Insights

The 20-dollar illusion

The cheapest thing about AI is the price. That is exactly what should worry you.

In one of my previous columns I wrote that the model your company uses is rented, that your competitors rent the same one from the same providers, and that the only advantage you can own is your knowledge. This year, the invoices began to show what the renting costs.

The meters have started

In June 2026, GitHub, where much of the world’s software is built, moved its AI assistant to usage-based billing. The subscriptions stay and include a monthly allowance of AI Credits. Consumption varies with the model and the volume of tokens. Simple suggestions while you type stay included. Longer agent sessions and advanced models can cost extra.

Anthropic is even more explicit with new Enterprise customers: $20 per seat per month buys access to the platform. Usage is billed separately, token by token, at standard API rates. The 20-dollar-per-person era will fade quietly, one new meter at a time.

Why the flat price broke

Flat seat prices were built for an assistant you interrupt a few times a day, and they break the moment that assistant works for hours on its own. Deloitte estimates that inference, the computing required to run trained models, will account for roughly two-thirds of all AI compute in 2026. The cost of a single operation keeps falling, but the volume grows much faster.

The early prices look, in hindsight, partly like an adoption strategy. The discount was part of the product. While they courted you as a customer, the discount also bought market share.

We have watched this film before

The transport analyst Hubert Horan, examining Uber’s accounts for the year to September 2015, estimated that passengers were covering roughly 41 percent of the true cost of their rides. Years of subsidized travel changed people’s habits. By 2021, driver shortages had pushed rideshare fares to record levels, while demand remained.

The cloud taught enterprises that the entry price and the living price are two different prices. The American software company 37signals reported reducing its annual cloud bill from $3.2 million to roughly $1.3 million after moving what suited it onto its own hardware. On the way out, AWS applied its exit waiver to what would otherwise have been a roughly quarter-million-dollar data-transfer charge. That leaving can carry such a bill is why exit terms matter.

The recurring pattern is low entry cost, then habit, then a fuller bill. AI can add technical lock-in and combine all three faster.

The same split is already in your Office license

The new Microsoft 365 E7 package costs $99 per user per month and bundles the familiar E5 suite with Copilot and controls for supervising automated assistants. Much ordinary Copilot and prebuilt-agent use is included for a fully licensed employee. Defined usage-based AI services and workloads are priced separately through a meter based on Copilot Credits.

The exact boundary depends on the license and the kind of work. In corporate budgets, two separate lines are already emerging: equipping the employee and running automated work.

The invoice is the smaller problem

Few companies budgeted for AI as infrastructure, because it arrived as a 20-dollar app. It entered through personal subscriptions, marketing tools and developers’ assistants, then quietly became load-bearing.

If the AI component of every task cost three times more tomorrow, what would you cancel, and what would you keep?

If you cannot answer, you do not know what you depend on.

We were late once already

I have watched this industry for twelve years and I build data foundations for a living, so treat me as an interested party. For years many mid-market companies postponed serious investment in data warehouses and foundations, and the logic was sound on paper: why pay for a foundation and governance when a good analyst with solid Excel skills can produce much of what you need manually, for less? That calculation held for a long time, and it is precisely why we are late with the foundations today. Now the same logic returns from the other side. A subsidized machine has become cheaper than a person with a spreadsheet, and it can postpone the same investment for another few years. If we miss this window the way we missed the last one, the damage will reach past reporting and into our ability to compete with companies that did not wait.

Use the honeymoon period while it lasts. Prices are still low, the tools are still generous and the mistakes are still cheap. That is the best moment to start building the foundations, and the worst possible reason to postpone them again.

Three moves before the real bill

First, map the dependency. Find where AI already carries daily work, officially and unofficially, because you cannot price what you have not mapped.

Second, stress-test at realistic prices. As a scenario, ask what still returns strongly at three times today’s price. That work is more likely to survive the next pricing change. Work that immediately becomes unjustifiable was probably a habit the subsidy created.

Third, build for efficiency. A machine that receives governed data and well-written organizational knowledge wastes less paid computing time searching, guessing and redoing, which is why the foundations I wrote about in my first columns now start paying in a second currency: accuracy first, and now cost.

The metered era adds a harder question to the one we learned during the discount era: what is AI worth to us, per use?

Once the real bill is on the table, a company will need to decide how much work it will continue to rent and what it will keep under its own control: its data, the knowledge and rules the machine works with, and the freedom to switch supplier when the price moves.

That is the next column.

Value Discovery

Bring the decision that is late.

A free Value Discovery, sixty to ninety minutes, built around your industry and your Monday morning.