China’s latest artificial intelligence champion is making a very specific claim about why the country still trails the United States in frontier AI. DeepSeek founder Liang Wenfeng argues that the real strategic advantage in AI is not better algorithms or superior talent, but sheer access to compute and the ability to use it efficiently.
Why Liang Wenfeng’s argument matters now
Liang recently spent nearly four hours with investors laying out DeepSeek’s strategy and his view of the global AI race, an unusually detailed look into one of China’s most closely watched labs. The timing is important. The United States continues to dominate high-end GPU supply and major cloud platforms, while China faces persistent constraints on advanced chips and relies on a fast-evolving domestic ecosystem that many outsiders still doubt. As IBM’s recent adjustments highlight, AI-focused data-centre infrastructure is becoming increasingly critical for companies looking to maintain competitive advantages.
Four hours of strategy reveal China’s AI race constrained by GPUs, not ambition
When a leading Chinese founder says the core gap with the United States is now primarily about resources rather than talent, it signals a shift in how Chinese AI leaders see their own trajectory. Liang frames the race not as a contest of ideas, but as a contest of how much compute can be deployed and how cleverly it can be used.
A long-standing compute gap between China and the United States
Liang describes the difference between Chinese and American AI labs in stark terms. In his view, all visible gaps in talent, model capability, and applications can ultimately be traced back to differences in compute resources. Talent is not the bottleneck, he says. Instead, fewer opportunities to run large-scale experiments mean fewer chances to develop top-tier researchers, which he calls a talent gap rooted in a compute gap.
He repeatedly summarizes the situation as China being roughly one to two years behind the United States while operating with only around one twentieth of the compute. That ratio is crucial. It suggests that Chinese labs are trying to match or approach frontier performance levels with an order of magnitude less hardware than their American counterparts.
DeepSeek itself has moved quickly to expand its infrastructure. Liang and related transcripts describe the company ramping to about twenty thousand H equivalent compute cards, mostly Nvidia accelerators, many of which were delivered only in recent months. Even with that expansion, he still sees a clear gap between what DeepSeek can access and what the largest United States labs deploy.
How DeepSeek quantifies its disadvantage
Liang does not treat the compute problem as a vague handicap. He gives specific multipliers. He estimates that even the best Chinese training setups are roughly twice as inefficient as the best overseas systems in terms of model structure and training dynamics. In practical terms, that means needing about twice the compute to reach similar performance.
He then adds a second layer. On data efficiency, he believes Chinese labs may also lag by a factor of about two, requiring roughly double the training data and associated compute to match the results of American frontier models. Combined, those two factors imply a fourfold increase in compute requirements for Chinese teams aiming at comparable outcomes.
Liang also points to architectural gaps. He describes the Llama style architecture as roughly two generations behind cutting-edge overseas designs when it comes to training efficiency and inference cost. Closing that kind of structural gap is not just about buying more chips. It requires new model designs and training methods that can squeeze more value out of every unit of compute.
Algorithmic efficiency as DeepSeek’s core strategy
Faced with these disadvantages, DeepSeek has embraced algorithmic efficiency as a central pillar of its strategy. Liang talks about narrowing the training and data efficiency gaps so that Chinese models no longer require four times the resources to keep up. That focus shows up in DeepSeek’s R1 reasoning model, which has become a benchmark for how far efficiency can be pushed.
Independent comparisons and industry analyses show that DeepSeek R1 can rival or surpass OpenAI o1 on several reasoning benchmarks while offering dramatically lower usage costs. One detailed comparison notes that DeepSeek R1 charges roughly zero point fifty-five United States dollars per one million input tokens and a little over two United States dollars per one million output tokens. By contrast, OpenAI o1 pricing is reported at around fifteen United States dollars per one million input tokens and sixty United States dollars per one million output tokens.
That puts DeepSeek R1 at roughly twenty to fifty times cheaper for large-scale deployments depending on caching and workload patterns. For startups, enterprises, and independent researchers, this cost profile changes who can realistically experiment with frontier-level reasoning models. It also illustrates Liang’s broader point. If you cannot win on raw hardware quantity, you try to win on how efficiently you turn limited compute into usable intelligence.
DeepSeek’s public posture aligns with this strategy. Liang emphasizes open-source releases and a technology-first culture, positioning the lab as more interested in structural constraints such as compute, data, and export rules than in short-term valuation or hype. That does not mean DeepSeek ignores commercial opportunities. It recently raised a fund exceeding fifty billion renminbi, which gives it more capital to buy chips and build infrastructure. Yet the narrative remains centered on efficiency and long-term advantage rather than quick monetization.
