China Doesn’t Need Better AI Models to Win the Market
I said I’d write this article, so here it is. I believe that in the race for AI, China already has a structural advantage. Not in raw model performance. Not in frontier research. In something far more consequential for long-term market capture.
I’m going to make this argument through the lens of Disruptive Strategy, a framework developed by Clayton Christensen at Harvard. I studied this framework extensively, and the more I look at what’s happening in the AI industry, the more I see a textbook disruption unfolding in real time.
The Disruptive Strategy Lens
For those unfamiliar with Clayton Christensen’s work, the core idea is deceptively simple. Incumbents in a market tend to improve their products along a trajectory that eventually overshoots what most customers actually need. They chase the high end because that’s where margins are. Meanwhile, a new entrant appears at the bottom of the market with a product that’s “worse” by every traditional metric, but cheaper, simpler, and good enough for a growing segment of users.
The incumbent ignores this entrant because the low end isn’t profitable enough to defend. The entrant improves. Hardware gets better. The “good enough” product gets better too. By the time the incumbent realizes what’s happening, the entrant owns the mainstream market.
This pattern has played out in hard drives, steel, smartphones, and most famously in how Linux, once dismissed as a hobbyist toy, now runs the vast majority of the internet’s infrastructure. What matters isn’t who builds the most powerful thing. What matters is who captures the market while the powerful thing overshoots.
In AI right now, the incumbents are Anthropic, OpenAI, and Google. The low-end entrants are Chinese open-source models. The disruption is already underway.
AI Hasn’t Saturated Yet, But Surplus Is Starting
At the moment, AI has not reached market saturation. We haven’t arrived at the point where the market has fully absorbed AI solutions and there’s nothing left to improve. There’s still enormous hunger for better models, better tooling, better integration. The race is very much on.
An analogy helps here. Think about the smartphone industry. There are smartphones that reach the majority of customers at very affordable prices: $150 Xiaomis, $200 Samsungs. Then there are flagships costing $1,200 or more. An iPhone with the thinnest body, five camera lenses, 100x zoom, and computational photography that nobody fully uses. That top tier represents surplus. Performance that exceeds what most people actually need.
The AI industry hasn’t reached full surplus yet, but certain niches are getting there. For most business use cases — summarization, content drafting, customer support, data extraction, code assistance — the frontier models are already overshooting. According to multiple enterprise surveys, roughly 70-80% of production AI use cases don’t require frontier-level capabilities. They need reliability, low latency, and reasonable cost.
This is exactly the gap where disruption enters.
Where Surplus Shows Up: “Good Enough” Tasks
Let’s take a concrete example. Consider generating stories for children. You could use Claude 3.5, GPT-4o, or Gemini Ultra. These are frontier models with extraordinary reasoning capabilities, nuanced style control, and massive context windows. But for a children’s story? DeepSeek handles it easily. So does Alibaba’s Qwen. So does 01.ai’s Yi. These models produce perfectly good, creative, engaging children’s content without breaking a sweat.
More than that, DeepSeek is open source. You can deploy it on your own infrastructure. You control the data. You control the costs. You don’t depend on an API that might change pricing tomorrow, or a terms-of-service update that restricts your use case.
The pricing difference tells the story. DeepSeek V3 API costs are roughly $0.27 per million input tokens. GPT-4o runs around $2.50 to $5.00 per million tokens, depending on the tier. That’s a 10-20x cost difference. For a startup building an educational app that generates thousands of stories per day, this isn’t a marginal consideration. It’s the entire business model.
This pattern extends far beyond children’s stories. Email drafting, basic translation, FAQ generation, product descriptions, internal documentation. The list of “good enough” use cases is growing faster than the list of tasks that genuinely require frontier reasoning.
The Real Wedge: Open-Source Weights and Deployment Economics
The market trajectory is straightforward. Frontier models from Anthropic, OpenAI, and Google will continue advancing. They’ll get better at complex reasoning, multi-step planning, and agentic workflows. China’s open-source ecosystem will continue producing models that trail in raw performance but cost a fraction and ship with open weights.
As these frontier labs push further into surplus territory, the open-source models capture clients who don’t want or need that surplus. They want acceptable performance at a lower cost with more control. Textbook low-end disruption.
So why won’t the top models and their performance surplus resist this in the long run?
One fundamental reason. Open source.
All top frontier models right now are closed source. Closed weights, closed training data, closed architecture details, closed everything. Their competitive advantage is built on secrecy. And this advantage is eroding.
The open-source community has more collective resources than any single company. Tens of thousands of engineers and researchers worldwide deploy their own models every day, run experiments with fine-tuning, quantization, distillation, and niche adaptations. These things happen constantly, across thousands of repos and projects.
Anthropic or OpenAI cannot possibly absorb all of this innovation. To maintain their edge, they would need to either hire every talented open-source contributor (impossible at scale), or somehow integrate the community’s output into their closed ecosystem. Both paths are structurally blocked. You can’t take an open-source community’s innovation and lock it behind a paywall — the community will simply fork and continue.
