Intelligence on tap, for everyone

A couple of weeks ago I wrote that intelligence on tap is addictive. That was about what it does to you once you get used to it. Then the US government switched Fable off, and the question turned outward.

The model that felt so addictive on a Tuesday was, by Friday, suspended for every customer on national security grounds. So instead of what this does to me: when intelligence gets cheap, fast and available around the clock to anyone with a connection, what happens to the world, and who gets it?

The cost is falling off a cliff

The cost of running a given level of intelligence has dropped roughly a thousandfold in three years. A GPT-4-class answer that cost around twenty dollars per million tokens in late 2022 now costs under forty cents, and the price war keeps going.

And it’s about to compound. This week OpenAI and Broadcom unveiled Jalapeño, a custom inference chip taken from design to manufacturing in nine months, the fastest cycle ever for a chip this advanced, and they used OpenAI’s own models to do it. Intelligence is now building the hardware that makes intelligence cheaper, which makes the next models cheaper still. When the thing getting cheaper starts compounding its own cost curve, you stop extrapolating in straight lines.

Sam Altman’s framing is intelligence as a utility, too cheap to meter, like electricity or water. Once it’s that cheap, the question stops being who can afford it and becomes what we do with the fact that nearly everyone can.

What it has already unlocked

The optimistic case isn’t a forecast. It already happened.

For about fifty years, working out the shape of a protein was one of the slow, hard problems in biology. Then AlphaFold solved it, predicted the shape of over 200 million proteins, handed them to more than two million researchers for free, and won the 2024 Nobel Prize in Chemistry. A half-century bottleneck became a lookup.

Dario Amodei calls what’s next the compressed 21st century, making fifty to a hundred years of progress in biology and medicine in five to ten. You don’t have to buy the timeline to feel the direction. The most hopeful thing happening isn’t chatbots writing emails. It’s cheap intelligence aimed at problems we’d given up on solving by hand.

Now run it forward

Take that and imagine it everywhere, with intelligence that keeps getting smarter and never sleeps. The kind of thing that stops being science fiction:

• A tutor for every child. One-on-one tutoring has always lifted struggling students to the top of the class. The only thing missing was supply, and supply just stopped being the problem.
• A doctor in every pocket. A triage nurse, a second opinion and a specialist who answers in the middle of the night, for the billions who have none of that today.
• A scientist for the unfunded problems. Rare diseases and niche materials that a few thousand patients could never justify become worth solving when the experiment costs almost nothing.
• Materials by simulation. New battery chemistries and better solar cells, found before anyone builds a prototype.
• Medicine tuned to you. Treatment shaped to your exact biology instead of the average patient.
• The language barrier gone. Every conversation translated the moment it happens.
• Someone to talk to at 3am. Patient, always there, when things are dark.
• A corner shop with a giant’s firepower. The legal, financial and marketing muscle a big company rents a whole floor of people for.
• A few people building what took a thousand. A small team shipping what used to need an army.
• An expert on call for anyone. A lawyer or a research assistant for the person who could never afford one.

When intelligence is cheap and always on, the question flips from can we afford to look into this to why wouldn’t we.

The honest counterweight

To stay optimistic without being naive, here are the people who disagree.

Cheaper per unit doesn’t mean cheaper overall. There’s an inference cost paradox: prices per token collapse while total bills climb, because an agent makes ten or twenty calls to finish one task.

And abundance doesn’t distribute itself. Cheap intelligence could entrench inequality rather than dissolve it, splitting the world into those who own the models, those skilled enough to wield them, and those who just receive whatever trickles down. A utility is only as fair as its grid.

There’s a humility note too. Both Altman and Amodei walked back their job-apocalypse predictions in May; Altman said he expected more white-collar jobs gone by now and was glad to be wrong. No one knows the future, so hold your predictions loosely.

The gatekeeping turn

Which brings us back to Fable. On June 13 the Commerce Department ordered Anthropic to suspend Fable 5 and Mythos 5 for every customer, citing national security. And just yesterday Altman told staff GPT-5.6 will ship only to government-approved partners, approved customer by customer, a rollout he admitted is “not our preferred long-term model.”

So at the exact moment intelligence becomes abundant, the top shelf is being put behind glass. I don’t think that’s the opposite of the story. Gatekeeping slows the ceiling, not the floor. The frontier gets fenced, but it keeps moving, and last year’s locked-down model becomes this year’s commodity anyone can run cheaply. You can ration the very best for a while. You cannot ration intelligence in general. The tap is already open for almost everyone, and it doesn’t close.

So do we change how things work?

Yes, and not at the edges. When intelligence is cheap, capability stops being the bottleneck and everything else takes its place.

For companies, it’s the lesson I keep coming back to: bolting AI on gets you a few percent, rebuilding around it gets you a multiple. When anyone can generate the output, value moves to what stays scarce, taste, trust, attention, real human connection, none of which a model hands you for free.

For society, the question isn’t capability, it’s distribution. We’re about to have far more than enough intelligence to go around. Whether that lands as broad abundance or a wider gap is a choice about access and who’s allowed to plug in, not a limit of the technology.

The next step

What I’m watching is intelligence that doesn’t just answer but acts. Agents that do the work, and physical AI that steps off the screen into the world. Cheap reasoning was the first half; cheap reasoning wired into things that move is the half that changes daily life.

I said before that what you do in the next few years matters more than what you did in the last ten. Abundant intelligence sharpens that. The people and places that put it to work first will set the pace everyone else is trying to match. The tap is open. The only question left is what we point it at.