For years, the global conversation about artificial intelligence seemed to have a fixed script: the United States as the undisputed leader, driven by technology giants with almost unlimited budgets, and China trying to close the gap based on scale and state investment.
However, 2025 is showing that the picture is much more nuanced (and much more interesting) than previously believed. We have previously tried to answer whether China will win the AI race, but now we take a deeper look at the landscape.
Either way, a profound change is underway: Chinese models already achieve between 80% and 90% of the performance of their American equivalents, but they do so at costs up to ten times lower.
And far from being an isolated phenomenon, it is a growing pattern that is forcing us to rethink how we compete, what “avant-garde” really means, and why efficiency could become the new measure of technological power.
A performance that no longer surprises… because it competes face to face
The quality gap between the AI models of both powers has shrunk rapidly. Recent reports show that the gap in key benchmarks has gone from almost ten percentage points in 2024 to just 2% at the beginning of 2025.
What does this mean for the general public? That models capable of writing complex texts, analyzing documents, generating code or summarizing information (previously the exclusive domain of companies like OpenAI, Google or Anthropic) now have Asian competitors that offer practically indistinguishable results in real tasks.
The key is a change of focus:
- The United States has opted for brute force: massive clusters of GPUs, gigantic models and budgets that easily exceed 100 million dollars per training.
- China, on the contrary, has opted to optimize, reduce and refine: get more out of the available hardware, innovate in architecture and use hybrid algorithms that reduce computational cost.
This change not only closes the gap: it threatens to redefine the pace of the technological race.
Efficiency as a driving force: doing more with much less
Perhaps the most striking fact is that several of the most advanced Chinese models cost between 5 and 10 million dollars to train, compared to hundreds of millions for their Western equivalents.
The DeepSeek-V3 case: when necessity becomes a virtue
This model launched at the end of 2024 was a turning point. Despite training on Nvidia H800 chips limited by export restrictions, it achieved performances comparable to GPT-4o and Claude 3.5 Sonnet. All for an investment of just 5.6 million dollars.
The model went viral not only for its performance, but for what it represents: proof that limits can drive creativity. Even Western experts, such as Andrej Karpathy, publicly praised the project as a brilliant demonstration of algorithmic engineering.
Other examples that show a clear trend
- MiniMax M2: achieves 90% of the performance of GPT-5, but reducing infrastructure spending by more than 80%.
- Doubao-1.5-pro (ByteDance): equals benchmarks of leading models, but at a cost 50 times lower.
- Ant Group models: they compete with those trained on Nvidia hardware, but use chips made in China at a 20% lower cost.
This is no coincidence. China is turning efficiency into a structural competitive advantage.
Silicon Valley adopts Chinese AI (although it does not say it very loudly)
The most striking paradox of this phenomenon is that American companies are integrating Chinese models into their own systems for a very simple reason: they are just as good, but much cheaper.
Some recent examples:
- Airbnb adopted Alibaba’s Qwen model for internal AI tasks. The reason? Its cost per million tokens is a fraction of that of Western competition.
- Social Capital openly praised the Kimi K2 model as “tremendously cheaper and surprisingly capable.”
- Cognition AI, creator of one of the most talked about AI agents of the year, is based on the GLM model of the Chinese Zhipu AI.
The message is clear: when quality is equal, price rules.
The geopolitics of silicon: efficiency gains ground over scale
Despite these advances, the United States remains dominant in one fundamental aspect: infrastructure. It controls around 75% of global supercomputer performance and has the largest GPU clusters in the world.
But this domain has a weakness: it is extremely expensive to maintain and expand. China, forced by export restrictions, has taken another path:
- Massive clusters of “non-premium” chips, like those from Huawei.
- Energy subsidies to make model training cheaper.
- Aggressive optimization at an algorithmic level, reducing the demand for advanced hardware.
What was designed to slow its progress has generated the opposite effect: it has promoted a frugal, more resilient and surprisingly effective innovation ecosystem.
What does this mean for the future of AI?
The outlook for 2025 leaves several clear and, at the same time, disturbing conclusions.
- Technological hegemony is no longer guaranteed: the United States continues to lead in scale, but China has shown that it can compete without needing to replicate that scale.
- Efficiency will be the new currency of power: In a more uncertain economic context, the systems that offer the best performance per euro invested will be the most adopted.
- The industry is becoming multipolar: Actors from India, the Middle East and Europe are already following the Chinese formula: optimize, do not oversize.
- The impact will reach the common user. Lower costs mean cheaper products, faster models and more accessible tools.
The advance of efficiency is not a story of distant geopolitics: it is the reason why tomorrow we will be able to use quality AI in more devices, more countries and more contexts.
And if one thing is clear, it is that this new phase of the technological race will not be a linear marathon, but rather a strategy game where efficiency—and not just power—can be the key to reaching the top.
This post is also available in: