A S$150,000 AI Workstation: Are You Paying for the GPU or the Technology Connecting Everything?
Both, and the connecting technology is a bigger share of the value than most quotations make clear. A DGX Station-class NVIDIA GB300 workstation costs around S$150,000 depending on OEM, storage and options. That is the ballpark this series uses: public US pricing of roughly US$93,000 to US$125,000 for most OEMs works out to about S$119,000 to S$160,000 before GST at USD/SGD 1.2774. NVIDIA doesn't publish a cost breakdown, so nobody outside NVIDIA and the OEMs can say how many dollars go to the GPU. What you can see is what the money buys. The Blackwell Ultra GPU sits on a package with 252 GB of HBM3e, is wired to a Grace CPU by a coherent 900 GB/s link, and connects to the outside world through an 800 Gb/s SuperNIC. All of that sits in a liquid-cooled, managed chassis with an OEM warranty. Take the connecting parts away and the same GPU would be far less useful for large models.
What is actually inside the price?
| Part of the system | What it is | Why it matters | Nearest thing you could buy separately |
|---|---|---|---|
| GPU compute | Blackwell Ultra, 20 PFLOPS FP4 sparse (about 15 dense) | The headline figure | None. Not sold separately |
| GPU memory | 252 GB HBM3e at 7.1 TB/s, on the GPU package | Sets token-generation speed and the largest fast model | An RTX PRO 6000 has 96 GB GDDR7 (S$29,100, listed 6 Nov 2026) |
| CPU and CPU memory | 72-core Grace, 496 GB LPDDR5X at 396 GB/s | Holds overflow, data pipelines and services | A server CPU and RAM. Not coherent with a PCIe GPU |
| CPU–GPU link | NVLink-C2C, 900 GB/s total, hardware-coherent | Lets the GPU use all 748 GB directly | None. PCIe Gen5 x16 is about 128 GB/s bidirectional |
| Networking | ConnectX-8 SuperNIC, 2x QSFP112 400 Gb/s, plus 10GbE and BMC ports | Links stations, storage and servers at up to 800 Gb/s | Not listed separately at SourceIT; ask us to source |
| Storage paths | 2x Gen5 M.2 off Grace, 2x Gen6 M.2 off ConnectX-8 | Fast model loading and GPUDirect Storage | Micron 7600 PRO 3.84TB Gen5 U.2 (S$5,183, listed 6 Nov 2026), for servers |
| Power and cooling | 1,600 W budget shared dynamically, closed-loop liquid cooling | Data-centre silicon in an office tower | None for this chip |
| Management and support | BMC (IPMI, Redfish), ECC memory, MIG up to 7 instances, OEM warranty | Treat it as shared infrastructure, not a PC | Varies by OEM |
Source: NVIDIA DGX Station datasheet and documentation; OEM listings on SourceIT (6 Nov 2026). The S$150,000 figure is an assumption based on public US pricing, not a SourceIT price. GB300 systems are price on request.
Why can't you just buy the GPU part?
Because the GPU's value depends on what it's connected to. Take memory first. HBM3e is stacked on the GPU package and runs at 7.1 TB/s. For chatbots and agents, token generation is mostly limited by how fast weights can be read from memory, so this bandwidth matters as much as the tensor cores. Our post Your GPU Has 20 Petaflops of Power. So Why Is Your AI Model Still Slow? works through that bottleneck.
Then the CPU link. NVLink-C2C runs at 450 GB/s each way, about 7 times a PCIe Gen5 x16 slot, and it is hardware-coherent, so the GPU can read Grace's 496 GB as if it were its own (slower) memory. That's what turns 252 GB of fast memory into 748 GB of usable memory. No PCIe-attached card can do the same at any price, which is why NVIDIA doesn't rely on PCIe alone.
What would it cost to approximate this with parts you can buy?
A thought experiment helps. To match 252 GB of fast GPU memory with discrete cards, you'd need three RTX PRO 6000 Blackwell cards: 288 GB of GDDR7 for S$87,300 (listed 6 Nov 2026), before the workstation that hosts them. The cards alone draw up to 1,800 W, against GB300's 1,600 W whole-system budget. You'd also have three separate 96 GB pools joined by PCIe, so any model over 96 GB has to be split across cards and pay for cross-card traffic. No CPU memory would be coherent with them, and there'd be no 800 Gb/s networking built in.
