AI compute is now a line item in your business

A plumbing company, a dental office, and a small law firm now share a strange planning problem with the world’s largest AI labs: two curves they do not control decide part of their cost structure. The first is the capability curve, which sets what AI can take over next: estimate prep, insurance verification, after-hours intake, document review. That one is a planning input. The second is the compute-price curve, which sets what a unit of that work costs and how the price moves. That one is a cost input. My claim in this essay is that both now deserve a row in the same underwriting you apply to labor, parts, and rent. The two curves answer different questions, and mixing them up is how owners end up either dismissing AI as vendor noise or buying “automation” with no idea what they are paying for.

The plateau needs a mechanism

Start with the capability side, because it carries the comfortable dodge. Someone points at a rising AI capability curve and someone else says it will flatten into a sigmoid before it matters. Scott Alexander’s Astral Codex Ten essay The Sigmoids Won’t Save You takes that dodge apart. Every exponential eventually flattens, but curves flatten because something makes them flatten: compute supply, data, algorithms, economics, regulation, product limits, user trust, physical infrastructure. Absent a named mechanism, “it will plateau” is a hope wearing graph paper.

Serious forecasters have made the mistake before. Fertility forecasts expected falling birth rates to level off too early. Solar deployment forecasts kept drawing flattening paths while actual deployment continued upward. One AI capability curve that looked ready for a bend was overshot by the next model release. So Alexander argues for a Lindy-style prior: a trend that has already run for years should be expected to run for a comparable span, unless the skeptic can name the force that stops it, when it binds, and why it beats continued spending, better chips, algorithmic progress, synthetic data, tool use, and market demand.

You may still believe in an early plateau. Fine, but the belief now owes a mechanism. A compute limit has to explain why your competitors cannot buy more compute through ordinary software subscriptions. A regulatory limit has to separate the regulated decision from the document assembly, evidence checking, and scheduling around it. Until you can fill that in, the honest planning posture is to treat continued capability growth as a live underwriting assumption.

What the assumption does to your numbers

Underwriting the assumption is concrete work, not futurism. Small companies live inside thin financial models, where a few labor hours per job decide whether a line of work is viable. If capability keeps compounding, intellectual labor becomes variable in places you have always treated as fixed: preparing estimates, inspecting job packets, comparing plan rules, chasing closeout evidence, writing customer follow-up. That moves the margin math whether or not you repriced it, and it changes what a buyer should pay for a business that still depends on one person’s memory.

So reprice around bottlenecks instead of job titles. Start where margin leaks today: owner approvals, rework, callbacks, claims rejected for missing proof, estimates waiting on private memory, handoffs that rely on one senior person. Then ask what happens to each loop when its clerical, comparison, drafting, and monitoring parts get two orders of magnitude cheaper or better.

Some bottlenecks will not move. Physical execution remains physical, compliance authority remains authority, customers still decide whether the promise was kept, and you still choose what the business refuses to do. Others move a lot. The scheduler’s scarce contribution shifts from typing notes to maintaining the dispatch rule. The estimator shifts from writing every proposal to defining the evidence standard that makes proposals trustworthy. Which points at the durable position for you and your best people: “I do the task” is getting weaker as a defense, because the task is exactly what gets cheaper. “I own the standard behind the task” is the stronger position, held by the person who knows the source material, the failure modes, and the compliance boundary, and who can supervise agents, correct the playbook, and decide when the system should stop.

The cost side is getting a curve

Now the other input. Two announcements in one week of May 2026 showed markets trying to put a forward price on AI compute. ICE and Ornn announced on May 19 that they plan to launch U.S. dollar denominated, cash-settled GPU compute futures based on Ornn’s Compute Price Index, pending regulatory approval; ICE framed the contracts around price discovery and hedging for global compute, and Ornn says the index tracks live-traded spot prices across major GPU types, built from printed transactions. A week earlier, CME Group and Silicon Data announced plans to launch compute futures later this year, also pending regulatory review, based on Silicon Data’s GPU benchmark indices. Ornn’s index has also been added to the Bloomberg Terminal, an announcement that makes the financing angle explicit: transaction-based benchmarks are meant to help lenders, operators, and capital providers evaluate compute with more discipline.