The geopolitics of GPUs and cloud infrastructure
Liang’s comments sit inside a broader geopolitical context. United States export controls and domestic procurement rules have made it difficult for Chinese companies to buy the most advanced Nvidia chips in volume, pushing them toward domestic options and older generations of hardware. In his telling, chips cannot simply be bought in China at the scale or specification that leading United States labs enjoy, and capital investment in compute is also smaller.
He argues that this resource gap expresses itself everywhere. Model capability, product quality, and even the development of local talent are all downstream of how much compute can be deployed. That is why he sees United States dominance in GPU supply chains and global cloud infrastructure as the foundation of American leadership in frontier models, with Chinese labs forced to innovate under persistent computational scarcity.
At the same time, Liang is cautiously optimistic about the domestic chip ecosystem. He has suggested that within about one year of real-world deployment, the perception that Chinese chips are unusable or immature will begin to fade. He still expects a significant performance and timing gap on the chips themselves, describing a fourfold difference plus roughly two years in lag, but believes the ecosystem gap can close faster through practical experience and iteration.
AGI as a tide and compute as the real moat
Perhaps the most revealing part of Liang’s philosophy is how he talks about artificial general intelligence. He describes AGI as an impersonal tide, something no single company can truly own, and argues that structural innovation is required to navigate that tide. In this view, lasting advantage does not come from one clever model or one breakthrough paper. He also insists that this future must be built with open source models, fair profit, and visible restraint, so that AI’s benefits are widely shared rather than tightly controlled. It comes from securing and efficiently exploiting the largest pools of compute over time.
That perspective reframes familiar debates. Many discussions about AGI focus on alignment, safety, or specific architectural innovations. Liang does not dismiss those issues, but he consistently pulls the conversation back to the infrastructure level. Who has the cards, how many experiments they can run, how fast they can iterate, and how efficiently they can turn compute into capability. For him, these are the real levers of long-term advantage.
Implications for technology, businesses, and society
For technology development, Liang’s argument implies that compute planning becomes a strategic skill on par with model design. Labs that can architect systems to achieve frontier performance with a fraction of the chips will be structurally advantaged in a world where hardware availability is uneven. This puts a premium on research into efficient architectures, training techniques that reduce waste, and data selection strategies that maximize signal per token.
For businesses, DeepSeek R1 illustrates how cost-efficient reasoning models can reshape competitive dynamics. When a frontier-level model is twenty to fifty times cheaper to use than a leading Western alternative, companies that previously could not afford large-scale deployment suddenly have options. That can expand access to advanced automation, decision support, and coding assistance, especially in markets where budgets are tight and foreign cloud use is politically sensitive or legally constrained.
For society, the compute-centric framing raises both opportunities and risks. On the opportunity side, it suggests that innovation is not locked to one country forever. If Chinese labs can continue to close efficiency gaps and scale domestic hardware ecosystems, the global AI landscape could become more multipolar, with different regions contributing distinct strengths. On the risk side, it implies that countries which fall behind in compute infrastructure may find themselves structurally excluded from participating in frontier AI, regardless of the talent or ideas they possess. That concentration of capability could deepen existing economic and political inequalities.
Risks, contradictions, and what to watch next
Liang’s theory is intentionally bold, and the detailed transcripts show that it occasionally contradicts itself. In one exchange, he declares that all differences in talent and applications trace back to compute, then in another acknowledges that capital investment is also smaller and that the chips themselves have a performance and time lag. The reality is that multiple structural factors intersect. Hardware, capital, policy, talent, and organizational culture all shape outcomes.
There is also a strategic risk in focusing too heavily on compute. If future breakthroughs reduce the importance of brute force scaling or rely on novel paradigms that demand different kinds of hardware, today’s infrastructure advantage might matter less than expected. Liang partially addresses this through his emphasis on originality and structural innovation. He warns that China’s true gap is not just measured in years but in the difference between originality and imitation, and argues that some exploration is inescapable if Chinese AI is to move beyond following.
For outside observers, several signals are worth watching. First, how quickly DeepSeek and other Chinese labs can narrow their estimated twofold training and data efficiency gaps. Second, whether domestic chips deliver on the promise of closing ecosystem gaps, even if raw performance still lags. Third, how sustained cost advantages for models like DeepSeek R1 influence adoption patterns, both within China and globally.
