$300 k capex each, $50 k / yr opex each, replace at year 3.
Nebius 1-yr commit · $122,640 / server / yr · no upfront capex.
BUY is the only HK-on-prem path (only H20 ships there). RENT routes via SG / TYO — faster per dollar, no capex, no replacement cycle. Cloud APIs run on H200 / B200 fleets that aren't legally importable at any price.
DeepSeek V5 drops Friday — pull the weights, spin up vLLM, serve it Monday. No vendor gate, no waiting for the hyperscaler to host it.
Continue-pretrain on the research archive. Build the CLSA-flavoured model that turns generic LLM output into something only we can produce.
Nothing leaves the HK datacentre. Compliance signs off without auditing every API egress — and stays signed off if regulators tighten.
Fixed capex + fixed opex. No surprise $80 k week if an agent loop runs hot. Finance gets a number that doesn't move.
Cloud APIs run on H200 / B200 fleets we cannot legally import to HK. The fastest tokens in the world arrive over HTTPS, not in a datacentre.
No racks, no drivers, no CUDA upgrades, no on-call for the 3 a.m. OOM. A small team can ship Foundry without standing up a GPU ops function.
Idle nights, weekends, public holidays — bought GPUs depreciate whether they run or not. Cloud only charges for the tokens you actually serve.
10× peak at earnings week, 1× the rest of the year. Cloud absorbs the burst without staring at a row of idle servers in the trough.
On cloud we're at the mercy of Azure · Alibaba · Bedrock — price moves, deprecations, region outages. Owning silicon makes us our own provider for the models we serve. But the MCP gateway means switching between the two is a YAML edit — start on cloud, watch the bill, buy later if it tells you to.
May 2026 list · NVIDIA OEM
Nebius leads · hyperscalers 3 – 4 ×
BIS export controls restrict H100 / H200 / B200 from HK + mainland China. Only H20 ships to HK; everything else routes via SG or TW.
Reconnecting…
Network looks slow — attempting to reconnect
Reconnecting…
Server isn't responding — attempting to reconnect