That figure came from the First Deputy Minister of Digital Technologies this month. New capacity is due in November.
It's worth sitting with what 90% utilization means for a founder. It means compute is not something you buy; it's something you queue for. It means the cost of one training run is a line item you actually have to think about, at a stage where you should be thinking about customers. It means the cheapest path is renting from abroad and paying in a currency you don't earn in.
Capital is arriving in this market fast. Compute is not arriving at the same speed.
For AI-native companies, that changes what a first check has to include. Money alone doesn't unblock a team that can't get GPU time. What unblocks them is knowing which problems don't need a training run at all, where an existing model plus good product judgment gets you to a working product, and where you genuinely have to build from scratch.
Most early AI companies here don't need frontier compute. They need someone to tell them that before they spend six months and their entire seed round finding out.
That's most of what we do at AIS Ventures. The check is the easy part.
The infrastructure will catch up - November is a start, and 2029 is a real target. But founders are building now on what exists now.
