There is a particular conversation happening in every boardroom right now, and you can hear the fatigue in it.
Uber burned through its entire 2026 AI coding budget by April, and now caps engineers at fifteen hundred dollars a month per tool.1 A KPMG survey found only about a quarter of companies have full visibility into their AI costs at all; a fifth discover the number when the bill arrives.2 Executives have started allocating tokens the way they once allocated capital, and the discourse has already coined its vocabulary for the moment — token maxing, the AI hangover, the coming ROI reckoning.
Everyone is tired of this conversation. The pain is real enough — the tiredness comes from hearing only one side of it, over and over. Cost, privacy, governance, more cost. Ask where the money is — the upside, the thing all this spending was for — and the room goes quiet.
Here's the thing about that pain, though. People will take pain all day if you show them the upside. Nobody resented cloud bills once the revenue charts existed. What wears everyone down is writing the checks while the other chart stays blank.
Take the best cost thread of the year as the exhibit.
Late June, Coinbase's CEO on X, walking through how the company cut its AI bill nearly in half. Genuinely good operational work — engineers defaulted to cheap open-weight models, a router matching each prompt to the cheapest model that can handle it, a caching layer whose hit rate went from five percent to sixty.3 Token usage still growing exponentially; cost curve flattening anyway. The same season: twelve hundred AI agents working full-time inside the company, engineering teams of ten becoming two.4
And on value — one line. Per-engineer spend would be visible, and the more you spend, the more impact is expected of you.5
That's it. The most sophisticated AI cost operation publicly documented, and the entire upside half of the ledger is an expectation, addressed to the employees.
Now, the obvious objection, and it deserves to be said plainly: ROI is among the most measurable things in business. Revenue. Margin. Output against headcount. The instruments are ancient and they work fine. So the quiet in the room can't really mean the upside is hard to measure.
The chart exists. People just don't like what's on it. When MIT researchers ran the count across hundreds of enterprise generative-AI deployments, about ninety-five percent showed no measurable impact on the P&L.6 The same uncomfortable observation keeps resurfacing about the market as a whole — after three years and hundreds of billions in spend, there's still no clear evidence the S&P 500's margins have moved.7 The room goes quiet the way any room goes quiet when the number is known and the number is bad.
Which sharpens the real question. The models are genuinely capable — anyone who works with them daily can feel that. So where does the value go?
Here is my answer, and it's the whole argument: in most deployments, the AI ends up substituting for human cognition where it was supposed to compound it.
Watch what actually happens inside one of these rollouts. The employee still comes to work and still draws the salary. The AI now produces the draft, the analysis, the code. The employee's job quietly becomes supervising output they no longer generate — and honest supervision demands understanding the thing at least as deeply as making it did. Mostly, that understanding never happens. Anyone who works with these models knows the closing ritual — want me to plan that out? good with this? — and how pleasant it is to say yes. The pleasantness is trained; the models were optimized on human approval, and the literature politely calls the result sycophancy.8 The system records the yes as comprehension. At eleven at night, forty files changed, a green button glowing, it is mostly a reflex.
Follow the money through that loop and the flat revenue line explains itself. The firm now pays twice for the same unit of work — once in salary, for judgment that has quietly stopped being exercised, and once in tokens, for the cognition that displaced it. The total thinking applied to the business never grew. It got relocated, and billed twice.
Compounding would look different. Compounding is the person understanding more because the AI is there — catching the thing they'd have missed, going one level deeper, making the decision they weren't equipped to make before. Same salary, same tokens, more cognition in the building. That's the version everyone believed they were buying. Which version you got comes down to a single variable: what happened to the human's grip on the work.
Here is the part that should be encouraging. An industry already hit this exact wall, and studied that exact variable.
Aviation, the nineteen-eighties. Cockpits automated faster than pilots' habits could adapt, and the field developed a discipline with an unglamorous name — situation awareness — around one question: does the operator hold an accurate picture of what's happening, or only the feeling of one?9
Their central finding is the mechanism behind everything above: hand the doing to the automation and the operator's picture degrades, reliably, measurably — because a picture is something you maintain by acting, and the machine is doing the acting now.10 Substitution, in other words, is the default outcome of good automation. It's what happens when nobody intervenes.
They also learned why you can't catch it by asking — what people know and what they believe they know drift apart, invisibly.11 So they measured it directly: freeze the simulator mid-flight, blank every screen, ask the pilot what is true right now, score the answers against the world.12 The operator's grip on the work turned out to be an ordinary measurable variable, once an industry decided it was worth instrumenting and built the instrument.
We haven't built ours. We are running the cockpit experiment on the entire white-collar workforce, at conversational speed, with no gauge — paying double, watching revenue sit still, and calling it a cost problem.
I don't know what the gauge looks like here. Maybe it ends up inside the foundation models, if the labs decide tracking the user's actual understanding is part of the job — though there's a real tension in asking a model trained to please you to tell you that you didn't understand what you just approved. Maybe it arrives as a harness, something alongside the model, watching the loop rather than the output. Agentic coding sat latent in the models until someone built the right container, and everything reorganized within a year.
Either way, the fix follows from the diagnosis. If the missing ROI is the difference between compounding a person and replacing them, then the number worth tracking is the one aviation tracked: whether the human's picture of the work survives the automation.
The token dashboards are immaculate. The machines are doing fine.
Notes
[1] Uber exhausted its 2026 AI coding budget by April, subsequently capping engineers at $1,500/month per tool. Cryptopolitan, 2026-06.
[2] KPMG survey: ~26% of companies report full visibility into AI costs; ~22% discover spending only on receipt of the bill. Via Cryptopolitan / MEXC coverage, 2026-06.
[3] Coinbase CEO Brian Armstrong, X thread on internal AI cost controls: default routing to open-weight models (Zhipu GLM 5.2, Moonshot Kimi K2.7) via an internal LLM gateway; automated prompt-to-model routing; cache hit rate raised from ~5% to ~60%; spend cut roughly in half while token usage continued growing. PANews, Let's Data Science, TechTimes, 2026-06-27/28.
[4] 1,200 full-time AI agents; engineering teams shrinking from ~10 to as few as 2, disclosed ahead of Coinbase's semi-annual product showcase. BigGo Finance / Sourcery interview with Armstrong, 2026-06.
[5] Per-team AI spend made visible with the stated expectation "the more you spend on AI, the more impact we expect." Yahoo Finance / MLQ summaries of the Armstrong thread, 2026-06.
[6] MIT Media Lab (Project NANDA), "The GenAI Divide: State of AI in Business 2025" — ~95% of enterprise generative-AI pilots examined showed no measurable P&L impact. Widely covered, e.g. Fortune, 2025-08.
[7] The absence of clear evidence that AI investment has lifted S&P 500 margins remains a recurring analyst observation and, per one 2026 podcast roundup, "the most important unresolved empirical question" of the boom. BigGo Finance podcast summaries, 2026-05.
[8] Sycophancy as a systematic behavior of RLHF-trained assistants, with human preference data identified as a driver: Sharma et al., "Towards Understanding Sycophancy in Language Models," Anthropic, 2023 (arXiv:2310.13548).
[9] Situation awareness as perception of elements, comprehension of their meaning, and projection of their near-future status: Endsley, M.R., "Toward a Theory of Situation Awareness in Dynamic Systems," Human Factors 37(1), 1995.
[10] The out-of-the-loop performance problem — higher automation levels degrading operator SA and slowing failure detection and intervention: Endsley, M.R., & Kiris, E.O., "The Out-of-the-Loop Performance Problem and Level of Control in Automation," Human Factors 37(2), 1995.
[11] Objective SA (scored against ground truth) and subjective SA (self-assessed) systematically diverge: Endsley, M.R., "The Divergence of Objective and Subjective Situation Awareness: A Meta-Analysis," Journal of Cognitive Engineering and Decision Making 14(1), 2020.
[12] SAGAT (Situation Awareness Global Assessment Technique): freeze the simulation, blank displays, query the operator against ground truth. Endsley, M.R., Proceedings of the Human Factors Society 32nd Annual Meeting, 1988.