Edison invented self-attestation in 1882. A jar of zinc plates in a chemical bath, read only in the utility’s own lab, on the utility’s own scale. The customer’s only options were to trust the number or dispute it with nothing to point to. That method never went away. Ask a frontier AI lab how many tokens you used this month and you get their number, produced on their infrastructure, checked by nobody but them. The meter still lives wherever the vendor happens to be standing.
There’s a version of usage-based pricing that claims to have solved this problem completely, by removing the meter altogether. It’s called unlimited.
Unlimited isn’t the absence of a meter. It’s an unreadable one.
The promise that always breaks
“Unlimited” sells. It’s a simpler pitch than a rate card, and it removes the one thing customers hate most about usage-based products: watching a number climb while they work. So vendors reach for it early, before anyone has done the math on what happens if usage actually is unlimited.
The math always catches up. A flat, unlimited price makes no claim about your specific usage at all, and for a while that’s the whole appeal. But some users cost the vendor many times what others do, and no flat price survives that spread forever. Something has to give, and what gives first is the word “unlimited,” quietly, usually after the marketing has already shipped.
Telecom hit this exact wall a decade ago. AT&T sold unlimited mobile data, then quietly throttled speeds past an internal, undisclosed threshold. The FCC proposed a $100 million fine in 2015, which AT&T disputed, and the matter was ultimately resolved through a separate FTC case that AT&T settled in 2019 for $60 million. The industry changes. The pattern doesn’t.
Perplexity, this spring
In May 2026, Perplexity Pro subscribers, $204 a year, reported hitting weekly caps on advanced models far faster than before, some after three to five queries in a single day, with no notice, no changelog, and no update to the pricing page. The wall arrived with a nudge instead: upgrade to the $2,004-a-year Max tier for “enhanced access to advanced models.” Only after complaints spread across Reddit and X did Perplexity answer: the accounts hitting new limits were tied to promotional codes, and enforcement had tightened against fraud and resale of those codes. Maybe that’s the whole story. No customer is in a position to check it, before the change or after the explanation. The meter, and the reasoning behind it, live entirely on one side of the invoice.
OpenAI, by design
OpenAI’s own help documentation shows the same shape, and is more candid about it than most. ChatGPT Business is sold as offering “virtually unlimited messages for eligible base models,” not unlimited. That same directness carries through: usage oversight “may occasionally involve a temporary restriction on your usage. We will inform you when this happens.” Give OpenAI credit for saying so plainly. The criteria are published too, in the Services Agreement: abusive usage, account sharing, reselling access. What isn’t published is the measurement. Whether your account met one of those criteria is a determination only OpenAI can see, on usage only OpenAI holds. You’ll be told a restriction applied. You won’t be shown the number it was applied against.
This isn’t a claim that Perplexity, OpenAI, or AT&T did anything dishonest. Real infrastructure needs real limits, and a company that never rationed anything would go broke proving a point. It’s also not a story specific to one vendor. The same shape is showing up anywhere a flat promise meets real usage variance, and AI inference billing is exactly the kind of category where the spread between a light user and a heavy one is largest, which makes it exactly where this pattern shows up first. The objection is narrower and older than any specific company: whatever decision gets made about who’s using too much, it’s still a claim only one side can see, made about you, by the party billing you.
What “unreadable” actually means here
Edison’s zinc-plate meter was self-attested, but at least it was labeled honestly. It said it was a meter, and everyone knew to distrust it accordingly. An unreadable meter is worse: it’s sold as the removal of measurement entirely, so the claim underneath it, the quiet decision about who crossed a line, arrives with no label at all. Nobody expects to argue with a threshold they were told doesn’t exist.
Self-attestation didn’t get removed when vendors started selling unlimited. It just stopped being readable.
What recordation restores
PacSpace doesn’t decide who’s using too much, and it isn’t the arbiter of fair use either; that judgment still belongs to the provider running its own infrastructure. What changes is what happens to the decision once it’s made: committed to a record on the provider’s own cadence, before the invoice that reflects it, rather than reconstructed after a customer happens to complain, and checkable by both sides rather than surfacing only in a support ticket. The provider still makes the call. The provider no longer gets to be the only one who remembers making it.
That’s not just a downside avoided. A checkable threshold is what lets a provider keep selling flat, simple plans and actually enforce them, without spending trust every time a limit gets applied. A documented threshold is a defensible one, a reason on record before anyone disputes it. An undisclosed one is a $60 million settlement waiting for someone to notice the pattern. Neither side is better served by a decision that only exists in one party’s memory.
Usage recorded before the invoice. Written by one side. Visible to both. Owned by neither, even when the bill says unlimited.
Related reading: Edison Invented Self-Attestation. He Called It a Meter., on where the modern meter came from. And The Meter You Quietly Tune, on what happens to the meter’s corrections once the record is shared.
Sources
Federal Communications Commission, Notice of Apparent Liability and Order, AT&T Mobility (June 17, 2015): apps.fcc.gov
Federal Trade Commission, “FTC settlement returns $60 million to consumers affected by AT&T’s throttling practices” (November 2019): ftc.gov
Android Authority, “Frustrated Perplexity Pro users claim advanced AI model usage limits have been drastically cut” (May 18, 2026, includes Perplexity’s official statement): androidauthority.com
PiunikaWeb, “Perplexity seems to have quietly reduced usage limits for Pro users” (May 15, 2026, updated May 19 with company statement): piunikaweb.com
OpenAI Help Center, “ChatGPT Business — Models & Limits”: help.openai.com
PacSpace, “Edison Invented Self-Attestation. He Called It a Meter.”: pacspace.io/field-notes-edison-invented-self-attestation