From Punch Cards to Tokens: Some Thoughts on AI Pricing
Palantir made a long post on X regarding “sovereignty” from AI:
The post is a little esoteric, and requires an almost Straussian reading, but the basic theme is clear: protect your IP from prying AI. Fools and their IP are soon parted. This is especially true if you are lured by the supposed cheapness of Chinese models like DeepSeek.
Palantir’s Alex Karp was more straightforward in an interview (see especially starting around 9:20):
Karp is essentially warning against model lock-in, more than having your data scraped and used for training: there are contractual safeguards for that. This brings to mind the interoperability debates involving operating systems and applications (e.g., Office) in the late-90s and early-00s. But the issue with AI is much deeper, given the complexity of the models. Palantir is offering interoperability. But won’t that get you locked in to Palantir?
One area where I part ways with Karp is over pricing. Karp criticizes pricing AI on a usage basis, i.e., through charging for tokens. Instead, he envisions deals in which the frontiers get a share of he value created by their AIs.
Great thought. But such contracts require measurability and verifiability to a third party (a court). Maybe this would be possible for, say, a drug created by AI. You can measure the sales of the drug. But for something like improving the efficiency of an internal business process, measuring the savings and perhaps more importantly proving the value to a court (in the event of a dispute between the parties) is likely impossible. So for many of the AI use cases, such contingent contracting is economically infeasible.
And even the drug case is hard. How do you verify the portion of a company’s AI that is attributable to that project? Companies would have incentives to try to attribute AI usage for other things to the project.
Charging for tokens reminds me of the old case of IBM requiring users of its machines to buy punch cards from IBM. It was a way for IBM to price discriminate–extracting more value from intensive users. IBM couldn’t measure the value an individual computer user captured–which basically was determined by where the user’s demand curve was located–but figured that more intensive users had higher demands and more surplus to extract. By tying cards and computers, and pricing cards at above marginal cost, IBM was able to extract more surplus from the bigger demanders. That is, it could price discriminate.
Charging for tokens works the same way. Indeed, the price discrimination is more subtle, because of the different pricing levels and token restrictions based on access to different models, usage intensity, etc.
The industry is new, and changing rapidly. No doubt pricing models will change as well. There will be attempts to implement Karp’s vision, and some of those will work, but it is not likely to be the default pricing mechanism. Instead, some sort of token-based pricing, likely tweaked from its current implementation, is likely to remain in place. This is especially true for non-siloed, general business purpose uses of AI.
The biggest innovations will occur on the buy side. I’ve referred recently to some of the sticker shock that AI users had experienced, and how they are cutting back on AI usage. Meta (surprise, surprise) was arguably the biggest idiot.
Facebook (Meta) burned through an astronomical 60 to 73 trillion AI tokens in a single month—costing an estimated $221 million—because employees intentionally wasted them to climb a gamified internal leaderboard. Rather than a software glitch or an accidental AI loop, this massive token burn was the result of a corporate culture trend known as “tokenmaxxing.” [1, 3]
To encourage its 85,000 employees to adopt AI tools, Meta launched an internal, voluntary dashboard called “Claudeonomics” (or Clawonomics). [1, 4, 5]
- The Metric: The system ranked employees solely on the raw volume of AI tokens they consumed.
- The Incentive: Meta mistakenly treated high token usage as a proxy for employee productivity. Top users were publicly showcased on the intranet and rewarded with gamified corporate titles like “Token Legend” and “Cache Wizard.”
Because token consumption was tracked as an “input metric”—measuring how much AI you used rather than what you actually accomplished—employees figured out how to fake productivity:
- Ghost AI Agents: Engineers left autonomous AI agents running continuously in the background for hours to execute completely pointless, infinite research tasks.
- Massive Waste: The top individual “Token Legend” at Meta managed to burn 281 billion tokens in 30 days entirely on throwaway work—enough data to reproduce the entirety of Wikipedia more than 33 times over.
- System Strain: The unchecked AI overuse became so extreme that it even caused internal system outages and tech disruptions (SEVs) inside Meta.
Once details of the massive, multi-million dollar waste leaked to the public, Meta quickly shut down the leaderboard. The company has since clamped down on “tokenmaxxing” by introducing a centralized dashboard called AI Gateway, which enforces strict spending controls, real-time tracking, and automated alerts for unexpected token spikes.
Incentives matter! Who knew?
Apparently Meta didn’t, despite the fact that its whole business model revolves around manipulating its users’ incentives.
Other companies (e.g., Uber) have experienced similar, though less extreme outcomes due to incentivizing token usage. They are now adapting.
Which means that the demand curves facing AI firms will change dramatically, as companies experiment with different internal incentive structures–the design of which will be anything but trivial, for similar reasons to those cited above. How can a company determine the value a given employ can generate with AI, and design a system that incentivizes realization of that value rather than mindless waste.
As companies experiment, AI firms will adjust their pricing. Which will induce companies to modify their incentive structures. This will take a long time before things settle down. There will be learning on both sides, and that learning takes time.
No doubt companies will use AI to try to design incentive structures. And the AI firms will use AI to adjust their pricing. The use of AI might accelerate the learning process.
But at the end of the day I predict some form of usage-based pricing, with structures that incorporate a substantial degree of price discrimination. A long way from IBM and punch cards, but it rhymes.
Your punchcard mention stirred old memories. No one had heard of an apartment building named Watergate, I was at U of Illinois where what passed for a DOD supercomputer used IBM punchcards for CS 101. Task, build a deck to compute SS withholding. I did it, sort of. Studentscran on some off schedule which meant hanging around until the operator called your name. You had cards with data for some number of mythical people, he ran the program cards, then the data cards—presto IRS! Can’t remember how I did in the course in the days when few had heard of grade inflation and TAs delighted in raking undergrads over the proverbial hot coals.
Comment by The Pilot — July 2, 2026 @ 1:28 pm
Compute pricing is in an early battleground, and yes, we’re likely to see a few different armies emerge. We need a futures market and futures curve to clear away the clouds, but development and agreement of an underlying index is where most of the blood & sweat is expended at this stage.
Comment by Patrick Rooney — July 2, 2026 @ 1:30 pm
“Transferring that data hands over access to your pre-existing winning plays and yields the means of production for new ones.”
Isn’t Palantir’s entire business model based on getting people to hand over their data to them..?
Comment by HibernoFrog — July 3, 2026 @ 1:26 am
It’s remarkable how rich Mr Suckerburg has become without, apparently, having the instincts of a businessman.
You wouldn’t trust him with a sweetie shop, would you?
Comment by dearieme — July 3, 2026 @ 6:20 am