Leaderboards Are the Wrong Scoreboard
Every week brings another model ranking, another benchmark headline, another frontier comparison between US and Chinese labs. A new CSIS report argues this framing misses the actual contest. The real race is not about which lab ships the smartest model — it is about who builds the global infrastructure that produces tokens: the energy, chips, data centers, cloud platforms, and models that together form what the report calls the “token supply chain.”
That is the core argument of Tokenpolitik: How the United States Can Compete with China to Build the Global AI Stack, written by Benjamin Jensen and Yasir Atalan of the CSIS Futures Lab. The paper names a shift that has been building for two years: AI competition has moved from single technologies to entire systems.
What Is a Token, Strategically?
A token is the atomic unit a neural network ingests and generates — roughly a word in text models, a fragment of sound in audio models. In engineering terms it is mundane. In geopolitical terms, CSIS argues, it is the new oil: the fundamental industrial output of the data centers and power grids now being built around the world.
The report lays out the stack in five layers:
- Energy — power generation and grid capacity for data centers
- Chips — accelerators, from frontier GPUs to lower-cost domestic alternatives
- Infrastructure — data centers, networking, cooling, cloud platforms
- Models — frontier closed models and open-weight models
- Applications — the government and enterprise systems consuming tokens
Every operational constraint in modern AI — inference cost, context window limits, latency — is ultimately a token economics problem. Whoever generates tokens most efficiently, at the largest scale, and wires the most regions into their supply chain, sits at the commanding heights of the next international system.
Infrastructure Was Always the Strategy
The framing is not new in kind, only in domain. The Suez Canal redefined trade geography between Europe and Asia; Britain initially opposed it, then bought Egypt’s stake once the canal became essential to imperial commerce. During the Cold War, the US-backed INTELSAT and Soviet Intersputnik competed to wire the world with satellite communications — a contest over standards, prices, technical assistance, and developing-country relationships.
AI is the Suez moment of this century, and the analogy matters: the strategic question is never just who builds the infrastructure, but who finances, governs, secures, and benefits from it.
China Plays Full-Stack; Washington Assembles Pieces
The report credits China with a coherent full-stack industrial policy: an $8 billion national AI investment fund followed by a reported $295 billion AI infrastructure plan, state-backed chip champions (CXMT’s blockbuster market debut, YMTC, Huawei), and — crucially — open-weight models (DeepSeek, Qwen, Kimi, GLM) that perform near frontier levels at lower per-token prices. CSIS and RAND researchers describe open-weight models as instruments of soft power: they shape developer communities, technical education, standards, and downstream applications, and many US enterprises now experiment with them (typically hosted on US platforms).
Washington, the report says, has assembled many of the pieces without a grand strategy: the AI Action Plan, the American AI Exports Program, Pax Silica, DFC and EXIM financing (including the new ExportAI initiative), and bilateral deals with the UAE and Saudi Arabia. What is missing is a cross-agency theory of victory that coordinates State, Commerce, Treasury, and development finance around one goal: making secure, affordable AI capacity available to strategic partners before China’s Digital Silk Road locks in the Global South.
What This Means for the Industry
Three implications matter beyond geopolitics:
- Open-weight models are now a diplomatic asset. Model releases are no longer just product launches; they are infrastructure plays. Expect open-weight roadmaps to be treated with the strategic weight of export policies.
- The Global South is the battleground. The contest shifts from export controls to financing, standards, and turnkey packages. Who wires Brazil, Indonesia, Kenya, and the Gulf will shape the default AI stack of the next decade.
- Capital is the weapon. The same dynamic we analyzed in AI’s $2 trillion off-balance-sheet debt wave now has a geopolitical counterpart: development finance is becoming AI strategy.
What to Watch Next
For builders and investors, the practical takeaway is to track infrastructure decisions, not benchmark tables:
- Which governments sign turnkey AI deals — and with whom?
- How fast open-weight adoption grows in enterprises (and what security teams learn from it, as with GLM-5.3’s DNS vulnerability discovery).
- Whether Washington moves from one-off programs to the coordinated financing architecture CSIS proposes.
Tokenpolitik is a policy proposal, but the underlying signal is structural: the unit of AI competition has changed from models to stacks, and from benchmarks to infrastructure. Companies that treat AI as a supply chain question — not a model selection question — are the ones positioned for what comes next.
