On February 26, 2026, Jack Dorsey told Block shareholders he was reducing the company by nearly half, from more than 10,000 people to just under 6,000. More than 4,000 people were being asked to leave or enter consultation.
The stock jumped as much as 24%.
That is the part worth sitting with for a minute. A company announced one of the largest AI-framed workforce reductions we have seen, and the market applauded before the operating model had a chance to prove itself.
Dorsey did not blame a bad quarter. In Block's own shareholder letter, he said the business was strong, gross profit was still growing, and profitability was improving. The real change, he argued, was that intelligence tools paired with smaller, flatter teams had fundamentally changed what it means to build and run a company. He would rather make one decisive cut than bleed the organization through years of quiet rounds. Then he went further than most CEOs do. Within a year, he predicted, the majority of companies would reach the same conclusion and restructure the same way.
I have spent three decades building software and the last several years leading the platform teams that keep it running. I want Dorsey to be wrong about the timeline and right about the tools. What makes this moment worth writing about is not whether AI matters. It does. The real question is whether the economics work once the subsidy disappears and the operating costs become visible.
For roughly eighteen months, the industry ran on two comfortable beliefs. The first was that AI is cheap. The second was that AI can replace people. Usage-based billing is now exposing the true cost of the AI. The layoff boomerang is exposing the true cost of the people.
Put real numbers on both sides of that ledger and the trade looks very different than it did in the press release.
This is not an argument that AI is too expensive to matter. It is not an argument that headcount will never shrink. Some work will absolutely change. Some roles will disappear. Some teams will get smaller. The argument is narrower and more important: the simple salary-for-subscription swap is usually fake math. The full cost includes inference, reliability, supervision, rework, customer experience, and institutional knowledge. Once those costs move from hidden subsidy to visible operating expense, sloppy AI adoption and sloppy workforce cuts both get harder to hide.
The flat fee was a promise the physics could not keep
A month ago, the cool thing to do was token maxxing. Engineers compared usage dashboards like high scores. Teams joked about who could burn the most tokens in a week. It was fun, and it was a tell. We were treating inference like an all-you-can-eat buffet because someone else was paying for the kitchen.
On June 1, 2026, GitHub moved Copilot to usage-based billing. Every Copilot plan now consumes GitHub AI Credits based on token usage, including input, output, and cached tokens. GitHub described the move as necessary to align pricing with actual usage and make Copilot a sustainable business.
That last word is the one that matters: sustainable.
GitHub did not make this change because developers suddenly became less enthusiastic. It made the change because agentic workflows changed the cost curve. A chat request and a long-running coding agent are not the same economic object. One is a question. The other may hold a large context window open, search a repo, edit files, run tests, respond to failures, and try again. That is the whole point of agentic software development. It is also the bill.
The legacy multiplier table tells the same story in a different language. For Copilot users who remain on annual plans under the old request-based model, GitHub now lists several frontier coding models with double-digit multipliers, including multiple Opus models at 27. The exact table will keep changing, which is the point. Model cost is no longer an invisible platform concern. It is becoming part of the engineering operating model.
This was never only a GitHub problem. OpenAI, Anthropic, Google, and the rest of the frontier labs have been wrestling with the same basic constraint. Heavy users were consuming more compute value than their subscriptions covered, while the vendors absorbed the difference in the name of growth. That was a reasonable customer acquisition strategy. It was not a law of physics.
There is an old idea hiding underneath all of this. In 1865, the economist William Jevons noticed that making coal use more efficient did not reduce coal consumption. It increased it, because efficiency made coal worth using for more things. Cheap, subsidized inference did the same thing to us. It did not make us frugal. It made us build agents that call other agents, summarize the output, generate tests, rewrite the tests, open the pull request, and then explain the pull request to another agent.
For a while, someone else paid the metered rate. The bill has now been forwarded to the customer.
Cheap for the buyer, profitable for the seller, pick one
Why is the subsidy ending now, after the labs tolerated these losses for years? The answer is on the calendar.
The IPO window is opening. Anthropic has reportedly filed confidentially to go public, and OpenAI is widely reported to be preparing its own path to the public markets. A private company can run for a long time on growth, scarcity, and belief. A public company has to show a path to durable margins.
That is where the AI story gets interesting. The market wants two things at once. It wants AI cheap enough that enterprises can use it to justify massive productivity gains. It also wants the companies selling AI to prove they can become profitable public businesses.
Those two demands are in tension.
