The US-China AI Race: Who Wins When Models Go from Scarce to Abundant?

Key insights:

• The AI gap between Chinese and US frontier models is closing: The performance gap between top US and Chinese frontier models has shrunk to just 2.7%, according to Stanford's 2026 AI Index.
• Cost vs capability: While US labs dominate in valuations, Chinese open-weight models are driving down inference costs drastically.
• The distillation factor: The rapid closing of this gap is highly contested, with US labs identifying industrial-scale "distillation attacks" used to siphon capabilities.
• Value migration: As models commoditise, investment value is poised to shift toward application layers, sovereign data infrastructure, and strict compliance environments.

A portfolio manager in Menlo Park, voice tight with something between excitement and dread may be asking his analyst something akin these days: "Have you seen the DeepSeek pricing?" Not something unusual today, while speaking of DeepSeek a year ago was a niche interest in the valley.

We have all spent eighteen months watching frontier AI valuations climb past the trillion-dollar mark, and now it looks to be the time when we are first hearing genuine uncertainty in a Silicon Valley voice.

When the quality gap narrows

In April 2026, after DeepSeek released DeepSeek V4 Pro: The model itself was highly captable although probably not a foundational leap, but priced at a fraction of a dollar per million output tokens, currently $0.87 per million output tokens via Openrouter. For context, that's roughly 70 to 90% cheaper than the latest top-tier American models. The quality gap? 2.7% according to Stanford's 2026 AI Index.

That single data point captures the strategic inflection reshaping private markets. This isn't about which lab builds the smartest chatbot. It's about whether good enough at a fraction of the cost can do to AI what it did to solar panels, batteries and electric vehicles. The answer may determine where billions in capital flow over the next five years.

The cost-curve pattern is well-known

We've seen this pattern before. Historically, the pioneer often builds the pathway and the fast follower then frequently industrialises the cost curve. America invented the solar panel; China now manufactures the majority of global capacity at significantly lower cost. The same logic appears to be playing out in AI, but faster and with vastly more capital at stake.

For anyone looking to justify OpenAI's $852 billion valuation or Anthropic's $965 billion (as of its May 2026 Series H), the narrow quality lead is just one part of the investment case.

If "good enough" proves sufficient for most commercial and consumer use cases, it could well be that massive foundation model valuations could face structural pressure

Table 1: US vs China Frontier AI

Note: Pricing Data from Openrouter as of mid-2026, pricing can change

The infrastructure question that’s ongoing

The capital flowing into AI infrastructure has reached a scale that is hard to ignore: Goldman Sachs estimated in May 2026 that AI-related capital expenditure might hit $765 billion this year, rising to $1.6 trillion by 2031.

While this massive $765 billion figure commingles both data centre infrastructure and model training costs, the pure model training expenditure is precisely where Chinese labs are achieving parity for less.

However, capital intensity and value capture rarely move in lockstep. Firms spending capex are not necessarily also the ones who stand to extract value if ever models fully commoditise. In many prior technology waves, value migrated away from the core component like the PC towards the integration and application layers like software.

The question remains: Is an AI foundation model more like an infrastructure commodity layer or service that can easily be interchanged by flipping an API switch, or is it a truly unique application that cannot be uncoupled?

China's strategy: give it away

Whilst Silicon Valley has locked its models behind API keys and enterprise contracts, Beijing has made its capable models free, open and globally accessible.

For a developer for example in Lagos, Jakarta, or São Paulo looking for a capable language model, they can either pay premium Western prices or use open-weight alternatives at a fraction of the cost.

This is where the AI race stops being about benchmarks and becomes about ecosystems. Every developer who builds on an open model, every start-up that integrates it, and every government that deploys it in public services can become a node in a network that Beijing does not need to own to influence.

However, this rapid advancement is highly contested. In early to mid-2026 leading US labs and a white house adviser, alleged that Chinese labs (including Alibaba's Qwen, Moonshot AI and DeepSeek) were running industrial-scale "distillation" attacks. By feeding millions of queries into American frontier models, they are allegedly using US outputs to quietly train their own domestic systems.

The strategic mirror-image happened when the US Commerce Department ordered Anthropic to suspend access to its most advanced models (Fable 5 and Mythos 5), leading American labs to suspend access to their most advanced models for all foreign nationals. which has since been allowed to be released.

The sovereignty countermove: walls up, models locked

China's diffusion strategy is colliding with the regulatory and trust barriers of the West's most lucrative markets. Cheap and open models cannot easily overcome sovereignty concerns when the stakes are sensitive data, critical infrastructure and national security.

The European Union's AI Act, which came into force in 2025, layers a risk-based compliance regime on top of the already stringent GDPR. High-risk AI systems that are covered by the act face transparency, accountability and auditability requirements enforced by both the new EU AI Office and national data protection supervisors.

This is China's mirror-image of the access problem that hobbles its robotics and EV companies in Western markets, a point we explored in our analysis of the US-China robotics race.

DeepSeek is the cautionary tale here. A wave of regulatory pushback emerged against the model across several jurisdictions.

For Chinese labs, this represents a structural ceiling. Low cost and open weights are challenging tools to use to overcome a sovereignty and privacy trust deficit in the West's highest-margin sectors. At present, Chinese models are positioned to lead in the Global South, whilst US models are positioned to anchour high-trust Western enterprise environments. Concurrently, European nations are aggressively integrating regional models into critical infrastructure, such as the French intelligence agency DGSI replacing Palantir with domestic data firm ChapsVision, which operates alongside sovereign AI providers like Mistral.

The UK, interestingly, has charted a third path. Britain established its AI Safety Institute in 2023 to research frontier risk and inform governance (it is explicitly not a regulator). By early 2026, the government's AI Opportunities Action Plan had pivoted from safety towards sovereignty, launching a Sovereign AI Unit backed by up to £500 million to invest directly in UK AI firms.

For more on Europe's sovereign AI investments, see our analysis of N-Scale: Europe's Billion-Dollar Bet on AI Sovereignty.

Notably, the UK has imposed no blanket bans, making its regulatory stance a critical space to watch.

Where value migrates

The US-China AI race is now being fought across three distinct but interlocking fronts, each with clear investment implications for those building portfolios in frontier technology.

First, the capability premium versus efficiency-scale trade-off: US labs are valued under the assumption that the quality moat is durable. The 2.7% gap suggests otherwise. If "good enough at one-tenth the cost" becomes the dominant paradigm, the premium could evaporate in unregulated sectors. However, sovereignty and compliance barriers are designed to keep the premium intact within regulated Western markets.

Second, the sovereignty premium: Export controls, model access restrictions and the GDPR/AI-Act compliance wall create structural, cycle-independent demand for sovereign, trusted, auditable AI.

Third, the diffusion-layer capture: If models commoditise history suggests value migrates to the integration, inference, orchestration, and compliance layers that sit on top of them. F

We have arrived at a moment when the critical commercial question is no longer just who builds the best model, but who captures value when every model operates at a high standard. With open-weight architectures driving down costs and distillation attacks blurring capability lines, the ultimate winners are positioned to be the platforms that integrate these models into auditable, sovereign, and highly compliant workflows.

For sophisticated investors, looking past the sheer compute capex and focusing on the application and compliance layers may prove to be the most prudent strategy for the decade ahead.