You Built on Cheap AI. The Cheap Part Is Ending.

For the past two years, the obviously smart move was to wire AI into everything. Coding, the support queue, the SOC, the sales pipeline, the quarterly board deck. At the prices on offer, it was hard to argue. Twenty dollars a seat for something that drafts, summarizes, triages, and never calls in sick. Companies didn't just adopt these tools. They restructured around them. Headcount plans, budgets, entire workflows now assume the AI is there and assume it costs roughly what it cost the day someone signed up.

Here's the problem. Those prices were never real. They were a land grab, funded by venture capital and hyperscaler balance sheets, priced to win your dependency rather than to cover their costs. And in the last twelve months, quietly and then not so quietly, the vendors have started correcting that. If your business now runs on AI the way it runs on electricity, you should care a great deal about what happens when the meter shows up.

The repricing has already started

This is not a forecast. It's a timeline.

The first tremor was Cursor, in June 2025. The AI coding tool swapped its predictable monthly plan for metered usage at API rates, users burned through a month's allocation in a few prompts, and the CEO was publicly apologizing within three weeks. It read at the time like a startup fumbling a pricing page. It turned out to be a preview.

Since then the pattern has repeated at companies that do not fumble pricing pages. In April 2026, Anthropic ended flat-rate enterprise pricing for Claude and moved large organizations to usage-based billing; one licensing analyst estimated the change would double or triple costs for heavy users. Weeks later, GitHub Copilot replaced its request allowances with token-metered credits billed at API rates, and reports followed of power users' bills jumping 10x to 50x. Microsoft raised consumer Microsoft 365 prices for the first time in twelve years, up 43% on the personal plan, with Copilot as the stated reason, and has commercial increases landing this month. Salesforce raised list prices 6% and pointed at AI value to justify it.

And at the top end, the bills have crossed into the surreal. Axios reported this spring that one consultant's client spent half a billion dollars in a single month after failing to put usage limits on employee Claude licenses. Microsoft itself cancelled most of its internal Claude Code licenses, reportedly in part over cost. When the company with the deepest pockets in software is doing cost containment on AI tooling, the era of not looking at the invoice is over.

The discount was the business model

None of this should surprise anyone who looked at the underlying economics, which is another way of saying it surprised almost everyone, because nobody wanted to look.

OpenAI is projected to burn around $14 billion this year, up from roughly $8 to $9 billion in 2025. By some reporting it spends nearly two dollars for every dollar it earns on inference. The hyperscalers are pouring somewhere north of $600 billion into AI infrastructure in 2026 alone, and that capital expects to be paid back. Gartner now says the quiet part plainly: vendors are shifting from subsidized growth to profitability. Writer's CEO put it even more plainly: these companies are going public, and they're going to raise prices because they have to.

The confusing part is that the headline price of AI keeps falling. The cost per million tokens has dropped dramatically and will keep dropping. Both things are true, and the trap lives in the gap between them. A chat question was a couple thousand tokens. The agentic workflows everyone deployed this year burn fifty thousand to five hundred thousand tokens per task, and they run all day. Per-token prices fell, consumption grew faster, and total bills went up. Gartner's June prediction is the one that should stop a CFO mid-sentence: by 2028, the AI coding costs for a developer will exceed that developer's average salary. The same firm expects half of generative AI projects to blow through their budgets by 2028, and notes that inference, the boring recurring part, is at least 70% of a model's lifetime cost. The expensive part starts after the pilot succeeds.

Why you can't just walk away

A price increase is survivable if you can walk away. Most organizations no longer can, and a lot of them haven't noticed yet.

A Zapier survey of 542 US executives this spring found that 90% believed they could switch AI vendors within four weeks. Of those who actually attempted a migration, fewer than half reported it going smoothly. The gap between those two numbers is where the sticker shock lives. Model-specific prompts, tuned guardrails, vendor APIs threaded through your workflows, undocumented dependencies nobody mapped. Even Microsoft, moving its own engineers off Claude Code and onto its own product, had to give teams a hard deadline and weeks of migration work.

