The Hardware Era of Dev: Your Machine Is Now a Small Cluster

For a decade the developer machine was a solved story: enough RAM, a decent CPU, a black editor, coffee. Hardware was furniture. Then the model moved into the room, and furniture became a factory.

Your laptop is now a small cluster whether you asked it to be or not. Local models on Apple silicon. NPUs on the cheap Windows box that suddenly matters. A Mac Studio under the desk doing the jobs you used to send to a vendor. A cloud agent in an isolated VM burning tokens while the fan on your actual computer stays quiet. The "dev environment" is a placement problem: what runs here, what runs there, what never leaves the building.

This is why hardware talk came back into engineering Slack like it was 2012. Memory is context. Bandwidth is how fast the agent can see the repo. The thermal limit is a product constraint — you cannot pair-program with a model that throttles every twelve minutes. Teams that used to fight about tabs now fight about unified memory and whether the on-device 30-something-billion-parameter thing is "good enough" for refactors that must not leak.

Cloud versus local is not a religion. It is a threat model plus a bill. Cloud agents are convenient, fast to upgrade, and a pipe to someone else's GPU. They are also a pipe. Source, secrets, the weird internal name of a customer — if it went into context, it went somewhere. Local inference is slower, fussier, and a way to keep the ugly data in the room. Serious shops mix: local for the sensitive loop, cloud for the long raid, a human for the merge.

The era of dev is physical again in another way. Edge devices, POS terminals, phones on bad African networks, the Windows box behind a counter that cannot "just" call a 400-billion-parameter API. The hardware the user holds still decides whether your clever stack is a product or a press release. Agents that only work on a well-fed laptop in Cape Town are not a strategy. They are a studio.

Buy machines like you buy infrastructure: for the workload you actually run. If your team lives in agents, RAM and local acceleration are not perks. If your product is offline-first, the constraint is the device in the store, not the demo on Wi-Fi. The keyboard got cheaper. The silicon around it did not stop mattering. It started mattering in a new dialect.

There is also the quiet politics of who gets the good machine. If seniors run local 70B-class models and juniors get a browser tab and a prayer, you have not adopted AI. You have created a caste. Budget the inference the way you budget laptops: enough for the people who actually ship, with a path for the people learning to.

On-device is not a vibe. It is latency, privacy, and a fan curve. Cloud is not a villain. It is someone else's cluster with a contract. Draw both. Then decide per workflow, not per tweet.

Treat hardware as part of the architecture diagram. If it is not on the diagram, it will show up in the incident.