At the World AI Conference in Shanghai, Xi Jinping staged AI as a global public good — with hard-nosed industrial policy behind it. But the substance is real: Kimi 3 beats the best U.S. models on some benchmarks for the first time and ships as an open-weights model, and Huawei's Atlas 950 SuperPod is running ahead of schedule. The question for CTOs isn't whether to believe the narrative — it's whether your architecture could even use a second complete AI stack.
July 17, 2026: Xi Opens the World AI Conference — a Challenge Dressed as Cooperation
Opening the World AI Conference in Shanghai, China's President Xi Jinping delivered a keynote on AI policy — and explicitly rejected any U.S. claim to leadership. AI development, Xi said, must not be a "solo performance by a single country." Instead: open, multilateral cooperation under the umbrella of the UN, and stronger support for developing countries.
The speech was orchestrated. A day earlier, 29 countries had signed an agreement to found a World AI Cooperation Organization (WAICO), headquartered in Shanghai. Chinese state media had prepared the ground by contrasting an "oil mindset" — data and compute as scarce, exclusive resources — with a "water mindset" of sharing, casting China as a provider of international public goods. Xi added a development program on top: 5,000 AI training and seminar programs plus international cooperation centers for the Global South. Golem has a detailed report on the conference (German).
Anyone tempted to read this as altruism should look at China's track record in Laos or the DRC: loans that end in dependency, resource deals whose promised infrastructure never materializes. This is hard industrial and trade policy. But — and this is the part you shouldn't cynicism away — the technological substance behind the narrative is real this time.
Klingt interessant?
Kimi 3 and the Atlas 950: The Stack Takes Shape
The day before the conference opened, Chinese AI lab Moonshot released Kimi 3 — the first model to beat the currently best published models from Anthropic and OpenAI on some benchmarks. The real coup isn't the benchmark win, though; it's the license: from July 27, Kimi 3 will be available as an open-weights model — worldwide, locally, on your own hardware. The comparable U.S. models are proprietary and subscription-only.
On the hardware side, Huawei is following through. The Atlas 950 SuperPod, announced last September, is being shown in Shanghai as an operational system for the first time — several months ahead of schedule. Huawei claims 6.7 times the compute and 15 times the memory capacity of Nvidia's upcoming NVL144 supernode generation. Vendor numbers, sure. But industry insiders expect Deepseek and Z.ai to move their training onto the Huawei supernodes fairly quickly — both have been optimizing their models for Huawei chips for a while, since U.S. export controls cut them off from Nvidia hardware. Supply shortages aren't expected; Huawei plans to offer the SuperPods in South Korea by the end of 2026.
For the first time, this adds up to a complete, end-to-end AI stack outside the U.S. ecosystem: home-grown accelerator hardware as a likely cheaper Nvidia alternative, open models optimized for it — Kimi, GLM, Deepseek, Qwen — and a state-backed distribution program for developing and emerging markets.
The Irony of Export Controls
The U.S. answer to all this is ever-broader export controls — on AI chips, on chip-making technology, and most recently even on Anthropic's advanced proprietary models, effectively declaring them an exclusive national resource. Xi countered in Shanghai with the charge that one country's security is being placed above everyone else's.
You don't have to share Xi's motives to see the strategic problem: every export control on proprietary U.S. models strengthens the case for open alternatives — and the strongest open alternatives now come from China. If you build a European product on a proprietary U.S. frontier model today, you carry a dependency that can be tightened by decree from Washington. Since the Anthropic export controls, that scenario is no longer theoretical.
What This Means for CTOs and Tech Leads
First: there is now a second complete pole — and it will compete on price. U.S. models on Nvidia hardware deliver top performance at top prices. The Chinese stack positions itself as the cheaper combination of open models and its own hardware. Even if you never deploy a Chinese model, you benefit: a second pole changes prices, terms, and negotiating positions across the entire market.
Second: open weights doesn't mean worry-free. Running an open Chinese model locally solves the infrastructure sovereignty problem — your data never leaves your data center, and nobody can switch off your access. It does not solve the governance questions: training data provenance, alignment behavior, EU AI Act conformity, security review. That evaluation is work — but with an open model it's at least possible. With a proprietary API, all you can do is trust the terms of service.
Third: the real question is once again an architecture question. Whether Kimi 3 lives up to its benchmarks, whether Huawei hardware ever becomes a realistic option in Europe, whether the next strong open model comes from China, France, or the U.S. — you can't know any of that today. What you can decide is whether your systems are built model-agnostic. If you're hard-wired to one vendor today, you can't react to any of these developments without kicking off a re-platforming project.
And This Is Exactly Where nopex Comes In
The World AI Conference confirms what we designed nopex for from day one: the model landscape has become geopolitical, and the only robust answer is vendor agnosticism in your own architecture.
Because nopex decouples application logic from the specific model provider, the rising Chinese stack becomes what it should be for you: an option you can evaluate soberly on performance, price, compliance, and risk — rather than a bet you're forced to take or refuse. Open models go where sovereignty and cost tip the scales; proprietary frontier models go where they deliver real value. And if the geopolitical weather turns — in whichever direction — that's a configuration decision, not an architecture decision.
Whether China's "public good" is a gift or a Trojan horse is not a question you need to answer today. You just need to make sure your architecture leaves you the choice.