Key takeaways
- Liang Wenfeng frames the United States advantage in frontier AI primarily as a function of compute resources rather than talent or ideas, with China operating roughly one to two years behind on about one twentieth of the compute.
- He estimates that Chinese labs currently face around a twofold gap in training efficiency and a similar gap in data efficiency, implying roughly four times the compute is needed to match top overseas results.
- DeepSeek has responded by making algorithmic efficiency its core strategy, reflected in the R1 reasoning model which offers frontier-level performance at twenty to fifty times lower usage costs than OpenAI o1 in many scenarios.
- United States control over advanced GPUs and cloud infrastructure remains a structural constraint on Chinese labs, but Liang expects domestic chip ecosystems to mature quickly even if performance gaps persist.
- His depiction of AGI as an impersonal tide leads to a simple conclusion. In the long run, durable AI advantage will belong to those who can secure the largest pools of compute and use them with the greatest efficiency.
Conclusion
America’s apparent edge in artificial intelligence looks overwhelming, yet DeepSeek founder Liang Wenfeng argues that the real difference is more mundane and more structural: the United States simply has much more computing power, and that is what shapes the global AI landscape today. His claim matters because it shifts the debate away from secret algorithms or unique genius and toward infrastructure, industrial policy, and hard limits on what even the most talented teams can build.
Why computing power has become the real fault line
In a recent four hour discussion, Liang described computing power as the biggest constraint he faces and framed the main gap with the United States as access to resources rather than superior technical know how. He has been unusually explicit about the numbers, estimating that DeepSeek suffers roughly a twofold deficit in training efficiency compared with leading international labs and a similar twofold gap in data efficiency. Put together, that means a four times computing disadvantage just to reach comparable outcomes, even if the underlying ideas are similar.
This is a familiar story to anyone who watched the field change after the release of GPT 3 in 2020, when it became clear that progress in large language models was governed by a combination of model size, training data, and sheer compute. Liang himself points to that moment as a turning point, noting that once the scaling direction was understood, massive computing power became the obvious requirement for competitive work at the frontier. In that world, the limiting factor is no longer whether a team can design clever architectures, but whether they can afford and physically assemble enough high end chips and supporting infrastructure to explore those ideas at scale.
Ultimately, his argument reframes the AI race as a contest over infrastructure rather than a search for a single breakthrough idea. In this telling, Americas lead rests on abundant and relatively unrestricted computing capacity, not on a mystical technical gap that other countries could never bridge. That perspective turns export controls, cloud access, power availability, and hardware policy into decisive levers of power, because each one directly affects who can train and deploy large models and who cannot. It also implies that algorithmic innovation alone cannot fully close the divide: without comparable compute and energy, rivals will remain constrained no matter how creative their models or ambitious their visions over the coming decade.
How the United States built its infrastructure advantage
The numbers behind Americas compute advantage are stark. Estimates from economic and technical analyses suggest that the United States controls roughly three quarters of global high end AI supercomputer capacity, while China holds about fourteen percent and the European Union under five percent. Looking beyond supercomputers, the United States hosts around forty five percent of all data centers worldwide and more than half of global hyperscale capacity, the very large facilities run by major cloud providers that are essential for training and serving frontier models.
This lead did not appear overnight. American technology firms began building large data center campuses years before most other regions recognized how central they would become for cloud computing and later for AI. Private investment flowed into both hardware and software, backed by deep capital markets and a regulatory environment that, despite its flaws, allowed rapid expansion of energy hungry infrastructure in multiple states. Policymakers now explicitly describe US strength in advanced AI as rooted in research and development, investment, and infrastructure, highlighting supercomputers and compute capacity as key elements of national advantage.
There is also an upstream factor that is easy to miss but increasingly important. Training and running frontier models requires dependable electricity at large scale, not just chips and racks. This has prompted early work in the United States on securing long term access to firm low carbon power, such as nuclear and advanced renewables, to support future AI data center buildouts while managing climate and grid reliability concerns. Infrastructure in this context means chips, networks, buildings, and power plants working together, which raises the bar for any country that hopes to catch up.
DeepSeek as a case study in constrained innovation
DeepSeek itself illustrates both the importance of compute and the ways a talented team can partially offset a hardware disadvantage through efficiency. Before export controls tightened, Liang and his associates accumulated significant resources, including the operation of one of Asias earliest clusters with around ten thousand Nvidia A100 chips and further access to tens of thousands of newer accelerators. Capital from his financial career funded major supercomputer purchases between 2019 and 2021, giving DeepSeek a foundation that many Chinese startups lacked.
