For years, the spec conversation was boring, and that was a good thing. A mid-range laptop, 16GB of RAM, a decent SSD. Enough for a browser, a code editor, Slack, and a dozen background processes. Nobody filed a ticket about compute.
That era is over for one specific group: teams building with AI.
The thing that broke the old spec
The shift is not about faster processors. It is about memory.
Running a model locally, whether to prototype, to keep proprietary data off the cloud, or to ship an on-device feature, means the model has to fit into fast memory before anything happens. The practical floor for useful local work is now 16GB, and 32GB is where it starts to feel comfortable once you move past the smallest models. A larger 70-billion-parameter model needs somewhere around 40 to 45GB just to load. If it does not fit, it does not slow down a little. It falls off a cliff, swapping to disk and dragging performance down by an order of magnitude.
So the engineer asking for 64GB is not being precious. They are describing the minimum viable machine for the work you hired them to do.
Why this lands on IT, not just engineering
Here is the operational problem. AI is no longer contained in one team.
A year ago, the people running heavy local workloads were a handful of ML engineers. Now it is backend engineers wiring up AI features, data scientists, product folks testing prompts, even support teams running classification models. The "who needs the powerful machine" question used to have a short, stable answer. Now it changes every month.
That breaks the standard procurement playbook. The default company laptop is suddenly wrong for a growing slice of your team, and you find out when someone is already three weeks into a job they cannot do properly.
What good looks like
The companies handling this well are not just buying more expensive laptops. They are doing three things:
- Spec by role, not by default. Engineers and data teams get the memory they need from day one. Everyone else gets the sensible standard. One catalog, clear tiers, no case-by-case negotiation.
- Keep the refresh cycle honest. AI workloads age hardware faster. A machine that was generous eighteen months ago can be the bottleneck today. The refresh decision has to track what the team actually runs, not a fixed three-year clock.
- Make the upgrade path fast. When someone changes roles or picks up an AI-heavy project, swapping them onto the right machine should take days, not a budget cycle.
The quiet cost of getting it wrong
Underspecced machines do not announce themselves. They show up as slightly slower everything. A model that takes too long to test, so it gets tested less. A developer who context-switches while a process grinds. None of it generates a dramatic incident. It just quietly taxes your most expensive people, every day.
If a meaningful part of your team is building with AI, the laptop is not an admin detail anymore. It is part of whether the work gets done.
That is the part we think about so you do not have to. The right machine, specced for the actual job, in their hands wherever they are. Boring, on purpose.