All of this is announced, not launched. Products may change, approval may lag, liquidity may take years. The operating point survives anyway: the market is working on making compute a benchmarked, financeable input. If the contracts launch as planned, AI stops looking like a mysterious vendor cost and starts looking like a priceable one, because your vendor’s monthly fee sits on top of a curve that lenders, vendors, and large buyers can see, quote, and hedge.

The curve reaches you through contracts

You will never touch a futures contract, and you do not need to. The consequences arrive through vendor pricing. Today the compute cost under your AI phone agent or claims-review tool hides inside a usage tier, a margin assumption, or a rate limit. A vendor that can see or hedge part of its input cost can offer firmer terms, and I expect four contract shapes to show up in SMB-facing software as these markets mature. Fixed-price contracts for bounded workflows: a set price for a certain number of calls, estimates, claims, or documents. Reserved tiers, where a franchise group or multi-location practice locks a volume of AI work at a known rate. Spot tiers, where low-urgency batch work (old-file cleanup, outbound follow-up drafts, research batches) runs cheaper when capacity is cheap, while live customer work sits on reserved capacity. And pass-through clauses, where a compute spike adds a surcharge referenced to a published index instead of a vendor story. Buyers will dislike pass-throughs; they will dislike them less when the clause points at a public number both sides can check. Benchmarks do not need to be perfect to change a negotiation. They give both sides a shared object to argue over.

Purchasing power concentrates upstream, too. A single office gets little from any hedge; a dental support organization, home-services franchise system, or rollup with centralized procurement can compare vendors on compute-risk handling and negotiate across locations. If you are a single-location buyer, your leverage is better questions, which is the next section.

Price the workflow, count the whole cost

The unit of analysis is the workflow, not “AI.” After-hours intake is a different economic input than photo review for estimate packets; insurance verification is different from appeal-packet assembly; dispatch support is different from customer follow-up writing. Each carries its own cost curve, risk profile, and sensible contract shape. The useful buying posture is to pay against the work record: per qualified booked call, per completed estimate packet, per reviewed closeout file, rather than per seat.

And compute is never the whole cost. A cheap phone agent that books bad calls fills the board and burns technician time. An estimate tool that misses property context creates callbacks and discount pressure. A document agent that produces more drafts than your reviewers can inspect converts compute savings into management load. So the workflow record has to show the full economics: AI cost, human review time, retries, customer corrections, rework, denied claims, callbacks, refunds, and manager interventions. A compute saving only counts when that record says the job state actually improved.

On the next renewal call, skip “do you use AI?” and ask:

• Which workflow unit are we paying for?
• What volume is included before the price changes?
• Which tasks run on reserved capacity, and which only when capacity is cheap?
• What happens if compute prices spike?
• Is any pass-through clause tied to a public index?
• Can we audit job counts, retries, failed runs, and human review burden?

Underwrite both directions

The honest close is that both curves are uncertain. The futures are announced plans that may stall or change shape; the capability trend may flatten sooner than the Lindy prior suggests; and none of this is financial, investment, or legal advice. But uncertainty is what underwriting is for. If you plan around a plateau, name the bottleneck you are betting on and when it binds. If you plan around the curves continuing, name the workflow unit you are buying and who carries the risk when prices or volumes move. Revisit quarterly. Owners do not get paid for guessing the shape of the curve in the abstract; they get paid for knowing which line item changes while it keeps going, and for moving themselves and their best people from doing the task to owning the standard behind it.

---

Source evidence used in this note: Scott Alexander, Astral Codex Ten, The Sigmoids Won’t Save You, published 2026-05-15, for the mechanism requirement, the fertility and solar forecasting precedents, and the Lindy-style prior. ICE and Ornn to Launch GPU Compute Futures Contracts, CME Group and Silicon Data Partner to Launch First Compute Futures, and Ornn Compute Price Index Added to Bloomberg Terminal for the compute-market announcements, all pending regulatory approval as of writing. Hadto interpretation: the two-curve underwriting frame, the bottleneck repricing, the four contract shapes, the workflow-unit pricing posture, and the buyer questions are operating judgments, and this note is business-design discussion, not financial, investment, or legal advice.