The affordable AI era for small business is ending because the subsidized usage pricing that made it cheap is running out, and the cost of the exact same AI task can already swing by as much as 30x from one run to the next. Canva cut its own revenue forecast after what it called an AI cost "blow out," Wix's earnings took a hit from "high AI related expenses," and Uber capped internal AI spending after burning through its 2026 AI budget in four months on Claude Code (Gene Marks, Forbes, September 17, 2026). Those are large companies absorbing a shock that has not reached most small businesses yet. It is coming, and the unpredictability is the harder problem, not just the price.
Key Takeaways- Canva, Wix, and Uber all reported AI cost shocks in 2026, with Uber burning through its full annual AI budget on Claude Code in four months (Gene Marks, Forbes, September 17, 2026).- Analyst Jack Gold says some businesses are seeing token costs exceed the cost of the employee doing the same work, within a month or two of heavy use (Jack Gold, J.Gold Associates, cited by Gene Marks, Forbes, September 17, 2026).- McKinsey senior partner Lari Hämäläinen says the cost of running the same AI agent task can vary by up to 30x from one run to the next, because agentic tasks are multistep and the same goal can be reached many different ways (McKinsey, reported September 2026).- A third of organizations already spend over 10% of their technology budget on AI, 60% plan to spend more next year, and about one in five say AI spending is starting to constrain operating costs (McKinsey State of AI Survey, 2026).- JPMorgan already capped some engineers' Claude usage at $2,000 a month and added extra security controls in response (JPMorgan, reported September 2026). Insight: A flat monthly retainer does not expose a client to a bill that swings 30x for reasons outside their control, because that variance sits with Maxpertise, not the client.
Key Takeaways
- Canva, Wix, and Uber all reported AI cost shocks in 2026, with Uber burning through its full annual AI budget on Claude Code in four months (Gene Marks, Forbes, September 17, 2026).
- Analyst Jack Gold says some businesses are seeing token costs exceed the cost of the employee doing the same work, within a month or two of heavy use (Jack Gold, J.Gold Associates, cited by Gene Marks, Forbes, September 17, 2026).
- McKinsey senior partner Lari Hämäläinen says the cost of running the same AI agent task can vary by up to 30x from one run to the next, because agentic tasks are multistep and the same goal can be reached many different ways (McKinsey, reported September 2026).
- A third of organizations already spend over 10% of their technology budget on AI, 60% plan to spend more next year, and about one in five say AI spending is starting to constrain operating costs (McKinsey State of AI Survey, 2026).
- JPMorgan already capped some engineers' Claude usage at $2,000 a month and added extra security controls in response (JPMorgan, reported September 2026). Insight: A flat monthly retainer does not expose a client to a bill that swings 30x for reasons outside their control, because that variance sits with Maxpertise, not the client.
Three companies with real AI budgets already got surprised by the size of the bill. Canva cut its own revenue forecast after an AI cost "blow out." Wix's earnings took a hit that it attributed partly to "high AI related expenses." Uber capped its internal AI spending after burning through its entire 2026 AI budget in four months, largely on Claude Code (Gene Marks, Forbes, September 17, 2026). These are not small businesses guessing at a budget line. They are large, sophisticated buyers who still got the number wrong, and analyst Jack Gold says the pattern already shows up at smaller scale too: some organizations are seeing the cost of AI tokens exceed the cost of the human employee doing the same work, inside a month or two of heavy use (Jack Gold, J.Gold Associates, cited by Gene Marks, Forbes, September 17, 2026).
Because the subsidy that made early usage look cheap is ending. HBR analyst Stacia Garr's framing, cited in the same reporting, is that vendors have effectively been paying companies to adopt AI through complimentary or bundled premium access, and that subsidy does not last once adoption is established (Stacia Garr, Harvard Business Review, cited by Gene Marks, Forbes, September 17, 2026). The pattern looks familiar: a service launches cheap to build a user base, the way Netflix and Uber both did, then raises prices once the habit is locked in. AI usage billing is following the same arc, and the businesses paying the least attention to their bill right now are the ones most exposed when the price moves.
