From RTX PRO 6000 to GB300 NVL72: When Does Your AI Workstation Need to Become an AI Factory?

Your AI workstation needs to become something bigger when your largest workload no longer fits the memory and links of the machine under your desk. That usually means a model too large to split cleanly across a few PCIe cards, more users than one box can serve, or training runs that need many GPUs working as one. For most teams the path has clear steps. One NVIDIA RTX PRO 6000 Blackwell (96 GB) handles development and models up to about 70B parameters at 8-bit. Two to four Max-Q cards add capacity but split memory over PCIe. A DGX Station GB300 puts 748 GB of coherent memory behind one GPU at your desk. Beyond that you're into rack systems: RTX PRO Servers for shared inference and visual computing, HGX B300 for NVLink-connected training, and GB300 NVL72, where 72 GPUs act as one NVLink domain.

What does each step on the ladder add?

Source: NVIDIA RTX PRO 6000, DGX Station, HGX and GB300 NVL72 pages (fetched 6 Nov 2026); OEM workstation specifications (HP, Dell, Lenovo); ServeTheHome (5 May 2025). DGX Station figure is this series' ballpark: public US pricing of roughly US$93,000–125,000 for most OEMs is about S$119,000–160,000 before GST at USD/SGD 1.279 (6 Nov 2026). Prices change; ask for a current quotation.

NVIDIA doesn't draw this whole ladder on one page. We've assembled it from its product pages and its April 2026 Enterprise Reference Architectures, which describe three data-centre tiers: an RTX PRO AI Factory, an HGX AI Factory and an NVL72 AI Factory.

What tells you one RTX PRO 6000 is no longer enough?

• The model plus its context won't fit in 96 GB. Llama 3.3 70B at BF16 needs about 141 GB of weights, so two cards. The same model at FP8 fits one card, but its KV cache at 128K tokens of context needs about 43 GB more, which won't sit beside 71 GB of weights.
• Too many people are waiting. One card shared through MIG gives up to four isolated 24 GB slices, which suits several small models but not several large ones.
• Single-user speed is too slow. Token generation is limited by memory bandwidth. At 1,792 GB/s, the ceiling for a 70B model at FP8 is about 25 tokens per second for one user.
• You're training, not just running, models. Full fine-tuning keeps optimiser state alongside weights and quickly outgrows one card.

Our posts on who needs 96 GB and whether more memory is always better go deeper into sizing.

Should the next step be more cards or a DGX Station?

It depends on whether your work splits. Two to four Max-Q cards suit teams that run several models or users at once, need x86 software, or want to grow one card at a time. Their memory doesn't pool, though: a model bigger than 96 GB has to be split, and the cards synchronise over PCIe. Our post on multi-GPU without NVLink covers where that hurts.

A DGX Station suits teams with one large model. Its single Blackwell Ultra GPU holds 252 GB at about 7.1 TB/s, and on the one like-for-like test available, StorageReview measured 9,280 output tokens per second on gpt-oss-120b at 128 concurrent requests on an ASUS ExpertCenter Pro ET900N G3, against 3,064 on one RTX PRO 6000. It also uses the same Grace Blackwell architecture and the same 900 GB/s NVLink-C2C link as each node in a GB300 NVL72. NVIDIA describes Station workloads "scaling seamlessly to GB300 in the data center or cloud" (31 May 2026). The catch is the Arm-based Grace CPU, so check that your x86-only tools have Arm or Linux alternatives. RTX PRO 6000 vs DGX Station GB300 and two RTX PRO 6000 cards vs one DGX Station compare them in detail. A Station can also take one RTX PRO card for displays and visualisation; a 300 W Max-Q keeps the GB300's power cap at 1,250 W, against 950 W with a 600 W Workstation Edition.

Most teams don't need an AI factory. They need to know which step their largest workload is standing on.

When does it become a rack, and which rack?

