AI doesn't need subsidies - it needs government restraint
Private money is already stampeding into AI. NVIDIA blew past a $5 trillion market cap on AI demand, and Microsoft, Google, Amazon and Meta invested roughly $230 billion in AI infrastructure in 2024, with plans for $325 billion this year. Venture capital is swarming the space. None of that signals a funding shortfall.
Yet federal policy is moving as if AI is starved for capital. The America's AI Action Plan and recent executive orders direct agencies to mobilize loans, grants and tax incentives, even for mega data centers drawing 100+ megawatts. Canada and the EU are doing their own versions with multibillion-dollar programs.
What this gets wrong
Subsidies make sense when markets miss something society needs - not when markets are already investing aggressively. Rigorous policy starts with a specific, measurable market failure: costs or benefits that private firms can't capture. It then targets that gap - nothing more.
Instead, we're seeing a scattershot approach: subsidize compute, data centers, training, you name it - without showing where private capital falls short. That's not strategy. That's spending.
A smarter test for public money
Before you back another AI initiative, run this quick test. If you can't answer "yes" to at least one of these with evidence, don't spend.
- Is there a clear market failure (national security, basic science, public goods)?
- Are benefits broad and hard for private investors to monetize?
- Will the program create knowledge, standards or security that industry won't fund on its own?
- Is this time-limited, with measurable outcomes and a real exit?
Good targets vs. bad targets
- Good targets: defense applications, basic research, safety and security testing, secure compute for classified work, open benchmarks and standards, privacy-preserving methods, evaluations that reduce systemic risk.
- Bad targets: commercial data centers and cloud buildouts that private firms are already racing to fund; generic workforce subsidies without proof of net public benefit.
Remember Solyndra
Industrial policy often spreads money based on geography and interest groups, not technical due diligence. We've seen how that ends. Solyndra's $535 million loan guarantees were justified as risk-taking for a worthy cause - then collapsed.
That rationale doesn't fit AI. Unlike early solar, AI is saturated with private capital. Piling taxpayer dollars on top of already-profitable projects won't move the needle on the public interest.
China is a security problem - commercial subsidies aren't the solution
Concerns about Chinese government progress in AI are real and should focus attention. But the pressure point is defense and high-end research - not commercial data centers that Big Tech and VCs already fund. If you want to strengthen the U.S. position, invest where private firms underinvest and where spillovers are national, not just corporate.
Practical guardrails for agencies
- Define the market failure: Put it in writing. Cite evidence.
- Target the gap: Limit awards to clearly public-benefit outputs (e.g., open evaluations, standards, secure testbeds).
- Use performance-based funding: Milestones, not lump sums. Tie payments to verifiable results.
- Insist on competition: Open solicitations, transparent criteria, independent review.
- Add time limits and sunsets: Every program gets an end date and an off-ramp.
- Protect taxpayers: Include clawbacks and equity/royalty provisions where appropriate.
- Publish outcomes: Share data, methods and lessons learned unless classified.
What to do this quarter
- Shift funds toward defense research, basic science and security evaluation programs with clear spillovers.
- Stop subsidizing commercial data center capex. If agencies need compute, buy it as-a-service and competitively bid it.
- Back open testing, benchmarks and red-teaming that reduce systemic risk and inform procurement.
- Streamline permitting and interconnection for critical infrastructure rather than paying firms to build what they already plan to build.
Standards and discipline help you say "no" with confidence
Use established playbooks. Benefit-cost principles and independent cost estimation are built for this. If a proposal can't clear those bars, it shouldn't clear your desk.
Upskill your workforce - without subsidizing industry
Agencies do need internal capability. That's different from subsidizing the commercial market. If your team needs practical training by role, start small, set outcomes and fund what improves mission delivery.
See curated AI courses by job role for planning internal upskilling.
AI may reshape the economy. That's the point - and the warning. In a capital-heavy boom, restraint beats subsidies. Put taxpayer money where markets won't go on their own: defense, basic research and security.
Your membership also unlocks:
