# The Math Under the AI Bet Author: Pramod Prasanth Published: 2026-04-23 Canonical: https://pramod.ch/essays/the-math-under-the-ai-bet Description: Uber's CTO told The Information that Claude Code blew through Uber's 2026 AI budget by April. The personal version of that math is already true for me. The macro version is the problem the industry has not priced in. --- I pay $200 a month for Claude Max. My usage on it is moderate: coding, research, writing, the things one practitioner does in a working week. If I moved that same workload to the API, my monthly bill would be in the low thousands. For one person. Scale that to a company of five thousand engineers and the arithmetic turns vicious. *Figure: Monthly cost for comparable AI usage at three levels. The jumps are order-of-magnitude, not linear. The hero bar on the right is what "blown budget by April" actually looks like arithmetically.* In April 2026, Uber's CTO Neppalli Naga told The Information[^1] that the company had already maxed out its 2026 AI budget, principally on Claude Code. "I'm back to the drawing board, because the budget I thought I would need is blown away already," Naga said. Four months into the fiscal year. The single largest engineering line item, gone. Uber is not a small company. Its engineering organisation is one of the largest in the world. If Uber's AI budget is blown by April, the question is not whether Uber budgeted badly. The question is what every CFO is about to discover, and what nobody has priced into the dominant AI investment narrative. That is not a budgeting error. It is the math, arriving. *Here is the argument. Frontier AI requires roughly two trillion dollars a year in revenue to justify its capex. That money can only come from replacing human labour, not from selling software. But regulated work requires a human signature the model cannot give, which is why the architecture that survives the math has a human at the top who signs.* ## The arithmetic only works if AI replaces labour Frontier-AI capital expenditure for 2025 alone runs to roughly four hundred billion US dollars across the major hyperscalers (Amazon, Microsoft, Google, Meta), with cumulative spend through 2028 projected in the low trillions. Assume a generous margin of sixty to eighty percent over a three-to-five year recovery window and the implied revenue base needed to justify the capex sits north of two trillion dollars a year. Two trillion dollars a year, in revenue, from selling AI services. Take that number and look around for where it can come from. Global software licence spend is roughly seven hundred to eight hundred billion dollars a year, per Gartner. Even if AI vendors captured every dollar of that market, they would not cover the bill. Software is too small a TAM (total addressable market). The capex story does not work as a software story. The only revenue base big enough to justify the capex is labour. Global wages run somewhere north of sixty trillion dollars a year. If AI replaces three to five percent of that labour spend, two trillion in vendor revenue is plausible. Without labour replacement at that scale, it is not. *Figure: The three pools against each other, at the same scale. Software is too small. The revenue required to justify the capex is only reachable if the AI industry takes a substantial bite out of global labour spend, the only pool large enough.* This is the unstated assumption beneath every bullish AI investment thesis. The macro bet is not that AI sells well as software. It is that AI replaces enough labour to bill for, at the price of labour. > The capex bet only works if AI replaces labour at the price of labour. That is a much taller order than the software-first framing admits. ## Accountability keeps most labour non-replaceable The first wall the labour-replacement thesis hits is regulatory. Here is what I mean concretely. In every regulated enterprise context I have worked with, pharmaceutical batch release, banking risk, payments, insurance underwriting, supply-chain compliance, the law requires a named human decision-maker for material actions. EU AI Act Article 14 mandates effective human oversight for high-risk systems. Sarbanes-Oxley Section 404 puts internal controls on a CFO's signature. MiFID II Article 17 requires algorithmic trading systems to have human supervision and kill switches. GxP requires a Qualified Person to sign off on every batch release. Basel III puts model validation on a named risk officer. The pattern is identical across regimes. You can automate the inputs. You cannot automate the signature. That ceiling has consequences for the labour-replacement TAM. The categories of work that require a regulated human signatory are not replaceable by a model output. They are augmentable. A model can draft, summarise, triage, score, prepare. A human still has to read it, accept it, and put their name on it. The productivity gain is real, but it does not free up the human's salary line. It frees up a fraction of their time. The labour spend that is genuinely replaceable by an unaccountable model output is a much smaller subset than the sixty-trillion-dollar headline. It is the slice of knowledge work that is low-stakes, low-accountability, and low-regulatory-exposure: research synthesis, code scaffolding, drafting, summarisation. Real, valuable, but small. Maybe five to ten trillion dollars if every replaceable role were fully replaced. Closer