The chip sanctions angle reinforces this. Because China faces US restrictions on high-end GPUs like the H100, Chinese teams have been forced to innovate on software efficiency and training optimization. DeepSeek’s V3 model was reportedly trained at a fraction of the compute cost of comparable Western models. Constraint breeds efficiency. While US companies rely on brute-force compute scaling, Chinese teams are learning to do more with less, and that efficiency translates directly into cheaper deployment for end users.
As hardware continues to improve and the cost of inference drops, models that were “good enough” yesterday become “surprisingly capable” tomorrow. The gap between open source and closed frontier narrows with every GPU generation.
Why China Benefits Disproportionately
China will succeed in this disruption because Chinese companies are structurally integrated into the open-source community. This isn’t a single company story. DeepSeek, Alibaba’s Qwen, 01.ai’s Yi, and others are all releasing competitive open-weight models. It looks less like a coincidence and more like a national strategy: build the ecosystem, benefit from the ecosystem, let the ecosystem compound.
When a Chinese company releases an open model, thousands of developers worldwide improve it, adapt it, and extend it. The company gets that value back for free. Compare that to Anthropic or OpenAI, which operate on a completely different framework. Their business model demands closed systems. They can’t embrace open source without undermining the very thing their investors are paying for: proprietary advantage.
The moment closed-source frontier labs enter what the open-source community offers, they’ve already conceded the game.
This is the classic Innovator’s Dilemma. OpenAI and Anthropic cannot pursue the low-end market, not because they lack the technology, but because their business model, their investor expectations, and their pricing structure make it irrational for them to do so. They’re structurally trapped in the high end. Just like premium smartphone makers couldn’t respond to Xiaomi capturing emerging markets with $200 phones that were “good enough” for most users.
What This Means for Frontier Labs
I want to be clear about the limits of this argument. For AGI research, complex scientific reasoning, drug discovery, and cutting-edge agentic systems, frontier closed models will likely maintain an advantage for some time. This isn’t about who builds the most impressive demo.
It’s about who captures the market.
And the market is being captured from the bottom up, through open weights, low costs, and local deployment. Frontier labs have limited strategic options: they can try to open-source selectively (Meta’s playbook with Llama), acquire key players, or compete on integration and ecosystem. But the structural dynamics favor the disruptor.
It’s a bit of a game over for closed-source-only strategies. Just a matter of time.
China Doesn’t Need Better AI Models to Win the Market
China Doesn’t Need Better AI Models to Win the Market
I said I’d write this article, so here it is. I believe that in the race for AI, China already has a structural advantage. Not in raw model performance. Not in frontier research. In something far more consequential for long-term market capture.
I’m going to make this argument through the lens of Disruptive Strategy, a framework developed by Clayton Christensen at Harvard. I studied this framework extensively, and the more I look at what’s happening in the AI industry, the more I see a textbook disruption unfolding in real time.
The Disruptive Strategy Lens
For those unfamiliar with Clayton Christensen’s work, the core idea is deceptively simple. Incumbents in a market tend to improve their products along a trajectory that eventually overshoots what most customers actually need. They chase the high end because that’s where margins are. Meanwhile, a new entrant appears at the bottom of the market with a product that’s “worse” by every traditional metric, but cheaper, simpler, and good enough for a growing segment of users.
The incumbent ignores this entrant because the low end isn’t profitable enough to defend. The entrant improves. Hardware gets better. The “good enough” product gets better too. By the time the incumbent realizes what’s happening, the entrant owns the mainstream market.
This pattern has played out in hard drives, steel, smartphones, and most famously in how Linux, once dismissed as a hobbyist toy, now runs the vast majority of the internet’s infrastructure. What matters isn’t who builds the most powerful thing. What matters is who captures the market while the powerful thing overshoots.
In AI right now, the incumbents are Anthropic, OpenAI, and Google. The low-end entrants are Chinese open-source models. The disruption is already underway.
AI Hasn’t Saturated Yet, But Surplus Is Starting
At the moment, AI has not reached market saturation. We haven’t arrived at the point where the market has fully absorbed AI solutions and there’s nothing left to improve. There’s still enormous hunger for better models, better tooling, better integration. The race is very much on.
An analogy helps here. Think about the smartphone industry. There are smartphones that reach the majority of customers at very affordable prices: $150 Xiaomis, $200 Samsungs. Then there are flagships costing $1,200 or more. An iPhone with the thinnest body, five camera lenses, 100x zoom, and computational photography that nobody fully uses. That top tier represents surplus. Performance that exceeds what most people actually need.
The AI industry hasn’t reached full surplus yet, but certain niches are getting there. For most business use cases — summarization, content drafting, customer support, data extraction, code assistance — the frontier models are already overshooting. According to multiple enterprise surveys, roughly 70-80% of production AI use cases don’t require frontier-level capabilities. They need reliability, low latency, and reasonable cost.
This is exactly the gap where disruption enters.
