We used to assume anything below a datacentre budget belonged in the cloud. Then we measured. For teams of five to fifty people, a workstation running a quantised 70B model beats a shared cloud queue on latency, privacy and — more often than not — cost. That measurement rewrote our device roadmap.
The numbers
A W5 workstation with 48 GB of unified memory runs a 70B model at 4-bit quantisation at roughly 12 tokens per second for a single user, and holds three concurrent sessions before the queue visibly forms. The same team on a shared cloud endpoint sees 40-second cold starts during peak hours and pays per token forever.
Break-even for a five-person team lands at roughly eleven months. After that, the workstation is pure savings — and the data never leaves the building, which for legal, medical and finance teams settles the debate before the spreadsheet does.
What changed in our roadmap
Two things. First, memory became the headline specification on every device we ship — bandwidth and capacity first, FLOPS second. Second, we stopped treating the workstation as a consolation prize for teams that could not afford a server. It is a first-class deployment target with the same agent runtime as the cloud.
The best place to run a model is as close to the decision as memory and thermals allow.
Where the cloud still wins
Be clear about the limits: bursty workloads, models above 100B parameters, and teams that need elastic scale for batch jobs still belong in the cloud. Our advice is rarely “one or the other” — it is usually a workstation for daily work and rented capacity for the monthly training run.