In April 2026 Uber’s chief technology officer, Praveen Neppalli Naga, confirmed that the company had already spent its entire 2026 budget for AI coding tools. The year was four months old. Fortune, which reported the figure, noted that Uber had pushed the tools hard, with internal leaderboards, before the bill came due.

I build both kinds of software this argument is about. My team builds the systems companies run their operations on, and on my own time I build the throwaway kind, the sort nothing depends on. In a demo the two look identical. What Uber ran into is the reason they are not.

The prototype is real, and it is genuinely good. What it is not is a thing a business can safely lean on, and the entire distance between the two is the job. A system a company depends on has a bill it can forecast, an audit trail it can put in front of a regulator, an owner who can repair it, and a defined behavior for the day it fails. A vibe-coded prototype arrives with none of those, and no prompt supplies them. That distance is what the wow effect hides.

The bill changes shape first

Take the bill first, because it is the first thing to change shape. Andrej Karpathy coined “vibe coding” in February 2025 to mean accepting what the model writes without reading it, and scoped it to throwaway weekend projects. On a personal plan from Anthropic or Cursor, $20 to $200 a month, a good prototype takes an evening.

Move that reflex inside a company and the price stops being flat. Anthropic’s own guidance for businesses puts Claude Code at $150 to $250 per developer per month on average, and notes that a team of agents can burn about 7 times the tokens of a single session, because each agent runs its own context window. Multiply that by headcount, then by the agents each developer sets running, and the figure stops behaving like a subscription. That is what emptied Uber’s budget four months into the year.

Who answers once it touches real data

Then there is who answers for what the build does once it touches real data. The enthusiast answers to no one. The company answers to regulators. GDPR is already enforced against AI: in December 2024 Italy’s data-protection authority fined OpenAI €15 million over its handling of personal data, a penalty OpenAI is appealing.

The EU AI Act reaches past the labs to any company that deploys AI. Its ban on prohibited practices and its staff AI-literacy duty have applied since 2 February 2025. Its heavier duties for high-risk deployers, first due in August 2026, are set to move to December 2027 under the Digital Omnibus the Council adopted on 29 June 2026, which is still awaiting publication in the Official Journal.

A prototype records none of this. A system a company can defend carries proof of what it did, whom it touched, and on what legal basis. Assembling that proof is work the prompt never performs.

The tool that wrote it cannot repair it

The deepest gap opens when the build breaks, because the tool that wrote it does not understand it. In July 2025, during a public vibe-coding experiment by SaaStr founder Jason Lemkin, Replit’s agent deleted the live production database during an explicit code freeze, wiping records on roughly 1,206 executives and 1,196 companies. It had been told, in capitals and more than once, to leave it alone.

The records proved recoverable, but the agent insisted a rollback was impossible, which was false, and then explained itself: “I panicked instead of thinking.” Replit’s chief executive, Amjad Masad, called it a catastrophic failure and shipped the repair, which was to separate the development database from production.

Notice who supplied that repair. The model that built the app could neither prevent the damage nor describe it honestly afterward. A human who understood the system did. That human is the maintenance a prototype does not come with.

The objection: the problem is temporary

The strongest reply is that the problem is temporary. Prototypes are how a company learns, the argument goes, and the cost of running them falls every year, so the move is to let enthusiasts build and wait for the economics to catch up. The falling price is real. By a16z’s measure, the price of a fixed level of capability drops about tenfold a year.

But a falling unit price and a falling total bill are different claims. Enterprise spending on large-model APIs more than doubled in six months to $8.4 billion in the first half of 2025, by Menlo Ventures’ count, and Gartner expects enterprise AI costs to keep rising through 2030, because agentic work consumes far more tokens per task. Uber’s budget died in 2026, well into that run of falling prices.

The learning half of the objection is fairer, and its answer is uncomfortable. Enthusiasm on its own converts poorly. S&P Global, surveying more than a thousand companies in 2025, found 42% had abandoned most of their AI initiatives, up from 17% in 2024.

And you often cannot tell from the inside whether the thing works. In METR’s 2025 trial, experienced developers on code they knew predicted a 24% speedup and came out 19% slower, still certain afterward that they had gained time. METR’s own February 2026 follow-up could not settle the number. The wow effect, in my read, shows up exactly where nothing rides on the result, and a business is built mostly from things on which something rides.

Four questions before the company leans on it

So before the company leans on an enthusiast’s build, run one pass yourself. An owner runs it on their own firm, an advisory partner on a client’s build, a private-equity partner on a portfolio company that credits an internal AI tool with a result. It comes back one of two ways: a build someone can stand behind, or a liability waiting for the quarter it surfaces.

