《B300逼近2000萬：DGX同HGX一字之差，差價可以差到百萬》B300 Nears 20 Million: DGX and HGX Are One Letter Apart, and the Price Gap Reaches a Million

[Zheng Ge on Compute]

Prices across the compute market have been climbing hard, and the spot price of the B300 has climbed with them. A single machine now sits just under the 20-million mark.

Over this stretch I have sat with more than a dozen data-centre buyers and investors, and the same mistake keeps coming up: they fix their eyes on "eight B300 cards" and never read the suffix after the model number. DGX or HGX is one letter apart, and the transaction price can open by several hundred thousand, sometimes a million. One careless moment and you have overpaid by a fortune.

One: A Misconception With Deep Roots

The trade has long repeated one line: HGX is the stripped-down contract-built version, and only DGX is a genuine factory machine.

After years in compute hardware delivery I will put it plainly: there is no cut to the GPU core in either one. DGX B300 and HGX B300 both carry eight SXM B300 chips, both come with the NVSwitch switch and the full NVLink interconnect channels. The core hardware that decides large-model training and cluster communication efficiency is identical, and there is no gap in raw compute.

What actually opens the price gap was never the card itself. It is the machine's provenance, the hardware configuration, the compliance quota, the after-sales system and the degree of standardisation.

Two: DGX B300 — Factory End to End, Specification Locked

Start with DGX B300. This is a closed machine: designed under Nvidia's control across the whole chain, delivered through the factory system.

Every hardware specification is locked. It ships with dual Intel Xeon 6776P processors, 2 TB of system memory as standard, factory high-speed NVMe cache drives, and Nvidia's native BMC management system. Before it leaves the factory Nvidia has already finished the NVLink link tuning and the full-machine stress test, and pre-installed the DGX software stack. You can build a large cluster the day you open the box; machine-to-machine consistency is total.

Government and enterprise supercomputing projects, and standardised AI training clusters, will almost always take DGX first.

But DGX's great weakness is supply: it is extremely tight. Export controls, factory allocation quotas, end-user review and end-use filing all bind it at once. An ordinary company struggles to get a stable source, approval thresholds are very high, and goods actually in circulation are few. Scarcity pushes the premium up directly, and that is the core reason DGX prices stay high.

Three: HGX B300 — Not a Cut-Down Model, a Platform

Now HGX B300. It is not a low-spec cut-down build. It is the standardised compute-board platform Nvidia opens to the outside.

Nvidia supplies only the B300 core modules and the high-speed interconnect design. Everything around it — chassis, power, cooling, CPU, memory, network cards — is integrated and customised by proper OEM manufacturers such as Dell and Supermicro.

The hardware configuration can be adjusted freely, and some builds support both air and liquid cooling, so it can be tuned to inference or to training as the business needs. Against the fixed DGX specification, HGX offers more freedom to customise, and it fits the infrastructure a domestic data centre already has.

For the great majority of commercial compute leasing and data-centre cluster projects, HGX is the more practical machine to land. Because an OEM integrates it, the qualification and review path for buying the whole machine differs from a factory DGX, and the bar for customising a build is relatively lower.

Four: Six Checks Before You Buy

Once you understand where the two models sit, the current B300 pricing mess becomes easy to read. DGX quotas are scarce, so it carries a factory compliance premium. HGX supply is ample, so it is better value. Some channels deliberately blur the two and quote an OEM-built HGX at DGX factory prices, pocketing the difference. Plenty of newcomers bow to the DGX factory label, chase the higher price, and hand over several hundred thousand in cost for nothing.

So the checks have to come down to individual items. Six of them:

1. Check the platform model number, and keep DGX and HGX strictly apart.
2. Check the GPU model and count (eight SXM B300), and whether the NVSwitch switch and the NVLink interconnect channels are complete.
3. Confirm the CPU (dual Xeon 6776P), memory and storage configuration.
4. Confirm the cooling design (air / liquid) matches the data centre's own cooling.
5. Settle driver and software-stack compatibility in advance.
6. Nail down the whole-machine warranty and the spare-parts after-sales chain.

I put the second item near the front on purpose: the model number is written on paper, while the GPU count and the interconnect links have to be seen on the machine. Paper is easy to agree with. What is inside the chassis is what you actually bought.

Five: Nothing Is Best, Only Fitting

• DGX fits: large supercomputing, unified standardised clusters, government and enterprise projects with strict factory-compliance requirements
• HGX fits: commercial compute leasing, data-centre retrofits, liquid-cooled custom deployments

Right now a B300 is hard to find at any price, the market keeps climbing, and phantom offers multiply. Rather than comparing prices blindly or fighting for expensive factory stock, lock in goods whose source is clear, whose configuration is transparent and whose after-sales can actually be delivered. That is the choice you do not lose on.

The compute industry is into its second half. What decides it is no longer who can pay the most for stock, but who avoids the purchasing traps, controls supply risk and guarantees delivery.

One letter apart. What differs is not only the money — it is whether you know what you bought.