OpenAI put its model on a chip as large as a plate. The result: up to fourteen times faster.

Yesterday, OpenAI opened a new tier in its developer interface. Its name: Ultrafast. A tier is a service plan: same product, different machine behind it. The model does not change. It is still gpt-5.6-sol, with the same answers and the same intelligence. Only the machine running it is different. And it can go up to fourteen times faster.

The dish of the day measures 21 centimeters across and costs the price of a car. The wine list does not display any prices either.

Here is an example provided by the company, because percentages mean nothing to anyone. The same request consists of generating a financial dashboard, twice, with the same model.

Standard    12 min 20 s
Ultrafast    1 min 50 s

Twelve minutes is enough time to go get a coffee, then come back having forgotten what you wanted to do. One minute and fifty seconds is enough time to stay in your chair. It is not the same job.

Why your AI spits out words one drop at a time

A short detour into the mechanics. This is where it gets interesting.

A model writes one token at a time. A token is a piece of a word. Figure on about three tokens for every two words. To produce each token, the machine has to reread all the model's weights, meaning the hundreds of gigabytes of numbers that make up the thing. It does this for one piece of a word, then starts over for the next.

On a graphics card, these weights do not fit inside the chip. They are stored in memory placed next to it. So they have to be fetched constantly. The result: your €30,000 card spends most of its time not calculating, but making trips back and forth to the fridge.

On the left, today's kitchen. The chef is excellent. He simply spends his whole life in the hallway.

It is counterintuitive, so many people get the subject wrong. The bottleneck, meaning what slows down the entire system, is not computing power. It is the pipe that brings data to the computing units. You can put a three-star chef in the kitchen. If he has to cross the warehouse for every onion, you will still be waiting.

A chip as large as a plate

That is where the idea from Cerebras, which makes the hardware used by this new tier, comes in. Instead of cutting the silicon wafer into small chips, the company keeps it whole. The processor is therefore 21.5 centimeters across. It covers 46,225 mm² and contains 4 trillion transistors as well as 900,000 cores. An H100 chip from Nvidia, the industry benchmark, measures 814 mm². It is fifty-seven times smaller.

The point is not to make a huge chip for the sake of it. This surface makes it possible to place 44 GB of ultra-fast memory directly on the chip. The model's weights therefore stay on the work surface. No more trips to the fridge. No more hallway. No more warehouse.

Cerebras claims an internal throughput of 21 petabytes per second, compared with around 3.35 terabytes per second for the memory of an H100. Let me add an honest caveat right away: these two figures are not measured in the same way. The first adds up the exchanges between 900,000 cores and their nearby memory. The second measures an external bus, meaning the link to the memory. So do not remember the exact ratio. Remember the order of magnitude. We are no longer playing in the same plumbing league.

Personally, what amuses me most about this whole story is that this bet was dismissed as an engineer's whim. A chip the size of a plate seemed impossible to manufacture without defects, impossible to cool, and impossible to sell. Ten years later, it powers OpenAI's model.

No, that's not the wafer you saw going around

A photo circulated widely, and it could easily confuse you. It confused me too. It shows Sam Altman, all smiles, holding a silicon wafer with a man in a suit. But it's not the one we're talking about here.

On June 24, Sam Altman and Hock Tan, CEO of Broadcom, present OpenAI's in-house chip. Photo by OpenAI and Broadcom. This wafer, however, is headed for the dicing saw.

This photo dates from June 24. The man on the right is Hock Tan, CEO of Broadcom. What they're holding is called Jalapeño. It's the first chip designed by OpenAI, manufactured with Broadcom, then assembled into racks by Celestica. A rack is a large cabinet that houses the machines. Nine months passed between the first sketch and the plans being sent for fabrication. For this kind of beast, that's very fast.

Now, take a close look at the wafer. See the grid and the hundreds of little rectangles? Each rectangle corresponds to a chip. This wafer will therefore be cut up. That's how processors have been made for fifty years: the chips are etched, cut apart, sorted, and then the defective ones are discarded.

That's why this isn't the same thing. Cerebras starts with a similar round wafer, but chooses not to cut it up. The same piece of silicon at the outset, the opposite decision at the end. The photo shows individual chips. This article is about a wafer used whole.

Three bets at the same time

That said, this photo tells us something else interesting about OpenAI. The strategy resembles that of a prudent family man: multiply the solutions so as not to depend on anyone.

