The Real AI Race

Private AI companies are suddenly calling on the federal government to regulate artificial intelligence. OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei recently told the United Nations Security Council that the threat is real and the time to act is now. Bill Gates has echoed the demand for legislation. On the surface this looks like responsible leadership. Look closer and a different picture emerges.

These firms are not operating as ordinary private businesses. They burn cash at extraordinary rates while capturing enormous public subsidies, the bulk of it directed toward military applications rather than the public benefits they advertise.


The Scale of Public Money

Federal AI contracting has exploded. According to analysis of contracts compiled by Leadership Connect and examined by Brookings researchers, the value of funds obligated for AI-related work rose from $261 million (through early 2022) to $675 million (through 2024) and then to $7.2 billion in the 2026 data. Potential award value jumped even more dramatically to $91.8 billion. Nearly all of it, 98.9 percent of potential value, is concentrated at the Department of Defense.

The Defense Department’s share grew from 76 percent of potential value in the earlier period to 95 percent and then to almost the entire pie. Other agencies, including HHS and NASA, became rounding errors by comparison. Large established vendors now dominate, though smaller firms still hold many individual contracts. The market is maturing rapidly around national-security priorities.

Recent awards illustrate the point. The Pentagon’s Chief Digital and Artificial Intelligence Office issued contracts valued at up to $200 million each to Anthropic, Google, OpenAI, and xAI for advanced AI capabilities. Separate agreements brought SpaceX, Nvidia, Microsoft, Oracle, Amazon Web Services, and others onto classified networks for “lawful operational use.” The Department of Defense has requested tens of billions for AI and autonomous systems, including $54 billion for autonomous weapons development alone.

Meanwhile, leaked financials show OpenAI recorded $20.9 billion in operating losses in 2025 even as revenue grew roughly 250 percent to about $13.1 billion. Losses continued into early 2026. These companies survive on investor capital and government contracts, not sustained commercial profits.

Rhetoric Versus Reality

Company leaders and advocates highlight potential upsides like cancer detection algorithms, productivity gains, new jobs. The historical parallel currently being regurgitated by dupe is the cotton gin: technology that multiplies output rather than simply displacing workers, the comparison gives me the ultimate ick and belies the fact that the dominant federal spending pattern is not public-health research or civilian infrastructure. It is defense.


When Anthropic resisted language allowing the military “any lawful use” of its models citing concerns about domestic surveillance and fully autonomous lethal weapons the Pentagon designated the company a supply-chain risk. The message was clear: take the money on the government’s terms or watch it go elsewhere. The company’s later public calls for regulation and slower development arrived after the checks had cleared and the strategic position was secure.

The race is framed as a contest with China for “AI supremacy.” That framing prioritizes strategic advantage over the everyday improvements sold to the public. Clean water, reliable baseload power, and affordable energy are treated as secondary constraints rather than primary goals. 

Data centers already consume water and electricity at rates that strain local grids. In Pennsylvania, the restart of Three Mile Island Unit 1 is being positioned to supply power for AI facilities while residents face sharp increases in electricity costs after a PJM capacity auction that jumped from $29.50 to $270.35 per megawatt-day, adding roughly $2.2 billion in costs for the state.

Nuclear power itself offers a useful comparison. After Three Mile Island in 1979, the commercial industry largely froze for decades despite the absence of detectable public health effects from the limited release of radioactive gases. Military nuclear propulsion has operated safely for generations. Civilian nuclear remains politically and regulatorily difficult even as AI data centers receive preferential access to restarted capacity.  The technology that could deliver zero-emission, 24/7 power is constrained; the technology racing toward more powerful autonomous systems is subsidized.


What Exactly Is Being Built?

The public is told AI will identify disease, create jobs, and solve hard problems. The contracts and budget lines point elsewhere: decision superiority across warfighting domains, autonomous systems, intelligence analysis, and enterprise tools that keep the military “AI-first.” Even if secondary civilian benefits appear, the primary subsidized direction is clear.

History is not destiny. Nuclear weapons produced both destruction and, later, nuclear power. Yet the United States has made it extraordinarily hard to expand civilian nuclear power while pouring public money into AI systems whose most well-funded applications are military. The question is not whether AI can produce useful tools. It already has. The question is whether the current allocation of tax dollars, energy, and water is building the future the public is being sold or something closer to the “Vishnu” image of overwhelming destructive capacity that its own developers sometimes invoke when they warn of existential risk.

Altman, Amodei, and Gates are right that the technology is powerful and that self-regulation alone is insufficient. They are less forthcoming about the degree to which federal procurement has already shaped the technology’s trajectory and the companies’ balance sheets. Once the money is taken and the military contracts are signed, calls for democratic oversight arrive late. The public is funding the race. It is entitled to a plain-English accounting of what that money is actually buying.


Next
Next

The Most Important Thing Wyndham Clark Won Wasn't the U.S. Open