Tokens vs Humans

A few weeks ago Alex Karp, the CEO of Palantir, went on CNBC and basically lost it on live TV. Not in a crazy way, in a "someone finally said what everyone's been thinking" way. He looked at the camera and said something has gone completely wrong with how AI companies charge for this stuff. His actual line to the room full of enterprise buyers was simple: you're going to burn through tokens, get basically nothing for it, and hand these companies your IP in the process.

And look, Karp has a business reason to say this, he sells the alternative. But that doesn't make him wrong. It's actually a really good question and nobody in AI wants to answer it honestly. If this technology is going to triple your revenue, rewrite your org chart, replace half your team, whatever the pitch deck said this quarter, why does the bill look exactly like an electricity bill? Why am I being charged per token like I'm renting bandwidth, if what you sold me was a strategic weapon?

Think about it from the sales pitch. Nobody sells AI to a company by saying "this is a utility, priced like water." They sell it as transformation. Then finance gets an invoice that says $3 per million tokens in, $15 per million tokens out. That's not how you price a miracle. That's how you price compute.

Here's why they actually do it that way though. Value based pricing means sitting down and agreeing on what "value" even means, which is a nightmare of a negotiation. Token pricing is just easy. You can meter it, scale it, bill it automatically. So the pitch says magic and the invoice says utility, and everyone just kind of accepts the gap between those two things because arguing about it is annoying.

the part that isn't on the invoice

You already know this feeling even if you've never touched an AI product. It's the same panic as watching a prepaid data bundle burn down faster than you expected, checking the balance three days before the month ends and doing quiet math about whether you'll make it. A token meter is just that feeling with better branding and a corporate logo on it.

Okay but here's the thing that actually worries me more than the price. A token isn't just compute. It's also information leaving your building.

Every prompt you send is describing something about you, your codebase, your customers, whatever problem you're actually trying to solve. Karp's sharper point, once you get past the TV theatrics, is that you're not just wasting money on tokens that don't help you, you're also teaching a company everything about how your business runs while you do it. That's a weird trade. You're paying someone to learn your business while you're not looking.

This isn't paranoid. Atlassian announced this year that starting in August they'll train their AI on customer data inside Jira and Confluence by default. Not opt in, opt out, and the free and standard tiers get flipped on automatically, admins have to go find the switch and turn it off themselves. This is a company that used to say, in writing, "customer data is the customer's data and we are custodians of it." Most AI companies, including the big frontier labs, say they don't train on your enterprise inputs by default. Fine. But a policy on a website isn't the same as something you can actually verify. Once your prompt leaves your building you basically have to take their word for it.

what actually happened to figma

If you want to see this play out for real, not hypothetically, look at Figma and Anthropic.

For most of last year these two were close. Figma built AI features on top of Claude. In February they launched a feature that turns AI generated code back into editable Figma files, real integration, real partnership energy. Mike Krieger, Anthropic's product chief and the guy who cofounded Instagram, sat on Figma's board. This looked like a healthy relationship between a design company and its AI partner.

Then on April 14th Krieger quietly resigned from that board. Three days later, Anthropic launched Claude Design. A tool that turns a prompt straight into a working prototype. Which is, more or less, Figma's entire job. Figma's stock dropped the day it happened, and it kept dropping, down around half its value by summer. At an event a few weeks later, a Figma board member asked the CEO, Dylan Field, to talk about the relationship in front of everyone. He didn't hold back. He said Anthropic had "not consistently candid" with them.

By June an activist investor was sending Figma's board a formal letter asking for an investigation into whether Anthropic had used insider knowledge from the partnership before launching a competing product. Worth noting too, two of Figma's own board seats belong to VC firms that have money in Anthropic. So the people who are supposed to police this conflict are sitting inside it.

Nobody's proven Anthropic did anything wrong here, that's literally what the investigation is supposed to figure out. But you don't need the smoking gun to get the lesson. A close AI partnership gave one company a front row seat into exactly how another company works, and three days after that seat opened back up, a competing product showed up. Intent aside, that's just what happens when you give someone deep access to how you operate, and that someone is also deciding what to build next.

A person who learns your business is bound by things like, you know, being an employee, social norms, maybe a noncompete if you're lucky. A model doesn't have any of that. It's not loyal, it's not disloyal, it's just a tool that happens to now know a lot about you, sitting inside a company that's also deciding what product to ship next across every customer relationship it has at once. That's the actual difference between tokens and humans. Not intelligence. Obligation.

nobody's actually paying full price, which is its own problem

Now here's the other side of "why are they charging for tokens." They're not even charging you the real price.

