Hardware failure. Cloud providers have SLAs. Your server closet does not.
Noise. Two RTX PRO 6000 Blackwells under full load exceed 50 dB — a loud dishwasher, sustained, all day. In a dedicated server room, fine. In a shared office, your colleagues will have opinions.
Availability. The RTX PRO 6000 Blackwell is a new, high-demand professional card with constrained supply and multi-week lead times. If one card fails, you are not buying a replacement over the weekend. You wait — potentially a month or more. Keeping a spare sounds prudent; that spare costs another ~8,000ドル and is equally hard to source. A single-point-of-failure setup with no redundancy and a six-week replacement window is not infrastructure. It is optimism.
Where the Argument Has a Point
Data sovereignty is real. GDPR compliance for third-country data transfers is genuinely complex, vendor terms change, and strategic dependence on external model providers is a risk that tends to get underweighted until it isn't. The upfront capital requirement is the actual barrier for most teams, not the long-run economics.
But the most important question gets skipped entirely: is the local model actually as good? Two Blackwells with 192GB VRAM can run serious open-weight models — this is not a toy setup. But if developers need two or three attempts to get what a frontier cloud model produces in one, the labour savings evaporate and the break-even never arrives.
The Bottom Line
Local AI infrastructure can make sense — for teams with heavy, sensitive workloads, strong in-house ops capability, and the capital to do it properly, including redundancy, cooling, and the realistic assumption that hardware will occasionally fail at inconvenient times.
What it is not is a simple 18-month arbitrage available to anyone with a GPU and a spreadsheet.
The sovereignty argument is the strongest card in the deck. Lead with that. The cost argument needs a lot more columns in the spreadsheet before it holds up.