This is the part that should make you a little nervous: when a tool becomes a utility, the people running it stop acting like a product company and start acting like infrastructure.
GPT-5.4 hitting “5 trillion tokens a day” in its first week is not just a flex. It’s a signal that we’re sprinting into a world where language—and by extension, thinking work—gets bought and sold the way we buy bandwidth. And once you’re in that world, the arguments change. It’s less “Is this app good?” and more “Who controls the pipe?”
From what’s been shared publicly, OpenAI is saying GPT-5.4’s usage shot up to 5T tokens daily almost immediately, beating what their entire API did the previous year. They’re also floating a projected annualized run rate of $1 billion in new revenue tied to this wave. And Sam Altman’s old framing is back in the spotlight: intelligence as a utility, like electricity or water, and the plan to “flood the market with tokens” so people build more things.
On paper, that’s exciting. Cheap, abundant “intelligence” should mean more experiments, more small teams shipping real stuff, more people who aren’t rich or well-connected getting leverage. If you’re a solo founder, or a teacher making materials, or a customer support lead drowning in tickets, the dream is obvious: suddenly you can do what used to take a whole department.
But there’s a darker reading that I think is more realistic: “flood the market with tokens” isn’t generosity. It’s market capture.
When you flood a market with something, you don’t just enable innovation. You set expectations. People start designing their day, their company, their workflows around the assumption that the tokens will always be there, always be cheap enough, always be allowed. That’s how dependence forms. Not with a dramatic lock-in contract—just with a thousand small decisions that make it painful to leave.
Imagine you run a small online shop. You wire GPT-5.4 into your product descriptions, your email replies, your ad copy, your supplier messages. At first it feels like you hired three great helpers. Then your costs creep, or policy shifts, or quality changes, or your account gets flagged by mistake. Now your “helpers” are gone overnight and you don’t even remember how you used to do it manually. That’s not a theoretical risk. That’s what happens when a core function becomes a rented service.
The token number also hints at something else: this isn’t just a bunch of people playing around. At 5 trillion tokens a day, a lot of that has to be systems talking to systems. Bots drafting, bots reviewing, bots summarizing, bots generating options for other bots. It’s speed, but it’s also noise. The more text we can produce, the more we’ll drown each other in plausible junk.
And the real question isn’t “Can the model write?” It’s “Who is responsible when it’s wrong?”
Say you’re a nurse using a hospital tool that drafts patient notes. Or a lawyer at a small firm using it to outline a contract. Or a manager using it to write performance reviews. If the model confidently invents something—or just nudges a tone in a way that hurts someone—the human is still on the hook. But the human is also being pushed to move faster, trust more, and double-check less, because the whole point of the utility is volume.
That’s the trap: abundance doesn’t just lower costs. It lowers care.
There’s also a power shift hiding in plain sight. When “intelligence” is metered in tokens, you create a world where the most valuable skill isn’t thinking. It’s budgeting thought. The winners are the people who can afford to run more attempts, more drafts, more tests, more exploration. The losers are the people who get rationed. If you’ve ever seen what happens when a school district can’t afford the same software as the richer district, you already know how this ends—except now it’s not just software. It’s the ability to generate and process information at scale.
To be fair, there’s a counter-argument I don’t want to dismiss: maybe this is exactly what we want. Maybe the utility framing is honest. Electricity is regulated (in many places), standardized, broadly available. If intelligence becomes that kind of layer, society could benefit. A lot of people who are currently blocked by poor education, weak writing skills, language barriers, or disability could get real mobility. That’s not a small thing.
But the utility analogy cuts both ways. Utilities can be abused. They can price discriminate. They can shape what’s possible for everyone downstream. And when they fail, the blast radius is huge.
I’m also not fully convinced the “5T tokens a day” story tells us what people think it tells us. Tokens are a weird yardstick. High volume could mean value, or it could mean waste. It could mean a few giant customers hammering the system. It could mean spammy automation. We don’t know how much of this is durable, healthy usage versus a launch rush where everyone tries the shiny thing. And “annualized run rate” is not the same as money in the bank—it’s a projection, and projections are easy to love when momentum is on your side.
Still, the direction is clear: the center of gravity is moving from “AI as a feature” to “AI as a base layer.” That changes what we should argue about. Not just accuracy and safety, but dependency, pricing power, and what happens when one company’s choices become everyone’s constraints.
If intelligence really becomes a utility, who do you think should get to set the rules for it?