Some analysts ran the numbers on this earlier in the year. They bought every subscription tier from Claude and ChatGPT and hammered them with heavy, realistic workloads until they hit the usage caps, then converted that into what it would've cost at normal API rates. The $20 a month plan delivered something like $400 worth of tokens. The $200 plan delivered around $8,000 worth. OpenAI's numbers were even more lopsided, up to 40 or 70 times the subscription price for their heaviest users. That's not a pricing mistake. That's the business model. It's a buffet, and someone else is covering the difference between what you pay and what it costs.

That someone is venture capital. OpenAI reportedly loses more than a dollar for every dollar it makes on inference. Sam Altman has said out loud that they lose money even on the $200 a month plan. Only about 5% of ChatGPT users pay anything at all. This is a strategy, buy market share while it's cheap to do it, but it's a strategy with a shelf life, because VC money isn't a permanent subsidy program, it's a bet that eventually wants a return.

And here's the twist that makes it genuinely strange. Prices per token have actually collapsed, something like 95% cheaper than two years ago for the same quality of output. You'd think that means the whole system is getting cheaper for everyone including the AI companies. Instead total spending on AI tripled in a year, because cheaper tokens didn't shrink usage, they exploded it. Agent style workflows can burn five to thirty times more tokens per task than a simple chatbot exchange. Cheaper per unit, way more units. The bill went up anyway.

the same thing is happening to the power grid, and it's your electric bill

This exact pattern, get more efficient, end up consuming more not less, isn't new. An economist named William Stanley Jevons noticed it in 1865 watching steam engines. More efficient engines meant cheaper coal per unit of work, which meant way more machines got built to use that cheap power, which meant England burned more coal overall, not less. Efficiency didn't lower demand. It unlocked a bunch of demand that wasn't worth paying for before.

If you've ever lived somewhere the power just goes off on a schedule, or kept a generator running out back because the grid couldn't be trusted to carry the load, you already understand this in your bones. The infrastructure never scales for free. Someone always ends up covering the gap, and it's rarely the people who asked for more power in the first place. Same story if you were in Europe through the winter of 2022, when gas and electricity prices tripled almost overnight and heating your flat turned into a monthly negotiation with your own bank balance. That wasn't AI, but it's the exact same mechanic: demand outruns supply somewhere upstream, and the bill lands on whoever's plugged into the same grid, whether they caused the spike or not.

Same thing is happening with AI and electricity right now, except this time it shows up on actual people's power bills. A guy in Manassas, Virginia opened his January bill and it had jumped from about $100 to $281. He told a reporter it was just so far beyond anything he'd ever seen. He hadn't changed a thing about how he lived. What changed was the data center that went up near him.

Data centers were about 2% of global electricity in 2024. That's projected to double by 2026, more power than all of Canada uses. A UN report this year said AI alone could hit 3% of the world's electricity and use more water than humanity needs to drink. And every time a lab announces a big efficiency win, more efficient chips, leaner models, the actual electricity used doesn't go down. It goes up, because cheaper compute just means more stuff gets built on top of it. Satya Nadella summed it up in four words on social media: "Jevons paradox strikes again."

Virginia is the extreme version of this. Data centers are now about 40% of the entire state's electricity use. Dominion Energy, the local utility, just raised base rates for the first time since 1992, adding about $8.51 a month to a normal household's bill, and they're saying they'll need almost 27 more gigawatts of power by 2039 just to keep up. In the mid-Atlantic grid region, capacity prices jumped 174% in one year. Baltimore residents saw their bills go up by double digits after a single auction. None of these people signed up for AI. They just live near a data center.

Some companies are trying to dodge this by building their own power plants next to the data center instead of using the grid. Sounds responsible on paper. Problem is, most of that power is currently coming from gas, not renewables, so it's not actually solving anything, it's just moving the mess somewhere the electric bill doesn't show it.

And it's not even that these companies are being sloppy about efficiency. Google actually cut its data center emissions per unit by 12% in one year through cleaner energy deals. In that same year its total electricity use from data centers went up 27%. Getting more efficient and getting a lot bigger at the same time aren't contradictions. They're literally the same story, which is exactly the trap Jevons described a hundred and sixty years ago.

so what's actually going on here

Put all three of these together and you get one honest sentence: nothing about this is free, someone is always paying, and right now it's just being spread out so you don't notice who.

You pay in tokens, sure, but that price is fake, propped up by investors betting on a future that hasn't arrived yet. You pay in data, whether or not you notice, because every prompt teaches the model something about you and there's no real way to check what happens to it after. And somewhere, a guy in Virginia is paying in his power bill for a data center he never asked for, so that the token price you're paying can stay artificially low a little while longer.

None of this means don't use the stuff. It's genuinely useful and it's genuinely cheap right now, which is a weird combination worth taking advantage of while it lasts. Just don't mistake "cheap right now" for "free" or "sustainable." Somebody's covering the difference. Sooner or later that bill comes due, and it usually shows up somewhere you weren't looking.