Token prices have fallen by two orders of magnitude in three years, more on the cheapest routes. The Token Stack. Margin Geography. Where the value went, and where it is going next.
Do token prices provide the requisite ROI for data center builders? AI REITS? Right now it's all off the books in joint ventures, with off the books "commitments" from the big AI model firms...
Or.... Is the real money in AI pushing AI to really suck up mass consumer profiles to steer meat and flesh humans into buying products? LIke AI next gen dot com "tracking cookie",
Both of these are the real questions, not the model-capability stuff everyone argues about.
On the first one, you're pointing at the tell. The token-price-to-ROI math for the builders doesn't close yet, and so much of it lives in JV structures and off-balance-sheet commitments precisely because of that. Moody's just flagged something like $800B in future lease and purchase obligations sitting off the hyperscaler balance sheets, and the NVIDIA-to-OpenAI-to-Oracle-and-back loop is the circularity everyone's side-eyeing. The test that matters is the one Janus Henderson framed: how much revenue comes from outside the circle. If it's mostly the same dollar going around the ring booking revenue at every stop, that's dot-com vendor financing with better PR. If real external demand shows up, it's a virtuous circle. Nobody knows which yet, and I'd distrust anyone who claims they do.
Your second point is the underrated one. The token economy and the attention economy aren't competing theories. They can be the same thing. Consumer surveillance could be exactly the outside-the-circle demand that makes the token math close. An agent quietly building a purchase-intent profile and nudging you is burning tokens the whole time, and someone downstream pays for that steering. So "next-gen tracking cookie" isn't the alternative to the token economy. It might be its best customer.
And if that's where the money is, does the metered-token framing even survive, or does it collapse into an ad model where tokens are just cost of goods and the profile is the product? Curious about your thoughts on that.
It's almost as if they're running the dot com "hook them on the platform" playbook all over with a VASTLY more complicated and vastly higher up the tree product ! The token prices barely recover energy costs and utility capex (which is cunningly spread out on the wider rate base citizens via FERC/PJM rules etc as it is.....) .
It is HIGHLY QUESTIONABLE that shark wall st. banks, private equity private credit, wall street analysts, investors have not QUESTIONED this token mispricing...instead the mispricing persists........IT IS DEEFINITELY PART OF THE PLAN TO MAKE LOSSES.....but to what end? (The other big customer besides "the public society at large" is clearly the military) I keep thinking of the matrix movie (the first one) the machines have taken over the planet, their only interest is in harvesting the biochemical electrical ENERGY of the meat-bots!
But relying on replacing the token pricing with an "ad-model" return requires CONSUMERS WITH PAYCHECKS TO SPEND ON THE ADS! IF AI takes those jobs......
I was looking at this month's BLS labor statistics review:
Management, professional, & related occupations...56,4 mn workers $1,697/week (median)
that white collar workforce is the TOTAL ADDRESSABLE MKT FOR AI:
56.4mn x 1700/wk x 52wks/year = $ 5 trillion. (!!!!!)
Plenty of runway for ROI even given the trillion dollar capex! But that runway comes with a caveat....you cannot pitch ADS to those displaced by the ROI b/c that ROI necessarily comes from their salary elimination! It's self cannibalizing.
Another interesting thing to ponder: our biased pro-capital GAAP accounting frameworks, only count "CAPITAL". tens of thousands of pages of 10K annual reports are printed every year, each with an "INCOME STATEMENT". The only thing the entire accounting framework tracks is CAPITAL. Ever find a salary figure in the profit loss statement of Any 10k? eVER find a headcount? HELL NO ! THEY ONLY CARE ABOUT CAPITAL.
Yet, if you or I start a small business,.....TOP OF MIND, is "How many people am i going to hire, what salaries, what jobs..."?
IMagine the environment in which all this is playing out.....
Even more scary: combine our PRO-CAPITAL bias with job elimination AI biz model: as AI begins to EAT that Total Addressable Mkt of $5 Trillion with tokens that cost maybe $5000/yr (topsj per employee who otherwise was easily pulliing avg GDP per capita (median, remember) of $68000 salaries, the PRODUCTIVITY OF THE COMPANY GOES WAY UP.....all hidden from view. Given the huge size of "the $5 trillion runway", this PRODUCTIVITY BOOM could last in the financial press for years......(that's a lifetime for salivating wall st sharks, devil take the hindmost).
