AI’s Shadow Output Gap
While Washington obsesses over debt and inflation, AI is already ushering in an age of abundance (Part 1)
⸻
The political and economic establishment can’t stop talking about deficits, debt, and the CPI. Capitol Hill hearings, FOMC minutes, and financial news all pulse to the same beat.
Yet this fixation ironically coincides with the arrival of the most powerful productivity engine in human history: generative AI. Its impact is creating a shadow output gap — an invisible but rapidly widening expansion of supply-side capacity. Policymakers, especially at the Federal Reserve, act as if the boom doesn’t exist.
The real risk is not inflation. It is a stealth supply shock that pushes prices, wages, and term premia down. Deficits may prove too small. Monetary policy may already be too tight.
⸻
Productivity Everywhere — Except in the Data
This is Solow’s Paradox, redux: “We see the computer age everywhere except in the productivity statistics.” Only this time the curve is ten-times steeper.
Previous tech waves required hardware diffusion—mainframes, PCs, smartphones. AI requires none of that; it arrives through an app. That frictionless uptake already generates latent productivity that never reaches GDP because it appears as:
•lower input costs (fewer billable hours),
•consumer surplus (time saved, spending skipped), and
•silent substitution (high-skill labor quietly displaced).
Illustrations abound:
•A patient triages symptoms with ChatGPT and skips four clinic visits.
•An analyst masters a new industry without three costly expert calls.
•A five-person start-up closes a seed round with no CFO, lawyer, or recruiter—AI fills those roles off the books.
Each case creates real value, but none is logged as “output.”
⸻
Counting the Invisible Token Economy
Tokens — the fragments of text an AI model processes — are the kilowatt-hours of knowledge work. Track them and you watch the shadow gap in real time.
•Google’s token throughput grew 50-fold year-over-year as usage soared and per-token cost collapsed.
•OpenAI’s models now sit in support desks, research departments, and legal teams worldwide.
•Rapidly falling costs are unlocking accelerating demand across every provider.
The data-center capex from Nvidia, Microsoft, and other hyperscalers is simply the physical expression of this surge.
(1/2). $NVDA $AMZN $GOOGL $MSFT $TSM $CRWV $NBIS
In our conversation about the challenges associated with measuring the economic impact of AI, @DratchCap suggested that we might be under-rating how popular AI with the people who use it.
Matt called out AI-assisted therapeutics as one category where public sentiment is likely to be very positive.
"You mentioned Retratutide... The other day I saw that one of the healthcare companies has got AI going on with cancer screening tools and the results are astounding."
"Things like that are why this is really exciting."
Some thoughts on the $GOOGL raise. First, I don’t think this has anything to do with their view of equity value. Larry / Sergey want to spend; Sundar / CFO said ‘our debt rating tho!’, and so here we are. Of the $80bn, $40 is for infrastructure. BUT I suspect that is the *equity check* for a META-like SPV structure that will be levered 5-6x. In other words, the spending implications are much greater than the headline…. Can’t believe it’s only Monday!
@TMTLongShort What could happen is JVs with the memory companies to speed up expansion. More prepayments, equity investments, which de-risk capex / depreciation of the fabs and smooth earnings. Would help the buyback story. I think Hynix is open that they are exploring all options.
They dont run b/s waiting for a response from rating agencies, they build in cushions. 2026 capex doesn’t matter, the SPV would not show up as traditional capex (that’s the point). If u decide to break ground now on something, when does the capex go in? Not this year… takes a bit. I expect the SPV structure to be a thing
Too absolute. There are always other constraints. In this case, they are just protecting the corporate b/s since already issued size in bond mkt. CFOs don’t run the balance sheet until the rating agencies are forced to act, they leave headroom. Especially if ‘27 capex is going UP. Re the ‘they reiterated ‘26 capex’ point mentioned elsewhere… so? If you decide to break ground today when does your capex go in? Not this year. And if they do the meta spv structure they wouldn’t have to take it up anyway.
@tropicalvalue They probably see the full stack as a durable moat with multiple ways to win (same). $googl suffering from some meh ad checks (if one believes them) and model disappointment short term. But there’s a deep bid below (hence $BRK). Good opportunity for patient folks.
@mcuban Agree, there is some potential over time that it all becomes table-stakes and the margin is competed away. For example, I use AI all day everyday running my portfolio. I can’t imagine ever going back. And yet.. it’s still just as hard to make PNL :). So, we’ll see…
@mcuban Agree, there is some potential over time that it all becomes table-stakes and the margin is competed away. For example, I use AI all day everyday running my portfolio. I can’t imagine ever going back. And yet.. it’s still just as hard to make PNL :). So, we’ll see…
Tokenmaxx first. Marginmaxx later. Some thoughts.
