From London to New York: Three themes that dominated every room
Three themes came up in every room on both sides of the Atlantic: control of data, the cost of inference, and who you are actually trusting.
Over the last two months we ran Options Technology's Modern Cloud Road Shows in London, Chicago, Toronto, and New York. Our clients were in the room. Three themes came up every time.
Breadth
The clients in these rooms came with problems that cross boundaries. Corporate IT teams need SSO and access guardrails to safely roll out frontier models. Engineering teams need GPU compute farms to train their own models. Organisations that need sovereign AI need the PrivateMind platform. Problems do not split by vendor, and clients do not want them to. They want one partner who spans the whole stack from IT infrastructure to sovereign inference.
Cost
The second theme in every room was cost, but not that cheaper is better. It was the recognition that token-based pricing is a variable cost with no ceiling.
A team doing one billion tokens a month through a hosted API spends between $15,000 and $30,000 on inference alone. At ten billion tokens the number is $150,000 to $300,000 per month. That is before model escalation, context window overhead, or the weekend queue nobody budgeted for.
More importantly, that pricing changes when the provider says it changes. When a model is retired and its replacement costs twice as much, your budget doubles on their schedule, not yours.
Sovereign inference flips the structure. Options provides the hardware, the power, the cooling, and the team that operates it. The cost is known and predictable, with no per-token meter and no vendor deciding next quarter's rate. PrivateMind serves open source models through a gateway the client runs. At sustained scale the math is straightforward, and increasingly clients in the room have already run it.
Who sees your data
The question in every room: who else sees the prompt and the completion? For client positions, patient records, legal documents, the answer must be "no one."
The terms of service for hosted inference products do not offer that answer. Data retention policies, training opt-outs, and processing agreements all say the same thing: your data leaves your control and is governed by someone else's rules.
PrivateMind is built on the opposite assumption. The platform runs open weight models on GPUs the client operates. Embeddings live in the client's own database. Logs stay on the client's own storage. There is no third party in the data path, and that property can be verified with a firewall rule, not a contract.
What is next
Hong Kong and Tokyo are next. The conversation there is sharpened by June's news: JPMorgan and Goldman cut off their HK staff from Claude, forced by licensing terms they did not control. The US government took down Anthropic's Fable 5 and Mythos. OpenAI's GPT-5.6 is rolling out only to government-approved partners.
These are not hypotheticals. A model you depend on can be turned off by a contract clause, an export rule, or an administration you do not elect. Sovereign AI does not solve geopolitics, but it removes your inference from that list. The models keep running because you own the hardware.
The organisations change. The problem does not. AI is now permanent infrastructure, and the cost of depending on someone else's is becoming visible.