If you count them, OpenAI is pursuing three hardware bets in parallel. The first is renting Nvidia cards, like everyone else. This is expensive and leaves the company at the mercy of a single supplier. The second is manufacturing its own chip with Broadcom, the one in the photo. Its deployment is scheduled by the end of the year, at gigawatt scale. Microsoft would take 40% of it, and one of OpenAI's coding models is already running on it in a trial version. The third bet, launched yesterday, is renting pure speed from Cerebras for situations where wait time becomes the real problem.

As for Jalapeño's energy efficiency, the company claims it is “significantly better than what's currently available.” No figure, no technical report. You know my position on this kind of statement: we make a note of it, then wait for the evidence.

The figures announced, and the ones I believe

The Ultrafast tier reaches up to 750 tokens generated per second. For comparison, a human reader reads about 250 words per minute, or four words per second. Here, we reach 560 words per second. The model therefore writes about one hundred and thirty times faster than you read.

The orange bar comes from the vendor. The gray ones were measured by someone else. That matters.

The figure of 57 tokens per second for the standard tier does not come from OpenAI. It comes from Artificial Analysis, which continuously measures the performance of available models. And it works out nicely: 750 divided by 57 equals 13. So the advertised “up to 14 times” holds up.

For the rest, you have to read the fine print. Cerebras also claims to be eleven times faster than Fable 5 and five times faster than Opus 4.8 in fast mode. The company also says it completed a supposedly difficult examination in 11 h 11, compared with 78 h 27. These are its measurements, conducted under its conditions and with its hardware. I'm not saying they're false. I'm saying Cerebras is not an independent referee.

The most interesting figure is also the most modest. Across a series of professional tasks, the total gain, from the start to the end of the work, drops to 5.6 times. That's logical. An agent at work doesn't spend its entire life writing. It compiles code, waits for the network, and reruns tests. The model may respond in the blink of an eye. If compilation takes two minutes, it will still take two minutes.

The price is missing, and that's not a detail

Oh yes, the price. There isn't one.

Not “it's expensive.” Not “it will be announced next month.” Nothing. No price, no general availability date, no indication of the cost. This is a closed preview, reserved for a handful of customers. The names mentioned are Jane Street, Podium, Basis, and Rogo. Everyone else can join a waiting list. And all of this remains limited to the developer interface. Nothing in ChatGPT.

When the menu doesn't show the prices, it's never to give you a pleasant surprise.

To give you an idea of the scale, the same model, at the standard tier, costs 5 dollars per million input tokens and 30 dollars per million output tokens. Input tokens correspond to what we send to the model. Output tokens correspond to its response. As for an entire silicon wafer etched at TSMC, it's not exactly a spare part picked up on sale. Will we pay twice as much for this speed? Five times as much? Will the service be reserved for people who trade at millisecond speed?

I have no idea, and that's precisely the problem. Speed you can't afford is called a demo.

So, what does this actually change for you?

If you don't code and have never opened a developer interface, the question is legitimate. Here's what's at stake.

Imagine the voice assistant in the kitchen. You ask it to convert grams into spoonfuls, and it takes two seconds to answer. Those two seconds are enough to ensure that no one uses it. We'd still rather take out the phone. Same problem with a customer service chatbot on a shopping site that waits twenty seconds before every response: you close the tab. Same thing with live translation while on vacation. With a delay, it's a gimmick. Without one, it becomes a conversation.

The cold coffee, the cobweb, and the beard: the real cost of artificial intelligence in 2026.

Now, let's talk about the timeline, without sugarcoating it. Today, it changes nothing for you. Zero. This is a closed preview, with no price, tested by four companies and accessible through a tool you will probably never open. The service could become widely available next year. It could also remain a luxury product for trading floors, or disappear if the price-to-speed ratio doesn't hold up.

The comparison that seems fairest to me is the transition from 56k modem to ADSL. On paper, we had merely sped up the same web. In practice, no one was watching online videos in 2000. It wasn't a matter of taste. The wait was unbearable. Speed didn't just improve an existing use. It created another one, and then YouTube arrived five years later. That's the same bet here. And it could very well fail.

What I think

I work with coding assistants every day. What no one measures in these announcements is the effect of waiting on your brain. Twenty seconds is nothing on paper. Except that in twenty seconds, you check your email. And you don't come back.

One of OpenAI's researchers sums it up better than I ever could: “it finishes before I even have the chance to move on to something else.” There you have it. The real issue isn't speed. It's attention. A tool that responds while you're still focused on it is a tool you use. A tool that responds after your coffee is a tool you work around.

Still, let's not get ahead of ourselves. Without a price, a date, or independent measurements, all we have for now is a nice promise and a pretty wafer. But if this speed becomes normal, we'll laugh when we think back to the time when we watched a cursor spin while a machine finished its sentence. A bit like the sound of the modem.