AI Progress Is Exponential, But Capital Investment Is Super-Exponential: Can This Growth Last?

The Hidden Risk Behind the Exponential Curve

Over the past few years, nearly every metric measuring AI progress — model capability, inference performance, benchmark scores — has tracked tightly along an exponential curve. This means AI capability has been improving at a steadily accelerating pace, following the familiar rhythm of "doubling every so often."

But one researcher raised a more sobering observation on social media: during the same period (2023 to 2026), capital expenditure (capex) in the AI sector hasn't grown at a merely exponential rate — it's grown at a slightly super-exponential rate. In other words, the growth rate itself keeps climbing. The acceleration of investment inputs is outpacing the acceleration of capability outputs.

This subtle but critical gap reveals a structural problem underlying the current AI boom — one that's easy to overlook.

Viewing AI as an Input-Output System

If we abstract the entire AI industry into a system — where external investment is the input and AI capability improvement is the output — then mathematically, this system's response is slightly sublinear.

The logic isn't complicated: when you need to inject resources at a super-exponential rate just to achieve exponential capability gains, it means the marginal output per unit of input is diminishing. The system doesn't exhibit the ideal property of "greater efficiency with greater investment." Instead, it's sustaining what looks like a smooth capability growth curve by piling on capital at an increasingly aggressive pace.

This perspective pulls AI development out of a purely technical narrative and places it squarely within an economic framework of resources and returns. Exponential capability growth is exciting, but if it's fueled by capital that's expanding even faster, the sustainability of that curve must be questioned.

Sublinear response is a concept from cybernetics and systems engineering: if the output growth rate consistently falls below the input growth rate, the system exhibits sublinear behavior. The opposite — superlinear response — is when each additional unit of input yields an even larger output increment, akin to increasing returns to scale. Current AI systems show the reverse: the marginal returns on compute and capital are narrowing. This is partly because the Scaling Laws underlying mainstream training paradigms (pre-training large language models) are not perfectly linear — once parameter counts and data volumes reach a certain scale, the rate at which loss curves decline gradually flattens, requiring exponentially more compute for equivalent performance gains. This is precisely why the industry has been actively exploring alternative approaches such as post-training, reinforcement learning, and inference-time scaling.

External Resource Injection Is Not Sustainable

The core of the problem is this: injecting external resources at a super-exponential rate is not sustainable over the long term.

External investment — whether from venture capital, tech giants' balance sheets, or public markets — is not infinite. Any system that consumes resources at an ever-accelerating rate will eventually hit a ceiling. Capital market patience, GPU supply, energy constraints, and investor expectations for returns will all form hard limits at some point.

In other words, the current exponential curve of AI progress is partly being propped up by a flood of capital that cannot be maintained indefinitely. Once the rate of external injection slows — even if only from "super-exponential" back to "exponential" or even "linear" — the system's inherent sublinear response could meaningfully drag on the pace of AI capability advancement.

It's worth noting that the arrival of "hard constraints" may not manifest as a dramatic collapse — more likely, it will appear as a gradual deceleration. Capital market patience for AI infrastructure is typically measured by the "window for profit realization": if major hyperscalers cannot demonstrate AI revenue growth commensurate with their capex within the next two to three fiscal years, market willingness to support the next wave of compute expansion will systematically decline. This tightening of expectations management can sometimes arrive before physical resource bottlenecks do, transmitting through the supply chain as valuation repricing and tighter financing conditions — ultimately affecting the survival space for smaller model companies.

Closing the Resource Loop Is the Key

So where is the way out? The original post's answer is: the AI industry must generate matching exponential profits and close the resource loop.

"Closing the resource loop" means shifting AI from dependence on external capital infusions to self-sustaining internal cash generation. If AI can produce exponentially growing real profits — not just exponentially growing valuations and fundraising rounds — those profits can be reinvested into the next cycle of R&D and compute, forming an internally self-sustaining growth flywheel.

This is the true condition for maintaining long-term exponential progress. Only when the economic value created by AI flows back into the industry at a rate matching its capability growth can the system stop depending on continuous external capital injections — and only then does the growth curve have a genuinely endogenous, sustainable source of momentum.

The resource loop concept is commercially equivalent to a "reinvestment flywheel that activates once unit economics turn positive." Historically, cloud computing offers an instructive precedent: AWS also required massive external capital to build infrastructure in its early days, but as its paying customer base scaled, cloud service profits began reinvesting into data center expansion at an exponential rate, forming a self-reinforcing growth loop. The challenge facing the AI industry today is that its primary monetization formats — API subscription calls and enterprise service contracts — have not yet grown fast enough to keep pace with the expansion of compute costs. If high-value application scenarios such as AI agents and automated workflows can achieve large-scale commercial adoption, it could accelerate the closure of this loop.

Implications for the Industry

This analytical framework offers a clear, quantified perspective for the ongoing debate about AI bubbles and sustainability. It reminds us to distinguish between two separate things: capability progress and business model viability.

Impressive curves on technical metrics do not equal a healthy business model. When capital pours in at a super-exponential rate while monetization has yet to form a loop of comparable magnitude, the industry sits at a fragile equilibrium. In the years ahead, the core metric for determining whether AI has truly entered a sustainable development phase may no longer be benchmark scores — it may be whether the industry can earn back what it burns, and earn it back faster.

For investors, entrepreneurs, and policymakers alike, this means paying far more attention to AI's path to profitability and unit economics, rather than just capability demonstrations. Whoever closes the resource loop first will be the one who truly holds the initiative in the next phase of competition.