When inference owns your margin.

The shift was not the model — it was the bill behind it

For three years the public story of frontier AI was a model story: parameter counts, benchmarks, release notes. The numbers that defined 2026 were not on a model card. They were on a balance sheet. The largest private financing round in history closed at around US$ 122B for a single model provider; a peer publicly committed US$ 200B+ in compute spending across two hyperscalers; a database company that few buyers thought of as an AI company queued up tens of billions to build capacity it can rent to model providers. None of this changes what a single API call looks like. All of it changes what a single API call costs to produce, and therefore what a vendor running on top of those APIs can sustainably charge.

The downstream effect on the document AI market is already legible. Per-page extraction prices for commodity OCR drifted to a floor near US$ 2 per 1,000 pages — unthinkable two years ago. Buyers who used to pay for a proprietary OCR engine now ask why they should not run an open-weight VLM on rented GPUs for one-fifth of the price. Vendors whose entire COGS sits inside a third-party API bill discovered that their pricing power lives at the intersection of someone else's gross margin and someone else's roadmap. The question stopped being "which model do we use" and started being "do we control any of the compute that runs our product?"

This is not the leaderboard race the press kept covering. It is a quieter, harder shift: compute is the moat, inference cost is the gross margin, and the buyer's procurement team learned the language faster than most vendors expected.

What changed under the hood

Three structural moves explain the squeeze that is now being passed down the stack.

Capacity got more expensive and longer-dated

The big rounds and the big commitments did not buy cheap compute — they bought the right to have compute. Multi-year reservation contracts replaced on-demand consumption as the dominant procurement shape for serious AI workloads. A model provider that signs a seven-year capacity deal can promise an API; a vendor that runs on top of that API inherits the model provider's commitment without any of the leverage. When demand spikes, the model provider rations to its largest accounts; the midmarket vendor finds out at the worst possible moment that "API as a commodity" was always conditional.

Open-weight got close enough

The gap between the best closed-weight model and the best open-weight model on a buyer's actual workload — not on a leaderboard — closed to a few percentage points in most extraction and classification tasks through 2026. For predictable, high-volume document workloads, that gap no longer justifies paying ten times more per token. The open-weight option is not free — somebody has to operate the stack — but it changes the comparison from "API price vs. zero" to "API price vs. fully-loaded inference cost on rented or owned hardware." Once the comparison is legible, the math moves a lot of cargo.

Buyers started asking unit-economics questions

The most consequential shift is procurement. The RFP for document AI in 2026 includes a line that did not exist in 2024: "what is your cost per processed page, broken down by inference, storage, validation, and human review, and how does it change at our committed volume?" Vendors who answer with a marketing number lose. Vendors who can show the build-up and explain which components are elastic and which are fixed move to the shortlist. The decision moved from finance to a joint procurement-plus-FinOps table, with the CFO present in the room.

The model is a feature. The compute is the business. A vendor that does not understand its own inference cost build-up is a vendor whose pricing power belongs to someone else.

The three decisions a CIO and a CFO have to make now

Three procurement and engineering decisions moved from "nice to have" to "load-bearing" in 12 months.

1. Negotiate reserved capacity as a multi-year contract — with a portability clause

Spot pricing on inference is the consumer experience. Enterprise pricing is committed capacity, multi-year, with the discount that comes with it. The mistake is signing the commitment without the exit. A capacity contract without a portability clause locks the buyer into the provider's future roadmap — including model deprecations, regional availability decisions, and the day the provider raises its own prices because their compute cost moved. The clause to insist on: prompts, evaluations, and routing logic must be exportable in a form that runs against a functionally-equivalent model from a second provider, with a defined transition window. Buyers who skipped this clause in 2024 paid for it in 2026; buyers who insist on it today get to keep their negotiating leverage.

2. Self-host open-weight for predictable workloads

Not for everything. For the workloads where the volume is predictable, the latency is forgiving, and the model gap is small — a large slice of document classification, bulk extraction, and routine validation — self-hosting an open-weight model on reserved GPUs is now cheaper by a factor of 5–10x on fully-loaded cost. The catch is the word fully-loaded: the cost includes the platform team that maintains the serving stack, the on-call burden, the eval pipeline, and the rollback plan when a model update breaks something. Buyers who modeled only the hardware bill and not the operations cost found that the math is real, but the staffing is real too. The right framing: self-hosting is a capacity bet, not a savings play; it makes sense when you have enough predictable volume to amortise the operations cost across.

3. Treat inference as a FinOps line with SLAs per intent

The legacy pattern was to bury LLM cost in a general "AI infrastructure" line and let it drift. The 2026 pattern is to model inference cost the way mature ops teams model cloud cost — per intent, per workflow, per customer cohort, with target unit economics and an SLO attached. extract_invoice_total has a budget of X cents per call and a target latency of Y ms; classify_kyc_packet has different numbers; the router enforces them. When a model update changes the cost curve, the FinOps system surfaces it before the monthly bill does. We covered the broader measurement gap in the AI ROI piece; this is the operating layer that closes it.

The unit-economics buyers now ask vendors to show

The single most useful artefact a document AI vendor can produce in 2026 is a per-intent cost build-up that the buyer can stress-test. Five lines, none of them optional.

