The ceiling isn't price — it's the grid.

The bottleneck stopped being the chip

For two years the shortage story was about silicon. Whoever held the accelerators held the capacity, and the buyer's anxiety was allocation — could the vendor get enough GPUs. In 2026 the constraint moved one layer down, to the thing that runs the accelerators. A modern AI data center does not fail to get built because there are no chips on the loading dock. It fails to get built, or gets built and cannot be fully energized, because the local grid cannot deliver the power the racks demand. The interconnection queue to attach a new large load to the transmission network stretches years in the constrained regions. High-voltage transformers — the unglamorous steel-and-copper units that step grid power down to what a facility uses — carry lead times measured in quarters and sometimes past a year. Transmission capacity itself is finite, and building more of it is a permitting exercise counted in years, not sprints.

That is why the operators with the deepest pockets stopped treating electricity as a utility bill and started treating it as a strategic reserve. Across 2025 and into 2026, the largest cloud and model operators signed long-term power purchase agreements for firm, continuous generation — dedicated nuclear output, direct allocation of reactor capacity, and forward commitments to small modular reactors that have not been built yet. You do not sign a contract for power that will not exist for years unless you believe the alternative is running out of it. Continuous power became a resource to be secured ahead of demand, the way an airline hedges fuel or a manufacturer locks a scarce input. The signal for everyone downstream is simple: the people closest to the compute concluded that energy availability, not chip supply and not token price, is the binding constraint on how much AI can actually run.

Three different conversations, all called "energy"

Part of why this risk hid for so long is that the word "energy" is doing three jobs at once in the AI conversation, and boards keep answering the wrong one. It helps to separate them cleanly, because a vendor can look excellent on two of these axes and be exposed on the third — and the third is the one that takes the service down.

The token price is the number on the invoice. Grid capacity is whether there is an invoice at all. A vendor can win the price conversation and the sustainability conversation and still be one regional interconnection freeze away from a capacity wall.

The reason this matters for a procurement team is that the first two questions are the ones your vendor volunteers. The pricing page shows the token economics; the ESG appendix shows the carbon posture. Almost nobody publishes their power supply position, because it is uncomfortable and because until recently it was nobody's job to ask. The gap between what vendors offer to discuss and what actually caps their capacity is exactly where this risk lives.

Why the chain reaches your SLA

It is tempting to file grid physics under "someone else's problem" — the hyperscaler's, the model lab's, the energy market's. The reason it is your problem is that the dependency chain is unbroken and it terminates at your contract. Your SaaS platform runs a document workflow. That workflow depends on frontier inference to classify, extract and reason. That inference depends on compute capacity in specific data centers. That compute capacity depends on continuous power delivered to those specific buildings. And that power is physically constrained by a grid that cannot be expanded on the timescale of your renewal. Every link is real, and the risk propagates up the chain the same way a supply shock propagates through any dependency: the tightest link sets the ceiling for everything above it.

When the compute market tightens — and it tightens every time a new wave of demand meets a grid that cannot grow to match — the squeeze does not announce itself as a power problem. It arrives as a quota on your account, a throttled rate limit during peak hours, a region that is "temporarily unavailable" for new capacity, a roadmap feature that slips because the inference budget it needed was reallocated to keep existing customers served, or a price step that the vendor blames on "infrastructure costs". The buyer who never asked where the power comes from experiences all of these as unrelated annoyances. The buyer who mapped the dependency sees them for what they are: the capacity ceiling reaching down through four layers to touch the SLA they signed.

The due-diligence frame that treats power as a supply input

The correction is not to become an energy analyst. It is to add power availability to the list of things you already diligence in a critical vendor — the same discipline you apply to a subprocessor, a data-residency commitment or a security posture. Five moves turn the abstract risk into a concrete procurement question.

Map where the inference actually runs

Ask the vendor the questions they are not used to hearing. Where does the inference behind this service physically run — which providers, which regions? What capacity guarantee sits behind the SLA: reserved capacity, best-effort, or a shared pool that gets rationed under load? And what is the plan if a primary region freezes new interconnection or hits a power cap — is there a second region already provisioned, or is the answer a hopeful "we would migrate"? The vendor who can answer crisply has thought about this. The vendor who treats the question as exotic is telling you they have assumed the power is infinite.

Place workload by energy availability, not only cost and latency

Workload placement has historically optimized for two variables: unit cost and latency to the user. A third variable now belongs in the decision — where firm power is actually available. A region that is cheap and close but sits behind a multi-year interconnection queue is a fragile place to anchor a critical workload. The mature posture spreads critical inference across regions chosen partly for their power headroom, so that a constraint in one place does not become an outage everywhere. This is the same instinct as diversifying any single point of failure; the input being diversified is electricity.

Reserve capacity in the contract, not in a roadmap wish

Hope is not a capacity plan. When the compute market rations, the customers who keep their capacity are the ones whose contracts reserved it in writing — a committed capacity floor, a defined priority tier, and a stated remedy if the vendor cannot serve the committed volume. The buyer who signed on list terms with no capacity language is, by construction, first in line to be throttled when the vendor has to choose who gets served. The clause does not need to be adversarial; it needs to name the capacity, the priority and the fallback, so that "we ran out of power in that region" is the vendor's problem to solve and not the customer's to absorb.

Move the sensitive and the steady to the edge

Not every document has to travel to a frontier data center to be processed. A meaningful share of steady, high-volume extraction — and nearly all of the most sensitive material — can run on compact models close to where the work happens, on air-gapped or on-premise infrastructure where local power resolves both latency and dependency. Shifting that slice of the workload to the edge does two things at once: it removes a chunk of demand from the constrained frontier supply, and it makes the remaining frontier calls the exception rather than the rule. The residency logic behind this is the same one that governs where data is allowed to be processed; here the driver is capacity resilience rather than jurisdiction.

