The Host Froze the First Night. Two Days Later, a Private AI That Saves ₽75,500 a Month
The first night with the new card, the host locked up solid. Not a process — the whole physical machine: ping alive, SSH dead, console unresponsive. ₽65,000 for a used Tesla V100 32GB, and within hours it nearly took six running services down with it.
Started digging. Found and fixed five things over two days — from power draw to the finance math.
One: two power cables at once, and that's it
The card runs off two separate connectors. Plug both in at once on a cold boot and the host hangs, every time. The seller had already admitted, unprompted, that he'd seen the same thing. Staggered the connectors, reran the full cycle — cold reboot, both GPUs under load at once, a formal health check. Clean. Never recurred.
Two: latency, in the wrong place
The model answered correctly, but 28 seconds passed before the first word — six times slower than the cloud subscription it was replacing. First guess: a narrow data lane on the board. Checked the numbers instead: almost no correlation with prompt length, near-perfect correlation with how long the model spent reasoning before answering. The latency wasn't hardware. The model was thinking out loud for too long before it started typing.
Three: the reasoning cap worked, halfway
Set a limit — 2048 tokens of reasoning. Worst cases dropped 85% (227 seconds to 34). But across every task, not just the worst ones, typical response time went up 13%. Given an explicit budget, the model started spending all of it instead of sometimes stopping early on its own. Better worst case, slightly worse median — a real tradeoff, not a free win.
Four: memory was still sized for the old host
Added RAM while I was at it — 64GB to 128GB. Found the disk cache was still pinned at its old limit from when the machine had half the memory. The disk was gasping while half the new RAM sat idle. Raised the cache limit — the disk calmed down immediately.
Five: the math checked out too
Card plus RAM: ₽154,000 one-time, on top of ₽284,000 already spent on the machine itself. Against renting comparable capacity: ₽75,500 a month, indefinitely, with nothing to show for it afterward. Payback: 6-7 months.
Same week: Anthropic's CEO publicly asked the whole industry to slow model development down — safety isn't keeping pace with capability. OpenAI paused sign-ups for its most expensive tier the same week — not enough capacity for the new model. Neither is a verdict that cloud AI is collapsing. Both point the same direction: capacity is getting scarce sooner than expected. Memory and VRAM prices are up 355-500% over the year — buying now turned out to be timing, not risk.
| Model score vs. the cloud subscription | 94% and 92% across two runs |
| Worst-case latency | −85% |
| One-time investment | ₽438,000 |
| Savings vs. renting | ₽75,500 / month |
| Payback | 6-7 months |
What I took from it:
→ Hardware fails in specific ways the seller already knows about and doesn't always volunteer — should have asked directly sooner → A limit isn't a free win — it has a real cost on the typical case, worth measuring instead of assuming → Settings sized for old hardware don't update themselves — upgrade one thing, check everything that depended on it → A private AI at home stops being a hobby once the numbers work on their own — and they worked out sooner than planned