You laid off your engineers to save money. Then spent it all on AI.

Here's the thing about a good irony: it takes a while to land.

In 2023 and 2024, tech companies cut hundreds of thousands of engineers. The stated reasons varied. AI productivity was always somewhere in the background, if not the foreground. The implied logic was tidy: AI makes engineers more productive, a smaller team with AI does as much as a larger team without it, reduce headcount, adopt AI, maintain output, improve margins. Clean. Simple. Slide-deck-ready.

In 2026, the results are coming in. Uber deployed Claude Code to roughly 5,000 engineers. Adoption hit 84-95%. The budget assumption did not survive contact with that adoption rate. Uber CTO Praveen Neppalli Naga put it plainly: "I'm back to the drawing board, because the budget I thought I would need is blown away already." The company exhausted its entire planned 2026 AI budget within four months. One unnamed enterprise company, reported by Axios on 28 May 2026, spent half a billion dollars on Claude API calls in a single month. (Half. A. Billion. In one month. With no spending cap.) Microsoft cancelled Claude Code licences for its Experiences and Devices division in May 2026 because the tool was, in the company's words, "perhaps a little too popular." GitHub changed its Copilot billing model from 1 June 2026 to usage-based pricing, because the flat rate didn't survive contact with how engineers actually use agentic tools.

I was cautiously optimistic about the productivity narrative, if I'm being honest. I use these tools myself. I've seen real gains from them. But there's a gap between the per-seat subscription model that enterprise procurement knows how to handle and the token-based usage model that AI tools actually run on. That gap is now producing invoices.

The engineers cost a salary plus benefits. The AI tools cost a salary plus benefits plus whatever they decide to consume tonight.

Mark Ajzenstadt

@mardehaym

None of this is satire.

→ A company spent $500,000,000 on Claude in one month because nobody set usage limits

→ Uber ran leaderboards ranking engineers by how much AI they used, not what they shipped

→ Uber burned their entire 2026 budget by April. Their COO said he can’t

22.6K

28 May 2026

The Replacement Cost Thesis

The argument wasn't wrong in principle. It went something like this:

AI makes engineers more productive. A smaller team, properly tooled, can do what a larger team without AI can do. Therefore: reduce headcount, adopt AI, maintain output, improve margins.

The unit economics looked reasonable at the individual level. A senior engineer at a major tech company costs $150,000 to $300,000 all-in per year when you count salary, benefits, equipment, and real estate allocation. An AI coding tool subscription at $500 to $2,000 per month works out to $6,000 to $24,000 per year at current rates.

On paper, you could replace twelve engineers with one AI tool at the worst-case pricing and still come out ahead on cost. The spreadsheet looked great.

Then introduce the Uber complication.

When you deploy at real enterprise scale, with engineers who are genuinely skilled and find the tools genuinely useful, you don't get modest adoption. You get 84-95% adoption. At $500 to $2,000 per month per head, across 5,000 engineers, you're looking at $2.5 million to $10 million per month. Annualised: $30 million to $120 million per year in tooling costs alone. That's before compute, platforms, and the rest of the AI infrastructure that's accumulated alongside the developer tools.

The comparison stops looking clean at that scale.

And here's the part that should appear in every budget discussion but rarely does: the better your engineers are at using the tools, the more the tools cost. If you have a team of highly skilled developers who get genuinely productive with Claude Code, they generate significant value, and they generate the highest token bills. The productivity and the cost are coupled. This is the inverse of how software pricing was supposed to work. Historically, productivity improvements don't increase your cost per productive hour. With token-based AI tools, they can.

The METR study, which we covered in a previous article, found that AI coding tools made developers 19% slower on certain complex tasks. The productivity dividend has terms and conditions, and they're not printed on the front of the box.

Three Ways the Bill Gets Out of Control

Not all AI cost spirals look the same. In 2026, three distinct patterns have emerged.

Pattern 1: The usage spike

This is what happened to Uber. You deploy a tool. Engineers love it. Adoption accelerates. The budget assumption was built on modest, exploratory usage. Actual usage is near-universal and intensive. By the time the quarterly review arrives, the number on the invoice doesn't match the number in the model.

