The power bill behind the token.

The number that moved — inference, not training

For most of the last cycle the mental model of AI energy was training-shaped: a giant one-time burn to build the model, then a comparatively cheap tail of serving it. That model inverted. As deployment scaled, the running system — every extraction, every classification, every validation pass — became the dominant consumer, and the accepted picture now puts roughly two-thirds of frontier compute energy on the inference side of the ledger, not the training side. The build is a spike; the serving is a plateau that never comes down while the product is live.

Three things happened at once, and they reinforced each other. Data-centre electricity demand for AI grew at double digits a year. The build-out concentrated in a handful of regions, and electricity prices in exactly those regions rose as the new load arrived. And the workloads that grew fastest were the high-volume, always-on kind — document processing squarely among them. The consequence is not a sustainability talking point. It is an arithmetic one: power per query stopped being a rounding error and became a line a CFO can see. The token-price reset covered the number on the invoice; this is the number underneath it, and the two do not always move together.

The price of the token is the number on the invoice. Power per query is the number under it — and in a pipeline that runs the same call a million times a day, the one underneath is the one that compounds.

Why energy landed in document AI unit economics

Document AI is unusually exposed to this shift, for the same structural reason it is exposed to the price reset. The workload is high-volume and steady — an accounts-payable pipeline, a KYB packet line, a claims-intake flow runs the same shape of call continuously, so any per-call energy cost multiplies straight through to the contract. There is no spiky, occasional usage pattern to hide the draw inside. Every efficiency or inefficiency in the energy line shows up at full volume in the denominator of cost per document.

The second reason is that document work increasingly runs on the heavier end of the model stack. A structured extraction with per-field rationale, a multi-pass validation, a reasoning step over a contract clause — these are not the cheapest calls, and when a team reaches for a cost-per-successful-task view of the pipeline, the energy behind the heavy calls is part of what they are actually paying for. The unit that matters is not cost per million tokens and not kilowatt-hours in the abstract. It is energy per decision — the power drawn to turn one document into one trustworthy output — sitting on the dashboard next to cost per decision, because the two can diverge and the divergence is where the margin leaks.

The efficiency frontier moved the other way

The story is not one-directional. While the demand curve climbed, the efficiency curve climbed with it, and a team that reads only the demand side misses half the picture. Several levers compounded across the cycle, and each one moves energy per decision independently of the token price:

The last row is the one a document AI team controls directly. The others are supplied by the provider and the chip vendor; routing is the lever that lives in the pipeline. A platform that sends a trivial classification to a frontier reasoning model is paying frontier energy — and frontier cost — for a decision a small model closes at a fraction of both. The compute-economics piece covers the supply side of why these levers exist; the point here is that they turn energy per decision into something you engineer, not something you accept.

FinOps of energy: the frame that answers it

The discipline that emerged to hold this is a widening of FinOps to include power, not just spend. The shape has five parts, and none of them is exotic — they are the same moves a mature cost function already makes, extended one axis.

Cost the workload in power, not only in dollars

The first move is to carry an energy figure alongside the dollar figure for each workload class. A document class that runs on a heavy model at high volume has an energy profile that the dollar line alone hides, especially when the token price falls but the power draw does not. Teams that measure both catch the case where a price cut masks a workload that is still energy-heavy — and is still exposed if the power line, not the token line, is the one that moves next.

Place the workload by energy intensity × SLA × jurisdiction

Placement stops being a latency-only decision. The same extraction can run in a region on clean, abundant power or in a region where the grid is strained and the marginal kilowatt is expensive and dirty. The mature posture places each workload against three axes at once — the energy intensity of the region, the SLA the tenant is owed, and the jurisdiction the data must stay in — rather than defaulting every call to the nearest or cheapest zone. The air-gapped and deployment-tier piece covers the data-path side of the same placement decision; energy is the axis that got added to it.

Route the commodity call to the efficient model

The routing lever from the table is the operational heart of the frame. A policy that maps a call's signature — document class, field, expected reasoning depth — to the smallest model that clears the quality bar keeps frontier energy for the calls that genuinely need it. This is the same router that captures a token-price drop; it captures an energy saving on the same policy change, because the two savings travel together when the routed model is both cheaper and lighter.

Reserve capacity where the power is clean

For the predictable base load — the steady, plannable volume — reserving capacity in regions with clean and available power is both a cost hedge and a sustainability position. The spiky exception can burst elsewhere; the base load, which is most of a document pipeline, is exactly the part that rewards a deliberate placement decision made once and held.

Report energy per decision as a first-class metric

The metric that makes all of this legible is energy per decision, reported next to cost per decision on the same dashboard. A team that reports only dollars cannot answer the procurement question that is now being asked, cannot see the workload where cost fell but power did not, and cannot show a regulated buyer the number that buyer's own disclosure regime increasingly requires. The metric is the difference between managing energy and hoping it stays small.

Sustainability moved from ESG footnote to RFP line

The commercial half of the story is that disclosure stopped being a report filed once a year and became a criterion scored inside procurement. Regulated buyers — banks, insurers, large enterprises with their own reporting obligations — now ask a document AI vendor where the inference runs, on what power, with what efficiency posture, and increasingly want it in a form they can put into their own scope-3 accounting. A vendor with a credible footprint and an energy-per-decision number to show clears a filter that a vendor with a hand-wave does not.

