AI bottlenecks: 3 structural constraints that will decide what this technology actually becomes
The cone is narrowing. Understanding the three concurrent pressures on AI development matters more right now than tracking model benchmarks.
In My Big Technology Framework, I describe the bottleneck phase as the moment when a technology's expanding cone of possibilities is compressed by structural constraints into a few viable market forms, and the output is almost always a more modest, more specific thing than what entered that cone. For AI, the bottleneck is no longer a distant projection. Three concurrent constraints are now pressing, each operating on a different timescale and with different implications for which players survive and in what form.
Philippe MEDA
The energy bottleneck
Sam Altman has been candid enough to say in several forums that OpenAI's expansion is primarily constrained by energy access, not by capital or talent. Microsoft's 2023 partnership with Constellation Energy to restart the Three Mile Island nuclear plant is an engineering acknowledgment that the GPU scaling curve cannot continue at its current rate without new power generation being built in parallel, and that energy is the key bottleneck for AI in the medium term. Energy infrastructure takes 5 to 10 years to plan, permit, and build (I'm being rather optimistic here, as AI and nuclear energy are key now).
Which means the scaling assumptions embedded in current AI investment theses are running materially ahead of what the physical grid can actually support. Of course, the large infrastructure players all know this, which is precisely why AWS, Google, and Microsoft are each pursuing dedicated nuclear and geothermal contracts with an urgency that makes no sense unless you believe the energy wall is real and close.
The reliability bottleneck
The second bottleneck is the one that gets discussed least honestly. LLMs have an incompressible floor of probabilistic error because transformer-based systems fundamentally work this way. Five perimeters of AI Financial auditing, clinical diagnosis, legal reasoning under liability exposure: these are not simply high-bar applications that will become accessible as models improve. They are applications where the liability architecture and regulatory environment make stochastic systems structurally inadmissible regardless of benchmark performance. This is worth stating plainly: a significant share of current AI revenue projections depends on access to exactly these markets.
The valuation bottleneck
Last, but not least, the current valuations of OpenAI and Anthropic (and, to a significant extent, Nvidia's forward multiples) imply revenue flows that require the addressable AI market to reach tens of trillions of dollars annually within a decade. The arithmetic demands exponentially more than just incremental productivity gains across existing workflows. It's really about a wholesale restructuring of how value is created across the global economy, and that's what every AI zealot is currently selling right now. That outcome is imaginable in principle, but realistically, it still feels like the big Metaverse fever dream from just four years ago. As an investment thesis, it is entirely dependent on scenarios that remain undemonstrated at scale — and in markets where the reliability constraint has already partially placed them out of reach.
The shape of AI under pressure
The second-order consequence of all three constraints pressing at once is already becoming visible in the market structure. A bifurcation is emerging between, on one side, large proprietary frontier models with enormous capital requirements and uncertain monetization, and on the other, smaller, task-specific, energy-efficient models designed to operate within the error tolerances of real operational contexts. The technology that emerges from the bottleneck in the strongest position may not be the one with the largest parameter count, but the one that identifies a specific perimeter where AI output is reliable enough to automate a decision — and is built there.
This is a key lesson for corporations currently trying to make sense of all this. There is no premium advantage in racing forward with OpenAI, Anthropic, or others just because they need their valuation to make sense. Smaller, cheaper, more energy-efficient models are certainly capable of doing everything you need for internal processes or customer-facing activities right now.
It's past time corporate customers regroup, reassess and decide for a more frugal approach.
