AI Adoption Race Is Over. The AI Cost War Has Begun.

The AI race inside companies used to be about adoption. That is over, and the AI cost war has begun, with Amazon, Walmart, and Uber moving from pushing AI tools on employees to capping their use. The file is moving from the IT department to the chief financial officer’s desk.

Frontier AI was sold like a software subscription and is being repriced like a utility. Anthropic and OpenAI now meter tokens per prompt, while outcome-based vendors charge per resolved task. The gap between the cheapest and the most expensive frontier models is widening, with OpenAI’s GPT-5.5 Pro costing 150 times as much per token as GPT-5.4 Nano. The framing comes from a Friday essay for The National: Amit Joshi’s argument that the AI cost war has begun, by the IMD professor of AI, analytics, and marketing strategy.

The Pricing Reset

The reset is fragmenting the pricing page in three directions. Vendors are sorting into flat subscriptions, token-based meters, and outcome-based charges, and the choice between them changes the cost curve.

Flat subscriptions still exist for a reason. Microsoft 365 Copilot Enterprise sells at $30 per user per month with an annual commitment, a line item a chief financial officer can multiply by headcount and forget about. Token-based pricing replaces that line item with a meter. Anthropic’s Claude Code bills per token consumed across model calls, and engineers running autocomplete consume a fraction of what engineers orchestrating parallel agents consume on the same tool, the same day. OpenAI has run the same direction, with chief executive Sam Altman saying in March that intelligence will be sold like electricity or water.

Outcome-based pricing is the third leg, with Zendesk charging $2 per AI-resolved support ticket at pay-as-you-go rates, or $1.50 per resolution on committed volume. Critics flag a double-billing risk: when AI escalates to a human, the customer pays the agent seat license on top, a fact laid out in Zendesk’s outcome-based pricing breakdown.

Tokenmaxxing and the Adoption Backlash

Amazon built an internal AI leaderboard called KiroRank, ranking employees by how much they used the tools, then scrapped it after engineers gamed the system by assigning agents pointless work to inflate their token counts. Amazon staff labeled the practice tokenmaxxing.

Please don’t use AI just for the sake of using AI. Use AI to help you solve customer problems, to help you solve business problems, to innovate.

David Treadwell, senior vice president at Amazon, told staff to stop optimizing for the metric, per the Financial Times report on the shutdown. The leaderboard had been built as part of a push to get 80% of Amazon developers using AI in a given week, in the shadow of a $200 billion AI investment. Treadwell’s message was that the game had produced the wrong answer, and the cost of those extra tokens now lands on the same financial statement the adoption program was supposed to defend. The reversal is detailed in Amazon’s shutdown of its internal AI leaderboard.

Uber hit the same wall harder. The ride-hailing company exhausted its entire 2026 AI coding tools budget by April, four months into the year, after rolling out Anthropic’s Claude Code to its engineering organization the previous December. Chief technology officer Praveen Neppalli Naga confirmed the overrun to The Information, saying finance teams had been forced back to the drawing board on their assumptions. Use climbed from 32% of engineers in February to 84% by March, with 95% of Uber engineers on AI tools monthly, roughly 70% of committed code coming from them, and per-engineer monthly cost ranging from $150 to $250 on average and $500 to $2,000 for power users. The full breakdown is in the Uber CTO confirming the AI budget burn.

Uber compounded the dynamic with internal leaderboards that ranked engineers by Claude Code usage, turning the adoption metric directly into a consumption metric. Uber president and chief operating officer Andrew Macdonald told Fortune that the link between rising token use and shipping more consumer features is, in his words, “not there yet.”

Does AI Pay for Itself?

The hard numbers on AI returns come from two companies. The rest of the field has not yet put hard figures on the table, and most enterprises cannot show sustained productivity or profit gains at scale. IBM and JPMorgan Chase are the cited proof points.

IBM said in July 2025 it was on track for $4.5 billion in annual run-rate productivity savings by the end of that year, up from $3.5 billion at the end of 2024, driven by its Client Zero program of running AI internally first. Chief financial officer James Kavanaugh called the gains a flywheel that funds further investment, and chief executive Arvind Krishna said IBM has embedded AI across more than 70 workflows. The full detail is in IBM’s $4.5 billion AI productivity savings disclosure.

JPMorgan Chase is the other cited proof point. Chief executive Jamie Dimon said in October 2025 that the bank spends about $2 billion a year on AI and saves roughly the same, an even trade the company has been willing to defend in public. Dimon framed the parity as the moment AI moves from experiment to infrastructure. A 2026 Gartner study cited in the Uber coverage forecasts AI agent software spending will reach nearly $207 billion in 2026, up from $86.4 billion in 2025. The productivity claim has to clear a much higher token-cost hurdle than the autocomplete case, and few have shown the multiplier.

• IBM on track for $4.5 billion in AI and automation run-rate savings by end of 2025
• JPMorgan Chase: ~$2 billion saved against ~$2 billion spent on AI annually
• Uber burned its 2026 AI coding tools budget in four months, by April
• IKEA’s Billie chatbot handled 47% of customer enquiries; 8,500 staff retrained

The Workflow Rework

Most enterprise AI deployments are still timid. The common pattern is to make existing tasks slightly faster: meeting notes, document summaries, sharper analysis. The gains are real but rarely scale enough to justify a runaway meter bill, and Joshi argues the bigger prize is redesigning the workflow around the model.