Where Surplus Shows Up: “Good Enough” Tasks
Let’s take a concrete example. Consider generating stories for children. You could use Claude 3.5, GPT-4o, or Gemini Ultra. These are frontier models with extraordinary reasoning capabilities, nuanced style control, and massive context windows. But for a children’s story? DeepSeek handles it easily. So does Alibaba’s Qwen. So does 01.ai’s Yi. These models produce perfectly good, creative, engaging children’s content without breaking a sweat.
More than that, DeepSeek is open source. You can deploy it on your own infrastructure. You control the data. You control the costs. You don’t depend on an API that might change pricing tomorrow, or a terms-of-service update that restricts your use case.
The pricing difference tells the story. DeepSeek V3 API costs are roughly $0.27 per million input tokens. GPT-4o runs around $2.50 to $5.00 per million tokens, depending on the tier. That’s a 10-20x cost difference. For a startup building an educational app that generates thousands of stories per day, this isn’t a marginal consideration. It’s the entire business model.
This pattern extends far beyond children’s stories. Email drafting, basic translation, FAQ generation, product descriptions, internal documentation. The list of “good enough” use cases is growing faster than the list of tasks that genuinely require frontier reasoning.
The Real Wedge: Open-Source Weights and Deployment Economics
The market trajectory is straightforward. Frontier models from Anthropic, OpenAI, and Google will continue advancing. They’ll get better at complex reasoning, multi-step planning, and agentic workflows. China’s open-source ecosystem will continue producing models that trail in raw performance but cost a fraction and ship with open weights.
As these frontier labs push further into surplus territory, the open-source models capture clients who don’t want or need that surplus. They want acceptable performance at a lower cost with more control. Textbook low-end disruption.
So why won’t the top models and their performance surplus resist this in the long run?
One fundamental reason. Open source.
All top frontier models right now are closed source. Closed weights, closed training data, closed architecture details, closed everything. Their competitive advantage is built on secrecy. And this advantage is eroding.
The open-source community has more collective resources than any single company. Tens of thousands of engineers and researchers worldwide deploy their own models every day, run experiments with fine-tuning, quantization, distillation, and niche adaptations. These things happen constantly, across thousands of repos and projects.
Anthropic or OpenAI cannot possibly absorb all of this innovation. To maintain their edge, they would need to either hire every talented open-source contributor (impossible at scale), or somehow integrate the community’s output into their closed ecosystem. Both paths are structurally blocked. You can’t take an open-source community’s innovation and lock it behind a paywall — the community will simply fork and continue.
The chip sanctions angle reinforces this. Because China faces US restrictions on high-end GPUs like the H100, Chinese teams have been forced to innovate on software efficiency and training optimization. DeepSeek’s V3 model was reportedly trained at a fraction of the compute cost of comparable Western models. Constraint breeds efficiency. While US companies rely on brute-force compute scaling, Chinese teams are learning to do more with less, and that efficiency translates directly into cheaper deployment for end users.
As hardware continues to improve and the cost of inference drops, models that were “good enough” yesterday become “surprisingly capable” tomorrow. The gap between open source and closed frontier narrows with every GPU generation.
Why China Benefits Disproportionately
China will succeed in this disruption because Chinese companies are structurally integrated into the open-source community. This isn’t a single company story. DeepSeek, Alibaba’s Qwen, 01.ai’s Yi, and others are all releasing competitive open-weight models. It looks less like a coincidence and more like a national strategy: build the ecosystem, benefit from the ecosystem, let the ecosystem compound.
When a Chinese company releases an open model, thousands of developers worldwide improve it, adapt it, and extend it. The company gets that value back for free. Compare that to Anthropic or OpenAI, which operate on a completely different framework. Their business model demands closed systems. They can’t embrace open source without undermining the very thing their investors are paying for: proprietary advantage.
This is the classic Innovator’s Dilemma. OpenAI and Anthropic cannot pursue the low-end market, not because they lack the technology, but because their business model, their investor expectations, and their pricing structure make it irrational for them to do so. They’re structurally trapped in the high end. Just like premium smartphone makers couldn’t respond to Xiaomi capturing emerging markets with $200 phones that were “good enough” for most users.
What This Means for Frontier Labs
I want to be clear about the limits of this argument. For AGI research, complex scientific reasoning, drug discovery, and cutting-edge agentic systems, frontier closed models will likely maintain an advantage for some time. This isn’t about who builds the most impressive demo.
It’s about who captures the market.
And the market is being captured from the bottom up, through open weights, low costs, and local deployment. Frontier labs have limited strategic options: they can try to open-source selectively (Meta’s playbook with Llama), acquire key players, or compete on integration and ecosystem. But the structural dynamics favor the disruptor.
It’s a bit of a game over for closed-source-only strategies. Just a matter of time.
Dan Gurgui | A4G
AI Architect
Weekly Architecture Insights: architectureforgrowth.com/newsletter
Archives
Categories
Archives
Recent Post
China Doesn’t Need Better AI Models to Win the Market
March 5, 2026Running Claude on Your Laptop: A Wardley Map of When Local LLMs Will Replace the API
February 11, 2026The Per-User Product: How LLMs Are Forcing a New SaaS Architecture
February 3, 2026Categories
Meta
Calendar