This is the reframe that matters for a platform team: sustainability is not a cost centre bolted onto the pitch. It is an asset in the RFP, in the same box as the audit trail and the SLA. The buyer's diligence packet — the same one the provider due-diligence piece maps for compute risk — now carries an energy-and-sustainability section, and the vendor who can answer it with a number wins the section instead of dodging it. The audit-evidence piece makes the parallel point on the governance side: the thing you can evidence is the thing that survives diligence.

Four ways teams get it wrong

The teams that lose margin here rarely lose it on a technical failure. They lose it on four habits carried over from the era when energy was genuinely a rounding error, and each stays invisible until a bill or an RFP makes it visible.

Optimising the token price and ignoring the passed-through power bill. A team that celebrates a token-price cut without checking whether the power draw moved with it is optimising one line and missing the one underneath. When a provider's cost structure is dominated by energy the provider itself is paying more for, the token price can hold or fall while the real cost — and the pass-through the vendor eventually eats or bills — is being set by power. Read both lines or manage neither.

Assuming the cloud is always greener. The reflex that a hyperscale region is automatically the lower-carbon choice does not survive contact with the grid mix of the specific region at the specific hour. Some on-prem or region-pinned placements, on cleaner local power, beat the default cloud zone on both carbon and marginal cost. The honest posture measures the actual placement rather than assuming the cloud averages out in your favour.

Having no disclosure when the regulated buyer already asks. A vendor selling into banks and insurers who cannot produce an energy-and-footprint answer is not neutral in the RFP — it is behind, against a competitor who can. The gap is not discovered in engineering; it is discovered in the diligence packet, at the worst possible moment to start building the number.

Running commodity workloads on frontier, energy-intensive models. The most common and most expensive habit: sending trivial, high-volume calls to the biggest model because it is the one the team trusts. That pays frontier energy and frontier cost for decisions a small model closes at a fraction of both — and at document-pipeline volume, the waste is not marginal, it is the largest controllable line in the energy budget. Right-sizing is not a nice-to-have; it is where most of the saving lives.

The trade-off, stated honestly

None of this is free, and the frame is dishonest if it pretends otherwise. The three main levers pull against real things. Efficiency versus latency: the lightest model on the cleanest distant region is not always the fastest, and a tenant on a tight SLA may need a heavier, nearer, less-efficient placement — the router has to respect the SLA, not just the energy score. Efficiency versus chip lock-in: the most energy-efficient inference silicon is often the most proprietary, and optimising hard for one vendor's part is the same lock-in risk the provider-portfolio discipline warns about, now wearing an energy label. Clean-region reservation versus flexibility: committing base load to a clean-power region for a cost-and-carbon hedge trades away some of the freedom to move that workload later.

The resolution is not to pick a side but to make the trade-off explicit per workload class. The high-volume commodity base load rewards aggressive right-sizing and clean-region reservation, because it is steady and latency-tolerant. The latency-critical exception justifies the heavier, nearer, less-efficient placement, because the SLA is the product. A team that applies one energy policy to every workload gets the trade wrong in both directions; a team that segments by workload class gets the saving where it is cheap and pays for latency only where it is owed.

What this means for document AI specifically

Document AI sits where the energy line and the commercial line meet, for the same three reasons it sits at the centre of the price reset. The workload is high-volume and steady, so energy per decision maps straight to contract margin. The schemas are stable enough that a right-sizing decision can be evaluated against a held-out set before it ships. And the buyers are regulated enough — banks, accounting firms, insurers — that the energy-and-sustainability question will be on the diligence packet whether the vendor prepared for it or not.

Energy per decision belongs on the dashboard. A platform that reports cost per document but not the power behind it is managing half the unit economics and cannot answer the procurement question at all. The metric is small to add and expensive to lack.

The route table is an energy instrument too. The same policy table that captures a price cut captures an energy saving, because the right-sized model is usually both cheaper and lighter. A team that already runs the router for cost gets the energy win on the same surface — it just has to measure it.

Sustainability is part of the product, not the pitch. The energy figure, the placement policy and the disclosure a regulated buyer can consume are as much a feature as the extraction accuracy. In an RFP that scores them, the vendor who treats them as product wins the section the vendor who treats them as marketing loses.

Closing thought

The teams that still reason "AI cost equals token price" are optimising the visible line and missing the one that now moves underneath it. Energy overtook training, power per query entered the unit economics of every high-volume document pipeline, and sustainability walked out of the ESG report and into the RFP. The platform that measures energy per decision, right-sizes the model that serves each call, places the workload against intensity and SLA and jurisdiction together, and can hand a regulated buyer a credible number is the platform whose margin and whose proposals survive the shift. The one that reasons in token price alone pays a quiet power tax on every document it processes — and loses the RFP section it never knew was being scored.

At Cogneris we treat energy per decision, the route table and the placement policy as part of the product — because the power line moves independently of the token line, and the vendor whose unit economics survive the move is the vendor whose customers stay through it. If you are sizing the energy-and-cost side of a document AI programme, or answering the sustainability section of a procurement packet 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; the power behind it is what the margin turns on.