Then there's the ratchet that makes this a business risk rather than a procurement annoyance. Companies didn't just buy these tools, they removed the human capacity the tools replaced. Tech firms have cut over a hundred thousand jobs in 2026 while telling investors AI writes most of their code. One CEO told Axios that when AI costs bite, workforce cuts may be "the only lever they can pull." Play that forward to renewal time.

The vendor raises the price, the manual process was deprecated eighteen months ago, and the people who ran it are gone. Whatever happens in that meeting, it isn't a negotiation.

In security, the version of this I worry about is specific. If AI triage let you run the SOC with fewer analysts, and the renewal comes in at three times the price, you cannot rehire the night shift by Thursday. A dependency you can't unwind, priced by someone else, is exactly the kind of single point of failure we'd flag in anyone else's architecture.

Let me argue the other side for a paragraph

The optimist's case is real, so it deserves a fair hearing. Inference keeps getting cheaper; Gartner expects the cost of running a given model to fall more than 90% by 2030. Open-weight models now deliver a large fraction of frontier capability at a fiftieth of the price, and DeepSeek's aggressive cuts this spring show real price competition exists. Roughly a third of enterprises already run five or more models in production. Cheap, good-enough AI will exist. But notice what every one of those escape routes requires: portability. The falling prices rescue the companies that kept the option to move. They do nothing for the ones welded to a single vendor's most expensive model with the fallback process deleted.

What I would do about it

Run the tabletop. Pick your most embedded AI tool and ask what happens if the price triples at renewal. Not rhetorically. Walk it through the way you'd walk through an outage, because that's what it is. If the honest answer is "we pay," you've learned your negotiating position before the vendor does.

Meter yourself before they meter you. Most organizations discovered their AI spend the way they once discovered their cloud spend: on the invoice. Know your usage per team and per workflow, set caps and alerts, and give the number an owner. The FinOps world has already pivoted; nearly all of the discipline's practitioners now manage AI spend, up from two-thirds a year ago.

Buy AI the way I keep telling you to buy an MSSP: skeptically, in writing. Price protection at renewal, notice periods for pricing-model changes, caps on usage-based charges, and clear data-export terms. A vendor who won't discuss any of that is telling you what the renewal will feel like.

Keep the fallback warm. The manual process you're about to deprecate is your leverage. Document it before the people who know it leave, and think hard before you cut the human capability you'd need to get back.

And keep the exit real. Multi-model isn't overhead, it's insurance, the same way multi-cloud was. If your prompts, guardrails, and workflows only work on one vendor's model, you don't have a supplier. You have a landlord.

The tools are real and worth paying for. That was never the question. The question is whether your business still works at a price you don't set. A lot of companies are about to find out. The ones that answer it now, on paper, before the renewal lands, will be fine.

The rest are going to learn what the discount was for.

Sources

*Gartner press releases on AI coding costs (June 2026), GenAI budget overruns (June 2026), and inference costs (March 2026); Axios, "AI sticker shock hits corporate America" (May 2026) and "AI may never be as cheap as it is today" (March 2026); TechCrunch on Cursor's pricing apology (July 2025); PYMNTS on Anthropic's enterprise usage-based billing (April 2026); GitHub blog and TechTimes on Copilot's move to AI Credits (April–June 2026); CNBC and Tom's Hardware on Microsoft 365 price increases (January 2025); Salesforce pricing announcement (June 2025); The Register on the Zapier vendor lock-in survey (April 2026); CX Dive on Gartner's cost-per-resolution analysis (March 2026); Forbes, "AI Costs More Than the People It Replaced" (July 2026); Artefact, "Is AI Really Getting Cheaper?" (April 2026); CIO Dive on the State of FinOps 2026 survey (February 2026); InfoWorld on DeepSeek's price cuts (2026); a16z enterprise AI survey (June 2025).*

You cannot fully predict the incident that will test your organization. You can absolutely control the quality of the people, processes, and authority structures that respond to it. That preparation is the difference between organizations that recover in days and those that recover in quarters.