The economics of artificial intelligence just shifted beneath everyone’s feet. Chinese AI laboratories are now producing tokens — the fundamental units of text, code, and data that power every large language model interaction — at a fraction of what Western providers charge, and the gap is growing as agentic AI pushes consumption through the roof.

The token economy nobody saw coming

For years, the AI competition narrative centred on model capability: who could build the most powerful system, pass the hardest benchmarks, generate the most convincing output. That race continues, but a parallel contest has quietly become just as consequential — the battle over token economics.

Since February 2026, Chinese AI models from groups including DeepSeek and MiniMax have overtaken US rivals in token consumption on OpenRouter, the widely tracked platform that monitors usage across large language models. This is not a marginal shift. MiniMax’s M2.5 model, now ranked among the most-used globally by token volume, has seen usage surge 476 per cent in a single month.

Critical Context: Token consumption serves as both a proxy for model adoption and a pricing battleground. When Nvidia’s Jensen Huang described token production as the driver of the AI economy this month, he was acknowledging that raw capability means nothing without affordable scale.

The pricing differential tells the story plainly. Chinese providers like MiniMax and Moonshot charge $2 to $3 per million output tokens. Anthropic’s Claude Sonnet 4.5 sits at roughly $15 for the same volume — a near sixfold gap that compounds dramatically at scale.

Why agentic AI makes token costs existential

The shift from chatbots to AI agents transforms token pricing from a line item into a strategic concern. A chatbot summarising Shakespeare’s Hamlet might consume around 30,000 tokens. An AI agent performing a minor coding task can burn through 20 million.

That hundredfold difference means the cost gap between Chinese and Western providers is no longer academic. As Will Liang, chief executive of Sydney-based Amplify AI Group, put it: “If your agent is burning through millions of tokens a day, even a small per-token price difference becomes a significant line item. That’s a structural tailwind for Chinese labs, and it only grows as agentic adoption scales.”

Strategic Reality: For a UK business running 50 AI agents processing 5 million tokens daily each, the annual cost difference between Chinese and Western providers could exceed £1 million. That figure grows linearly with adoption.

The behavioural shift is already visible among developers. Terry Zhang, a Hong Kong-based developer, described moving from exclusive use of Claude to a mixed approach — spending roughly $50 a day on Moonshot’s Kimi model for 80 per cent of his workload and reserving Claude for complex tasks. The alternative, using Claude for everything, would cost him $900 daily.

The structural advantages behind China’s pricing

China’s token cost advantage is not the result of a single factor but a convergence of policy, infrastructure, and engineering decisions that Western competitors will struggle to replicate quickly.

Energy economics form the foundation. The Chinese government designated “computing-electricity synergy” a national priority in its 2026 work report, explicitly tying energy policy to AI competitiveness. China’s massive investment in renewable energy translates directly into cheaper compute — and cheaper tokens.

Architectural efficiency compounds the advantage. Chinese labs have embraced mixture-of-experts designs that activate only portions of a model for any given query, dramatically reducing computational demand per token. This engineering push was partly born of necessity: US export controls on advanced chips forced Chinese groups to extract more performance from less silicon.

Implementation Note: Mixture-of-experts architectures trade some accuracy for significant efficiency gains. Organisations evaluating Chinese models should test them against their specific use cases rather than relying on general benchmark comparisons.

Corporate strategy is accelerating the trend. Alibaba’s creation of Alibaba Token Hub — a new business group led by chief executive Eddie Wu — signals that China’s largest technology companies view token economics as the defining battleground. “Billions of AI agents are poised to take on an ever-greater share of digital work, each powered by tokens generated by models,” Wu wrote in an internal memo.

Where the advantage breaks down

The cost story is compelling, but it comes with caveats that UK organisations should weigh carefully before restructuring their AI procurement around Chinese providers.

Infrastructure reliability remains unproven at scale. Zhipu AI’s GLM-5 model briefly topped OpenRouter charts in February before usage surged beyond its compute capacity, causing delays and service degradation. The company had to apologise and raise prices. Its shares fell 22 per cent in a single day, wiping more than $10 billion in market value. As one veteran Google developer observed: “The model’s capability matters, but stable compute and service are equally indispensable.”

Warning ⚠️: A cheaper token that arrives late or not at all is more expensive than a premium token delivered reliably. Service-level agreements and uptime guarantees should carry equal weight to per-token pricing in any procurement decision.

