Posted in AI Security
At Cyberr, we’re already experimenting with a hybrid AI setup: flagship models for demanding tasks and our…
Gheorghe B. — Senior Software Engineer
At Cyberr, we’re already experimenting with a hybrid AI setup: flagship models for demanding tasks and our own RTX 5090 GPUs for internal workloads.
Now NVIDIA is bringing Blackwell Ultra to Windows with a $95,000 AI workstation capable of running massive open-weight models locally.
Sounds expensive, until you look at what companies are spending on AI inference.
Revolut reportedly cut its AI coding inference costs by around 90% by self-hosting Chinese open-weight models instead of relying primarily on proprietary APIs.
Not every task needs a frontier model, and not every token needs an API bill.
With 748GB of coherent memory, NVIDIA’s new DGX Station makes the same approach increasingly interesting for smaller companies, too.
Of course, model quality, concurrency, electricity and maintenance still matter. But if you’re already spending $10–20K a month on AI, the maths starts looking pretty different.
At what point does owning your AI compute make more sense than renting every token?
https://www.nvidia.com/en-us/products/workstations/dgx-station-for-windows/
6 reactions · 4 comments
Comments
Victor Font: Gheorghe, I think the hybrid approach is where this conversation becomes particularly interesting. The economics are compelling, but I'd evaluate the decision on more than inference cost alone. Workload predictability, utilization, model quality, data sensitivity, operational support, and the cost of maintaining the infrastructure all influence whether owning compute makes financial sense. There's
Gheorghe B.: Victor Font Exactly, and I think utilization is probably the biggest deciding factor here. A $95K machine sitting idle half the day can quickly make those API bills look reasonable again 😄
Victor Font: Gheorghe Botnaru Gheorghe, exactly. And utilization introduces another consideration: capacity planning. An organization might justify the investment based on today's inference volume, only to discover that workload patterns, model requirements, or hardware capabilities change before the equipment is fully depreciated. Cloud inference converts much of that uncertainty into variable operating expen
Ilias Mavropoulos: Only thing I’d like to research/plan really carefully is future proofing, because owning and maintaining a setup that does the work today doesn’t mean it will keep doing it for the next 5 years. Looking at the way big AI corps are racing I would try to lay out financial planning considering cost for relatively frequent hardware updates.