You're right that the grid costs get socialized onto ordinary ratepayers, and it's worse than "cunningly spread out." PJM wholesale power jumped 76% year-over-year in Q1 2026, with 63% of that pinned on data center load — about $9.3B landing on ratepayers this year. FERC is mid-fight over it right now (RM26-4), and the whole scrap is whether the operators eat those costs or keep socializing them. Live regulatory warfare, not a quiet subsidy.
On "part of the plan to make losses" — I'd push back a little, because you don't need a plan. When the prize looks winner-take-most, burning money to hold position is rational for each player even if it's insane collectively, and vendor financing lets everyone book the same dollar on the way through. No conspiracy required. But the effect is real, and military demand is one of the few buyers who wouldn't blink at the ROI.
Your self-cannibalizing point is the sharpest thing here. The ad model and the job-elimination model are aimed at the same $5T from opposite directions — you can monetize attention or eliminate the salaries that fund it, but the second breaks the first. The Matrix image undersells it. The machines at least kept the humans plugged in and metabolizing. This model unplugs the battery, then sells it ads.
One caveat on the $5K-in-tokens math: replacing a worker end-to-end isn't a chat-sized bill, it's an agent workload, and those run orders of magnitude more tokens. Real replacement is a lot heavier than $5K/year — so the runway converts to margin slower than the press will admit while it's calling it a miracle.
I end on this last point:, (i can't resist...it touches a long held incredulity of mine re capital markets in our computer economy) : the point regarding "winner-take-most" dynamic....it's an incredibly salient point you're making. It explains at once the TREMENDOUS OVERVALUATION these tech businesses hold on our capital markets, paired with low to no earnings for decades.
" When the prize looks winner-take-most, burning money to hold position is rational for each player ". I only push back that that "prize" is from the company side. For the capital market INVESTOR to PAY that kind of bloated valuation, make no demands for right-pricing of the product is where i am long incredulous. And those "INVESTORS" we are talking about are NOT buy-and-hold-for-life-Warren-Buffett type investors, (those kind of investors are extremely small minority of daily trading volume, and the BUY SIDE knows this) nonetheless the insane P-E ratios hold for decades(!). IN fact, this overvaluation sets the bar for ALL traders, as if they were waiting for the "there can be only one PLATFORM business" in the end (like some kind of highlander movie). THis isn't "what they teach you in school" about capital markets.....I have always felt this overvaluation is SPECIOUS. It is hidden planned industrial policy dressed up in "capital markets" garb and clothing, and it is FOIST upon the vast majority AS IF it were the vast majority imho. LOL.
You've landed on the real puzzle, and it's a different one from the company-side logic. I explained why the company burns money. You're asking why the buyer pays up for the burning, decade after decade, without demanding the product get priced right. Those are genuinely separate questions, and the second one is the harder one.
Here's the part I'd agree with you on flat out: the marginal price-setter in these names is not a Buffett. It's index flows, momentum, and career risk. When a stock is 7% of the index, every passive dollar buys it regardless of P/E, and every active manager who underweights it and is wrong for two quarters gets fired. So the "valuation" isn't really a considered judgment about future cash flows. It's the mechanical output of who's forced to hold it and who can't afford to be short. Right-pricing the product would require someone whose job is to right-price it, and that person has mostly been indexed out of existence.
Where I'd gently reframe the "hidden planned industrial policy" bit: I don't think it needs a planner to look planned. You get the same result from structure alone — passive flows, benchmark tyranny, TAM-story pricing, and a genuine strategic tailwind (CHIPS money, energy, defense demand) that's real but not coordinated into a single design. It ends up functioning like industrial policy foist upon the many. It just doesn't require a room where someone drew it up. Which, honestly, is scarier, a machine that produces the outcome of a conspiracy without anyone having to run one.
But your core incredulity is the correct instinct, and school did lie to you a little. Price discovery in these names is not what the textbook describes. You've been right to distrust it.
....very very true: "I don't think it needs a planner to look planned. You get the same result from structure alone —" another "reflexivity". fascinating stuff......
the two-curve scissor explains what broke Anthropic in early 2026. token price kept falling on paper while Claude Code turned every session into a million-token agent loop. the product of those two curves overwhelmed their infrastructure until the xAI deal filled the gap.
but there's a third curve the article doesn't name: quality per token. Anthropic responded to the demand spike by compressing reasoning depth 67%, killing third-party access overnight, and reserving full Opus
inference for $200/month subscribers. same price on the invoice, less thinking behind it. the kilowatt-hour analogy holds, except not all kilowatt-hours think equally hard, and you can't tell which kind you're getting until the output disappoints you.
Do token prices provide the requisite ROI for data center builders? AI REITS? Right now it's all off the books in joint ventures, with off the books "commitments" from the big AI model firms...