to one or two trillion if you adjust for capability, deployment friction, and the trust gap. That is enough to justify a healthy AI industry. It is not enough to justify two trillion dollars a year in revenue. ## The concentration problem The second wall is structural. Suppose, against the accountability ceiling, that AI vendors do find their two trillion. The mechanism by which they find it is concentration: revenue accrues to a small handful of frontier labs and the hyperscalers that host them. Productivity gains in the customer base accrue to shareholders, not workers, because the work that got cheaper had labour costs attached and now does not. The problem with this end-state is that it eats its own demand. Concentrated wealth is well-documented to circulate less than distributed wealth. Capital piles up in asset prices instead of being spent on goods and services. The enterprise customer of the AI vendor is itself an enterprise that sells things to people. If those people have lower aggregate purchasing power because their labour income has been compressed, the enterprise's revenue falls. Which means the enterprise's AI budget falls. Which means the AI vendor's revenue falls. The post-war European trajectory illustrates the inverse: distributed wealth, sustained broad demand, four decades of growth across a wide middle class. The 1990s-onward US trajectory is the cautionary side: financial growth strong, broad-based prosperity weak, demand increasingly concentrated. Apply the AI thesis at scale and the trajectory bends toward the second outcome at speed. A March 2026 working paper by Brett Falk and Gerry Tsoukalas formalises the dynamic[^2]. They model the labour-replacement decision as a strategic game between firms competing for the same consumer base. Their result: "individually profitable automation is collectively destructive because each firm's cost saving erodes the revenue base all firms share." Rational firms, aware of the destination, still race toward it. Each firm's incentive is to automate before its rivals do; the equilibrium is one in which every firm automates, collective demand contracts, and the aggregate loss is borne by both workers and firm owners. Their phrase for it: "firms automate their way to boundless productivity and zero demand." The real-world signal is already on the wire. In February 2026, Block cut nearly half of its ten-thousand-person workforce; Jack Dorsey publicly said AI had made many of the roles unnecessary and predicted that "within the next year, the majority of companies will reach the same conclusion." Salesforce replaced four thousand customer-support agents with agentic AI. Over one hundred thousand tech workers were laid off in 2025, with AI cited as a primary driver in more than half the cases. The firms racing toward the cliff are the same firms whose customers they are replacing. > The capex bet assumes a world its own success would prevent. The bet contains its own contradiction. Massive labour replacement to recover the capex shrinks the demand base that pays for the capex. The trillion-dollar AI revenue line and a hollowed-out middle class cannot coexist for very long. ## Why "give everyone ChatGPT" is not the answer The dominant enterprise response to all of this is the seat-licence strategy. Buy ChatGPT Enterprise or Microsoft Copilot for the whole organisation, thirty to sixty dollars per user per month, let productivity bloom. Uber's experience is not an outlier. Three structural failure modes recur in every seat-licence rollout that has run long enough to generate data. The seat economics get away from CIOs faster than expected. The most active users, the engineers and analysts, blow through usage tiers and end up in API territory. The least active users sit on unused licences. Usage is bimodal, not normal, and bimodal pricing models lose money relative to fixed-seat assumptions. The accountability dimension is worse. Decisions made with a chatbot's help leave no audit trail that survives scrutiny. A regulator asking how a decision was made is not satisfied by "I used the company's enterprise chatbot." The compliance team builds a shadow-governance regime around the seat fleet, and the productivity gains erode against the audit overhead. The capability fit is wrong as well. Most enterprise decisions are not language problems. Supplier risk scoring, claim adjudication, fraud detection, demand forecasting, route optimisation, batch classification: these are pattern-recognition problems on structured data. Machine learning is purpose-built for them. A chat interface is the most expensive way to ask a model to do them. This is why I have not deployed enterprise chatbot seats across my own workflow. I keep my Claude Max account at two hundred dollars a month for the language work where a model genuinely earns its place. ## The architecture that survives is human-accountable, not LLM-first The architecture that survives the math is not LLM-first. It is human-accountable. The system drafts the action. The human reads, accepts, and signs. This is the principle behind every regulated enterprise decision system I have designed over twenty-five years of enterprise work. It is also the principle I am building into ChainAlign. By design rather than by accident, it is meant to be the architecture that survives the economics. Auditability is part of the design, not bolted on. > The system carries the load. The