I think people are missing a necessary first phase of enterprise AI adoption. Of course early enterprise AI use looks wasteful. Companies encourage usage (leaderboards) and everyone starts brute forcing workflows. But I think the mistake is treating early token burn as steady-state COGS when a lot of it is really “workflow R&D”.
One of the first things I built when we got AI fully connected to work systems was my own AM Brief. There were many iterations / fails / re-dos / permission issues / tool issues / formatting issues / etc. The process was more tenuous than I thought it would be. And certainly different than the experience I had using my personal AI tools.
In the end, it came out great. But if I had done it the final way first, I probably could have saved ~90% of the tokens. Obviously, that’s not realistic for a first pass and the workflow has to be discovered. The process though was extremely valuable. We got to see which systems mattered, what permissions broke, which tools were needed for what, where things timed out, and then of course what “good output” means to me / my judgment.
The result was that my next automated project (podcast scrape + weekender + risk / chart review) was done in 1/2 a day and not over the course of multiple days. I.e. the investment paid off.
This all brings me to the “harness” side of AI, which I think will be a large part of successful “goodput” over time. Claude Code showed us that innovation extends beyond parameters / FLOPs to “system coherence” and tool use, among other things. And these agent harnesses are still very early. Like the rest of the space, they are going to get better fast…. The same goes for things like $NBIS token factory which is focused on improving token efficiency.
So I think the path is:
tokenmaxx => experimentation / workflow discovery => better harness => better goodput / token => MARGINMAXX.
And once goodput per token goes up, we know what follows (Jevons blah blah).
The other interesting thing that is just starting is that enterprise AI is going to give the AI stack exposure to corporate decision traces for the first time.
I saw this point somewhere else and wish I knew who to credit, but machine learning worked insanely well for ad/reco engines because companies like Meta had rich behavioral data / knowledge graphs / decision traces around users. AI doesn’t really have that yet for corporates.
Most corporate data is the final output of deterministic software. But the interesting stuff is the messy middle… why did we choose xyz? Why did we reject? Why did the process break? What context mattered? What judgment call did the human make?
That all sounds a lot like… Tokenmaxxing.
Marginmaxxing is next. And, luckily, the cost per token curve keeps bending down as we move fwd :)
Happy Monday!
🦔Tech companies that pushed employees to maximize AI usage are now realizing the math does not work. Microsoft, Meta, and Amazon all set internal targets that pressured workers to use AI tokens aggressively to hit productivity scores. The problem is agentic AI burns up to 1,000 times more tokens per task than a standard LLM query because it loops through multiple steps and self-checks.
OpenClaw's creator Peter Steinberger said his team spent $1.3 million on OpenAI tokens in a single month. Nvidia CEO Jensen Huang told his engineers they should be consuming AI tokens worth at least half their annual salary every year. The behavior has its own name now, "tokenmaxxing."
My Take
The cost trajectory works backwards from how the labs sold it. Per-token prices have fallen, but the number of tokens each task consumes has climbed faster, and the all-in spend keeps going up release after release. Agentic AI is the worst offender because the model talks to itself, second-guesses itself, and runs the same logic three times before landing on an answer. Goodhart's Law also shows up clearly here. When AI usage became the performance review metric, employees started using AI to inflate the metric, not because the task needed AI.
OpenAI and Anthropic are losing roughly $2 for every $1 of revenue, and the only way the math fixes itself is by raising prices or capping consumption per enterprise contract. Both moves slow the revenue growth the labs need to show on the IPO roadshow. Goldman Sachs and the underwriters know this, which is why SpaceX's S-1 came out before OpenAI's. Whichever AI lab files first gets the cleaner narrative, and whoever files second has to explain why their largest enterprise customers just started rolling back token consumption. The companies pushing tokenmaxxing internally are now the same companies signaling cost pressure externally, and that contradiction is going to show up in earnings the moment these labs start reporting publicly.
Hedgie🤗
Agree w/ the concern, but I’d frame it differently. These sectors aren’t simply “broken” (though of course some parts like drug pricing are, and h/t @mcuban for taking on this fight w/ his new bestie DJT). They’re also where rising real incomes go after things like goods / software / telecom go down in price. Wealthy societies spend marginal dollars on health, education (I.e. to some extent it’s a feature not a bug).
Thankfully, AI is directly attacking these “last frontiers of services inflation” in a way other technology break throughs haven’t. *This* should be the “pro AI” pitch from the labs and politicians instead of job loss doom and gloom.
Of course the big question in the short term is whether AI bends those service cost curves before the surplus is capitalized into the remaining bottlenecks. But tutors for everyone, ai curing disease, legal help for the little guy is the upside we’re chasing and those angles don’t get enough attention. Instead we get regulatory capture / Tower of Compute (babel) attitudes and green new deal 2.0s to save us from the machines.
This is an interesting theory, but one may worry that the inefficient and broken sectors of the economy will simply eat up any dark economic surplus, the same way they did during the computer revolution. https://t.co/VLGWQUSoiZ