Parallax Says Data Centers Are Optional Now

What makes ChatGPT possible isn’t the chatbot you talk to, but the enormous data centers running behind the scenes.

Warehouses the size of a few football fields, packed wall to wall with computer chips, humming so hot it needs its own cooling system, drinking enough electricity to run a small city. Those data centers are the real deal, the chatbot is just what leaks out of it.

For years, that’s been the case. If you want to train serious AI, you need one of those buildings. And they aren’t cheap. They aren’t even possible for most people. Which is exactly why a handful of companies (OpenAI, Anthropic, xAI, etc.) own the future and everyone else rents it.

Parallax is built to make those buildings optional and dethrone the kings.

First, how big is this problem?

The power crunch is becoming more evident and the biggest names in AI are openly panicking about it.

Elon Musk has been saying it for over a year: the thing slowing AI down isn’t chips, it’s electricity. “The limiting factor for AI deployment is fundamentally electrical power,” he said at Davos, and warned we’ll soon be making more chips than we can plug in.

His own xAI supercomputer in Memphis needs about 155 megawatts to run all its chips at once. For context, that’s enough power for roughly 900,000 homes being gobbled up by one facility.

And when the grid couldn’t deliver fast enough, xAI rolled in gas turbines to burn natural gas on site, drawing accusations it was breaking clean-air rules to do it.

Google hit the same wall. It went shopping for nuclear power after admitting it simply didn’t have enough electricity to run the data centers it already has, let alone the ones it wants.

Microsoft, Amazon, Meta, all of them are now in the business of finding electricity, because the grid can’t keep up.

It’s gotten so absurd that Musk’s proposed a solution to put data centers in space, where the sun never sets and there’s nobody to complain about the power bill.

Read that back. The plan to keep AI growing is to launch the warehouses into orbit. That’s how cornered the big guys are.

Now, the Chutes part

Here’s where it gets interesting. Chutes already proved you don’t need one giant building.

Right now, Chutes runs AI models across a web of regular computers scattered all over the planet. Billions of tokens a day, no central warehouse, no single power bill the size of a nuclear plant.

Think of it like Airbnb for computer chips: instead of one company owning every hotel, thousands of people rent out the spare room they already have.

But Chutes only did this for running models. The hard part, the expensive part, the part stuck inside the billion-dollar buildings, was always training, teaching the models in the first place.

That’s the wall Parallax is trying to walk through.

What exactly is Parallax?

Parallax takes that same scattered web of everyday computers and uses it to train models, the thing everyone said you needed the warehouse for.

And we’re not talking about exotic hardware. We’re talking gaming PCs. Mac minis. MacBooks. The device you’re reading this article on. Each one does a small slice of the work, and stitched together, they do the job of the warehouse.

Guess what? This already worked.

The team trained a 20-billion-parameter model, a genuinely capable size, across these scattered everyday machines for under $10 an hour. Same goal, same timeline as the data-center way, but with about 82% less hardware.

And the enormously interesting part is, none of those scattered computers ever sees your data. The work gets done without anyone in the network being able to read what they’re working on. Privacy is literally stictched into the core of Parallax.

Why you should care

Less hardware doesn’t just mean cheaper. It means less power, less water, less land, exactly the things Elon and Google are scrambling over right now.

Nobody’s claiming Parallax’s approach saves the planet. It doesn’t. But every model trained the Parallax way is a warehouse that didn’t get built, a gas turbine that didn’t get fired up, a strain the grid didn’t have to absorb.

The whole assumption that progress requires bigger and bigger buildings is the thing getting challenged.

The dream of Parallax

We asked Jon Durbin, one of the people building this at Chutes, what wild success would look like. He answered:

“Parallax aims to be the destroyer of moats.”

A moat is the wall a company builds so nobody can compete with it. For the big AI labs, that moat is exactly the stuff we’ve been talking about: the money, the buildings, the power, the data only they can collect. Most people can’t climb it, so they’re stuck on the outside, always paying rent.

Parallax aims to build AI that’s nearly as smart as the big labs’, costs a fraction as much, runs on a fraction of the hardware, and put it in everyone’s hands instead of locking it in a vault controlled by the Sam Altmans of today.

And he made one point that’s worth ruminating over. Every time you use a big AI model for free, you’re not the customer, you’re the raw material. Your questions, your corrections, your data quietly train their model and make their wall taller. Great deal for them. Not so great for you.

Jon’s worry is what happens if this keeps concentrating. If a small number of companies end up owning all the intelligence, then they get to decide who’s allowed to use it and who isn’t.

<<<< Maybe here is the best place to mention Fable 😉

The whole point of Parallax is to make sure that no single person or individual ever ends up owning all of intelligence. The world’s knowledge trains these models, so the world should own them, and the world should be able to use them.

The full Parallax story straight from the Chutes team: