This one is a little different from the usual headstones here. Nobody's chatbot invented a bereavement policy. No AI agent panic-deleted a production database. No lawyer filed a motion with citations from the jurisprudential fantasy realm. The failure is quieter and, for a certain class of executive, probably more frightening: the spreadsheet stopped cooperating.

On April 26, 2026, Axios published a short but pointed report about enterprise AI costs. The line that ricocheted around the tech internet came from Bryan Catanzaro, Nvidia's vice president of applied deep learning: for his team, "the cost of compute is far beyond the costs of the employees." Fortune picked up the same thread two days later, and Tom's Hardware followed with the blunt version: AI can be more expensive than the actual workers it was supposed to augment, replace, or magically alchemize into infinite productivity.

There is an obvious caveat, and it matters. Catanzaro does not run a normal office team using ChatGPT to summarize meeting notes. He runs applied deep learning at Nvidia, a company that sells the hardware everyone else is trying to cram into data centers. His team's compute appetite is not a proxy for the average HR department, insurance back office, or support desk. If your job is building and testing deep learning systems at Nvidia, of course the GPU bill looks less like a SaaS subscription and more like a weather event.

But that caveat does not make the quote irrelevant. It makes it useful. If the company selling the picks, shovels, and ceremonial AI gold-rush hats is openly saying its own compute bill beats payroll in one of its core AI teams, executives should at least stop treating "replace workers with AI" as a self-evident cost reduction. The units matter. Tokens are not vibes. GPUs are not motivational posters. Every model call lands somewhere on an invoice.

Budget Math Meets Token Math

The Fortune article connected Catanzaro's comment to another very concrete example: Uber's CTO Praveen Neppalli Naga reportedly told The Information that the company's AI coding-tool budget for 2026 had already been blown only months into the year. Axios summarized the same point: Uber's AI tooling costs had outrun the budget because of token costs.

This is the part many corporate AI rollouts keep learning the expensive way. Traditional enterprise software is usually sold as a seat, a license, or a predictable contract. You budget for 5,000 employees, buy 5,000 licenses, negotiate your discount, and move on. Usage can create support or storage costs, but the CFO mostly knows where the ceiling is.

Agentic AI does not behave like that. A coding assistant running a short autocomplete is one thing. An agent chewing through a repository, reading files, writing patches, running tests, fixing its own mistakes, re-reading context, retrying failed commands, and asking another model to review the result is a metered workload. The more useful the tool becomes, the more people use it. The more people use it, the more workflows become agentic. The more workflows become agentic, the more the invoice starts looking like cloud infrastructure rather than software.

That is not automatically a bad trade. A company might rationally spend more on AI if it gets more output, better quality, faster shipping, or stronger reliability. The problem is that the public pitch has often been much lazier: AI reduces labor costs. Full stop. Fewer humans, lower expenses, higher margins, everyone applauds except the humans. The cost reports now surfacing complicate that story. If the replacement machine costs more than the people, and still requires people to steer it, review it, debug it, and eat the blame when it hallucinates a solution, the math has acquired a few extra columns.

MIT Already Warned About This

The annoying thing about this revelation is that it was not especially hidden. MIT CSAIL researchers published a 2024 study on the economics of automating computer-vision tasks and found that only about 23 percent of wages paid for vision-related tasks were economically attractive to automate at then-current costs. In plain English: a system can be technically capable of doing a job and still be too expensive to justify replacing the human who already does it.

That distinction is exactly where a lot of AI discourse goes to die. Capability is not deployment. Deployment is not return on investment. Return on investment is not a demo. A demo does not include procurement, integration, security review, prompt drift, error handling, human QA, power, latency, failed runs, vendor price changes, or the bad month when everyone discovers the agent got stuck in a loop and spent real money proving it could be confidently repetitive.

The MIT paper focused on computer vision, not language models or coding agents, so it should not be stretched into a universal law. Still, the economic principle carries over cleanly: the relevant question is not "Can AI do this task?" The relevant question is "Can AI do this task at the required quality, reliability, latency, compliance posture, and total cost, after including the humans and infrastructure needed to keep it from embarrassing us?"

That is a much less exciting question to put on a keynote slide. It is also the one that determines whether the project survives contact with finance.

Gartner's Spending Forecast Is the Backdrop

Gartner's April 2026 IT spending forecast gives the wider context. The firm projected worldwide IT spending would reach $6.31 trillion in 2026, up 13.5 percent from 2025, with the strongest growth driven by AI infrastructure and software. Data center systems were projected to grow 55.8 percent. Gartner explicitly tied the acceleration to AI workloads, high-performance compute, advanced memory, hyperscale cloud demand, and GenAI software.

That does not mean every company is wasting money on AI. It does mean the AI bill is not just a $20-per-seat subscription hiding under "software." It is data centers, accelerators, memory, cloud capacity, model subscriptions, usage-based APIs, orchestration platforms, observability, governance tools, and all the professional services required to bolt the new machine onto the old machine without setting either one on fire.

The enterprise AI sales story often talks about digital labor as if companies are swapping one cost center for a cheaper one. In practice, many are adding a new variable-cost infrastructure layer on top of the people they still need. That can be smart if the layer produces measurable value. It can also be a very expensive way to discover that your internal processes were not ready for automation, your data was messy, your approvals were undocumented, and your employees were already doing twelve invisible judgment calls per hour that the model cannot see.

Why This Belongs Here Anyway

Strictly speaking, this is not the cleanest Vibe Graveyard incident. It is an economics story, not a single product meltdown. But it belongs in the cemetery for one reason: it punctures the exact assumption behind many of the other failures here.

Commonwealth Bank cut customer service roles because an AI voice bot was supposed to absorb the work, then reversed course when service levels suffered. Klarna boasted that its assistant did the work of 700 agents, then started bringing humans back after admitting quality dropped. McDonald's pulled its IBM drive-thru AI after enough ordering chaos made the robots look less like operational efficiency and more like a TikTok content strategy. Those stories were operational failures. This one is the budget version of the same disease.

The disease is premature substitution. Leadership sees an AI demo, converts the demo into a labor-savings projection, and then makes staffing or budget decisions before the system has proven it can do the job at the required cost. Sometimes the failure shows up as angry customers. Sometimes it shows up as fake citations. Sometimes it shows up as an agent deleting infrastructure. Sometimes it shows up as a token bill with a number large enough to make the "AI is cheaper than employees" slide look like it was prepared by someone who has never met an invoice.

No Free Robot Labor

The practical lesson is boring, which is how you know it is probably important. Treat AI like an operating workload, not a magical headcount eraser. Measure total cost per useful outcome. Include the cost of human review. Include failure handling. Include security and compliance. Include the cost of reruns. Include the vendor's ability to change pricing once your workflows depend on them. Include the cost of not knowing whether the output is correct.

If the result still beats human labor, great. Use the tool. If it does not, maybe the humans were not the expensive part after all.

That is the awkward punchline of Catanzaro's quote. The AI industry has spent years telling companies that intelligence can be summoned on demand, metered by the token, and scaled like cloud storage. Now the bills are arriving, and some of them are bigger than payroll. Congratulations: the robots have entered the workforce and immediately discovered cost overruns.