The giants sell everything at once - and that's why they fall behind. The smaller players sell one single thing, clean access to a chip, and that's exactly why they get there first.
Hyperscaler is the word for the three companies holding most of the computers on the planet: Amazon (owner of the cloud service AWS), Microsoft (with its cloud Azure) and Google (with Google Cloud). Your mail runs on their servers, your photos, the accounting software of half the companies in the world - and on top of that, the chips for training AI models. This is exactly where the crack is. If thousands of companies want the newest AI chips at once, even a company with an unlimited budget can't build data centers fast enough. The line grows faster than the warehouses.
That's why the price difference isn't random. The same computation, bought from a hyperscaler, comes out up to three times more expensive than the same computation bought from a neocloud, because the hyperscaler spreads the cost of email, security, and databases across every client, even the one who wants only a chip and nothing else. A neocloud doesn't carry that baggage. It sells one thing, and sells it clean.
CoreWeave is the company that showed how that crack turns into a business. It started as a crypto miner, turned entirely to renting out graphics chips for AI, and today it supplies power even to Microsoft itself - the hyperscaler that admits it can't build fast enough on its own. No email, no databases - just clean computing power, sold fast, at a clear price per hour.
Sharon AI is the Australian version of the same idea - a company listed on the NASDAQ exchange that doesn't build from scratch, but installs rows of chips in other people's buildings and rents them out by the hour. It recently signed a deal with NVIDIA (the company that makes the AI chips themselves) for more than 40,000 of its newest processors - a quantity enough to power models for an entire country. No email of its own, no storing of customer photos, just chips lined up in rows.
Firmus attacks the problem from a different angle - heat. AI chips give off so much heat that ordinary air cooling can't keep up, so the company builds modular, liquid-cooled data centers it calls AI factories. Their combined target is 1.6 gigawatts across several sites in Australia - as much as a small city consumes around the clock, aimed entirely at AI computing. NVIDIA came in as an investor in the company, because Firmus sells more of its chips, faster than NVIDIA itself does.