Most forecasts about AI capability start from the same place: curves went up, so they will keep going up. What would have to be true for the scaling trajectory to stall?
A March 2026 inversion analysis of the scaling narrative names four candidate constraints: energy physics, data exhaustion, capital misallocation, and regulatory intervention. The ranking most people assume is backwards.
The three constraints that are engineering problems
Energy. The claim is that training energy demand grows faster than efficiency gains, and that grid capacity or per chip power delivery hits physical limits before algorithms catch up. The constraint is real, and it also has the clearest set of available responses. Power is a scheduling and siting problem well before it is a physics wall. Operators can move training to where interconnect queues are shorter, shift load in time, or accept worse utilization in exchange for cheaper power. Lead times on new generation and transmission run in years, so energy can bind on that horizon even where the underlying physics is not the limit. On that timeline energy most likely shows up as margin compression and delayed buildouts rather than an immediate stop.
Data. The estimate in circulation is that high quality text data is depleted somewhere around 2027 to 2028. Treat the date as soft. Depletion estimates depend heavily on what counts as high quality, how much licensed and non public corpus exists, and how much multimodal and interaction data substitutes. Synthetic data is the standard answer, and the standard counter is model collapse, where training on generated output narrows the distribution. This is unresolved. Filtered synthetic data with verifiable rewards appears to work in domains where correctness is checkable, such as code and mathematics. It is less clear that it works for taste, judgment, or open ended reasoning. So data probably changes the shape of progress, faster in verifiable domains and slower elsewhere, without imposing a hard ceiling.
Regulation. The EU AI Act, US executive actions, and Chinese restrictions are all real, and they are priced differently across jurisdictions. Some frameworks look negotiable. The EU AI Act sets fixed compliance deadlines. Where the mandates are fixed, the main effect is a transfer of advantage toward whoever can absorb the legal overhead, because compliance cost works as a barrier to entrants.
The constraint lacking an engineering response
That leaves capital.
The analysis cites roughly 189 billion dollars of AI investment in a single month, February 2026, as evidence of bubble dynamics. These totals are sensitive to whether they include infrastructure commitments, multi year contracts, and private rounds counted at announcement rather than at close. The distribution is the more informative part.
On the pattern the analysis describes, capital appears concentrated at two extremes: a handful of mega scale labs and a long tail of very small startups. That suggests a thin middle. The analysis borrows Taleb's barbell, which describes a deliberate strategy for an individual portfolio combining safe assets with small convex bets. Applying it to market structure is a loose analogy, since frontier labs are not safe assets in any traditional sense. But if market wide capital is deployed this way, it raises a structural problem. Technology commercialization has historically depended on a mid sized layer that turns research results into cash flow through integration and customer retention.
The failure mode is that scaling continues while the layer that monetizes it never forms. Capability keeps improving, benchmark numbers keep rising, and enterprise adoption stays slower than the capital deployed against it. In that scenario the correction arrives as a funding contraction rather than a technical wall, and it hits hardest at the layer that was already underweight.
That also explains why the four constraints are not independent. As a conjecture about correlated stress, a capital contraction could make energy contracts harder to sign, make expensive data licensing look optional, and make regulatory compliance a larger share of a smaller budget.
What to watch
A few leading indicators would test the hypothesis.
- Enterprise adoption depth rather than logo counts. Seat expansion and renewal rates matter more than pilot announcements.
- Series B and Series C activity in applied AI. Measuring that volume directly is the way to check whether the missing middle is real or just an artifact of the analogy.
- Power purchase agreements and interconnection queue positions. These reveal buildout intent years before capacity exists.
- Data licensing deals and their pricing. Rising prices are evidence that the depletion story has teeth. Flat prices suggest substitution is working.
- Whether frontier gains concentrate in verifiable domains. If code and math keep improving while open ended judgment plateaus, that is the synthetic data constraint showing up in results.
Infrastructure, tooling, and the applied layer have different failure timing than frontier model providers.
This is a claim about where the primary risk sits. The technical constraints have engineering responses. Capital structure does not. The market has priced a monetization layer that, pending data on applied mid stage funding, it may not have funded.