科创板) on 5 December 2025 as ticker 688795.SH, raising close to RMB 8 billion (≈US$1.1 billion / HK$8.7 billion).

- MetaX, ticker 688802.SH, listed on the same exchange in late 2025 and reported a first-day gain of several hundred percent.

- MetaX disclosed cumulative GPU shipments above 25,000 units and, in its listing documents, revenue concentration of 97.28% in 2024 and 97.87% in the first quarter of 2025 from a single training-and-inference product family, the XiYun C500 series.

- Moore Threads has reported that its top five customers account for more than 80% of revenue in each reporting period.

Those numbers describe a business that is being pulled by a policy-driven demand for domestic compute, not yet by developer preference. That is not a criticism; it is how every semiconductor platform starts. But it changes what "success" looks like in the near term. A domestic GPU can win a data centre contract because it is the option that can be procured, and still lose the developer once the contract ends.

## Where the switching cost actually lands

For a large model team, the expensive part of switching is not the initial port. It is the regression testing, the retuning, and the operational knowledge that has to be rebuilt.

Consider what a single training run depends on: a distributed launcher, a collective communication library tuned for the cluster's topology, a checkpoint format, a mixed-precision recipe, and a set of fused kernels that someone spent months optimizing. Move to a new accelerator and every one of those becomes a project. The chip may be 20% slower and the migration may cost a quarter of engineering time — which is the better trade depends entirely on how long the hardware will be in the building.

That is why the interesting competition is not at the top-end benchmark but at the middle of the market: inference. Inference workloads are narrower, more standardized, and easier to validate, and a serving stack that works out of the box on vLLM is a far cheaper migration than a research training cluster. It is also where volume lives.

## Honest limitations

- **I did not benchmark any of these chips.** Every performance and compatibility claim in this piece is the company's own or comes from Chinese financial media reporting, and none of it has been independently reproduced here.

- **Operator-count claims age quickly.** The 1,050-operator figure for Torch-MUSA is from November 2024 and almost certainly no longer current in either direction.

- **Revenue concentration has two readings.** It signals early-stage product portfolios, but it also means a single customer decision can move a quarter's results, which is a real risk rather than a badge of focus.

- **Legal risk around translation layers is genuinely uncertain.** Commentators argue about whether source-level translation sits comfortably within the original platform's licence terms. I found no court ruling or definitive public determination, so I have not asserted one.

- **"Software ecosystem" is not measurable in one number.** Anything resembling an ecosystem score here would be an artefact of the metrics chosen.

## What readers can do now

- **If you are evaluating a domestic accelerator, run your own kernel, not the demo.** Port one real workload with a profiler attached and measure where time goes. A vendor demo tells you the platform works; your kernel tells you whether it works for you.

- **If you are building a product on top of one, budget the profile-and-tune phase explicitly.** The port is the cheap part. The retuning is what eats the schedule.

- **If you are watching this market, watch library releases rather than launch events.** The cadence of updates to the math, communication, and serving libraries is the clearest public signal of whether a software stack is actually alive.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-cuda-moat-software
Free to quote with attribution and a link to the original.