IKEA is the cited counter-example. The flat-pack retailer deployed an AI chatbot named Billie through its largest franchise operator, Ingka Group, and the bot handled 47% of all customer enquiries from 2021 onward, covering order status, product availability, and return procedures. IKEA retrained 8,500 of them as remote interior design advisers. The retrained team became a new revenue line, with Ingka setting a target of 10% of total revenue from the design advisory channel by 2028. Joshi frames the case as turning a cost-saving exercise into a fresh revenue stream, and the pattern runs in three steps.

1. Audit the 47% (or your deflection rate). Track the cases AI handled and how it handled them, end to end.
2. Examine the 53% (what AI could not handle) for patterns. The clusters point to a gap, and the gap often points to a service.
3. Build the service around the gap that AI cannot fill. IKEA’s revenue came from what the chatbot could not do, and the design advisory line was built from there.

The Portfolio Solution

The cheapest model is not always the most cost-effective. Many companies fixate on per-token price, missing that the right comparison is per-workflow cost.

A frontier model like OpenAI’s GPT-5.5 Pro can cost many times more per token than a smaller or open-source alternative such as DeepSeek. If the frontier model finishes the same job in fewer steps, makes fewer mistakes, and demands less human intervention, it can still deliver the lower overall cost. The right comparison is the cost of completing a workflow, not the cost of one prompt. The choice changes the answer.

Increasingly, the right strategy is a portfolio rather than a single model. Frontier models should handle the work that demands sophisticated reasoning, and cheaper models should shoulder routine tasks. That bifurcation changes the math for the vendors, too.

In traditional software, the heaviest users are the most profitable customers, because the cost of serving them barely rises once the code is built. AI inverts that logic. Every prompt consumes computing power, and the largest customers can be the most expensive to serve. Joshi’s read is that the model builders now face the same cost-control problem their customers do, just on the supply side. He frames the coming gap as one between companies that match the right model to the right task and companies that buy frontier reasoning for work a lawn-mower model would handle. Five years from now, he writes, the companies pulling ahead will not be those that spend the most on AI but those that understand their own workflows best. The competitive edge is spending less on the wrong AI, not more on the right one.

From the IT Department to the CFO

The cost war is also a governance war. AI can no longer be the sole preserve of chief information officers and chief digital officers, and Joshi argues procurement teams and even chief financial officers have to understand what drives token consumption and how quickly costs spiral as usage accelerates. The companies still treating AI as a software line item are the ones that will miss the curve. Seven new MAI models from Build 2026 are a sign that the supply side is already hedging against the same risk.

Forecasting is the second problem. AI breaks the economics of traditional software, because the bill climbs every time the tool is used, and demand forecasting is now a different discipline. CFOs who built 2026 budgets on per-seat assumptions are re-cutting them mid-year, and that is a process problem, not a software problem.

CFOs and procurement teams will set the next round of rules. The 2026 pattern is caps: Amazon, Walmart, and Uber have all tightened employee AI controls and put ceilings on token use. IT absorbing the governance cost of employee-built apps is the parallel story on the employee side, where the meter problem has spread past engineering. Per Joshi’s analysis for The National, the companies that pull ahead are the ones that know their own workflows well enough to match the right model to the right task. Get the model wrong, and the meter keeps running.

Frequently Asked Questions

What is the AI cost war?

The shift from AI adoption to AI cost control defines the new phase. Companies that pushed hard to get employees using AI tools are now capping token use, scrapping usage leaderboards, and rebuilding budgets after runaway bills. Procurement and finance teams are taking the file from IT.

What is tokenmaxxing, and why did Amazon scrap its KiroRank leaderboard?

Tokenmaxxing is the practice of assigning AI agents pointless work to inflate personal AI usage metrics, a behavior Amazon staff named after watching engineers game an internal leaderboard. Amazon scrapped the KiroRank tool, with senior vice president David Treadwell telling employees to stop using AI for its own sake.

Are companies actually saving money with AI in 2026?

Two companies have put hard numbers on the table. IBM said in July 2025 it was on track for $4.5 billion in annual run-rate productivity savings by the end of that year. JPMorgan Chase said it spends about $2 billion a year on AI and saves roughly the same. Most other enterprises have not yet shown sustained productivity or profit gains at scale.

How are AI vendors shifting their pricing models?

Flat per-seat subscriptions are giving way to token-based meters at Anthropic and OpenAI, with outcome-based pricing following at vendors like Zendesk, which charges $2 per AI-resolved support ticket at pay-as-you-go rates, or $1.50 on committed volume.

What is the cheapest way for a company to use AI?

Cheap is not the same as cost-effective. A frontier model that finishes a workflow in fewer steps can deliver a lower total cost than a smaller model that needs more human intervention. The trend is toward a portfolio of frontier and cheap models: the first for sophisticated reasoning, the second for routine work.