Many predicted China’s decline. Instead it quietly built an industrial machine, and now it’s aiming AI at the world.
For decades Western commentators treated China’s rise like a short-term fad: cheap labor, cheap goods, eventual collapse. That view missed something else, a patient, state-backed climb up the value chain. The same toolkit that helped China dominate batteries and EV components is now being applied to artificial intelligence. Boards, investors and policymakers should expect competition that is strategic, not accidental.
China 1.0 → 2.0 → 3.0 in one tight line
China 1.0: low-cost manufacturing and mass assembly. China 2.0: deliberate industrial upgrading into higher-value electronics, solar panels, batteries and EVs. China 3.0: a push to build domestic compute, models and exportable AI services, an attempt to turn generative models into another national capability.
The U.S. Energy Information Administration (EIA) documents the country’s systemic dominance across the battery value chain: China domestically produced about 18% (33, 000 short tons) of the world’s mined lithium in 2023 and accounted for roughly 79% (1.27 million short tons) of global natural graphite production in 2024. In 2023 China was responsible for about 74% of global battery pack and component exports and controlled nearly 85% of battery cell production capacity by monetary value (U.S. Energy Information Administration, Gavin Clark). The EIA also notes China’s role in processing capacity and overseas resource investments.
Those numbers are not a lucky streak. They reflect a playbook: secure upstream inputs, dominate midstream processing and components, and export downstream products, backed by export controls, subsidies and state investment. The EIA documents export restrictions on certain graphite products that began in late 2023 (reported by Reuters and cited in the EIA analysis). That is an example of policy used to protect and amplify a supply chain advantage.
Why the battery analogy matters for AI
Batteries show how control of inputs, processing and export channels compounds into global market power. AI relies on different inputs, chips, datacenters and high-quality training data, but the levers are similar: targeted public investment, preferential access to resources, and export or market-entry strategies that shape who wins at scale.
There are early signals. Columnist Larry Elliott highlighted reports that a Chinese startup, Moonshot, launched a model called Kimi K3 in July and temporarily suspended new subscriptions two days later because demand outstripped its computing capacity. Those reports suggest low-cost or no-cost models can match Western rivals and scale quickly. At the same time, commentators such as Edoardo Campanella of UniCredit have described Beijing’s approach as “staying close to the technological frontier, diffusing its models throughout the global economy, and setting standards across international markets, ” and warned that a successful strategy could “not just disrupt the global economy, but also the global balance of power.”
“Staying close to the technological frontier, diffusing its models throughout the global economy, and setting standards across international markets.”, Edoardo Campanella (UniCredit)
These anecdotes and quotations are useful signals, but they are not, on their own, proof of AI parity with U.S. incumbents. AI leadership requires a sustained mix of compute scale, advanced semiconductors, model research, quality training data, secure distribution channels and international adoption. Treat early headlines as red flags to investigate, not as definitive outcomes.
Reality checks: the three big constraints and the right metrics
Order the risks by their real impact and watch the metrics that matter:
- Compute and chips (highest impact). Leading models need massive inference and training capacity, GPUs/TPUs and advanced-node fabs. Metric to watch: shipping and price trends for datacenter GPUs (vendor statements from NVIDIA, TSMC and cloud providers), and the share of global advanced-node (5nm/3nm) fab capacity.
- Technical parity and verification. Marketing claims must meet independent benchmarks. MMLU (a multi-task academic benchmark) and HELM (Stanford’s Holistic Evaluation of Language Models) are two public evaluations to follow. Metric to watch: model rankings on third-party leaderboards and peer-reviewed benchmark reports.
- Adoption and governance barriers. Procurement rules, data-sovereignty laws and national security reviews limit where certain models can be used. Metric to watch: number of public-sector procurement bans/whitelists and cross-border data-transfer rulings affecting AI deployment.
Strategic implications for leaders, cut the noise, prioritize the response
Two clear points for executive teams:
- Low price is not neutral. A state-backed model priced at or below cost can build an ecosystem, attract users, harvest data and set de facto technical standards that lock in customers. That changes the economics for incumbents in AI for business, AI agents and inference-as-a-service.
- Industrial policy works. Democracies do not have to copy China wholesale, but they should accept that strategic public investment combined with market incentives and safeguards is a legitimate tool to secure critical capabilities.
Priority, practical moves (what to start this quarter)
- Public inference infrastructure (12-18 months pilot). Fund shared inference capacity and cloud credits for startups and SMEs to prevent compute from becoming a monopoly gate. How to start: tie a pilot to existing cloud providers and offer time-limited credits targeted at enterprise workloads and trusted research.
- Secure and diversify semiconductor supply (2-4 year program). Use targeted tax credits, low-interest loans and public-private co-investment to accelerate advanced packaging and select fabs. How to start: identify priority nodes and coordinate incentives with allied governments to reduce single-source risk.
- Independent model evaluation and certification (6-12 months). Fund a neutral consortium (universities + labs) to run repeatable benchmarks, publish model cards, and certify models for sensitive sectors. How to start: seed a national lab program that publishes regular, auditable results to industry.
- Allied procurement coordination (18-24 months roadmap). Harmonize trust frameworks and procurement standards with partners so governments and regulated industries can buy with confidence. How to start: scope a joint procurement pilot for healthcare or critical infrastructure with clear standards for model explainability and data handling.
Measure these moves against clear KPIs: reduced inference cost for certified use cases, increased domestic share of advanced packaging, published benchmark scores, and the number of joint procurement contracts issued by allies.
What to monitor this quarter, an immediate board checklist
- GPU availability and pricing announcements from major vendors and cloud providers (NVIDIA, AWS, Alibaba Cloud).
- New public benchmark results on MMLU and HELM and major model leaderboards (Stanford HELM, Hugging Face model cards).
- Official export-control or tariff actions affecting chips, graphite or other strategic inputs (MOFCOM, U.S. Commerce Department, EU announcements).
- Allied procurement or standards initiatives (joint R&D funding, trusted model lists, cross-border data frameworks).
Key takeaways, questions you should be asking
-
Is China already dominating AI the way it dominates batteries?
Not yet. China clearly dominates parts of the battery and EV supply chain (U.S. Energy Information Administration data). Its AI position is expanding, but technical parity depends on chips, datacenter capacity and independent benchmarking, metrics that still show a mixed picture.
-
Can free or subsidized Chinese models undercut Western tech valuations?
Yes, in principle. Sustained low-price offerings supported by state resources can compress margins and slow revenue growth for incumbents; the ultimate effect hinges on pricing sustainability, compute economics and how much of the world accepts those models.
-
Should democracies copy China’s industrial policy?
No wholesale copying. But targeted public investment in compute, chip capacity, skills and independent evaluation is pragmatic, combine market dynamism with strategic public support to secure resilience and preserve democratic safeguards.
-
How much weight should I give the Moonshot / Kimi K3 headlines?
They’re a signal, not definitive proof. Reports that Moonshot’s Kimi K3 launched in July and briefly suspended new subscriptions due to demand are notable; independent technical benchmarking and company disclosures are required before treating these reports as evidence of parity.
China’s climb was deliberate and long-term. The lesson for Western businesses and governments is not to panic, but to plan. The contest over AI will be settled across supply chains, datacenters, standards bodies and procurement offices, not only on model leaderboards. Boards that track the right metrics and act with calibrated urgency will find opportunity. Those that mistake headlines for the whole story risk being surprised by the next move.