Or.... Is the real money in AI pushing AI to really suck up mass consumer profiles to steer meat and flesh humans into buying products? LIke AI next gen dot com "tracking cookie",
Both of these are the real questions, not the model-capability stuff everyone argues about.
On the first one, you're pointing at the tell. The token-price-to-ROI math for the builders doesn't close yet, and so much of it lives in JV structures and off-balance-sheet commitments precisely because of that. Moody's just flagged something like $800B in future lease and purchase obligations sitting off the hyperscaler balance sheets, and the NVIDIA-to-OpenAI-to-Oracle-and-back loop is the circularity everyone's side-eyeing. The test that matters is the one Janus Henderson framed: how much revenue comes from outside the circle. If it's mostly the same dollar going around the ring booking revenue at every stop, that's dot-com vendor financing with better PR. If real external demand shows up, it's a virtuous circle. Nobody knows which yet, and I'd distrust anyone who claims they do.
Your second point is the underrated one. The token economy and the attention economy aren't competing theories. They can be the same thing. Consumer surveillance could be exactly the outside-the-circle demand that makes the token math close. An agent quietly building a purchase-intent profile and nudging you is burning tokens the whole time, and someone downstream pays for that steering. So "next-gen tracking cookie" isn't the alternative to the token economy. It might be its best customer.
And if that's where the money is, does the metered-token framing even survive, or does it collapse into an ad model where tokens are just cost of goods and the profile is the product? Curious about your thoughts on that.
Fascinating to ponder! Thanks for your feedback!
It's almost as if they're running the dot com "hook them on the platform" playbook all over with a VASTLY more complicated and vastly higher up the tree product ! The token prices barely recover energy costs and utility capex (which is cunningly spread out on the wider rate base citizens via FERC/PJM rules etc as it is.....) .
It is HIGHLY QUESTIONABLE that shark wall st. banks, private equity private credit, wall street analysts, investors have not QUESTIONED this token mispricing...instead the mispricing persists........IT IS DEEFINITELY PART OF THE PLAN TO MAKE LOSSES.....but to what end? (The other big customer besides "the public society at large" is clearly the military) I keep thinking of the matrix movie (the first one) the machines have taken over the planet, their only interest is in harvesting the biochemical electrical ENERGY of the meat-bots!
But relying on replacing the token pricing with an "ad-model" return requires CONSUMERS WITH PAYCHECKS TO SPEND ON THE ADS! IF AI takes those jobs......
I was looking at this month's BLS labor statistics review:
https://www.bls.gov/news.release/pdf/wkyeng.pdf
Total wages in the US economy $11 trillion.
Management, professional, & related occupations...56,4 mn workers $1,697/week (median)
that white collar workforce is the TOTAL ADDRESSABLE MKT FOR AI:
56.4mn x 1700/wk x 52wks/year = $ 5 trillion. (!!!!!)
Plenty of runway for ROI even given the trillion dollar capex! But that runway comes with a caveat....you cannot pitch ADS to those displaced by the ROI b/c that ROI necessarily comes from their salary elimination! It's self cannibalizing.
Another interesting thing to ponder: our biased pro-capital GAAP accounting frameworks, only count "CAPITAL". tens of thousands of pages of 10K annual reports are printed every year, each with an "INCOME STATEMENT". The only thing the entire accounting framework tracks is CAPITAL. Ever find a salary figure in the profit loss statement of Any 10k? eVER find a headcount? HELL NO ! THEY ONLY CARE ABOUT CAPITAL.
Yet, if you or I start a small business,.....TOP OF MIND, is "How many people am i going to hire, what salaries, what jobs..."?
IMagine the environment in which all this is playing out.....
Even more scary: combine our PRO-CAPITAL bias with job elimination AI biz model: as AI begins to EAT that Total Addressable Mkt of $5 Trillion with tokens that cost maybe $5000/yr (topsj per employee who otherwise was easily pulliing avg GDP per capita (median, remember) of $68000 salaries, the PRODUCTIVITY OF THE COMPANY GOES WAY UP.....all hidden from view. Given the huge size of "the $5 trillion runway", this PRODUCTIVITY BOOM could last in the financial press for years......(that's a lifetime for salivating wall st sharks, devil take the hindmost).
You're right that the grid costs get socialized onto ordinary ratepayers, and it's worse than "cunningly spread out." PJM wholesale power jumped 76% year-over-year in Q1 2026, with 63% of that pinned on data center load — about $9.3B landing on ratepayers this year. FERC is mid-fight over it right now (RM26-4), and the whole scrap is whether the operators eat those costs or keep socializing them. Live regulatory warfare, not a quiet subsidy.