That build has real strengths: three GPUs for parallel jobs, graphics output and x86 software. But it isn't the same machine. Our comparison NVIDIA GB300 vs traditional RTX workstations goes through where each one wins. The thought experiment shows that a large share of what a GB300 buyer pays for is a single, large, fast, coherent memory system, and that comes from the interconnects, not the GPU die alone.
You're not paying for 20 petaflops. You're paying for 20 petaflops that can actually reach your data.
How much of the value is in the networking?
For one standalone workstation, less than the spec sheet suggests. For a team that plans to link systems or pull data from fast storage, quite a lot. The ConnectX-8 gives two 400 Gb/s ports, 100 GB/s each way in total, and its built-in PCIe Gen6 switch feeds two of the M.2 slots. Linking two stations back to back takes two QSFP112 cables; the HP ZGX 0.4 m QSFP112 400G DAC is S$365 each (listed 6 Nov 2026). Connecting several needs a 400G switch. The MikroTik CRS804 has four 400G ports for S$2,255 (listed 6 Nov 2026), which suits a small lab but is not a cluster fabric.
If you never plan to use those ports, you're paying for capacity you won't use. Our posts on 400 Gbps networking on a desk and DGX Station's networking help decide. At data-centre scale, networking becomes a much larger line item; see the hidden cost of AI networking.
What about the enterprise features?
These are easy to overlook on a quotation and expensive to add later. They include a BMC for remote power control and monitoring, ECC on both memory pools, MIG to split the GPU into as many as seven isolated instances, closed-loop liquid cooling built to the OEM's design, and the OEM's warranty and support terms. Those terms differ. HP's ZGX Fury G1n listing states 3-year onsite cover, Supermicro's states 3-year next-business-day onsite, and Dell's states 3-year ProSupport. Other OEMs confirm warranty at quotation. NVIDIA AI Enterprise software is a separate licence. Our guides on whether the cheaper quotation tells the whole story, liquid cooling maintenance and who manages the system cover the ownership side.
When is the connecting technology not worth paying for?
- Your models fit in 96 GB: a workstation with one or two RTX PRO cards may give better value and adds graphics and x86 compatibility.
- You run many small, independent jobs: several discrete GPUs can beat one large GPU for throughput per dollar.
- You need production scale: if the plan is a rack of GPUs, budget for an NVLink server or rack system rather than several workstations. See Can NVIDIA GB300 replace your multi-GPU AI server?
- Your usage is low: renting cloud GPUs can be cheaper; see our three-year ROI comparison and productivity ROI model.
Buying in Singapore
SourceIT lists six GB300 workstations, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01, Supermicro ARS-511GD-NB-LCC and Dell Pro Max with GB300. Accessories mentioned above: the HP ZGX QSFP112 DAC, MikroTik CRS804, Micron 7600 PRO 3.84TB and RTX PRO 6000 Blackwell. Our DGX Station GB300 buyer's guide compares the OEMs. When you request quotations, ask each OEM to itemise storage, display card, warranty and software so you can compare like with like. 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
How much does a GB300 workstation cost in Singapore? Around S$150,000 is a reasonable ballpark, based on public US pricing of about US$93,000 to US$125,000 for most OEMs. Actual prices depend on the OEM, storage, display card and warranty, so ask for a quotation.
How much of the price is the GPU? NVIDIA and the OEMs don't publish a breakdown. The GPU, HBM3e and Grace CPU come as one superchip, so the GPU can't be priced on its own.
Why not buy three RTX PRO 6000 cards instead? Three cards give 288 GB of GPU memory for S$87,300 (listed 6 Nov 2026) before the host system, but as three 96 GB pools joined by PCIe. That suits parallel jobs and graphics better than one large model.
Do I need the 800 Gb/s networking? Only if you plan to link systems or use fast network storage. A single workstation used on its own needs little more than its 10GbE port.
Is NVIDIA AI Enterprise included? No. NVIDIA describes it as a separate software licence. The system ships with Ubuntu and NVIDIA's AI developer tools.
