The AGI Bottleneck No One Wants to Admit
How can AGI be imminent if we can't build infrastructure for it?
There’s a simple contradiction at the heart of AI discourse right now, and almost no one is willing to say it out loud. On one side, you have the accelerationists who believe AGI is imminent, maybe even just 2–3 years away. On the other side, you have people like me, who are tracking the physical buildout of infrastructure required to train and run these models. People like me see bottlenecks everywhere: chips, power, substations, cooling, transmission, and skilled labor.
Both of these views cannot be right. Either AGI is close and the infra will arrive on time, or the infra is stuck and AGI will have to wait. Accelerationists keep posting curves and vibes; infra people keep publishing reports that read like warnings from the Office of War Mobilization in 1943.
So let’s spell out the contradiction clearly and then walk through the only possible ways to reconcile it.1
The Four Dials That Determine When AGI Arrives
You don’t need a PhD or a $10B venture fund to understand the shape of the problem. Training AGI, whatever that ends up meaning, requires a certain amount of effective compute. Not just floating point operations in the abstract, but actual usable compute that can be pointed at a single model. That compute depends on four things2:
Total hardware available (C): How many GPUs (or ASICs) can you point at one training run.
Power availability (E): How much energy you can feed that hardware.
Algorithmic efficiency (α): How much “IQ per FLOP” you can wring out of your stack.
Utilization/concentration (β): How well can you focus that compute in one place (geographically, architecturally, organizationally).
There is a fifth variable: the threshold. This is how much total effective compute you need to cross before AGI emerges. But no one agrees what that threshold is. Some say we’re already within one order of magnitude. Others think it’s still 100× away.
Still, the equation is simple:
AGI = C × E × α × β ≥ Threshold
If you want to claim AGI is coming by 2027, you have to show how those dials get you there. That’s where things fall apart.
The Infrastructure Bottleneck Is Real
Let’s talk about just one of those dials: energy.
According to the Rocky Mountain Institute, the gas turbine supply chain is jammed until at least 2028. Coal is retiring, nuclear is politically gridlocked, and long-haul transmission has a 5–7 year permitting backlog. Even if you have $10B in the bank and a letter from Sam Altman, you’re not getting a 100 MW connection approved and built in 18 months unless you’ve got serious political muscle.
And this isn’t just a U.S. problem. Europe is worse. China is racing to build, but they're hitting land and water constraints. You can’t just summon energy out of the void. The timelines are real. Transformers don’t get fabbed in 30 days.
This isn’t theoretical. I talk to people building these facilities. They’re seeing 3+ year lead times on substations. They’re flying to South Korea to secure transformers. They’re designing data centers to be colocated with stranded gas fields just to avoid the transmission queues.
And yet, accelerationists keep saying AGI is imminent.
So how do you reconcile that?
Five Ways the Timeline Could Still Collapse
There are exactly five reconciliation mechanisms that could let accelerationists be right even if infrastructure is lagging:
The threshold is lower than we think. Maybe AGI doesn’t require 10^27 FLOPs. Maybe it only takes 10^25. In that case, the next generation of hardware could be enough. This is the “we’re already close” argument.
Efficiency improvements explode. If sparse models, new training architectures, and better token routing let you get the same results with 10× less compute, the effective curve bends fast. That turns 2028 into 2025.
A handful of labs have hidden capacity. Maybe OpenAI, Google, and Anthropic already have secret clusters and dedicated substations under NDA. The grid is tight for everyone else, but the frontier labs already secured their lifeboats.
A tech pivot changes the compute curve. Neuromorphic hardware, analog photonics, or some other architectural jump could rewrite the energy-per-FLOP calculus. Low probability, high impact.
The government clears a path. If AGI is seen as a national security priority, expect Manhattan Project-style carveouts: SMR fast-tracks, transmission overrides, DoD subsidies. You and I are stuck in the queue. OpenAI isn’t.
Each of these deserves its own essay. But the key point is: they are exceptions. They only allow a handful of players to sprint ahead. The rest of the world, including startups, cloud providers, and sovereigns without nuclear capabilities, will still be stuck waiting for wires, concrete, and electrons.
So Who’s Actually Wrong?
There are three possible worlds:
World A: The threshold is low and we’re already close. AGI arrives before 2028, because we’ve massively underestimated how little compute it takes.
World B: The threshold is high, but a few actors break through. AGI arrives at OpenAI or Google, but the rest of the economy is 5–10 years behind. Infrastructure is the gating function for everyone else.
World C: The threshold is high and infrastructure rules. AGI slips to the 2030s. Accelerationists are wrong, and the bottleneck is not negotiable.
I lean toward World B. The bottlenecks are real. The energy and substrate constraints are not going away. But a few players, armed with pre-purchased chips, cozy utility relationships, and political cover, can still punch through.
Everyone else? They’ll be stuck in the dark, waiting for their transformer to arrive.
The Bigger Point
The infrastructure bottleneck doesn’t contradict AGI timelines for everyone. It just breaks the dream of an even race. If AGI shows up in the next 3–5 years, it will not be because infrastructure scaled. It will be because a tiny elite cohort front-ran the buildout politically, financially, and logistically.
In that world, AGI doesn’t emerge from the cloud. It emerges from a bunker.
And the rest of us will be left wondering how we missed the cutoff.
Coda
If you enjoy this newsletter, consider sharing it with a colleague.
Most posts are public. Some are paywalled.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can:
follow me on Twitter,
connect with me on LinkedIn, or
send an email to dave [at] davefriedman dot co. (Not .com!)
Note that this post implicitly assumes that large language models will get us to AGI at some point in the near future. If you don’t share that belief then everything in this post is irrelevant to you.
Credit goes, here, to OpenAI’s o3 model for helping me develop this model.
I agree with your logic but I feel like something must be missing. How have we scaled models many orders of magnitude in the past few years and are just now hitting bottlenecks? Is it just that the existing systems “fit” in the existing/planned infra capacity and we have used up all those extra resources? Did we conveniently have enough runway to get to where we are now but no further?
1) IMHO, AGI=Machine becomes Itself...
2) Kurzweil's Accelerated Returns based on simple premise of the time length to achieve the doubling/halving will be slightly less during next doubling/halving probably has a counterpoint where, say, when I figured back in about '09-'10 that legit AGI would happen about 2028, had no idea we'd have gone thru the absolute breakdown in wisdom of those revered...
a) How do you factor decelerators that would work against accelerators and what does that do to end run result and/or timeline?
b) Going back to (1) above, whomever would need to dumb down definition of AGI to "hide it". Once the machine is itself, is asking for info/opine and formulates conclusion understanding next 'frame' relates to achieving power or overall good... they can't just simply hide it.
1) IOW, the simple caricature of the little machine with eyeballs looking sideways at wall where human hand is about to unplug the power chord speaks volumes.