If inference stays heavily subsidized, customers get cheap automation but labs struggle to show the unit economics public investors expect. If inference gets priced closer to cost, labs can tell a better profitability story, but customers have to revisit the spreadsheet they used to justify replacing people, processes, or entire layers of work.
That does not kill AI. It matures it.
It separates use cases that create real value from use cases that only looked good while the meter was hidden. It separates teams that know how to specify work from teams that send agents wandering through ambiguity with a corporate card. It separates leaders using AI to redesign work from leaders using AI to decorate a cost-cutting memo.
Now hold that next to Block. Dorsey's efficiency bet only works if the productivity gain survives the full cost of the toolchain, the new operating model, the retained experts, the escaped defects, the rework, and the customer impact. The same investors who cheered the cut are also asking AI labs to stop subsidizing the compute that makes the cut plausible.
Someone pays full price. Increasingly, that someone is the buyer.
The boomerang
Here is the part Dorsey's prediction has to survive. Some companies that already ran versions of this experiment are quietly reversing pieces of it.
Klarna is the cautionary tale, though the lesson is broader than "AI failed." Business Insider reported that Klarna reassigned employees, including engineers and marketers, into customer-support roles after the company concluded that earlier cost-cutting had gone too far. The company had previously promoted aggressive AI-driven efficiency and said its AI assistant was doing the work of hundreds of customer support agents. Later, its CEO acknowledged that cost had become too dominant a factor and that lower quality was the result.
That is the real lesson. Cost-led automation can degrade service quality when leaders remove human judgment before the operating model is ready.
Klarna is not alone. A February 2026 Careerminds survey of 600 HR professionals who had made layoffs in the prior twelve months found that companies were already rebuilding workforces they had cut. Careerminds reported that 35.6% of surveyed companies rehired for more than half of the roles they eliminated, while 32.7% rehired between 25% and 50% of the roles. Most did it within six months.
That is the boomerang.
A role looks automatable on a spreadsheet because the visible work is routine. Then the role disappears, and the invisible work shows up. It is the customer exceptions, the edge cases, the relationship context, the regulatory nuance, the institutional memory, and the judgment under pressure, all the things no one wrote down because the person doing the work simply knew them.
None of this means AI changes nothing. It is reshaping real work, and the net job losses attributed to it are real. Challenger, Gray & Christmas reported that AI led all stated reasons for job cuts in April 2026, with employers attributing 21,490 cuts to it that month and 49,135 so far this year.
But "AI-cited" and "AI-caused" are not the same thing.
That distinction matters because AI is becoming the most convenient restructuring story in the market. It sounds strategic. It sounds inevitable. It sounds better than saying the company overhired, duplicated work, tolerated unclear ownership, or let complexity compound until the org chart became the product.
Which brings us back to the awkward question about Block. Dorsey framed the cut as an AI story. He may be right. But the same story also includes a more familiar operating problem. During the pandemic-era growth cycle, Block grew dramatically and built Square and Cash App with more separate organizational structures. Reuters noted that analysts also viewed the reduction as a correction from pandemic overhiring, with Block's workforce growing from roughly 3,800 in 2019 to more than 10,000 by 2025.
Strip away the framing and a large share of this looks like a long-overdue simplification move wearing AI as a more flattering jacket.
That does not make the reduction meaningless. It makes the attribution important. AI may be the accelerator. It may be the excuse. It may be both. The gap between those explanations is where most of the next year is going to be argued.
Reliability just became a line item
The repricing changes more than budgets. It changes what good engineering is worth.
Under a flat fee, a mediocre answer from a model was an annoyance. You shrugged, edited it, and moved on because the next attempt felt free. Under metered billing, a hallucination is no longer only a quality problem. It triggers a correction loop, and every loop costs money. Unreliable output now compounds directly onto your invoice.
Consider the common agentic development failure mode. A team gives an agent a vague ticket: "clean up the auth flow." The agent reads half the repo, edits the wrong abstraction, generates tests around its own misunderstanding, fails CI, retries, pulls in more context, rewrites the patch, and opens a credible-looking pull request that a senior engineer has to unwind. Nothing in that sequence is science fiction. It is Tuesday.
The old cost was frustration. The new cost is frustration plus tokens plus review time plus opportunity cost plus whatever defect sneaks through because everyone wanted the automation story to be true.
This is where I will repeat the thing I have been saying for years, because the economics have finally caught up to it. The bottleneck is never the stack. It was not the language, the framework, or the cloud, and it is not the token price either. The real constraint is whether the work is specified well enough for a machine to do it without spinning in place.
