The RTX PRO 6000 is the card we drop in a Lenovo when the model has to live on the desk.

NVIDIA’s RTX PRO 6000 Blackwell is the workstation card people mean in 2026 when they say “a 6000”. It is not the Ada 6000 from two years ago. It is not a GeForce 5090 with a different sticker. It is 96 GB of GDDR7 with ECC, about 24,000 CUDA cores, fifth-generation Tensor cores, up to 600 watts, and a feature called MIG that can split the card into four smaller cards so two engineers do not trip over each other.

We put these in Lenovo ThinkStations. That is the sentence. The rest of this post is when that sentence is a good idea, and when you should buy a quiet Mac or a rack instead.

01What changed from the Ada 6000

The last generation, RTX 6000 Ada, had 48 GB. Fine for a lot of CAD. Tight for a 70 billion parameter model, or a document job that wants the whole PDF in GPU memory at once. Blackwell doubles the memory, moves to GDDR7, and roughly doubles training speed on the workloads NVIDIA’s own customers published. GE Healthcare said 2.5 times faster than Ada for their AI training. You do not have to be GE. You do have to notice that 48 GB was the ceiling a lot of plants hit without knowing the name of the ceiling.

MIG, multi-instance GPU, is the grown-up feature. Up to four isolated slices, each with its own memory and a quality-of-service promise. One card, four people, or one card running a small shop copilot and a CAD viewport without one killing the other. Ada could not do that in a way we would sell to a plant manager.

02What we actually run on it

Local models that have to stay on Windows, because the rest of the plant is Windows. A document intelligence job that turns a packing slip into a row, on the same PC the clerk already sits at. SolidWorks, Inventor, simulation, a render that used to go overnight. Fine-tuning a small plant model on last year’s tickets, then serving it to three people. NVIDIA AI Enterprise, CUDA, the stack the ISV already certified. That last line is why this card beats a Mac for a lot of our shops. The software they paid for already knows NVIDIA.

The Max-Q version exists for a reason. The full workstation edition is a 600 watt guest. Lenovo also ships a Max-Q 6000 for denser boxes, up to four cards, when you want more than one GPU in a tower without melting the carpet. We would rather one honest 600 watt card in a ThinkStation that has the right PSU, than two bargain GeForce cards with no ECC and a warranty that ends at the gaming forum.

03What it is not

It is not a rack. A ThinkSystem SR675i holds more GPUs, shares them, and has a contract. The 6000 is one person’s desk, or a small shared workstation in engineering. It is not a 5090. The 5090 is faster at some games and has less memory than you think once the model and the viewport both sit down. It is not silent. 600 watts is a hair dryer with a driver. If the office is a glass box next to accounting, ask about the Max-Q and about putting the box in the IT closet with a long Thunderbolt or a thin client, instead of under the desk.

Price: the card itself has been selling in the neighbourhood of 8,000 to 10,000 US dollars, before the ThinkStation, the 1600 watt PSU, and the three year on-site. Anyone quoting “an AI PC” at 2,000 dollars is quoting a laptop with a sticker. Anyone quoting a four-GPU tower without measuring the circuit in the room is quoting a fire.

04A decision that fits on one page

If the work is CAD plus a local model, and the people already live in Windows: ThinkStation, RTX PRO 6000, ECC, MIG if two people share it. If the work is “keep the PDF in the parish” and the people already live on Mac: Mac Studio Ultra, maybe two, not four. If several teams will share one model all day: a ThinkSystem inferencing server, not a GPU under a desk. If you just need Copilot in Word: you do not need this card. You need SharePoint that is not a junk drawer.

We are a Lenovo shop. We will size the PSU, the warranty, the SentinelOne agent, and the human gate on whatever you run on the GPU. The card is the easy part. The boring part is why plants still call us.