Read the PPA as a resilience signal, not an ESG line

When a provider announces a long-term power agreement, the instinct is to file it under sustainability marketing. Read it instead as a supply-resilience signal. A vendor whose inference runs on a provider that has secured firm generation ahead of demand is a vendor with a lower probability of being rationed in the next squeeze. A provider with no such position, buying spot power in a constrained market, is more exposed. You will rarely get the full picture, but the direction of travel — has this provider secured continuous power, or is it hoping the grid keeps up — is a legitimate diligence input, sitting right next to financial health and security posture.

Four ways it goes wrong

The buyers who get surprised in the next compute crunch tend to make the same four mistakes. Each one is silent right up until the moment it is not.

Assuming compute is infinite and elastic when the grid is the real limit. The cloud sold elasticity so convincingly that a generation of buyers internalized it as a law of nature. It is not a law of nature; it is a business model that works until the underlying physical input runs short. Capacity that "just scales" scales only as far as the power behind it, and in the constrained regions the power stopped scaling first. A plan that assumes bottomless, on-demand capacity is a plan with an unpriced tail risk.

Ignoring geographic concentration that creates systemic region risk. If a vendor — and its provider — concentrate inference in one or two power-favorable regions, a single constraint in that geography becomes a correlated failure for every customer served from it. Concentration looks efficient on a cost sheet and reads as a single point of failure the moment power in that region is capped. The buyer who never asked about geographic spread inherits the vendor's concentration as their own continuity risk.

Treating energy as a sustainability topic, not an availability one. Carbon and cost are real and worth managing, but they answer a different question than "will the service run." A vendor can have an immaculate sustainability report and no firm power position. Filing energy entirely under ESG means the availability question never reaches the risk register, and the first person to notice the gap is whoever is on call when the region goes capacity-constrained.

Signing a contract with no capacity clause for when the market rations. Most SaaS and AI contracts say a great deal about uptime and nothing about capacity — the implicit assumption being that capacity is always there to be uptime on. When power tightens and the provider has to allocate scarce compute, the contract that named a committed capacity floor is the contract that gets honored first. Silence in the contract is not neutrality; it is an unwritten agreement that the customer absorbs the shortage.

The trade-off nobody states out loud

Guaranteed capacity is not free, and pretending otherwise is how these conversations go wrong. Reserving a committed capacity floor costs more than best-effort access, and it usually ties the workload to a specific region and a specific provider — which is a form of lock-in, and lock-in has its own risks. The buyer who reserves everything, everywhere, pays for headroom they may never use and constrains their own ability to move when a better option appears. The buyer who reserves nothing keeps maximum flexibility and maximum exposure.

The honest position is that this is a portfolio decision, not a binary. Reserve firm capacity for the workloads whose interruption is genuinely costly — the regulated pipeline, the customer-facing extraction with an SLA, the volume the business depends on daily. Leave the elastic, deferrable and experimental work on best-effort terms where flexibility is worth more than certainty. And keep enough of the workload on the edge that a frontier squeeze degrades the service rather than stopping it. The point is not to eliminate the trade-off; it is to make it a deliberate allocation instead of an accident of whatever the vendor's default terms happened to be.

What this means for document AI specifically

Document AI sits in an unusually exposed spot, for three reasons, and each one has a mitigation that a serious vendor should already be running.

The workload is steady and high-volume. Unlike bursty consumer traffic, an extraction pipeline runs a predictable, continuous load — which is both a liability and an advantage. A liability because a capacity constraint hits it every hour, not occasionally; an advantage because a predictable load is exactly the kind that can be placed deliberately, reserved sensibly and partly shifted to the edge, rather than left to chase spot capacity.

Much of it does not need the frontier at all. A large share of classification and structured extraction clears the quality bar on compact models running close to the data. Routing that share off the constrained frontier supply and keeping frontier calls for the genuinely hard exceptions is the same discipline as the routing table that reacts to price — here the routed variable is capacity resilience, not just unit cost.

The buyers are sophisticated and regulated. Banks, insurers and accounting firms already run vendor continuity diligence; adding a power-availability question to the existing framework is a small extension, not a new program. The vendor that can answer where the inference runs, what capacity is reserved and what happens if a region caps — and that has already moved the sensitive and the steady to the edge — is the vendor that passes the diligence the incumbent without a power position quietly fails. The provider-selection discipline sits alongside the broader due-diligence frame; energy availability is one more axis on the same scorecard.

Closing thought

The buyer who still evaluates an AI vendor on price and features alone is running a due-diligence process built for a world of elastic, infinite compute — a world that stopped existing the moment the grid became the binding constraint. In the next compute crunch, that buyer discovers that energy availability was the silent ceiling on the AI they contracted for, and that no amount of favorable token pricing helps when the tokens cannot be produced at their volume. Meanwhile the vendor that secured firm power, spread its regions, reserved capacity in writing and moved the sensitive work to the edge delivers something the price-only competitor cannot: the predictability of a service that keeps running when the market tightens.

At Cogneris we treat power availability the way we treat any other supply input to the platform — mapped, diversified across regions and deployment shapes, and partly resolved at the edge where the sensitive and steady document work belongs, so that a constraint upstream degrades gracefully instead of going dark. If you are running vendor diligence on a document AI programme, or rewriting the continuity section of a procurement template before the next renewal, see our product page, the pricing page, or talk to our team. The token price is what the headline writes about; whether the power exists to produce the token is what the SLA turns on.