At MetaX's WAIC 2026 booth, the chip was not the centerpiece. Ten open-source community logos were—Anolis (龙蜥), vLLM, PyTorch Foundation, SGLang, arranged across the most prominent central position, with a dedicated "open-source community stamp trail" for visitors to walk.
A GPU company handed its best real estate to software it does not own. That is the story of this year's show.
Last week, this column tracked Chinese labs going vertical on silicon—wafer allocation, in-house architectures, the scramble to secure fab capacity. This week the frontier has visibly moved up the stack. The question at WAIC 2026 was no longer how many transistors, but everything sitting above them: the compiler, the operator library, the migration cost, the watt.
Walk to Qingwei Intelligence's booth and you see the same confession staged differently. A reconfigurable chip, a 4K super-node, and the RAISA software stack sit together as a single answer. The chip alone is not the pitch; the system is.
The thing this staging admits out loud is simple. Releasing a chip only buys you the entrance ticket. The game is whether a customer will dare to migrate, can actually run their workloads, and will keep using your hardware for years.
And the vendors themselves are the ones saying it. The wall is not NVIDIA's GPU. The wall is CUDA.
The wall is not NVIDIA's GPU. The wall is CUDA—the developer trust it bought slowly over many years, and the migration cost that stops most teams.
The framing recurs almost verbatim across booths: process nodes, architecture, and memory are visible, measurable things that Chinese vendors can narrow with effort. What is truly hard to chase is the moat NVIDIA hides beneath the hardware—the CUDA software ecosystem, where millions of developers' project code, habits, and self-built modules have accumulated over many years.
Switching cards, in this telling, means tearing down years of accumulated engineering. The migration cost is what stops most teams. So the real target of domestic compute is not a single GPU—it is the developer trust CUDA bought slowly over many years.
China's answer at WAIC is not to out-transistor NVIDIA. It is to route around CUDA through three disciplines, each attacking a different layer of that moat.
The first move is community capture, and MetaX is the clearest case. Its self-built automated test suite runs roughly 5,000 popular GitHub projects on every MXMACA release; 92% of them run with zero code changes. For PyTorch 2.8, MetaX claims full compatibility across all 2,410 GPU operators—and where rivals need months to adapt to a new PyTorch version, MetaX says it finishes in a week.
The strategy has a name inside the company: upstream first. Instead of pulling a community release, patching it privately, and recompiling, MetaX contributes its support code directly back into the upstream projects. When a new model ships, the support is often already there by default—pip install, run.
That is why MetaX became the first chip company outside the United States to sign an official cooperation agreement with vLLM, and why its hardware sits on the Red Hat plus vLLM official shelf. The relationship is no longer "I adapt to you." It is co-construction.
MetaX CTO Yang Jian gives the ambition an anchor: the Android of the AI era. Keep the full stack open, from operating system to training and inference to private deployment, until other vendors and customers build on it—the way Android's activated-device and developer counts, not its kernel, carried the ecosystem.
The second move is operator-level Day-0 adaptation as repeatable engineering, and this is where Qingwei's RAISA lives. When DeepSeek-V4's preview dropped, Qingwei and BAAI's FlagOS completed full adaptation and verification of the Flash version's 67 operators on launch day.
That specificity matters more than any "we support model X" claim. Whether a model runs at all depends on whether the underlying stack covers every core operator it calls. RAISA now reportedly supports nearly a thousand mainstream operators and has adapted more than 200 large models.
The third move reframes efficiency itself, and here SenseTime is the aggressor. Its 算电协同 Agent explicitly proposes replacing PUE with TPW—Tokens Per Watt.
PUE only measures whether a data center wastes energy on cooling and overhead. TPW measures how many valuable tokens each unit of electricity cost actually produces, factoring in time-of-day pricing. At SenseTime's Lingang AIDC, the Agent reportedly delivers an 80% increase in token output per unit of electricity cost, 96% load-forecast accuracy, and an estimated 24,000 tons of annual CO2 reduction per 10,000 PFLOPS.
Damao Technology is pushing the same reframing with its 算电协同 2.0 platform, an "energy operating system" that folds compute scheduling, renewable generation, storage, and electricity trading into one decision space. The pitch: turn electricity from a cost line into a value lever for tokens.
Notice what all three moves share. They shift the metric of competition from something sanctions can touch—transistors, process nodes, advanced packaging—to something sanctions cannot: open-source commits, operator coverage, migration cost, and watts.
This is elegant. You cannot embargo a pull request into vLLM. You cannot restrict the price of electricity in western China. If the axis of competition becomes community gravity and energy arbitrage, the export-control apparatus is aiming at the wrong layer.
But the same vendors who staged this confession also flagged its weakest link, and to their credit they did so candidly.
Qingwei's own framing warns that Day-0 adaptation must become "stable, replicable software engineering," not a one-time launch-day event. A truly mature software ecosystem is not "the new model runs on release day." It is staying compatible across version updates, resolving customer issues quickly, and holding stable performance under real production load.
That is a far harder and unproven bar. Launch-day demos are visible and photogenic. Stability under version churn and real traffic accrues quietly, over years.
MetaX's own numbers hint at the gap between demo and durability. In the vLLM ecosystem, MetaX deeply tuned only fifty or sixty mainstream models itself and handed the rest to undergraduates to test on its cards—most worked out of the box, but around ten hit adaptation obstacles. Impressive breadth; also a reminder that "runs" and "runs at target throughput, cost, and precision under load" are three different tiers.
That is the crux of the wager. CUDA's moat is not any single capability—it is fifteen years of accreted trust, the confidence that when you update, things still work. Community gravity and energy arbitrage are real forces, but they are being asked to substitute for time.
MetaX says its ecosystem holds 500,000 developers today and targets five million by 2029–2030. Even Yang Jian's own metaphor concedes the point: Android needed years of activated devices and developer accumulation to become inevitable.
So the strategic picture is this. China has correctly diagnosed where the wall actually stands, and has moved its offensive from the layer it can be blocked at to the layer it cannot. That reframing is the smartest thing on the WAIC floor.
Its weakest link is precisely the one MetaX names without flinching. Software ecosystems accrete over years, not quarters—and no amount of open-source enthusiasm or clever electricity pricing has yet proven it can compress that clock.
The chip was the entrance ticket. Whether China can win the game above it will not be settled by launch-day operator counts. It will be settled by the far quieter question of whether the stack is still compatible, and still stable, on the tenth version update under real load.