The AI Race Isn't One Race
Trade friction cooled a bit, but the tech sprint is heating up. Export rules on chips and minerals loosened, yet both Washington and Beijing rolled out new models, deals, and use cases. World leaders are in New Delhi for the India AI Impact Summit. The big topic: how to compete without tripping over short-term fixes.
The common question-who wins, the U.S. or China-misses the point. What counts as "leadership"? Are we optimizing for the best models, widest diffusion, or strategic depth? What can be de-risked in global supply chains, and what can't? The wrong answers lead to fragile plans and wasted time.
The Many Races Inside AI
Open vs. Closed
There's a race to build the top closed-source model and another to build the best open-source model. American frontier labs mostly keep weights and methods closed and monetize well. China leans open, pushing models that spread fast and cheap, which appeals across the Global South. Some leading U.S. open-source teams are moving to closed approaches to capture value.
Performance gaps are narrowing in places, but the scoreboard isn't just model quality. Reach, price, and distribution matter. Open models can swarm markets; closed models can capture margin and trust in regulated settings.
AGI vs. Diffusion
U.S. labs aim at artificial general intelligence-systems beyond human-level cognition. China prioritizes mass deployment: driverless logistics, IoT, consumer apps, factory robots. These aims compete at times but also reinforce each other. As models spread, they learn from scale; as frontier models advance, they embed into daily work, classrooms, and clinics.
Where the U.S. Leads-and Where China Scales
Chips and Compute
The U.S. holds the edge in chips, models, and sales, supported by a deep bench of partners. China still dominates legacy semiconductor volume but lags at the frontier. Top U.S. accelerators remain far ahead on performance, and as larger training runs resume with newer architectures, scaling laws are likely to reward labs with the biggest compute pools.
In short: more advanced chips, larger runs, and tighter infrastructure loops mean outsized gains for those with access to frontier compute. That advantage currently leans West.
Capital vs. Energy Constraints
U.S. capital markets are funding the AI buildout at unmatched levels. The bigger constraint is energy and grid capacity: generation, siting, and transmission for data centers. Federal plans call for a step-change in energy build, but state and local rules can slow execution. Some investors are pushing builds in the Gulf, where land and energy are plentiful.
China faces fewer local planning barriers and is adding coal, nuclear, and renewables, with sizable spare capacity projected by 2030. The takeaway: Finance is ready; electrons and permits decide the pace.
The Global Stack Is Interdependent
AI isn't local. Nvidia designs chips in America; TSMC fabricates them in Taiwan; extreme ultraviolet lithography comes from the Netherlands' ASML with German and Japanese components. Etching tools are led by the U.S. and Japan. That chain won't be fully replaced by any single country.
- ASML EUV lithography is a critical step in leading-edge manufacturing.
- TSMC's advanced nodes remain the bottleneck most labs optimize against. See their process overview at TSMC Technology.
The Gulf states are buying stakes and building energy-ready campuses. India is stepping up as a convening hub and talent engine. Israel, Japan, France, South Korea, and the UK are pushing frontier research and standards. Talent remains global: the U.S. attracts and educates at scale; China trains and collaborates widely and is courting more STEM professionals.
Meanwhile, both blocs are rebuilding supply lines: minerals, lithography, packaging, memory, and cloud. Export controls have pinched China's frontier path, but Beijing is closing gaps and still leads in key minerals. Security stakes are rising: dual-use AI is being tested, cyber tools are getting smarter, and some non-U.S. models have produced code with vulnerabilities.
What This Means for Leaders
For Government and Policy
- Compete across every track: models, chips, energy, data, and deployments. Avoid single-metric thinking.
- Clear the energy backlog: fast-track siting and transmission for data centers; align incentives for flexible load and firm generation.
- Lock in coalition supply: minerals, lithography, advanced packaging, HBM, and trusted cloud. Stress-test for single points of failure.
- Procure for diffusion: secure, open interfaces and shared datasets that small agencies and local partners can use.
- Talent mobility: visas, research grants, and public-interest fellowships tied to safety, audits, and evaluation.
Need a focused starting point? See AI Learning Path for Policy Makers and our collection on AI for Government.
For IT Leaders and Developers
- Optimize for memory and bandwidth, not just FLOPs. HBM, interconnects, and storage IOPS will be the next choke points.
- Design for energy intensity: heat management, location strategy, and demand response. Model lifecycle costs with energy first.
- Multi-model, policy-driven inference: route by task, cost, and sensitivity across closed and open options.
- Security by default: model gating, prompt and output filtering, SBOMs, and continuous red-teaming for agents.
- Observability for AI: dataset lineage, run tracking, evals, and rollback plans. Treat models like living systems.
For deeper implementation guides, explore AI for IT & Development.
For Business and Finance
- Allocate like a portfolio: frontier bets (long-horizon) plus diffusion plays (near-term ROI). Measure against cost to serve, not vanity metrics.
- Index to energy certainty: prioritize regions with predictable rates, clear permits, and grid headroom.
- Back open where it compounds distribution; back closed where trust and safety win contracts.
- Insure supply: second sources for compute, memory, and packaging; negotiate pre-build reservations.
For Security and Defense
- Dual-use by design: shared rails for civilian and defense workloads with strict access control.
- Adversarial testing: red-team models against data poisoning, jailbreaks, tool abuse, and agent autonomy failures.
- Cyber with AI on both sides: expect more automated recon and exploit chains; invest in detection and rapid patch pipelines.
Metrics That Actually Matter
- Compute access: available accelerator hours, queue times, and cost per trained token.
- Energy reality: MWh per model lifecycle, PUE, and time-to-permit for expansion.
- Memory and I/O: HBM capacity per node, memory bandwidth, and storage throughput.
- Latency and quality: end-to-end inference latency, task-level accuracy, and safety eval scores.
- Supply resilience: dual-sourcing rates and lead times for chips, HBM, packaging, and critical tools.
- Security: incident rates tied to model misuse, patch time, and red-team coverage.
The Next 12-24 Months
- AI agents with lighter oversight move from demos to workflows; expect productivity jumps and new failure modes.
- Energy becomes a gating factor for growth; siting strategy beats headline funding.
- Memory (HBM) and cloud fabric outpace pure compute as the binding constraints.
- Software efficiency gains stretch older hardware further; don't write off "last gen" too soon.
- Military use accelerates in cyber and autonomy, with lessons flowing back into commercial stacks.
- Chokepoints migrate: from export controls on chips to quotas, memory allocation, and interconnect availability.
Compete in Every Race, For the Duration
There won't be a single finish line. Leadership will be partial and temporary, shifting with new chips, better training runs, and new ways to deploy. Some labs and countries will surge; some will stall.
Practical strategy is simple to state, hard to execute: keep optionality across open and closed models, lock in energy and supply, invest in security, and train people faster than the tech changes. Software advances will squeeze more from the same hardware, and that will keep resetting the game. Plan on change, and win by building systems that benefit from it.
Your membership also unlocks:
