Desk-side AI hardware has a new top tier. ServeTheHome published a long hands-on review of the Gigabyte W775-V10-L01, a tower that packs NVIDIA's full GB300 into a case you can roll under a desk. It is liquid-cooled, draws up to 1.6 kilowatts, and costs about a hundred thousand dollars, according to the review.
This post explains what the Gigabyte W775 GB300 workstation is, who should consider one, and how it stacks up against the DGX Spark, the Mac Studio M5 Ultra and an RTX PRO tower. One disclosure up front: Gigabyte loaned the unit to ServeTheHome and the outlet labels the review sponsored. The numbers below are the outlet's measurements, attributed as such. We have not touched the machine ourselves.
TL;DR: the W775 in one table
| Question | Short answer |
|---|---|
| What is it? | Gigabyte's build of NVIDIA's DGX Station GB300 |
| Chips | 72-core Grace CPU (Arm Neoverse V2), Blackwell Ultra (B300) GPU, ConnectX-8 SuperNIC |
| Memory | 496GB LPDDR5X at 396GB/s for the CPU, 252GB HBM3E at 7.1TB/s for the GPU |
| Networking | Two 400Gbps QSFP112 ports, 10GbE, BMC port |
| Power | 1,600W PSU, effectively a dedicated 20A circuit |
| Size and weight | 245 x 500 x 531mm, about 29kg (64 lb), two-person lift |
| Price | About a hundred thousand dollars, per ServeTheHome |
| Best at | Many concurrent agents, large open models, GB300 development |
| Weak at | FP64 science (1.3 TFLOPS), video output, cheap experimentation |
What is the Gigabyte W775-V10-L01?
NVIDIA defines the DGX Station platform and its partners build it. ServeTheHome explains that NVIDIA supplies the motherboard with the processors pre-installed, so every DGX Station system shares the same built-in features and rear ports. Partners such as Gigabyte get room to choose the case and cooling details, and little else. The W775 is Gigabyte's take, in a standard full-tower case with what the outlet calls minimal customization.
Inside sits a GB300-style "superchip": a Grace CPU, a Blackwell Ultra GPU and a ConnectX-8 SuperNIC. ServeTheHome stresses that this is a different shape from the server version. A server GB300 module carries two Blackwell Ultra GPUs. The desk version carries one, in what the outlet calls a slightly detuned B300 with seven HBM3E stacks enabled, for 252GB of GPU memory and 7.1TB/s of bandwidth.
The reviewer's key framing: "This is not a slice of the Grace Blackwell ecosystem like GB10-based Spark boxes, but it is the complete and unrestrained Grace Blackwell experience."
How does a 1.6kW tower stay quiet?
Liquid cooling is mandatory. ServeTheHome says NVIDIA requires it for all DGX Station systems because of the heat from the GPU, CPU and SuperNIC. The loop runs through cold plates on all three chips and then to two 360mm radiators, one on top and one on the side. A pump and reservoir carry a CoolIT label. A removable "fan wall" supplies intake air, and you must pull it when installing a long, open-air RTX PRO Blackwell card.
Gigabyte also added something the reviewer had not seen in a desktop before: leak detection. Sensor strips sit below the cold plates and at the front below the distribution hub and pump. A leak triggers a shutdown before damage spreads.
On noise, the outlet measured 36 to 38 dBA at lower loads and 44 to 46 dBA at full load. It calls that impressive for a 1.6kW machine and says Gigabyte made it deployable on a working floor. The case is still big and heavy, at 29kg. Gigabyte advises two people to lift it and shows where to place your hands, since the plastic front panel cannot bear the weight.
The power supply is the practical limit. ServeTheHome says NVIDIA is at the limit for one North American outlet with a 1,600W supply, which needs a high-capacity C19 cable and, in effect, a dedicated 20-amp circuit. Idle power did not fall below about 250W in the outlet's testing.
How fast is it on local models?
Here the numbers are striking, but they are aggregate throughput, not single-user speed. ServeTheHome reports:
- Qwen3.6-35B-A3B Flash in NVFP4: well over 26,000 output tokens per second at 512 concurrent users on short prompts, which it converts to about 2.25 billion tokens a day.
- Nemotron 3 Super 120B-A12B: around 7,000 tokens per second at the same concurrency, and about 86.5 million tokens a day at "very reasonable" concurrency levels.
- GLM-5.3 Flash (321B MoE): run locally at speeds the reviewer says often beat cloud providers, who were serving it at 20 to 40 tokens per second on OpenRouter while the article was being written.
- DeepSeek V4 Flash (284B A13B): slightly better throughput and lower latency than GLM-5.3 Flash in the outlet's runs.
The agentic coding test is the most useful for builders. The reviewer asked both GLM-5.3 and DeepSeek V4 to build a browser kart-racing game with no existing artwork. DeepSeek ran about 20 percent faster in tokens per second, around 120 versus 100. Yet GLM-5.3 finished the main coding tasks in roughly 2.1 hours versus about 4.7 hours. Faster tokens did not mean faster completion. That is a reminder to measure time to a finished task, not raw speed, and it echoes how we frame tests in our agent skills guide.
Two cautions. A figure like 26,000 tokens per second requires 512 concurrent streams and short prompts. A single developer chatting with a model will see far lower per-user speeds. And the outlet's tests are vendor-sponsored, so repeat the ones that matter on your own workload.
What is unusual about the networking?
ServeTheHome spends a lot of the review on this, and it is the part most buyers will miss. The ConnectX-8 chip acts as both a 400Gbps-per-port NIC and a PCIe Gen6 switch. The outlet found that both the CPU and the GPU connect to it as host devices, which Gigabyte's block diagram does not make clear.
- The GPU link is PCIe Gen6 x16, about 968Gbps on paper. With two 400G optics, the outlet measured 784.3Gbps of payload between Blackwell Ultra GPUs.
- The CPU link is only PCIe Gen5 x8, about 252Gbps on paper. The outlet measured 228.9Gbps, matching NVIDIA documentation.
So the GPU can send traffic out of the box at near full line rate, while the CPU is the slower path. The reviewer reads this as proof that NVIDIA designed the box to scale out, so several of them can work as a small cluster over the network.
The CPU and GPU also connect by a 900GB/s NVLink-C2C link. In a quick test, the outlet found that using LPDDR5X over that link to feed the GPU beat running on the CPU alone. It sees that as a way to expand memory for very large models, while noting that HBM is the right place for the decode phase.
Who is this for?
The honest answer is a narrow group. ServeTheHome calls it "a high-end developer workstation, and one of many that may serve the local AI needs of a workgroup or office." We see four fits.
- A team running many agents. Monitoring, triage and heartbeat agents are small models run at high concurrency. The throughput numbers suit that shape of work.
- Companies with data they cannot send out. Legal, health and defense groups that need frontier-class open models on premises.
- GB300 developers. Anyone writing kernels or inference code that will later run on GB300 servers gets matching hardware at the desk, with a BMC and OpenBMC management like a server.
- Small inference-serving shops. Sharing one box across a team is cheaper than renting equivalent capacity for a steady workload, though that depends on utilization and is for your own spreadsheet.
It is a poor fit for a solo hobbyist, for heavy FP64 science (the outlet lists just 1.3 TFLOPS), or for anyone who needs a normal graphics output. The Blackwell Ultra has no display output, so the only video comes from the BMC unless you add an RTX PRO Blackwell card.
How does it compare with DGX Spark, Mac Studio and RTX PRO towers?
We keep this comparison to explainx.ai's own coverage for prices and specs outside the review.
| Gigabyte W775 (GB300) | DGX Spark (GB10) | Mac Studio M5 Ultra | RTX PRO tower (2-4 GPUs) | |
|---|---|---|---|---|
| Class | Workgroup deskside | Personal desktop | Personal or small team | Workstation |
| Fast memory | 252GB HBM3E, 7.1TB/s | 128GB unified | Up to 512GB unified, 1.2TB/s | Per-card VRAM |
| Slow or CPU memory | 496GB LPDDR5X, 396GB/s | Shared pool | Same pool | System RAM |
| Network | 2 x 400Gbps | 200Gbps class | Thunderbolt 5 | Varies |
| Power | Up to 1.6kW | Low | Low | High |
| Rough price | About $100,000 (STH) | About $4,679 to $4,999 | About $18,299 at 256GB (HN reports) | Varies |
| Strength | Throughput, scale-out, GB300 parity | Entry to CUDA at low cost | Largest single memory pool | Familiar workstation |
Our DGX Spark guide and the Spark 64GB pricing post cover the small end. The Mac Studio M5 coverage covers Apple's side: more single-box memory, less CUDA and less bandwidth. For laptop-class decisions see MacBook vs dedicated GPU for local LLMs.
The outlet's own view of the RTX PRO comparison is worth noting. Many readers will think of the W775 as a scale-up from a workstation with two to four RTX PRO 6000 Blackwell cards. The reviewer says that is only part of the story, because the Grace CPU with large LPDDR5X and the 800Gbps of networking make it a different kind of machine.
What are the catches?
- Price. About $100,000, per the review. Memory costs are part of that, and they are rising, as we explained in RAM prices and local AI builds.
- Locked-down memory. The LPDDR5X sits in removable SOCAMM modules, but they are not user-serviceable, and Gigabyte gives no instructions to reach them.
- Limited ports. The top panel has two 5Gbps USB-A ports and one 10Gbps USB-C. Anything faster goes over the 400Gbps network.
- Power draw. A dedicated 20A circuit and a floor location. No front I/O.
- No integrated graphics. Plan to add an RTX PRO card for video.
- Sponsored review. Weigh the tone accordingly.
- Supply. HBM3E and LPDDR5X sit in the same constrained market we covered in Samsung's HBM4 yield. Lead times are unknown from the review.
Should you buy one, or rent GB300 capacity?
Run the math on utilization. A machine this size pays off when it stays busy: a team of agents running through the working day, or an inference service for a department. If your load is bursty, renting cloud GPUs or using a Spark for development will likely cost less. If your data cannot leave the premises, the choice is made for you.
A good test before you commit: take your real agent workload, set the concurrency you expect, and compare tokens per task and time to completion across a rented GB300 instance and a cheaper box. Our workshops include sessions on sizing local AI hardware for teams.
What we could not confirm
We read the ServeTheHome review across its pages and did not get hands-on time with the hardware. We did not find an official NVIDIA price for DGX Station systems or a Gigabyte price list, so the "hundred thousand dollars" figure is ServeTheHome's description. We did not verify the Mac Studio price beyond what our earlier coverage cites from social and Hacker News reports. We could not find a Hacker News thread with comments on this review at the time of writing.
Specs, benchmarks and prices were accurate as of October 9, 2026, and vary with configuration and supply.
Related reading
- NVIDIA DGX Spark: the best setup for local LLMs in 2026
- DGX Spark 64GB: what $4,999 actually buys
- Mac Studio M5 Max and M5 Ultra for local AI
- MacBook vs dedicated GPU for local LLMs
- RAM prices and local AI builds
- Nvidia's first US-made GB300 chips
- AI chip supply chain this week
- Source: ServeTheHome: Gigabyte W775-V10-L01 hands-on