Even so, DeepSeek has been public about building models under tight constraints. For a widely discussed reasoning model, the company reports using just 2,048 of the high end chips now restricted for the Chinese market, compared to tens of thousands of similar devices often used by leading American firms. The training budget was around six million dollars in computing costs, an order of magnitude less than the figures associated with some recent Western frontier models. Through architectural choices and training tricks, DeepSeek claims to deliver results comparable to competitors at much lower inference cost, with its latest reasoning system being up to twenty seven times cheaper per query than a rival American model in certain tests.
These efficiency gains are not just a marketing story. They are a deliberate institutional response to a reality where DeepSeek cannot count on indefinitely adding more chips to solve each technical problem. In an environment of constrained computing resources, the company has been forced to improve the efficiency of its model architecture, training procedures, and inference stack so that each unit of compute goes further. Lower cost per query allows the same hardware to support larger models, more experiments, and broader deployment, narrowing the distance to better resourced American laboratories without eliminating it.
Yet the underlying gap remains. Export controls on high end chips limit DeepSeek’s ability to scale clusters and maintain parity over time, and Liang is candid that embargoes, rather than a lack of funding or talent, are his primary challenge. His team continues to explore building its own large computing clusters inside China and may eventually invest in custom chips, but even those plans operate within a global supply chain where American and allied policy choices shape which components are available and at what price. The case underlines how policy, not just technical creativity, now defines the frontier of AI capability.
Infrastructure, policy, and the shape of the AI competition
Viewed through this lens, the AI competition between the United States and China begins to look less like a race to invent a magic algorithm and more like a struggle to build and maintain enormous digital and energy infrastructure. Export controls on advanced semiconductors directly reduce the total compute available to Chinese labs, forcing them either to innovate around scarcity or find ways to circumvent restrictions. Access to foreign cloud providers matters because it can temporarily soften these limits, although that access itself can become a target of future policy.
Cloud infrastructure, domestic data centers, and cross border network connectivity all determine how widely and reliably large models can be deployed once trained. Power and permitting rules influence where new AI facilities can be built and how quickly they come online. For American policymakers, this turns decisions about grid expansion, transmission lines, and nuclear licensing into strategic choices that affect national AI competitiveness, not just local energy markets. For China and other countries, similar questions arise around how much to invest in domestic chip capacity, how to secure long term energy, and whether to prioritize a few national champions or a broader ecosystem.
At the same time, it would be a mistake to dismiss algorithmic innovation and research culture. Compute is necessary but not sufficient. Efficient architectures, better training objectives, and smarter data pipelines can significantly reduce the amount of hardware needed to hit a given performance level, which is exactly the path DeepSeek has tried to pursue. Countries and firms with limited hardware can still influence the field through open research, clever methods, and standards for safety and evaluation, even if they cannot match the very largest models in raw scale.
The reality is that infrastructure and ideas now coevolve. Breakthroughs in algorithms can change how valuable a given cluster is, while new hardware designs and energy sources reshape what kinds of models become feasible. Liang’s comments highlight that when hardware access is asymmetric, even globally shared ideas do not translate into equal capabilities. That has implications for economics, security, and governance since it means the ability to execute at scale is concentrated in a few jurisdictions with the right combination of chips, data centers, and power.
What to watch next
Several practical takeaways emerge from this debate.
First, compute should be treated as a strategic resource, much like advanced manufacturing or energy, rather than an invisible background assumption. Businesses that depend on large models need to understand where their compute comes from, how exposed it is to policy changes, and whether they should diversify across providers or regions.
Second, efficiency will remain a central competitive dimension. DeepSeek’s experience suggests that constrained teams can achieve surprising results by focusing relentlessly on training efficiency, data quality, and inference optimization, even when they face a hard ceiling on total available hardware. That focus is likely to spread as more governments and companies confront the financial and environmental costs of ever larger clusters.
Third, the geopolitics of chips, cloud services, and power grids will increasingly shape who sets the pace in AI. Export controls, investment in domestic fabrication, and long term energy policy are now intertwined with questions about safety, openness, and alignment, since they determine which actors can realistically operate at the frontier.
Liang Wenfeng’s claim that Americas main AI advantage is access to computing power is therefore more than a complaint about embargoes. It is a reminder that beneath every headline model sits a vast physical base of chips, cables, buildings, and power plants that takes years to build and is not easily copied. Over the next decade, the shape of the AI race will be decided not only in research papers and product launches but in the quieter work of permitting, grid planning, data center construction, and strategic choices about who gets to tap into that compute and on what terms.