Because agentic AI tasks are not a single lookup, they are a multistep process, and the same end goal can be reached through many different paths with very different costs. McKinsey senior partner Lari Hämäläinen describes it directly: "Imagine like you're running an operation, but every day there's a 30x difference in the cost," for the exact same task (Lari Hämäläinen, McKinsey, reported September 2026). That is not a slow price increase a business can plan around. It is run-to-run unpredictability on top of a rising baseline, and McKinsey's own 2026 State of AI Survey found a third of organizations already spend over 10% of their technology budget on AI, 60% plan to spend more next year, and roughly one in five say AI spending is starting to constrain operating costs (McKinsey State of AI Survey, 2026). McKinsey senior partner Tanguy Catlin, director of the McKinsey Global Institute, put it plainly: "the spend is now becoming quite material and visible" (Tanguy Catlin, McKinsey, reported September 2026).
JPMorgan gives an early, concrete example of a large buyer reacting to exactly this pressure. The bank capped some engineers' Claude usage at $2,000 a month and rolled out extra security controls limiting direct model access, a direct response to the same usage-cost exposure smaller businesses are about to face (JPMorgan, reported September 2026). A small business does not have JPMorgan's ability to absorb a surprise bill or the internal team to build spending caps and monitoring around usage-based tools. Most small businesses running AI tools today are billed the same way JPMorgan's engineers were before the caps went in: by usage, with the bill only visible after the fact.
It moves the variance to the vendor instead of the client. Why Maxpertise doesn't charge per token, credit, or seat covers the reasoning directly: the client buys a function that runs, quoted per scope in a written proposal, not metered by how many tokens a task happened to burn on a given day. That structure does not eliminate the underlying cost variance McKinsey describes. It moves who absorbs it. A 30x swing on a single task is a rounding error inside a flat monthly number. It is a surprise line item on a usage bill.
This matters across every business function, not just software development. A clinic running AI-assisted intake, a logistics operator running AI-assisted dispatch, or a law firm running AI-assisted document review faces the same underlying token-cost variance McKinsey is describing, just inside a different workflow. See the industries pages for how the embedded model handles that exposure by vertical, from HVAC to accounting.
The shock has landed on large companies first, Canva, Wix, Uber, and JPMorgan among them, because they run enough volume to see the cost variance clearly (Gene Marks, Forbes, September 17, 2026; reported September 2026). Small businesses running less volume have not felt it as sharply yet, but the same subsidized pricing and run-to-run cost variance apply to any usage-billed AI tool, regardless of size.
Because agentic AI tasks are multistep, and the same end result can be reached through very different sequences of steps, each with a different cost (Lari Hämäläinen, McKinsey, reported September 2026). A simple lookup does not have this problem. A multistep agent task usually does.
No. It protects a client from usage-based billing volatility specifically, the kind McKinsey and JPMorgan are both describing. It does not replace the need for a written spec, human review of what ships, or an honest conversation about what a given engagement will cost, which is why Maxpertise still quotes every engagement in a written proposal rather than a published rate card.
Ask how the tools it already uses are billed, usage-based or flat, and get a written answer on what a worst-case month could cost under the current model. Send a written intake describing what you are running today, and a written proposal comes back inside one business day, no call required, structured as a flat retainer rather than a usage meter.
Canva, Wix, Uber, and JPMorgan all found out the hard way that usage-based AI billing can move fast and unpredictably. A small business does not need to find out the same way. Send a written intake and get a number in writing before the exposure becomes a live problem.
Internal links to add from older posts within a week: why-we-dont-charge-per-token-or-seat (anchor: "why metered AI billing is about to get worse"), how-much-cost-hire-ai-engineer-2026 (anchor: "the cost comparison against a direct hire"), smbs-piloting-ai-before-roi-proven (anchor: "why SMBs are still cautious about full AI adoption").
Maxpertise is an AI-native engineering company. We embed native AI engineers inside your team, live in about 10 days. Please enable JavaScript to view the site, or email contact@maxpertise.net.