Lenovo’s new AI server is not a cloud. It is a box that answers on the plant.

On 6 January 2026, at CES, Lenovo announced a family of machines built for inferencing, not for training. ThinkSystem SR675i. ThinkSystem SR650i. ThinkEdge SE455i. Ashley Gorakhpurwalla, who runs Lenovo’s infrastructure group, said enterprises need AI that turns data into a decision the moment it is created. That is a keynote sentence. In a plant it means: the slip lands, the model reads it, a person still posts it. The box that does the reading sits in your rack, next to the FortiGate, not in someone else’s region.

We are a Lenovo shop. We have been putting ThinkPads and ThinkStations into Mauricie for years. This post is about the new server, the SR675i, and why it is the first Lenovo AI box we would actually discuss with a plant manager instead of a data-center brochure.

01Training versus inferencing, in plant language

Training is how a model learns, once, on a pile of examples, usually in a rented hall of GPUs you will never visit. Inferencing is how a trained model looks at a new packing slip, a new photo of a weld, a new line in Visual, and answers. Lenovo’s January machines are inferencing machines. That is the pivot they are selling: you already spent the capital on AI. Now make it do a job at 10 a.m. on a Tuesday.

The SR675i is the big one. AMD EPYC on the CPU side, NVIDIA Blackwell GPUs, built to run full large language models and to accelerate simulation. Lenovo named manufacturing, healthcare, and finance. The SR650i is the one you slide into a rack you already have: dense GPU, less drama. The SE455i is the edge box, closer to the line, when the closet is 40 metres from receiving and you do not want the image of the part going upstairs first.

02What we would run on it, this year

A packing-slip model that has outgrown the clerk’s PC. Document intelligence that several people share, with permissions, not a demo on a laptop. A shop copilot grounded in your procedures, served to the floor, with a human gate. Simulation that engineering currently sends to a cloud queue and waits on. Overnight jobs: classify yesterday’s PDFs, draft the exceptions, leave them for a person at 7:30. The SR675i is the first time Lenovo has handed us a box whose job description matches the work we already sell.

Neptune cooling is the unglamorous line on the spec. These GPUs dump heat like a small kitchen. Air-cooled is fine until the closet hits 28 degrees in July. Liquid, Lenovo’s Neptune, is how you keep the same box from becoming a noise complaint. We would rather talk about the closet temperature on day one than explain a thermal shutdown during inventory.

03What we would not sell you

A GB300 NVL72 rack-scale monster. That is a cloud in a crate. It is real, Lenovo will sell it, and it does not belong behind the lunch room. A server with no SentinelOne, no backup, and no owner. An inferencing box whose first job is “chat with the company,” with no list of what it may read. A quote that skips the electrical panel. Blackwell GPUs are not a USB stick.

If the work is one engineer and a local model, a ThinkStation with an RTX PRO 6000 is cheaper, quieter to buy, and already certified for the CAD you run. If the work is “keep the PDF in the building” and the people live on Mac, a Mac Studio Ultra may be enough. The SR675i is when several teams share the model all day, the PC under the desk is the bottleneck, and you want a contract, a rack unit, and a phone number that is us.

04A week in the plant, not in a lab

Monday: pick one job that already hurts. Packing slips, AP, or the copilot that keeps asking the unofficial Copilot person. Tuesday: we size SR650i versus SR675i versus a workstation. Honest power, honest noise, honest who owns the keys. Wednesday: the model runs on a loaner or on a GPU we already have, with the human gate written on a sticky note. Thursday: if the answers are useful and boring, we talk rack date. Friday: if they are not, you learned it on a workstation, not on a 40,000 dollar slide.

Lenovo did the right thing in January. They stopped pretending every customer is training the next foundation model, and they shipped boxes that answer. We will still watch the FortiGate, the backup, and the person who clicks post. The server is a very fast clerk. The plant still signs.