Data sovereignty concerns are intensifying. For UK businesses operating under UK GDPR and sector-specific regulations, processing data through Chinese infrastructure raises questions that no pricing advantage can resolve. “Regulators are asking harder questions about where data is processed and under whose jurisdiction it falls,” noted Amplify’s Liang.

Quality trade-offs exist. The efficiency gains from mixture-of-experts architectures sometimes come at the expense of accuracy. For high-stakes applications — legal analysis, medical decision support, financial compliance — the cheapest token is rarely the right token.

Geopolitical risk is structural, not cyclical. The relationship between the UK and China on technology matters is not trending towards stability. Organisations building dependencies on Chinese AI infrastructure are making a bet about geopolitics, not just technology.

What this means for UK organisations

The token price war creates genuine strategic tension for UK businesses. The cost differential is real and growing. Ignoring it means paying a premium that compounds with every agent deployed. Embracing it means accepting risks that may be difficult to quantify.

SME Advantage: Small and medium enterprises have more flexibility than regulated industries to experiment with hybrid approaches — using lower-cost providers for non-sensitive workloads whilst keeping critical processes on trusted infrastructure.

For organisations at different maturity levels:

Early-stage AI adopters should focus on understanding their token consumption patterns before optimising for cost. The difference between 100,000 and 10 million daily tokens changes the strategic calculus entirely.

Active AI deployers should consider tiered approaches: routing high-volume, lower-sensitivity tasks through cost-efficient providers whilst maintaining premium providers for complex or sensitive workloads. This mirrors the approach developers like Terry Zhang are already using.

Organisations at scale need to model their token economics explicitly. At agent-scale consumption, the pricing gap represents a material budget line — one that boards and finance teams will eventually scrutinise alongside traditional infrastructure costs.

The challenges nobody is talking about

Vendor lock-in through token economics. As organisations optimise workflows around specific pricing structures, switching costs grow. The cheapest provider today may not be the cheapest provider in six months, but retooling agent architectures is neither quick nor cheap.

Hidden Cost: Migration between AI providers is not like switching cloud storage. Agent prompts, system instructions, and workflow integrations are often model-specific. Budget for a 3-6 month transition period when evaluating provider changes.

The benchmark problem. OpenRouter data, whilst widely cited, captures only a fraction of global model consumption. The true competitive picture may look quite different from what publicly available metrics suggest. Decisions worth millions should not rest on incomplete data.

Token deflation and margin pressure. If Chinese providers continue driving token prices down, Western providers will face pressure to follow. This is good for consumers but raises questions about the sustainability of the models — both economic and technical — that produce those tokens.

Regulatory divergence. The UK’s approach to AI regulation is still forming. Organisations that build deep dependencies on Chinese AI infrastructure may find themselves on the wrong side of future regulatory decisions they cannot currently predict.

The bottom line

The AI industry’s centre of gravity is shifting from raw model capability towards token economics — who can produce intelligence most cheaply, most reliably, and most compliantly. China has established a clear cost advantage that will persist as long as energy remains cheap, architectures remain efficient, and demand for agentic AI continues to grow.

For UK businesses, the strategic response is not binary. The organisations that will navigate this well are those that understand their own token consumption, maintain provider flexibility, and make deliberate choices about which workloads justify premium pricing and which do not.

Take Action: Audit your current AI token consumption and costs this quarter. Map workloads by sensitivity level and volume. The pricing landscape is moving fast enough that decisions deferred by six months may cost significantly more than decisions made now.

Three factors that will determine whether this price gap persists:

1. Energy policy execution — China’s computing-electricity synergy programme either delivers sustained cost advantages or encounters the scaling limits that renewables deployments often face
2. Regulatory response — UK and EU data sovereignty requirements either create hard barriers to Chinese AI infrastructure or settle into a workable compliance framework
3. Architectural convergence — Western labs either adopt similar efficiency techniques (narrowing the gap) or maintain their accuracy-first approach (preserving the gap but at a premium)

Next steps for UK decision-makers:

• Map current token consumption across all AI tools and agents
• Calculate annual cost differential between current and alternative providers
• Classify workloads by data sensitivity and regulatory requirements
• Evaluate hybrid approaches for non-sensitive, high-volume workloads
• Monitor UK regulatory developments on foreign AI infrastructure

Source: “The rise of China’s hottest new commodity: AI tokens” by Zijing Wu, Financial Times, 26 March 2026. Data visualisations by Haohsiang Ko.

Strategic analysis by Resultsense — Making sense of AI in the UK.