For most of the last decade, the binding constraint on artificial intelligence was intellectual and then financial. First the field needed better methods; then it needed capital large enough to train at scale. Both constraints were resolved with remarkable speed. The constraint that replaced them is neither, and it is proving considerably harder to buy your way out of: the physical capacity to deliver electricity to a building full of accelerators.
Key points
- The binding constraint on AI capability has migrated from algorithms to capital to grid interconnection, and the third is the slowest to relieve.
- Interconnection queues have lengthened faster than data-centre construction timelines, making power, not silicon, the critical path.
- Chips depreciate on a three-to-five year cycle; substations do not. The mismatch in asset lives is the core financial problem.
- The cost is currently absorbed by hyperscaler balance sheets. Whether it can be passed to customers is the question that determines returns.
This is not a story about chip shortages, which are a supply problem and therefore self-correcting on a manufacturing timescale. It is a story about infrastructure that takes seven to eleven years to permit and build, sitting underneath an industry that revises its capacity plans every two quarters. The mismatch between those two clocks is the single most underappreciated variable in the sector.
The constraint moved, and the models did not follow
Consider how capacity planning actually works now. An operator decides it needs a facility. Land is available; capital is available; accelerators can be ordered with a lead time measured in months. Then the operator applies to connect to the grid, and enters a queue whose length is determined by regulatory process, transmission planning and the availability of high-voltage equipment with global order books stretching years out.
The queue is the product. Everything else in the chain has been optimised aggressively; interconnection has not, because it is a public planning process rather than a market, and public planning processes do not respond to demand signals at the speed private capital does.
Chips depreciate over three to five years. Substations last forty. The industry is financing a forty-year asset with a three-year revenue assumption and calling the difference growth.
Zimba Capital Research
Why this is a pricing problem, not an engineering one
Engineers will solve the technical elements of this. More efficient cooling, higher-density racks, better silicon per watt and genuinely clever siting decisions all reduce the power required per unit of useful computation, and all are improving quickly. What none of them changes is the asset-life mismatch sitting at the centre of the capital structure.
An accelerator is a fast-depreciating asset whose economic life is bounded by the next generation of accelerators. The electrical infrastructure that serves it is a slow-depreciating asset with a multi-decade life and no meaningful technological obsolescence. When you finance both out of the same capital budget and justify both with the same demand forecast, you have implicitly assumed that demand persists long enough to amortise the slower asset, an assumption that is rarely stated because stating it invites the obvious question.
| Layer | Economic life | Lead time | Obsolescence risk |
|---|---|---|---|
| Accelerators | 3–5 yrs | 3–9 mo | High |
| Servers & networking | 5–7 yrs | 3–6 mo | Moderate |
| Cooling plant | 15–20 yrs | 12–24 mo | Low |
| Building shell | 30–40 yrs | 24–36 mo | Minimal |
| Substation & transmission | 40+ yrs | 60–130 mo | Minimal |
Read down the lead-time column and the critical path is unambiguous. The item with the longest lead time is also the one with the least obsolescence risk, which means it is simultaneously the hardest to acquire and the safest to own. That combination is the definition of a scarce asset, and scarce assets accrue economic rent to whoever controls them.
Who ends up absorbing the cost
At present, the cost is absorbed by hyperscaler balance sheets, which are large enough to do so without obvious strain and are rewarded by equity markets for capital intensity that would be punished almost anywhere else. This arrangement is stable while the demand narrative holds, and the demand narrative is currently holding.
The interesting question is what happens at the margin, and there the picture is less comfortable. Application companies building on top of this infrastructure are price-takers on inference. Their gross margins are a direct function of a cost they do not control, cannot hedge, and generally cannot forecast beyond the term of whatever commitment they have negotiated. A software business with a 45% gross margin and an input cost set by someone else’s power procurement is not, structurally, a software business. It is a reseller with better branding.
We would expect this to resolve in one of three directions. Application pricing rises, which tests demand elasticity that has never genuinely been tested. Model efficiency improves faster than demand grows, which is possible and has happened before, though the historical record on efficiency gains reducing total consumption is discouraging. Or the industry consolidates toward players who own their power, at which point vertical integration stops being a strategic preference and becomes a condition of survival.
The investable observation
If the constraint is power rather than silicon, then the assets worth owning are the ones positioned across the constraint rather than downstream of it. That points toward generation with firm capacity, transmission-adjacent land with existing interconnection rights, high-voltage equipment manufacturing with genuine order-book visibility, and the increasingly specialised engineering and procurement firms that can actually deliver a substation on a schedule.
It also suggests treating announced capacity with appropriate scepticism. A gigawatt announced is not a gigawatt energised, and the gap between the two is measured in years and in permits. We would weight announcements by their position in the queue rather than by their headline number, and note that very few operators disclose enough to allow anyone outside to do that with confidence.
What would change our mind
Three developments would materially weaken this argument. Regulatory reform that genuinely compresses interconnection timelines, not pilot programmes, but structural change to how queues are processed, would relieve the constraint at its source. Sustained deployment of behind-the-meter generation at scale would let operators bypass the queue rather than wait in it. And an efficiency step-change large enough to reduce absolute power draw, rather than merely improving power per unit of computation, would dissolve the problem entirely.
Each is possible. None is currently visible in the data at the scale required. Until one of them appears, the sector’s binding constraint is a civil engineering problem wearing a technology valuation, and that is an uncomfortable thing for a portfolio to own without noticing.
