The AI Race Is Being Timed in Months and Built in Years
Washington is counting down an AI race in months while the power and talent it needs arrive in years — and China is the only one compressing that timeline.
Anthropic told Washington in May that it has 12 to 24 months to secure a decisive lead before China reaches near-parity by 2028 [1]. PJM Interconnection, the grid operator for 13 states, is running a different clock: its emergency plan to add 15 gigawatts of power doesn't target generation until June 2031 [2]. The two numbers sit side by side and refuse to line up. The fast clock is real, and it is busy. The SAFE CHIPS Act, introduced in December, would deny AI chip export licenses for 30 months — a wall that runs to mid-2028 [3]. The legislative energy around the race has gone almost entirely into that wall: export bans, enforcement, the question of which chips may cross which border. Anthropic's warning framed the contest as a matter of plugging leaks — chip smuggling and offshore data centers as the ways China closes the compute gap [1]. On this clock the problem is diplomatic, and the answer is a ban. The slow clock doesn't care about any of it. PJM has told federal regulators that current load growth cannot continue, with a projected 6-gigawatt shortfall [4]. Pennsylvania's governor has tightened the rules so data center developers must procure their own power [4]. The only NRC-approved small modular reactor design won't have its first operational project online until 2033, in Romania [5]. And the people who build all of it are missing: the chip industry faces a talent shortage exceeding 100,000 jobs a year, a gap that runs on the generational timeline of who studies engineering and who is allowed to immigrate [6][7]. None of these units are months. They are years and decades. China is solving the same constraints on a different clock. The state offers 50% electricity subsidies to domestic AI chip users in Gansu, Guizhou and Inner Mongolia, pushing power costs down to 0.4 yuan per kilowatt-hour [8]. The Ministry of Industry and Information Technology is coordinating a 31-province strategy with 2,185 EFLOPS of computing capacity, up 177% in a year, with provinces like Xinjiang and Inner Mongolia hosting compute on clean energy [9]. Zhipu AI has stood up a 1-gigawatt data center running entirely on Chinese-made silicon [10]. These are levers a federal system cannot pull — no American agency can order 31 states to allocate compute or fix the price of a kilowatt-hour. The American grid operator, facing the same surge of demand, can only describe what it cannot supply.
The present trajectory of rapid load growth, tightening supply and rising capacity costs is not sustainable. — PJM Interconnection
The SAFE CHIPS Act promises a wall by mid-2028. PJM promises power by June 2031. Between them sit three years in which the diplomatic clock will have declared the race decided and the physical clock will still be waiting for its first megawatt.
- 1. Anthropic Warns U.S. Faces 24-Month AI Race Window Amid Trump-Xi Summit
- 2. PJM Interconnection Launches Plan for 15 Gigawatts of Power
- 3. Senators Introduce SAFE CHIPS Act to Block AI Exports
- 4. Governor Josh Shapiro Tightens Pennsylvania Data Center Regulations
- 5. NuScale Power Develops SMRs to Power AI Data Centers
- 6. Chip Engineer Warns of AI Energy and Talent Gaps
- 7. Economists and Tech Leaders Warn of AI Skills Gap
- 8. China Offers 50% Power Subsidies to Boost Domestic AI Chips
- 9. China Deploys Coordinated National AI Industry Strategy
- 10. Zhipu AI Launches 1-Gigawatt Data Center Using Domestic Silicon