AGI's Physical Limits: The Blind Spot in the Balance Sheet
AI accelerationists ignore three systemic bottlenecks that will define the real AGI timeline
The AGI Gold Rush Has a Problem No One Wants to Talk About
Most conversations about artificial general intelligence (AGI) take place in an epistemic vacuum, in which accelerationists float narratives about cognitive thresholds, software models, and recursive self-improvement. The silicon dreams are seductive: scale up the model, increase context length, chain reasoning steps, and wait for the magic to emerge.
But the dream skips a step. Actually, three.
AGI is not just a software problem. It’s a capex problem. A civilizational logistics problem. A trade war problem. In other words, AGI is a convergence of constraints that accelerationists and financial backers treat as afterthoughts, if they think of them at all.
1. Silicon Capex: The Hardware Arms Race With a Burn Rate
We are now in the capital-intensive phase of AI, where progress is less about code and more about access to chips. Unfortunately, chips are a commodity with half-life decay. Nvidia’s H100s, Blackwell B200s, and whatever’s next, each have a lifecycle of 12–18 months before obsolescence. The result? Billions in annual capital expenditure for compute clusters just to stay in the race.
This is a treadmill, not a flywheel.
An H100 costs around $30,000. You need 10,000 or more for a GPT-5 scale run. That’s $300 million in chips alone. This is before you even power them, cool them, or house them. Multiply that by a lifecycle of 18 months, and the depreciation curve looks like burning cash at scale.
And those chips don’t exist in a vacuum. They need power. Cooling. Land. Fiber. Integration. Permits. Labor. Construction. And none of that is instantaneous.
2. Grid & Civil Infrastructure: The Timeline No One Can Accelerate
A single large data center can draw 100 megawatts or more. This is comparable to a small city. To support this load, you need high-voltage transmission lines, substations, transformers, cooling towers, and water rights. Each of these things has its own permitting timeline measured in years, not quarters.
Here’s the civilizational drag coefficient:
Substation transformers are on 36-month backorder.
Grid interconnect queues are years long in many U.S. regions.
Local water battles have already derailed data center projects in Arizona, Oregon, and elsewhere.
Urban resistance to large-scale builds is rising, especially when the benefits accrue to elite firms rather than local residents.
AGI won’t be throttled by model architecture. It will be throttled by kilovolts and kiloliters.
Moore’s Law doesn’t apply to infrastructure. If anything, the trend is in reverse: longer lead times, more permitting friction, higher NIMBY resistance. If you want AGI, you don’t just need a lab. You need a permitting office, a municipal water board, and a long-term power purchase agreement.
3. Geopolitics & Trade Policy: The Squeeze from Above
Even if you solve 1 and 2, there’s a sword dangling over it all: the state.
Export controls. Licensing ceilings. Sanctions. Investment restrictions. One policy memo in Washington or Beijing can choke off a critical input overnight.
Nvidia has already been forced to spin up China-specific SKUs to comply with export controls. The U.S. CHIPS Act is reshaping global semiconductor flows. The next Taiwan crisis, or the current administration, could shut off a critical supply chain link with zero notice.
As Ray Dalio recently wrote:
Many exporters to the United States and importers from other countries that trade with the U.S. are saying they have to greatly reduce their dealings with the United States, recognizing that...radically reduced interdependencies with the U.S. is a reality that has to be planned for.
AGI is deeply exposed to this fragility. Fabs are geopolitical. So are rare earths. So is advanced packaging. And no one doing AI timelines seems to be pricing in a significant chance of hard decoupling within five years.
The Hidden Convergence: Where These Constraints Collide
Each of these constraints, capital intensity, infrastructure lag, geopolitical exposure, is occasionally acknowledged in isolation. What’s not acknowledged is their joint constraint geometry. They don’t just add up. They multiply.
A chip fab without a grid is inert.
A data center with no transformers is dark.
A GPU with a banned export license is scrap metal.
A power-hungry model with no cooling water is a fire hazard, not a breakthrough.
Yet you’d never know it from AI investor decks or AGI timelines tossed around on Twitter. There’s a balance-sheet blind spot here: venture capital is underwriting the option value of AGI without modeling its real-world input constraints.
This is the kind of blind spot that leads to mispriced capital, vaporware timelines, and massive overcommitment to a trajectory that is, if not doomed, at least dragged by layers of friction no optimizer can route around.
What This Means for the AGI Timeline
If your AGI forecast doesn’t account for:
substation and transmission buildouts,
grid interconnect waitlists,
chip lead times and export license ceilings,
and the physical limits of capital amortization…
…it’s not a forecast. It’s a hallucination.
In Dalio’s language:
The United States’ role as the world’s biggest consumer of manufactured goods and greatest producer of debt assets...is unsustainable.
In other words: show me your model. And then show me where in your model you’ve accounted for these constraints.
Accelerationism vs Thermodynamics
Accelerationist fantasies envision recursive self-improvement loops, cognition bootstrapped into superintelligence, and emergent minds leaping ahead of humans in years, not decades.
But the real world has friction. It has heat. It has permits. It has fiscal cliffs. Moore’s Law might bend abstraction, but thermodynamics still rules matter.
We are not in a clean software regime. We are in a high-entropy hardware-and-sovereignty regime.
AGI doesn’t emerge from clever prompts. It emerges, if at all, from a grid-connected, debt-financed, politically sanctioned, geopolitically aligned, physically cooled, supply-chain-validated infrastructure base.
That’s not pessimism. It’s infrastructure realism.
A Call for Realism
None of this is to say AGI is impossible. But it is physically gated. And until those gates are expanded, staffed, powered, and politically sanctioned, the path to AGI will look more like trench warfare than blitzkrieg.
What’s needed now is not more parameter scaling or clever prompting. What’s needed is a grounded, systems-level approach to modeling real AI progress. This is one that treats silicon capex, infrastructure drag, and geopolitical squeeze not as edge cases but as first principles.
Otherwise, we’ll keep throwing money at dreams and be shocked when reality doesn’t compute.
This was in the back of my head before I started reading your pieces on this topic, but you have very eloquently and persuasively articulated what almost everyone on LessWrong seems determined to ignore.
Thanks so much for sticking to this beat. The doomiest of doomers need someone to push back against their timelines when they ignore the laws of physics.