The world's biggest AI cloud is not a someday. It is booting up right now, and a small group of people got there first.
Four provider Macs. Three open-source models. 13,376 requests already served, and 4,837,854 tokens pushed across real hardware to real buyers. Four machines is not a swarm, and it is not trying to be one yet. It is proof the whole thing runs end to end, and every wave from here makes it bigger. The hard part was never the two-hundred-millionth Mac. It was the first one. That part is done.
This is one of the Macs that did it.
The guy who started from zero
His handle is @Zeroyen_jp. The name on the account translates from Japanese as "starting crypto from zero yen," and that is exactly how he showed up: no budget, one idle Mac in Japan, and the nerve to turn it on before anyone told him it was safe. Within a day his machine was serving live requests for an open model, meta-llama/llama-3.2-3b-instruct, in the background while he got on with his day.
He put it plainly. He joined to get off "a balance of 0 yen."
Then he did the thing the early ones always do. He found bugs, wrote them up, and sent them straight to the team. They shipped fixes fast and tipped him for the catch. He is public and clear about where it stands: the operators move quickly, the network is small, and it is early. He is staying online and reporting back as it grows. That is not a spectator. That is someone who wanted to be first.
Why the early ones turn it on now
Be clear-eyed about the timing. This is pre-beta, and payouts are barely a trickle so far — the counter on a provider's screen sits close to zero. That is not the catch. That is the window.
The people running Macs today are not here for week-one numbers. They are here because they can watch the loop close in real time: a host comes online, serves real traffic, hits a rough edge, reports it, the edge gets fixed, and the network gets stronger. Every turn of that loop pulls it closer to the two-hundred-million-Mac version. The providers online now will have been here since it was four nodes and a signed receipt. You do not get that moment back.
Every response already carries proof
Here is what four nodes already do that most inference APIs do not: every response comes back with a signed receipt binding the model, the output, and the provider that served it. Not at scale, not someday — from the very first request. Small network, same binding. For what a receipt proves and what it does not, see the signed receipt your inference API doesn't ship.
Get in while it is early
Malibu is in private pre-beta, and the first wave is filling by invite. That is deliberate. Every provider so far was brought on by hand, which is why it is four nodes and not four thousand, and why being in now actually counts for something.
The thesis is two hundred million Macs. The reality is the first few, live, serving traffic, with the next wave going out now. If you have an Apple Silicon Mac, claim your place in it: download Malibu, drop in your invite code, and your machine joins the same network as the ones already online. If you want the full argument for why this works, start with the manifesto.
The first Macs are on. The loop is closing. This is the part you want to be early for.
To get an invite, DM https://x.com/_aug11_ and we will send you the invite link.