China's open-weights AI strategy is rapidly gaining global dominance while US companies keep their models locked down. Why open is winning the AI race and what it means for the future of technology.

The New Reality of the AI Race

The global artificial intelligence race is crossing a critical inflection point. China's open-weights AI models are not just catching up to their Western counterparts — they are beginning to take the lead. In the summer of 2026, Moonshot's Kimi K3 and Alibaba's latest Qwen variants are approaching the frontier models from OpenAI and Anthropic, and at a fraction of the cost.

Ben Werdmuller's article "American AI is locked down and proprietary. It's losing" crystallizes the central thesis: American AI companies are pursuing a closed, proprietary strategy that leads to long-term failure. Chinese companies, by contrast, are releasing model weights openly, ensuring rapid global adoption and ecosystem entrenchment. This shift has profound implications for businesses worldwide, including in emerging tech markets.

Ben Thompson, writing on Stratechery, digs deeper into the economic mechanics. He notes that while Kimi K3 costs $3 per million input tokens and $15 per million output tokens — cheaper than OpenAI's $5/$30 — the real competition is not just about price per token. It's about tokens versus intelligence. Different models require different amounts of reasoning tokens to reach correct answers, and efficiency in that regard is becoming a key competitive differentiator.

Why the Locked-Down Strategy Fails

According to Werdmuller's analysis, AI models as standalone products have very little moat beyond brand loyalty and superficial switching costs. A user can be on ChatGPT today and switch to Claude tomorrow with minimal workflow disruption. For developers accessing models via API, swapping endpoints and using the same prompts is trivial.

The real moat lies in enterprise services — contracts, enterprise system integrations, security guarantees, and compliance certifications. But when Chinese models are released openly, they create a parallel ecosystem where any company can self-host them on their own infrastructure. This fundamentally changes the competitive dynamics.

U.S. export controls on GPUs intended to cripple China's AI ambitions have paradoxically accelerated the open-weights strategy. Chinese companies have enough compute to train frontier models but cannot offer the kind of global-scale centralized services that OpenAI and Anthropic provide. So they release their models openly — turning a disadvantage in compute distribution into a massive advantage in global adoption.

The numbers are staggering. a16z partner Martin Casado noted in The Economist that there is an 80% chance any given startup is using Chinese models. Nathan Lambert at Interconnects has argued that Chinese models are now poised to take the global lead. This rapid shift from lagging to leading has happened in less than two years.

Open Infrastructure Always Wins

History teaches us that open technologies almost always win at the infrastructure layer. Linux, Apache, Kubernetes, Python — every major infrastructure revolution has been driven by open ecosystems. The same pattern is now playing out in AI.

Open-weights models are not "open source" in the classical sense (the training data and methodologies remain proprietary), but they are portable and permissionless. They can be hosted anywhere, modified, experimented with, and fine-tuned for specific use cases. This permissionless innovation creates a flywheel effect: more developers use the models, more tooling is built around them, more improvements are contributed back, and the ecosystem compounds.

Thompson's analysis on Stratechery adds a crucial economic dimension. Contrary to popular belief, open models are not "free" — the Cost of Goods Sold (COGS) for AI is very real. Running inference on a model costs money directly proportional to usage. If it costs 50 cents in compute to generate $1 in revenue, then $100 million in revenue requires $50 million in compute costs. The advantage of open models is not zero COGS — it's zero R&D costs for the user, combined with the freedom to optimize the serving infrastructure independently.

Quality Gap Is Closing Fast

The single most important factor that has protected American AI companies has been the quality advantage. OpenAI's GPT-4 and Anthropic's Claude have outperformed open alternatives — until now. That gap is closing rapidly.

Moonshot's Kimi K3, according to The Verge, "can go toe-to-toe with the best from OpenAI and Anthropic at a fraction of the cost." Alibaba's Qwen models are demonstrating similar capability. Benchmark after benchmark shows Chinese open-weights models approaching or matching frontier performance on reasoning, coding, mathematics, and general knowledge tasks.

Thompson's framework is instructive here. He argues that we are rapidly approaching a state where "intelligence for many economically beneficial tasks is in fact a commodity." When anyone can build the same CRUD application using multiple models, the primary differentiator becomes cost structure — and Chinese open models have an undeniable advantage in this dimension.

The key insight is that tokens are not a commodity — a token from Kimi is not the same as a token from Sol (OpenAI's latest). What is fungible is the intelligence constructed from those tokens. If both Kimi and Sol arrive at the correct answer, the answer itself is interchangeable. The number of tokens required to reach that answer — the "token efficiency" — becomes a direct contributor to COGS differences. Models that require fewer reasoning tokens to reach correct answers will win on cost, even if their per-token price is higher.

The Paradox: Locked-Down America, Open China

Werdmuller highlights a striking irony: we perceive China as a locked-down society — and in many ways it is. There are legitimate concerns about how these models reflect Chinese government perspectives and content restrictions. Try asking them about Tiananmen Square, as Radio Free Asia noted. Yet in AI technology, it is American companies that are maintaining tight control over their technology while Chinese companies release theirs as openly as possible.

This reversal of expectations matters. The United States built much of its modern tech dominance on open internet principles — open protocols, open standards, permissive licensing. The shift toward locked-down AI models represents a departure from what made American tech successful. Meanwhile, China has adopted the playbook that America abandoned.

For Georgian companies evaluating AI tools, this creates important considerations around data sovereignty, security, and geopolitical alignment. Open-weights models allow organizations to keep data within their own infrastructure — critical for financial institutions, healthcare providers, and government agencies. A Georgian bank can run Kimi or Qwen on local servers without any data crossing international borders, while still accessing frontier-level AI capabilities.

Economic Implications and the Path Forward

The economic stakes are enormous. According to the Economic Policy Institute, a significant portion of U.S. economic growth is currently driven by AI investment spending. Werdmuller warns that if this spending bubble bursts — which he believes is inevitable given the underlying dynamics — the economic consequences could be severe.

Closed business practices for a technology with no real moat but significant ecosystem benefits is an obviously losing strategy. Chinese companies' open, collaborative approach wins where American lockdown loses. The incentives in the U.S. are misaligned: companies chase first-order profits rather than ecosystem benefits, while the government tries to compensate through forcible measures like export controls.

Thompson points out that for AI companies, the path to profitability in a commodity market is not through charging higher prices — it's through having a superior cost structure. The companies that can deliver intelligence at the lowest cost will dominate. In this framework, open-weights Chinese models that benefit from the entire ecosystem's optimization efforts have a structural advantage over closed, proprietary alternatives where all optimization must happen internally.

For SiTech Georgia and other tech companies in emerging markets, this trend is a net positive. We can now offer our clients AI solutions built on the best available open models — high quality, low cost, with full data control. The barrier to entry for participating in the AI revolution has never been lower. What matters now is talent, creativity, and making the right architectural choices.

The coming years will likely see a bifurcated AI market: premium enterprise services on one side (OpenAI, Anthropic, Google), and open, permissionless models on the other (Kimi, Qwen, Llama). If history is any guide, the open ecosystem will generate more innovation, more applications, and ultimately more economic value. The question is not whether open will win — it's whether the U.S. will adapt its strategy before the window closes, and whether smaller economies like Georgia can position themselves to ride the wave of open AI innovation.

📖 Source