The interesting line is the last one. Vendors who publish only "API cost passed through" implicitly promise zero overhead allocation — which is fine when they are a wrapper around a sub-processor and bad when they are a platform that needs ongoing investment. Buyers who do not ask end up funding the overhead through the renewal increase a year later. Buyers who ask early get to pick a vendor whose unit economics actually close.

Hedging GPU scarcity — the playbook that emerged

Capacity is not evenly distributed; demand spikes are not evenly forecast. The four-part hedge that ships in 2026 looks roughly like this.

Multi-vendor routing for the same intent. Two or more providers wired behind a single intent, with a live router that selects on cost, latency, and availability. The first time a primary provider hits a rate limit, the buyer's workload does not page on-call; it cuts over. The router carries the policy: the same prompt is evaluated against each candidate model on the buyer's eval set monthly, and the routing weights move on eval results, not on vendor pitches.

Capacity reservations sized for the trough, burst on demand. Reserved capacity sized to baseline traffic; on-demand burst on top. The reservation discount pays for itself below the trough; the burst is the price of refusing to be locked into a single provider's spot market. Vendors that reserve for peak end up renting the trough; vendors that reserve for trough get caught short on peak day. Neither extreme survives a CFO review.

Open-weight as the third leg. Even a small open-weight footprint changes the negotiating posture with the closed-weight providers. A buyer or vendor who can credibly point at an internal open-weight capability gets renewals at terms the buyer who cannot does not. The cost is operating that stack; the value is the leverage it produces upstream.

Eval as a first-class operating signal. The eval set runs on a cadence — daily on the small set, weekly on the full corpus — and the results feed the router. A model update that quietly degrades on the buyer's workload (which happens) is caught before a customer notices. The multi-agent observability piece covers the broader stack; compute hedging just adds cost and capacity as first-class signals on the same dashboard.

What this means for document AI specifically

Document AI sits in the slice of the AI market most exposed to the compute squeeze. The workloads are high-volume; the per-document margin is thin by construction; the buyer's procurement function is mature and tends to push prices toward the floor. Three moves separate a document AI vendor that survives the squeeze from one that does not.

Own the routing layer, not just the model. The model is replaceable; the routing layer, the eval set, the prompt library and the policy enforcement is the intellectual capital. A vendor whose product is "we wrote a good prompt for GPT-X" is a vendor whose product is one model deprecation away from a rebuild. A vendor whose product is "we route intents across N models, evaluate them continuously against your data, and enforce per-intent budgets and accuracy SLOs" survives a model change without a customer-facing event.

Build for verifier cost, not just inference cost. The next gross-margin lever in document AI is not cheaper tokens — it is fewer tokens spent re-verifying the same answer. The verifiable-reasoning architecture covered in the verifiable-reasoning piece looks expensive at first read and pays back through human-review savings and downstream rework that disappears. Buyers who model only the API line miss this; vendors who can show both lines in one chart win the deal.

Be honest about the unit economics in the pitch. The vendors quietly losing in 2026 are the ones who quote a per-page price below their fully-loaded cost on the first contract and plan to make it back on the renewal. The buyer who runs the build-up notices; the procurement team that did not notice in 2024 notices now. The differentiator is no longer "we are cheaper" — it is "we can show the build-up and you can stress-test it." We publish ours on the pricing page precisely because the alternative is the conversation we do not want to have in year two.

Where the squeeze hits hardest

Two vendor profiles are most exposed in the next 12–18 months.

Pure-API wrappers without their own inference leverage. A product whose entire COGS is a single upstream API bill, with no routing layer, no open-weight fallback, and no capacity reservation, is a product whose gross margin is set by a vendor it does not control. When that vendor raises prices, the wrapper has nowhere to go; when the wrapper raises prices to compensate, the buyer leaves. The exit is to build the routing and eval layer before the squeeze, not after — but the engineering cost of doing it under deadline is materially higher than doing it as a quarterly investment.

Single-vertical OCR specialists at the price floor. Specialists who priced at the old OCR margin face a floor moving against them faster than they can re-architect. The defensible move is up the stack — verification, write-back, workflow orchestration — where the value is higher and the cost is sticky. Specialists who try to defend the OCR margin alone end up consolidated into a larger platform on terms set by the acquirer. The adjacent extraction-to-execution piece describes the architectural direction.

Closing thought

The 2026 capital story is not a curiosity for AI watchers. It is a procurement signal. The buyers who read it correctly are asking unit-economics questions in the RFP; the vendors who read it correctly are showing the build-up before being asked. The vendors who did not read it correctly are about to have a year of compressed margins, awkward renewal conversations, and consolidation offers from larger platforms. Compute did not become a moat because the technology shifted. It became a moat because capital made it one.

At Cogneris we run a multi-vendor routing layer with per-intent budgets, reserve capacity sized for predictable workloads, and keep an open-weight option for the cases that justify it — not because we wrote a strategy slide about it, but because that is the only shape that lets us give a regulated buyer a stable per-page price and mean it. If you are mapping inference cost against your document workflows and want to see the build-up the way we expose it to procurement teams, see our pricing page, the product page, or talk to our team. The model is the feature; the compute is the business; the unit economics are the part that has to close.