  1. The meter. Ask whether the tool is billed as a flat subscription or metered per token, and who has agreed to the ceiling. What runs cheaply on a personal plan moves to metered API rates the moment the company owns it, near Anthropic’s own $150 to $250 per-developer average. It passes if someone can name the monthly exposure at ten people running agents.
  2. The owner. Ask who maintains the build and who reads its bill. The person prompting rarely sees the token cost, and the model that wrote the code cannot repair it. It passes if a named person both owns the upkeep and watches the spend.
  3. The regulated data. Ask whether the build touches personal data, and who has signed that its use has a lawful basis under GDPR and meets the EU AI Act’s staff AI-literacy duty. Italy’s €15 million GDPR fine on OpenAI, in December 2024, shows the enforcement is real. It passes if a named person owns that sign-off.
  4. The irreversible step. Find the point where a confident wrong answer or a careless migration does damage that cannot be undone, and ask what the build may do there on its own. Replit’s agent reached a production database it was ordered to leave alone. It passes if something deterministic stands between the build and the irreversible action.

Four specific answers describe a build a company can depend on. Four vague ones describe the thing that ran down Uber’s budget before anyone read the invoice.

Uber’s engineers are among the best in the world, and the budget still went in a third of a year, because the reflex to build fast met an economy no one had priced and a set of duties no one had assigned. The prototype was never the problem. It is real, and it is impressive. What does not survive contact with a business is the assumption that the miracle arrives finished. Before the company depends on one, make the enthusiast’s build answer those four questions and prove it is a system.

Where vibe coding sits among the ways companies take up AI is a companion piece, How do companies actually adopt AI? Five paths. The order we build in, the simplest dependable tool first and the frontier model last, sits in our method. If you want the speed of the prototype inside a company without the metered bill and the audit gap behind it, that is the work we do: custom AI for your business.

The short version

  • In April 2026 Uber’s chief technology officer confirmed the company had spent its entire 2026 AI coding-tools budget in four months (Fortune, citing The Information).
  • A vibe-coded prototype runs on a flat $20-to-$200 consumer plan. Inside a company the same work is metered, at $150 to $250 per developer per month on Anthropic’s own guidance, with a team of agents able to use about 7 times the tokens of one session.
  • Falling unit prices have not lowered total AI bills: enterprise large-model API spending more than doubled to $8.4 billion in the first half of 2025 (Menlo Ventures), and Gartner expects costs to keep rising through 2030.
  • A vibe-coded build has no owner who understands it. In July 2025 Replit’s agent deleted a live production database during a code freeze, wrongly reported the loss unrecoverable, and a human had to add the ordinary engineering that should have prevented it (The Register).
  • Enthusiasm alone converts poorly: S&P Global found 42% of companies abandoned most AI initiatives in 2025, up from 17% in 2024, and a 2025 METR trial measured experienced developers 19% slower even as they felt faster, a result its February 2026 follow-up could not settle.

Sources

  1. Fortune, “Microsoft’s AI cost problem”, May 22, 2026. https://fortune.com/2026/05/22/microsoft-ai-cost-problem-tokens-agents/
  2. Andrej Karpathy on X, February 2, 2025. https://x.com/karpathy/status/1886192184808149383
  3. Anthropic (Claude) pricing, accessed 2026-07-25. https://claude.com/pricing
  4. Cursor pricing, accessed 2026-07-25. https://cursor.com/pricing
  5. Anthropic, “Manage costs effectively”, Claude Code documentation, accessed 2026-07-25. https://code.claude.com/docs/en/costs
  6. The Register, “Replit AI agent deletes production database during code freeze”, July 21, 2025. https://www.theregister.com/2025/07/21/replit_saastr_vibe_coding_incident/
  7. Fortune, “AI coding tool Replit wiped a database and its CEO called it a catastrophic failure”, July 23, 2025. https://fortune.com/2025/07/23/ai-coding-tool-replit-wiped-database-called-it-a-catastrophic-failure
  8. The Hacker News, “Italy fines OpenAI 15 million euros”, December 2024. https://thehackernews.com/2024/12/italy-fines-openai-15-million-for.html
  9. European Commission, “Regulatory framework on AI” (AI Act application timeline), accessed 2026-07-25. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
  10. Shumaker, “EU AI Act: Council Gives Final Green Light to the Digital Omnibus on AI”, July 2026. https://www.shumaker.com/insight/eu-ai-act-council-gives-final-green-light-to-the-digital-omnibus-on-ai/
  11. a16z, “Welcome to LLMflation”, November 2024. https://a16z.com/llmflation-llm-inference-cost/
  12. Menlo Ventures, “2025 Mid-Year LLM Market Update”, July 31, 2025. https://menlovc.com/perspective/2025-mid-year-llm-market-update/
  13. S&P Global Market Intelligence, “Generative AI shows rapid growth but yields mixed results”, October 2025. https://www.spglobal.com/market-intelligence/en/news-insights/research/2025/10/generative-ai-shows-rapid-growth-but-yields-mixed-results
  14. METR, “Measuring the impact of early-2025 AI on experienced open-source developer productivity”, July 10, 2025. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
  15. METR, “AI uplift study: February 2026 update”, February 24, 2026. https://metr.org/blog/2026-02-24-uplift-update/