human carries the accountability. ## What I am watching I should be honest about where this argument can break. If inference costs collapse another order of magnitude over the next eighteen months, the seat-licence economics start to work and the urgency of the architectural argument softens. Open-weights models running on commodity hardware can change the curve. There are early signs of both, and either would change the picture without changing the accountability ceiling. If a major vendor moves its pricing model from token-based to outcome-based, paying per resolved decision rather than per token consumed, the alignment between vendor revenue and customer value improves and the budget-blow-up dynamic eases. Nobody is doing this at scale yet. It would be the most interesting development of the year. And if frontier model capability advances far enough that autonomous decisions can be trusted in regulated contexts, the accountability ceiling itself moves. I do not see this happening within the current architecture, but I have been wrong about capability ceilings before. What I am confident about is the direction of travel. Enterprises that build their AI strategy on the seat-licence model are buying a budget surprise. Enterprises that build on the architecture I have described are buying a system that costs less, audits better, and survives the next round of vendor consolidation. ## What this is actually a proposal for The Uber CTO's line is going to be repeated in CFO offices around the world over the next two quarters: the budget I thought I would need is blown away already. The first reaction will be to negotiate harder with vendors. The second will be to set internal usage limits. The third, eventually, will be to ask whether the architecture itself is right. The third reaction is the one that matters. Negotiating with vendors does not change the math. Setting usage limits does not change the math. Re-architecting around what AI is actually good at, what regulated accountability requires, and what the unit economics will tolerate over a five-year horizon does change the math. The bet that AI replaces labour at the price of labour is unlikely to pay. The bet that AI augments accountable humans at a fraction of labour cost will. The architecture follows from the bet. Build for the system that survives the second derivative, not the headline of the moment. --- *I design decision systems for regulated enterprises. The same architecture that makes them auditable also makes them economically durable. This essay is the longer version of why I do not deploy frontier chatbots at scale, even when I personally use one every day.* --- ### Postscript, same day After this piece went out, two details from the same week surfaced that belong on the record. They are different companies, same mistake. Uber was ranking engineers on Claude and Cursor token consumption. Token burn as a productivity proxy.[^3] Accenture, in February 2026, began tracking weekly logins to its AI tools across senior staff, and made "regular adoption" a factor in who gets promoted into leadership.[^4] These are Goodhart's Law in plain sight. The metric that was meant to track *adoption* became the objective function. Engineers optimised for token volume. Senior consultants optimised for logins. At Uber, eleven percent of backend code is now AI-generated, and the app, by the same reporting, has visibly degraded — AI-written restaurant blurbs all read as "authentic" and "rich", indistinguishable from the next. At Accenture, employees were already describing some of the tools as "broken slop generators" while simultaneously being told to log in more. Treating token usage, or login count, as a performance metric is the exact anti-pattern the accountability architecture is meant to prevent. Tokens and logins are input signals, not decision outcomes. A company that OKRs on either has confused cost with value. The accountability ceiling and the measurement trap are the same story from two angles. The bet only survives the second derivative if the system is designed so that the thing being measured is the thing that would also hold up to audit. --- ### Sources [^1]: [Laura Bratton, "Uber CTO Shows How Claude Code Can Blow Up AI Budgets", The Information, April 2026](https://www.theinformation.com/newsletters/applied-ai/uber-cto-shows-claude-code-can-blow-ai-budgets) [^2]: [Brett Hemenway Falk and Gerry Tsoukalas, "The AI Layoff Trap", March 2026 (arXiv:2603.20617)](https://arxiv.org/abs/2603.20617) [^3]: Hacker News discussion on the Uber CTO interview, including the detail that Uber ranked engineers on Claude / Cursor token consumption and that 11% of backend code is now AI-generated: ["Uber's Anthropic AI push hits a wall", April 2026](https://news.ycombinator.com/item?id=47826328). [^4]: [Fortune, "Last year, Accenture trained 550,000 workers in AI — now it's warning senior staff to use it or don't get promoted", February 2026](https://fortune.com/2026/02/23/last-year-accenture-trained-550000-staff-use-ai-now-promotions-hinge-on-putting-that-into-practice/). See also The Decoder's coverage of the employee pushback, which includes the "broken slop generators" quote: ["Accenture ties promotions to AI tool usage while some employees call the tools 'broken slop generators'"](https://the-decoder.com/accenture-ties-promotions-to-ai-tool-usage-while-some-employees-call-the-tools-broken-slop-generators/).