• RTX PRO Server: NVIDIA positions it for small to medium model inference, fine-tuning, generative AI, visual computing and industrial AI. Its reference architecture covers clusters "from tens to hundreds of GPUs". It uses air-cooled racks and PCIe cards, and the Server Edition supports vGPU for shared virtual workstations.
• HGX B300: eight Blackwell Ultra GPUs on NVLink in one server, for training and high-throughput inference. This is the step when tensor parallelism across eight GPUs matters. One published test found NVLink-connected H100 and H200 systems 3 to 4 times faster than eight PCIe RTX PRO 6000 cards at eight-way tensor parallelism.
• GB300 NVL72: 72 GPUs that NVIDIA describes as "a single, coherent compute domain", with 1.44 exaFLOPS of sparse FP4. It's for frontier-scale training and serving very large models to many users. See why NVL72 needs NVLink and scaling from DGX Station to NVL72.

The facility changes too. OEMs quote about 132–142 kW for an NVL72 rack, and it needs liquid cooling, so in Singapore it means a data centre. CBRE put Singapore colocation at US$330–475 per kW per month (reported by W.Media, 15 July 2026). At 142 kW that's roughly US$46,900–67,500 a month for power capacity alone. That's an illustrative figure: real contracts vary with size and terms.

How do you keep your work portable as you scale?

• Stay on the CUDA stack and containers. NVIDIA's pitch for its desktop systems is to prototype locally and move the same software to the data centre. NGC containers are built for both x86 and Arm.
• Expect parallelism settings to change. A model on one card needs no splitting; on four PCIe cards it needs tensor or pipeline parallelism; on HGX or NVL72 the best split is different again.
• Watch kernel support. Workstation Blackwell (the RTX PRO 6000, compute capability SM120) and data-centre Blackwell are different GPU designs. In April 2026, vLLM users reported the native NVFP4 mixture-of-experts path broken on the RTX PRO 6000's SM120 architecture, so a format that runs well in the data centre may need a different kernel on the desk.
• Plan for x86 to Arm if your step goes through a DGX Station or NVL72, since both use Grace CPUs.
• Rent before you buy. Several cloud providers offer RTX PRO 6000 instances, listed at about US$1.69 to US$5.50 per GPU-hour on 3 November 2026 in Thunder Compute's comparison, and renting is a cheap way to test the next step first.

For desk-side economics, see DGX Station vs an RTX PRO 6000 workstation build and four RTX PRO 6000 cards, 384 GB.

Buying in Singapore

SourceIT covers the desk-side steps. The RTX PRO 6000 Blackwell Workstation Edition (S$29,100, listed 6 Nov 2026) is here and the Max-Q (S$29,100) here. DGX Station GB300 systems are price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01, Supermicro ARS-511GD and Dell Pro Max with GB300. For servers, the HPE ProLiant DL385 Gen11 lists the RTX PRO 6000 Server Edition among its GPU options and is also price on request; more are in our AI workstations and AI servers collection. HGX and NVL72 systems aren't listed, so ask us to source them. Tell us your largest model and user count, and we'll suggest the right step. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

What is the next step after one RTX PRO 6000? Either two to four Max-Q cards, if your work splits into several models or users, or a DGX Station GB300, if one large model needs more than 96 GB in a single GPU memory space.

Does code written on an RTX PRO 6000 run on GB300 NVL72? CUDA software and NGC containers carry over, but expect to change parallelism settings and, for some formats, kernels. The RTX PRO 6000 uses a different Blackwell architecture from data-centre GPUs, and NVL72 uses Arm-based Grace CPUs.

How much power does a GB300 NVL72 rack need? NVIDIA's reference figures give up to 142 kW per rack, fully liquid-cooled. That normally means a data centre rather than an office server room.

Is an RTX PRO Server the same as an HGX server? No. An RTX PRO Server holds up to eight PCIe RTX PRO 6000 Server Edition cards without NVLink. HGX B300 links eight GPUs with NVLink at 1.8 TB/s each, which suits training and large tensor-parallel jobs.

Does SourceIT sell GB300 NVL72 racks? They aren't listed on our store. We list RTX PRO cards, DGX Station GB300 systems and GPU servers, and we can help source rack-scale systems on request.