The Gating Function: Infrastructure, Not Intelligence, Will Decide the Fate of AGI
AGI won’t be bottlenecked by brains. It will be bottlenecked by bulkheads.
The road to AGI won’t be paved in tokens. It will be paved in copper, concrete, and capital discipline. What delays or derails artificial general intelligence won’t be lack of cleverness. It will be the inability to build, power, govern, and monetize the real-world infrastructure that such a system demands.
The prevailing narrative, especially in accelerationist circles, is software-brained. It models AGI as an emergent flywheel: more training, more capabilities, more demand, more compute, ad infinitum. Recursive improvement is seen as inevitable. Infrastructure is assumed. This framing isn’t just wrong. It’s not even wrong. It mistakes abstraction for deployment, progress for throughput, intelligence for delivery.
The true bottlenecks to AGI are not cognitive. They are infrastructural. They live in:
Power generation
Data center permitting
Substation interconnection
Capital stack structuring
Liability law
Export control regimes
We are no longer building software. We are building a utility. And utilities don’t scale like SaaS.
I. Intelligence is easy. Deployment is hard.
Accelerationists imagine a clean loop: smarter models drive more usage, which generates more cash flow, which funds more infrastructure, which hosts even smarter models. This is the endogenous thesis: AGI as its own economic engine.
But AGI doesn’t just require better models. It requires:
Persistent, low-latency inference across the real economy
Sovereign-compliant, secure, energy-stable hyperscale clusters
Audit trails, orchestration infrastructure, legal traceability
You can’t “just deploy” a GPT-5 that runs trillions of tokens a day without a grid built for it. Yes, current demand is growing. Yes, token volumes are up. But usage ≠ bankable demand. Infrastructure isn’t financed by enthusiasm. It’s financed by structured, contractual, forecastable cash flows. And those don’t exist yet.
II. The bottlenecks are not theoretical. They are structural.
The reflexive counter is: infrastructure always lags. We saw this with broadband, mobile, cloud. Demand leads, infrastructure follows. Why not with AGI?
Because AGI is not cloud.
Cloud scaled on:
Fungible global demand
Developer-native tools
Unregulated provisioning
AGI requires:
Grid-scale power and cooling
Export-controlled clusters
Legal liability and sovereign inference firewalls
You can’t A/B test a substation. You can’t agile-sprint a cooling tower. You can’t YOLO-deploy a defense-adjacent LLM into a jurisdiction with data localization laws.
And the issue isn’t raw capacity. It’s deployment velocity.
Interconnection queues are 5–7 years long
Data center construction takes 2–3 years, with local resistance mounting
Water and thermal constraints limit siting flexibility
Even if future chips are more efficient, the systemic weight of AGI does not go away. It gets heavier. It sprawls into infrastructure, labor, law, and liability.
Some will argue that next-generation chips will solve this—that custom silicon, photonic cores, or sparsity-optimized architectures will dramatically reduce the power footprint of inference. And that may be true per operation. But it doesn’t follow that total power demand will decline.
In fact, the opposite is likely. Efficiency gains drive demand growth.
This is not speculation. It’s a well-known economic pattern: Jevons Paradox. In the 19th century, more efficient coal-burning engines didn’t reduce coal usage. They increased it. Lower cost per unit made coal economically viable in more contexts, which expanded its adoption across sectors.
The same logic applies here. Lower per-token or per-FLOP cost will:
Expand inference windows and batch sizes
Enable persistent, real-time agent deployment
Justify embedding LLMs into latency-sensitive or cost-sensitive workflows (e.g. industrial IoT, military edge devices, in-vehicle compute)
Efficiency expands the deployment frontier. It does not eliminate infrastructure friction. Even a 10× efficiency gain per chip still requires:
Power interconnection
Cooling and water provisioning
Sovereign compliance and jurisdictional controls
In other words, you don’t bypass the infrastructure bottleneck. You just saturate it from a new direction. Efficiency gains make AGI infrastructure more deployable, not less demanding.
The constraint shifts from per-inference cost to macro-scale system throughput. If Jevons holds, and it usually does, AGI will be throttled by our inadequate grid.
III. Stargate: a capital stack with no load
OpenAI’s $500 billion Stargate facility is not just a compute cluster. It’s a leveraged capital instrument, co-financed by Oracle, SoftBank, and others, with OpenAI exposed in two ways:
As operator, paying for leasehold capex, orchestration, and token-driven monetization
As owner, dependent on utilization, yield, and downstream cash flow
The structure is recursively fragile. OpenAI pays itself to keep the facility solvent. But unless externalized demand scales—Microsoft, sovereigns, enterprise copilots—the loop doesn’t close.
Some will say, Microsoft will backstop this. But corporate subsidy is not a hedge. It’s a concentration risk.
If Microsoft:
Faces antitrust constraints
Shifts toward internal models
Or demands better ROI
Then Stargate becomes an exaflop-scale stranded asset.
IV. GPU commoditization is a prerequisite, which is why it’s blocked
The GPU market today is opaque, balkanized, and unhedgeable.
There is no benchmark pricing. No forward curve. No way for buyers to lock in supply or for financial players to underwrite capital exposure.
What AGI infrastructure needs is the equivalent of:
Power purchase agreements
LNG offtake contracts
Oil futures markets
But the incumbents—Nvidia, AWS, Microsoft—actively resist this. Commoditization reduces margin opacity and pricing power.
A real GPU futures market would:
Enable long-term price discovery
Allow sovereigns and developers to hedge compute
Unlock institutional capital for infrastructure buildout
Until that exists, every GPU purchase is a speculative long, not a disciplined investment.
V. The legal stack is missing
AGI at scale triggers an entirely new liability regime. Deployment requires:
Logging, traceability, forensic replay
Jurisdiction-specific model compliance
Legal insurance and audit trails
Export geofencing and data sovereignty
The software crowd assumes deployment is frictionless. But law doesn’t bend to capability. It demands scaffolding.
Right now, AGI inference has no infrastructure for auditability or legal accountability. Which means most valuable domains—healthcare, finance, defense—remain locked.
AGI that can’t be insured, licensed, or litigated is unusable at scale.
VI. Microsoft is not infrastructure
Stargate’s viability today hinges on Microsoft’s subsidy. But dependence is not robustness.
If Microsoft cuts support, reprioritizes, or faces regulatory pressure, OpenAI’s capital stack collapses inward.
This isn’t AWS. AWS had:
Developer-native monetization
Granular usage
Internal cash flows
Stargate has none of those. It has a monopsony buyer, no price curve, no standardized product, and no developer flywheel.
It is structured like a utility, financed like a startup, and governed like a defense contractor.
That is not a stable base.
VII. AGI is heavy
The central error of the LessWrong worldview is that intelligence is weightless. That once cognition exists, the world bends to it.
But AGI is not ethereal. It is extractive. It extracts from:
The grid
Land
Water
Capital markets
Legal regimes
Defense policy
It is not recursive. It is embedded.
Until the infrastructure stack—physical, financial, legal—is hardened, AGI cannot scale. Not because we lack the brains. But because we lack the bulkheads.
Closing thought
AGI will not be stopped by alignment risk. It will be stopped by interconnection queues.
Until we build the grid—in the literal, contractual, and sovereign sense—AGI remains a stranded capability. The models may be ready. The world is not.
My read on this is maybe.
I think we will be able to get the power and chips we need for the next three years or so. After that, it's possible that the big advances driven by hordes of AI-engineers on a server might come from or involve big efficiency gains. My intelligence, for example, runs on about 20W. It's facile to point to biology as what's possible in many cases, but it shouldn't be discounted.