The CTO of Uber described it as having his budget "blown away." That's a measured description of what is, structurally, a governance failure. Not a malicious one. The tool worked. The adoption was exactly what you'd want. The budget model just didn't account for what "it works and engineers use it constantly" actually costs.

Pattern 2: The governance gap

The $500 million. One month. No spending cap.

I'll be honest: I don't know what governance process, or absence of one, produces that outcome. But the Axios report from 28 May 2026 confirmed it happened. An AI consultant revealed that one enterprise company spent half a billion dollars on Claude API calls in a single month with no limit in place. Nobody sets a spending cap until they've had a month that explains why spending caps exist.

For context: $500 million in one month is more than many mid-size listed companies earn in a year. The fact that token-based pricing can reach that number before anyone notices suggests some organisations are treating AI API access the way they treat internal email. Which is to say: as a cost that doesn't need watching. Email doesn't run at $6 per thousand tokens.

Pattern 3: The preference-to-mandate reversal

Microsoft brought Claude Code in. Engineers preferred it. They used it heavily. "Perhaps a little too popular" is the corporate communications way of saying the cost was untenable. Microsoft cancelled the licences for the Experiences and Devices division. Engineers are moving to Copilot CLI by 30 June.

This is, in its own way, the most instructive pattern of the three. Because the solution to "our engineers love this tool and it costs too much" wasn't finding a better tool. It was mandating a cheaper one.

The outcome is specific: you now have engineers using a tool they didn't choose, for reasons they understand completely (cost), with a productivity trade-off that's difficult to measure but real. If AI tools produce productivity gains proportional to how well engineers use them, and engineers use tools they prefer better than tools they don't, then the cost saving and the productivity saving are working against each other. You controlled the budget. You partially sacrificed the gain that justified the original headcount reduction.

One analyst on X tracked the three patterns (Uber, the $500M case, Microsoft) and noted that the end of AI inference subsidies was creating exactly this dynamic: costs that look sustainable at small scale become unsustainable as the pricing realities of agentic workflows hit enterprise procurement.

And a developer on X put the community's view on the whole thing more bluntly: "Remember when they competed on who could do the biggest layoffs? Now they're competing on who burns the most tokens." The developer's recommendation was to stop worrying about AI costs and just rehire the engineers. That's probably not the advice CFOs are looking for, but it captures the irony accurately.

One other voice put the trapped position plainly: AI spending has gotten so high that layoffs wouldn't make a meaningful dent in the bill. Can't cut AI. Can't cut people. Can't cut both. That's the honest version of the presentation nobody gives to the board.

What the ROI Numbers Actually Show

McKinsey's State of AI in 2025 found that for most organisations, AI had not yet significantly affected enterprise-wide margins. Our GenAI paradox research put the figure at around 80% of AI projects showing no significant bottom-line impact. The METR study found a 19% slowdown in certain complex coding tasks.

These findings don't mean AI tools don't work. They mean the productivity dividend isn't universal, isn't automatic, and isn't evenly distributed. Which is exactly what you'd expect from any tool category where outcomes depend heavily on the skill of the person using the tool.

Here's the structural problem that makes the cost story particularly awkward.

The engineers who get the most out of AI coding tools are the most skilled engineers. They prompt better. They structure tasks better. They catch the errors the AI makes faster. They generate the most value from the tools. They also generate the highest token bills.

Traditional productivity improvements scale with effort and are bounded by the number of engineers you have. If you work harder and smarter, you produce more. The cost per hour stays roughly the same. With token-based AI tools, the productivity and the cost curve together. Your best engineers get the most out of the tools and cost the most to enable.

You can't separate the capability from the consumption. They come as a package.

There's also the pricing paradox that @RaytoneHQ pointed out on X: per-token prices dropped by roughly 90% over the past two years. And yet the bills went up anyway. The reason is straightforward. Cheaper tokens plus higher usage plus agentic workflows, which consume vastly more tokens per task than simple completion requests, equals higher total cost. The vendor's price per unit went down. The number of units consumed went up faster. This is the part of the model that procurement teams who were used to per-seat SaaS pricing didn't anticipate.

