AI Policy Without Compute Policy Is a Shit Show
Every government on Earth is playing in a shit show while compute manufacturers owning AI.
And nobody notices because AI policy sounds smart: regulate the models, mandate safety testing, ensure alignment, guard against AGI. Serious business. Serious constraints. Serious conversations in serious buildings about serious risks.
But it’s all a show on a set someone else owns.
The game; the one that determines who wins the AI economy is happening in fabrication plants, power grids, and supply chains. And governments aren’t even looking at it.
The Policy Inversion Nobody Wants to Name
Here’s what’s actually constraining AI deployment:
Not regulation. Not safety. Not ethics frameworks.
Compute.
Specifically:
NVIDIA’s production capacity (the only company that matters in high-performance chips for five more years)
Electricity availability (you can’t run a 10,000-GPU cluster in a jurisdiction that can’t spare 50 megawatts)
Fab capacity and geopolitical access to TSMC and Samsung
Cooling infrastructure that doesn’t exist in 90% of the world
Supply chains for rare materials
These are the actual, hard limits on who can build models, deploy inference, run agents, and capture AI economics.
But governments write nothing about this.
Instead they write papers about:
Token regulations (meaningless)
Bias testing requirements (slows deployment by weeks; doesn’t change who owns the compute)
Transparency mandates (security theater)
AGI risk frameworks (for products that don’t exist yet)
Meanwhile, the US has 85% of the world’s GPU capacity. China is frantically trying to catch up. Everyone else is renting compute from US companies or locked out entirely.
That’s not policy.
The Perverse Incentive Loop
Here’s where it gets ugly:
When governments focus on model policy instead of compute policy, they accidentally:
Entrench US dominance. Only companies with capital and supply-chain access (read: NVIDIA customers with US backing) can run the training loops and inference clusters that matter. European AI champions? Renting GPUs from AWS. Chinese competitors? Blocked from the latest chips. Everyone else? Playing with last-year’s capacity.
Outsource actual power to private companies. NVIDIA doesn’t need government permission to set chip prices, allocate capacity, or decide which markets get served first. Neither does OpenAI or Anthropic. But they feel the weight of model-policy compliance. So who actually holds the pen? The people controlling compute.
Create a regulatory moat for the few. If you’re already big enough to afford compliance costs. hiring red teams, running safety audits, sitting through regulatory review, model policy hurts your competition more than it hurts you. It’s regulatory capture disguised as safety.
Make the policy physically impossible to enforce. You can’t “regulate” a model that runs entirely in a datacenter in another country. You can strangle access to the chips that make those datacenters possible. But nobody writes policy that way because compute policy requires telling NVIDIA what to do. And democracies have trouble telling capital what to do.
The Economic Geometry
Let me spell out who wins in a compute-constrained world:
Tier 1: Compute monopolists (NVIDIA, TSMC, Samsung, major cloud providers)
They set prices, allocate capacity, decide margins
Government policy literally can’t touch them because we all depend on their infrastructure
They capture 40-60% of the economic value created by AI, before models even exist
Tier 2: First-movers with capital and access (OpenAI, Anthropic, Google, Meta, Microsoft)
They buy compute at scale early
By the time second-tier competitors finish compliance paperwork, the hardware is gone
They own the training data, the trained weights, the distribution channels
They capture another 30-40%
Tier 3: Everyone else
Rents compute at spot prices that reflect scarcity
Cannot compete on model training; must operate at inference scale
Cannot own moats; competes on UX and integration
Captures maybe 5-10% of value and is rapidly compressed toward zero
Tier 0: Governments making policy about the wrong thing
Spends billions on regulation that affects nobody
Has zero levers on the actual constraint
Watches economic power consolidate in private hands while issuing press releases about “ensuring responsible AI”
What Real Compute Policy Would Look Like (But Won’t)
If a government actually wanted a seat at the AI table, it would:
1. Secure chip supply first, regulate models second (or not at all)
Fund TSMC equivalents, ASIC design, power infrastructure
Build compute at cost; make it available domestically at scale
Then you have leverage to do everything else
2. Stop writing model regulations until you can enforce them
A safety standard that only applies to models trained on chips you can’t access is just virtue signaling
Real policy would be: “You can only deploy models running on infrastructure we can audit.”
But that requires owning the infrastructure first.
3. Tie compute access to reciprocal commitments
Europe could say: “You want our 2B person market? You build compute here and hire here and train locally.”
It didn’t; instead it wrote GDPR and AI Acts that just slow down American companies slightly while Europeans use their products anyway.
4. Weaponize power grid policy as AI policy
A 50-megawatt datacenter needs 50 megawatts of electricity. Most jurisdictions can’t spare it.
Real policy is power-first: “Here’s how we build grid capacity. Compute follows.” Instead governments talk about model cards and safety testing while China builds the power plants that make AI deployment possible.
5. Build public-interest compute capacity
Not everything needs to be private. Some compute should be state-owned and made available to researchers, startups, and public-sector agencies at cost.
It would immediately level the playing field.
It’s also something governments can actually do instead of regulating products nobody can build yet.
But none of this will happen because:
It requires spending real money on infrastructure (not polling well)
It requires admitting you’re behind (politically hard)
It requires making decisions today with payoffs in 5 years (incompatible with election cycles)
It means telling NVIDIA and cloud providers that they need to participate in a capacity-sharing ecosystem (they’ll lobby against it)
So instead: more white papers. More task forces. More summits. More regulation of the symptom while the disease runs wild.
The Real Cost: Who Gets Left Behind
The developing world, the 6 billion people not in the US, Europe, or China, gets exactly what’s left over:
Last-generation chips at commodity prices
Inference-only access rented from US clouds
No ability to train anything novel
No local AI industry that captures value locally
Meanwhile, US companies literally just export America’s competitive advantage globally through software and cloud access. It’s genius. It’s also why AI will be a tool for concentrating wealth and power in exactly the places that already have it.
And governments had every chance to change this. They chose instead to argue about whether models should refuse hate speech.
The Truth
You want to know why AI safety and alignment and responsible AI are so talked-about?
Because it’s the one regulatory domain where governments can make noise without actually constraining the companies winning.
It’s safe. It’s future-oriented. It doesn’t step on anyone’s current profit margins. And it sounds moral.
Compare that to compute policy, which would:
Require massive government spending
Threaten private-sector supply chains
Admit we’re behind
Force uncomfortable conversations about economic power
So it won’t happen. And the outcome is locked in:
AI economics will be determined by compute access. Compute access is determined by semiconductors and power grids and capital. Those are owned or controlled by six companies and three countries. Everything else, all the policy, all the regulation, all the earnest conversations about AI safety is just decoration on a fait accompli.
If your government is writing AI policy that doesn’t start with “how do we secure compute capacity,” your government is writing policy for a world that doesn’t exist.
And losing in the one that does.
[Share this if you think compute policy matters more than model policy. Share it louder if your country is currently doing neither.]