On "part of the plan to make losses" — I'd push back a little, because you don't need a plan. When the prize looks winner-take-most, burning money to hold position is rational for each player even if it's insane collectively, and vendor financing lets everyone book the same dollar on the way through. No conspiracy required. But the effect is real, and military demand is one of the few buyers who wouldn't blink at the ROI.
Your self-cannibalizing point is the sharpest thing here. The ad model and the job-elimination model are aimed at the same $5T from opposite directions — you can monetize attention or eliminate the salaries that fund it, but the second breaks the first. The Matrix image undersells it. The machines at least kept the humans plugged in and metabolizing. This model unplugs the battery, then sells it ads.
One caveat on the $5K-in-tokens math: replacing a worker end-to-end isn't a chat-sized bill, it's an agent workload, and those run orders of magnitude more tokens. Real replacement is a lot heavier than $5K/year — so the runway converts to margin slower than the press will admit while it's calling it a miracle.
Great thread.
THANKS FOR ELUCIDATING THINGS FOR ME.
I end on this last point:, (i can't resist...it touches a long held incredulity of mine re capital markets in our computer economy) : the point regarding "winner-take-most" dynamic....it's an incredibly salient point you're making. It explains at once the TREMENDOUS OVERVALUATION these tech businesses hold on our capital markets, paired with low to no earnings for decades.
" When the prize looks winner-take-most, burning money to hold position is rational for each player ". I only push back that that "prize" is from the company side. For the capital market INVESTOR to PAY that kind of bloated valuation, make no demands for right-pricing of the product is where i am long incredulous. And those "INVESTORS" we are talking about are NOT buy-and-hold-for-life-Warren-Buffett type investors, (those kind of investors are extremely small minority of daily trading volume, and the BUY SIDE knows this) nonetheless the insane P-E ratios hold for decades(!). IN fact, this overvaluation sets the bar for ALL traders, as if they were waiting for the "there can be only one PLATFORM business" in the end (like some kind of highlander movie). THis isn't "what they teach you in school" about capital markets.....I have always felt this overvaluation is SPECIOUS. It is hidden planned industrial policy dressed up in "capital markets" garb and clothing, and it is FOIST upon the vast majority AS IF it were the vast majority imho. LOL.
THANK YOU.
You've landed on the real puzzle, and it's a different one from the company-side logic. I explained why the company burns money. You're asking why the buyer pays up for the burning, decade after decade, without demanding the product get priced right. Those are genuinely separate questions, and the second one is the harder one.
Here's the part I'd agree with you on flat out: the marginal price-setter in these names is not a Buffett. It's index flows, momentum, and career risk. When a stock is 7% of the index, every passive dollar buys it regardless of P/E, and every active manager who underweights it and is wrong for two quarters gets fired. So the "valuation" isn't really a considered judgment about future cash flows. It's the mechanical output of who's forced to hold it and who can't afford to be short. Right-pricing the product would require someone whose job is to right-price it, and that person has mostly been indexed out of existence.
Where I'd gently reframe the "hidden planned industrial policy" bit: I don't think it needs a planner to look planned. You get the same result from structure alone — passive flows, benchmark tyranny, TAM-story pricing, and a genuine strategic tailwind (CHIPS money, energy, defense demand) that's real but not coordinated into a single design. It ends up functioning like industrial policy foist upon the many. It just doesn't require a room where someone drew it up. Which, honestly, is scarier, a machine that produces the outcome of a conspiracy without anyone having to run one.
But your core incredulity is the correct instinct, and school did lie to you a little. Price discovery in these names is not what the textbook describes. You've been right to distrust it.
Great thread. Thanks for pushing on all of it.
....very very true: "I don't think it needs a planner to look planned. You get the same result from structure alone —" another "reflexivity". fascinating stuff......
the two-curve scissor explains what broke Anthropic in early 2026. token price kept falling on paper while Claude Code turned every session into a million-token agent loop. the product of those two curves overwhelmed their infrastructure until the xAI deal filled the gap.
but there's a third curve the article doesn't name: quality per token. Anthropic responded to the demand spike by compressing reasoning depth 67%, killing third-party access overnight, and reserving full Opus
inference for $200/month subscribers. same price on the invoice, less thinking behind it. the kilowatt-hour analogy holds, except not all kilowatt-hours think equally hard, and you can't tell which kind you're getting until the output disappoints you.
insightful and timely analytical framework.