What This Means for Australian Businesses

Australia is sitting in an uncomfortable position in this story. According to the ACS and Information Age, Australia ranks second globally for tech job losses attributed to AI in 2026. WiseTech cut 2,000 staff, which is close to 30% of its workforce. Atlassian cut 1,600. Telstra cut 650. "AI blamed for every Australian tech sacking this year," was how Information Age summarised it.

Most of those cuts aren't being made by organisations with the enterprise procurement teams, CTO-level governance structures, or CFO visibility that the Ubers and Microsofts of the world have. They're being made by mid-size technology companies and corporate teams who made a reasonable calculation about productivity tools replacing headcount, and are now discovering the governance requirements that make that calculation actually work.

For smaller Australian businesses that have reduced their developer teams and leaned harder on AI tools, the question isn't philosophical. It's operational. Do you know what your AI tooling actually costs per month? Do you have spending limits set at the API level? Are you measuring what productivity you're getting versus what you assumed? Most organisations aren't doing all three.

I should be transparent here: we use these tools at Webcoda daily. I've watched our usage grow. And I'll put my hand up and say I didn't model the usage curve properly when we expanded AI tool access to more of our team. I checked the cost per seat. I didn't build a model for what the bill looks like at 80% team adoption with high-engagement users. That conversation happened after the fact rather than before it, which is exactly the pattern I'm writing about. (It wasn't $500 million. But the principle was the same.)

The governance checklist is genuinely short. Know your current spend to the dollar. Set usage limits before they're needed, not after the invoice teaches you why they're needed. Measure productivity gains rather than assuming them. Most of the organisations now dealing with unexpected AI bills failed on at least one of these three.

For the deeper detail on the $500 million case:

Related Article9 min read

How do you accidentally spend $500M on AI?

An unnamed enterprise company spent $500 million on Claude AI in a single month because nobody configured a spending limit. Uber burned its 2026 AI...

Read full article

For the Microsoft story:

Related Article9 min read

Microsoft told its engineers Claude Code was too expensive. Then started selling Anthropic's AI at a 65% markup.

Microsoft cancelled Claude Code licences for thousands of its own engineers because the tool was too popular and cost too much. Meanwhile, the...

Read full article

The Bet Isn't In Yet

The companies that made the layoff-plus-AI trade-off in 2023-2025 weren't being cynical. Some of them were making a genuinely reasonable calculation about where productivity tools were heading. The engineers who were laid off weren't replaced by tools that didn't work. Some of the tools work extremely well.

The issue isn't whether the tools work. The issue is whether "the tools work" and "the unit economics work" are the same thing. In 2026, at production scale, with real adoption rates and token-based billing and agentic workflows running overnight, the honest answer is: sometimes yes, sometimes no, and almost nobody ran the numbers properly before they needed to.

@EvanKirstel put a harder version of this on X: the "AI replacing workers" layoffs were mostly cover for cutting payroll to fund the AI bill. That's probably too simple for most cases. But it's not entirely wrong for some of them.

The productivity gains are real for some engineers on some tasks. The costs are real for every engineer who uses the tools enthusiastically. The governance frameworks are, in most organisations, still catching up to both.

I'm still using these tools. I'm just logging the bill now.

Key Takeaways

• The tech layoff cycle of 2023-2025 was partly justified by AI productivity gains. In 2026, those gains are arriving at the same time as AI tool invoices that weren't in the original model.
• Usage-based token pricing at enterprise scale doesn't behave like per-seat SaaS. At 84-95% adoption, costs can reach $2.5M-$10M per month across a 5,000-engineer organisation.
• Three patterns are driving AI cost overruns: usage spikes from genuine adoption (Uber), governance gaps with no spending controls (the $500M case), and preference-to-mandate reversals where engineers are moved to cheaper but less-preferred tools (Microsoft).
• The productivity dividend is coupled to cost. Your best engineers get the most out of AI tools and generate the highest bills. You can't have one without the other under current pricing models.
• Australian businesses are second globally for AI-attributed tech job losses in 2026. Most lack the enterprise governance structures to catch cost overruns before they compound.
• Three basics that most organisations aren't doing: knowing their AI tooling spend to the dollar, setting usage limits proactively, and measuring actual productivity gains rather than assuming them.

---