The case for replacing humans with AI has, until this week, mostly lived on earnings calls and Twitter threads. The case against it is finally landing on the page. In the seven days before July 6, 2026, a long Forbes analysis laid out the P&L line items an enterprise forgets when it swaps a salary for an API key, Amazon confirmed it will stop signing new customers for Mechanical Turk on July 30, and the cost arithmetic behind Jack Dorsey’s February cuts at Block caught up with the press cycle. None of the three settles the question alone. Together, they produce the first clean ledger of what AI-driven headcount reduction actually does to a balance sheet.
What the Forbes piece puts on the table
Jemma Green’s July 2 piece is the most concrete accounting to date, assembling line items rather than slogans: inference, integration, supervision, and error correction. The headline inference figure is that OpenAI “spends ~$2 for every $1 earned” on running its frontier models, with ~$14B in projected losses for 2026 and ~$44B cumulative losses before any profit appears in 2029. The same piece reports one unnamed company ran a $500M Claude bill in a single month after forgetting to set a usage cap, and that the OpenAI-led chat layer lost money on its $200-per-month subscription tier in a statement Sam Altman made on the record. Andrew Macdonald of Uber gave the operational version of the same complaint, quoted in the Forbes piece: “token usage didn’t seem to correlate directly with useful features shipped to users.”
Integration is where most companies underestimate the bill. Green reports that enterprise AI bills are projected to rise 30 to 50 percent above current levels once subsidized model pricing normalizes. Bryan Catanzaro, Nvidia’s VP of applied deep learning, told the piece that “the cost of compute for his team now far exceeds what the company spends on the employees using it,” and Forbes’s framing - “If the person selling the compute calls the spending ‘the most fair criticism right now, there is a ton of waste,’ what does the person paying the bill call it?” - implies the same concession coming from the compute-supply side.
Supervision and error correction are the easiest lines to underestimate and the hardest to walk back. Green cites Faros AI’s finding that code churn - lines deleted versus lines added - increased more than 800 percent under high AI adoption, and that scaling doesn’t reduce the defect rate, only the volume a human reviewer must vet. She cites a Gartner forecast of $207B in AI agent software spending in 2026, up 139 percent year over year, and a Sequoia estimate from partner David Cahn that AI companies need roughly $600B in annual revenue to justify the infrastructure spending being committed. Green quotes Jensen Huang’s standard line that “a $500,000 engineer should be consuming at least $250,000 worth of AI tokens annually” - per Forbes’s “Jensen Huang tells the industry that” framing - with Nvidia itself working toward a $2B annual token budget for its engineering force. Green cites “tokenmaxxing” as the industry term for the management incentive now spreading through corporate KPIs that rewards raw AI usage over measured productivity - Amazon’s internal “KiroRank” engineering leaderboard is the named example; an MIT study cited in the piece puts AI as economically viable in only ~23 percent of roles.
Why the human-in-the-loop pipeline is winding down
The Mechanical Turk closure is the matching signal from the other end of the training pipeline. Amazon told TechCrunch on July 5 that, effective July 30, 2026, Mechanical Turk “will close to new customers,” and the AWS statement, quoted verbatim in TechCrunch, drew the line at effort, not at the service: “Existing customers can continue to use the service as normal. AWS continues to invest in security and availability improvements for Mechanical Turk, but we do not plan to introduce new features.”
Mechanical Turk launched in 2005 as the human-intelligence-task crowdsourcing primitive and was rebilled under SageMaker beginning in 2018 as a clean data-annotation interface. It is the originating platform of the modern human-in-the-loop annotation industry - the same industry that trains and evaluates the models executives now claim will replace humans. A 2023 analysis cited by TechCrunch estimated that between 33 and 46 percent of Mechanical Turk workers were already using large language models to complete tasks. The read is that the paid human-in-the-loop pipeline is being wound down precisely as the corporate message is that AI removes the need for that pipeline: either AI no longer needs paid human labels at scale, or the major US labs have moved annotation spend to synthetic data and internal contracting. Either way, the requester pipeline is shrinking while the deployment narrative expands.
The Block cuts in context
The third line item is the corporate case study. On February 26, 2026, Jack Dorsey’s Block cut roughly 4,000 employees, nearly half of its workforce - our coverage at the time framed it as the largest single AI-attributed layoff on record - and CFO Amrita Ahuja framed the cut on the record as positioning Block to “move faster with smaller, highly talented teams using AI to automate more work.” Dorsey’s framing, also quoted in TechCrunch, was that within a year “most companies will arrive at the same place,” and that he’d rather “get there honestly and on our own terms than be forced into it reactively.” Block’s stock rose more than 24 percent in after-hours trading on the announcement.
That press release is what the Forbes analysis now puts under a cost microscope. Block is a financial product company with regulatory and trust obligations that scale non-linearly with errors. The same four lines - inference, integration, supervision, and error correction - apply point-for-point: any reduction in human review raises supervision and error-correction spend, and any reduction in deploy-and-test engineers raises integration spend. The math Block is betting on - smaller team, higher AI usage per person - is the math Forbes says is producing “a ton of waste” inside Nvidia’s own engineering organization per Catanzaro’s own concession that compute cost “now far exceeds what the company spends on the employees using it.” The Block press release does not have a number that responds to Catanzaro’s complaint; the Forbes piece does.
What This Means
For executives, the practical read is that AI-replaces-a-salary is a slogan and AI-augments-a-smaller-team is a more expensive slogan, because the four cost lines - inference above subscription revenue, integration, supervision, and error correction - do not, in aggregate, fall below a fully loaded salary at the same task. For workers whose companies are using the Block template, the framing is most credible when announcing cuts and least credible when budgeting for what comes next. The Anthropic, OpenAI, and Nvidia price list is itself loss-making at the model layer, and the supervision and error-correction costs downstream are absorbed by a smaller, more stressed team rather than by a different category of spend. The cleaner frame for the next quarter is that AI capex (Sequoia’s $600B revenue figure, Gartner’s $207B in agent software, the ~$740B in 2026 Big Tech AI capex cited in the Forbes piece) and AI opex are running in opposite directions, and the recurring cost base is what this week’s reporting finally enumerated line by line.
The Bottom Line
Two analyses this week - the Forbes “AI costs more than the people it replaced” piece and Amazon’s quiet closure of Mechanical Turk to new customers - say the same thing in different dialects: the AI cost curve does not run below the human salary it was supposed to replace, it runs above it. The receipts are on the page, and they describe a P&L the AI-replaces-humans narrative does not yet have a response to.