Bryan Catanzaro, Nvidia’s VP of applied deep learning, told Axios on April 26: “For my team, the cost of compute is far beyond the costs of the employees.” That’s the VP of the company that sells the chips saying his input cost beat his payroll. When that sentence comes from Nvidia, not from a complaining customer, the conversation moves.

The diagnosis is on the table. The addiction architecture Alex Prompter described this Saturday is real: useful product, consumption-based billing, internal gamification, inevitable invoice. Where Prompter falls short is in the conclusion. He says it isn’t a bug, it IS the business model. Eric Ries this same week, in a separate piece, showed why that doesn’t hold: structural problems get structural fixes, and the organization that designs its selectivity gate before the rollout avoids the cost Microsoft, Uber, Meta, and Catanzaro’s own team have already paid.

That’s the IQ Source thesis this week, and the concrete piece is AI Maestro’s Go/No-Go gate. The rest of this post takes apart why Prompter is right on the diagnosis and wrong on the ending.

The crisis isn’t about price. It’s about design.

The cluster this week is dense. Marc Andreessen, on Joe Rogan episode 2501 released on May 19, admitted that AI is making people less efficient in many practical cases. Tom Warren broke the scoop at The Verge: Microsoft started canceling Claude Code licenses for the Experiences + Devices team, forcing a transition to GitHub Copilot CLI before June 30. The detail that matters is from Deirdre Bosa at CNBC: “Microsoft’s own engineers liked Claude code best…and they’re cutting it anyway.” The decision came out of finance, not engineering.

Uber is the case we already covered in April from a different angle. CTO Praveen Neppalli Naga told The Information: “I’m back to the drawing board, because the budget I thought I would need is blown away already.” The 2026 AI budget Naga had planned for got torched by April. Uber doesn’t publish the exact figure for that sub-budget; the public number is total R&D for 2025, $3.4 billion, up 9% year over year. Full coverage of the case lives in the post on Uber’s governance bottleneck.

The layer that makes everything worse is the internal gamification. Aakash Gupta did the math on Meta’s “Claudeonomics” leaderboard: 60 trillion tokens in 30 days, roughly $180M per month at Sonnet’s public pricing. Meta shut down the dashboard two days after The Information published the story. Amazon runs MeshClaw, an internal tool to inflate consumption and climb the ranking. Uber built its own per-team usage leaderboards. Gergely Orosz put it plainly: “Token usage is part of perf evaluations at Meta. This is just smart people hitting targets they assume leadership wants them to hit.”

Prompter synthesized the architecture: you build something so useful that the only way to sustain it is to stop people from using it freely. And he closed it like this: “that contradiction isn’t a bug in the business model. It IS the business model.” That’s where he stops short.

Eric Ries says the opposite this same week

Eric Ries published a piece on May 25 in the Foundr newsletter promoting his new book, Incorruptible (Authors Equity, 2026), titled “Beyond The Lean Startup: Eric Ries on the Corporate Design Flaw Costing Millions.” The exact quote:

“Corruption in business is rarely a problem of bad actors or sudden moral weakness. It is structural. If your legal and financial architectures allow your business to profit while betraying its original purpose, success itself will eventually turn your creation into a monster.”

Ries calls that force financial gravity. His full argument is that standard corporate charters have design trapdoors: 80% of founders get removed within three years of an IPO because the internal board acts according to gravity. The examples he gives are concrete: Novo Nordisk’s industrial foundation that insulates research from quarterly pressure, perpetual purpose trusts, benefit corporations, the model Saul Price built with Costco after Fed-Mart pushed him out. Companies with industrial-foundation ownership are six times more likely to reach age 50 than conventional peers.

The operational point is singular. If the architecture allows the behavior, the behavior happens. And if the architecture allows it because nobody designed it consciously, the behavior is the default. The fix isn’t asking people to behave. It’s redesigning the architecture.

That’s what Prompter left out of his tweet. Addiction is architecture. Architecture gets redesigned.

Applied to AI: the selectivity contract

Ries was talking about charters. The same logic applies to AI consumption, with a useful twist. In 1995, I lived through the shift from metered dial-up internet to flat-rate broadband. In 2010, I lived through the shift from packaged software to monthly SaaS. Thirty years of technology taught us that the marginal cost of using the tool was effectively zero. Then per-token billing showed up and nobody updated the habit. The organization discovers that “using it for everything” gets expensive when the invoice arrives, not before.

Selectivity can’t be requested. “Please use AI thoughtfully” is the corporate equivalent of “drive carefully” before handing the car keys to a teenager. What does work is designing the usage contract before the rollout: which processes return positive value per token at current price, which don’t, which sit on the boundary and need continuous measurement. Without that contract, the default behavior is Meta, Amazon, and Uber’s: every team assumes more is better because the leaderboard rewards consumption, not net result.

It’s the same structural origin a previous post on where your AI bill actually comes from covered from a different angle. There the problem was exposure surface (dormant credentials, MCP bloat, leaked AIza keys). Here the problem is the usage contract. Same structural origin: the organization didn’t design the governance before opening the tap.

What AI Maestro does with that

AI Maestro is a two-month discovery program. The concrete output, in order, is: the Process Reality Map, the AI Opportunity Score, and an Opportunity Ranking with Go/No-Go Recommendation.

The Go/No-Go gate at the end of the second month is the operational equivalent of a mission-locked charter for AI. The organization walks into the gate with a real inventory of processes (not the vendor’s marketing catalog, not the innovation team’s piloting wishlist) and walks out with three categories: processes where the economics close at current token price and scaling is worth it, processes where the economics don’t close and won’t be deployed, and processes where the decision depends on a price change or model improvement worth tracking. Selectivity gets designed before the rollout, not discovered afterward by finance.

The cross-check matters. Two months of AI Maestro has a known and capped cost. The alternative is what Uber paid by torching its 2026 AI budget in four months, what Microsoft paid canceling licenses to engineers who wanted them, and what Catanzaro’s team pays with a compute bill that exceeds payroll. The question isn’t whether your organization will go through the consumption-governance conversation. It’s whether you have it before the rollout or after the rollback.

The parallel track is Technology Partner, IQ Source’s other service positioned as the alternative to AI runtime commoditization. If your moat sits in the workflow rather than in token consumption, the structural redesign runs through that too: build owned capability that doesn’t evaporate when the vendor raises prices or when the team gamifies the dashboard.

Three things to look at on Monday. Start with the leaderboards. If a team in your org has a dashboard ranking people by AI usage and it doesn’t subtract correction overhead, you already have a small Claudeonomics in the making. Then look at who signs the deploy authorization for a new agent: when only the consuming team signs, Ries’s financial gravity has nothing pushing back against it and the invoice grows quietly. The last one is the question most CFOs don’t have a clean answer to, which is how recently any of your production processes had a measured per-token return number that wasn’t from a vendor demo. If the answer rounds to “we haven’t,” the selectivity contract isn’t real yet, and the architecture is running on default settings.

Design the selectivity contract before the rollout