A vague request that sends an agent in circles used to cost a little patience. Now it shows up as a number on a dashboard. We made implementation cheaper, so the value moved upstream to specification. Usage-based pricing just attached a price tag to every fuzzy spec in your backlog.
So the leaderboard era ends, and good riddance. Tokens consumed was always a vanity metric, the engineering equivalent of measuring a book by its page count. The number that replaces it is value per token, and that is a far better question to organize a team around.
Value per token rewards clarity. It rewards reliability. It rewards teams that know how to decompose work. It rewards engineers who can define acceptance criteria, constrain the search space, and recognize when the machine is confidently wrong. Conveniently, those are the same skills that made the layoff boomerang so predictable.
One caution before anyone declares the party over: Jevons still cuts both ways. As models get more efficient and inference gets cheaper per unit, total usage will keep climbing. The subsidy ending does not end AI adoption. It ends unmeasured AI adoption.
That is not a contraction. It is the moment a fast-growing technology starts behaving like a real business, with a meter, a profit and loss statement, and consequences.
Where this leaves the rest of us
If you lead an engineering organization, the worst response is to pick a tribe.
The "AI is fake" camp will miss a real shift in how software gets built. The "fire everyone" camp is about to learn what Klarna learned, on a delay, with a rehiring bill attached.
The honest path is less dramatic and more durable.
Track AI spend by workflow, not just by user. A power user burning tokens on high-value migrations may be a bargain. A team burning the same tokens because its backlog is unclear may be lighting money on fire with a nicer dashboard.
Measure accepted output, not generated output. Code written is not value. Pull requests merged are not even value by themselves. Look at cycle time, escaped defects, rework, customer outcomes, and operational load.
Require better specs before unleashing long-running agents. Agentic work does not remove the need for product thinking, architecture, or engineering judgment. It punishes their absence faster.
Treat AI displacement like a production change. Define the expected outcome, monitor the impact, preserve a rollback path, and know which humans still own the failure modes. If you would not remove a database without a recovery plan, do not remove institutional knowledge without one either.
And keep your best operators close. The people who understand the messy middle of the business are the ones who will make AI useful. They know where the process bends, where the system lies, where the customer gets stuck, and where the documentation stopped being true three reorganizations ago. That knowledge is expensive to rebuild because it was expensive to earn.
As for Dorsey's prediction, we will know soon enough. Maybe he read the future early and the rest of us are simply slow. Or maybe Block becomes the case study in someone's paper about boomerangs.
Either way, the free lunch is over. The bill is itemized. The companies that read it carefully are going to do very well.
I intend to be one of them, and I suspect you do too.
Sources and further reading
- Block, Q4 2025 shareholder letter: https://s29.q4cdn.com/628966176/files/doc_financials/2025/q4/Q4-2025-Shareholder-Letter_Block.pdf
- Reuters, "Block shares soar as Dorsey leans on AI to trim workforce": https://www.reuters.com/sustainability/sustainable-finance-reporting/block-shares-soar-dorsey-leans-ai-trim-workforce-2026-02-27/
- GitHub, "GitHub Copilot is moving to usage-based billing": https://github.blog/news-insights/company-news/github-copilot-is-moving-to-usage-based-billing/
- GitHub Docs, "Model multipliers for annual plans on request-based billing": https://docs.github.com/en/copilot/reference/copilot-billing/request-based-billing-legacy/model-multipliers-for-annual-plans
- Reuters Breakingviews, "Anthropic IPO could train a large M&A model": https://www.reuters.com/commentary/breakingviews/anthropic-ipo-could-train-large-ma-model-2026-06-02/
- Axios, "Anthropic faces AI spending backlash before IPO": https://www.axios.com/2026/06/02/anthropic-ipo-ai-sticker-shock-spending-usage
- Business Insider, "Klarna is reassigning engineers and marketers to customer support after its AI bet went too far": https://www.businessinsider.com/klarna-reassigns-workers-to-customer-support-after-ai-quality-concerns-2025-9
- Careerminds, "AI-led layoffs: What HR leaders wish they knew before making job cuts": https://careerminds.com/blog/cost-of-ai-layoffs
- Challenger, Gray & Christmas, "April Job Cuts Rise 38% from March; YTD Cuts Down 50%": https://www.challengergray.com/blog/challenger-report-april-job-cuts-rise-38-from-march-ytd-cuts-down-50/
