Nvidia's Jensen Huang says Blackwell GPU to cost $30,000 - $40,000, later clarifies that pricing will vary as they won't sell just the chip
Nvidia's GB200 costs $6,000 to make, according to one analyst.
Nvidia CEO Jensen Huang said in an interview with CNBC that Nvidia plans to sell its Blackwell GPU for AI and HPC workloads for $30,000 to $40,000. However, that is an approximate price since Nvidia is more inclined to sell the whole stack of datacenter building blocks, not just an accelerator itself. Meanwhile, a Raymond James analyst believes it costs around $6,000 to build one B200 accelerator (via @firstadopter).
The performance of Nvidia's Blackwell-based B200 accelerator with 192 GB of HBM3E memory certainly impresses, but those numbers are enabled by a dual-chipset design that packs as many as 204 billion transistors in total (104 billion per die). Nvidia's dual-die GB200 solution with 192 GB of HBM3E will cost significantly more than a single-die GH100 processor with 80GB of memory. The Raymond James analyst estimates that each H100 costs around $3,100, whereas each B200 should cost around $6,000, according to the market observer.
Development of GB200 was also quite an endeavor and Nvidia's spending on its modern GPU architectures and designs tops $10 billion, according to chief executive of the company.
Nvidia's partners used to sell H100 for $30,000 to $40,000 last year when demand for these accelerators was at its peak, and supply was constrained by TSMC's advanced packaging capacities.
It should also be noted that Nvidia's B200 (assuming Jensen Huang talked about B200) is a dual-die solution, whereas Nvidia's H100 is a single-die solution. To that end, it makes sense to compare the price of the B200 to the H100 NVL dual-card product aimed at training large language models. Meanwhile, H100 NVL is not exactly sold in retail, which makes analyzing Nvidia's costs much more complicated.
Another thing to consider about Nvidia's B200 is that the company may not really be inclined to sell B200 modules or cards. It may be much more inclined to sell DGX B200 servers with eight Blackwell GPUs or even DGX B200 SuperPODs with 576 B200 GPUs inside for millions of dollars each.
Indeed, Huang stressed that the company would rather sell supercomputers or DGX B200 SuperPODS with plenty of hardware and software that command premium prices. Therefore, the company does not list B200 cards or modules on its website, only DGB B200 systems and DGX B200 SuperPODs. That said, take the information about the pricing of Nvidia's B200 GPU with a grain of salt despite the incredibly well-connected source.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
There's a budget GeForce GPU selling in China that not even Nvidia knew it made — RTX 4010 turns out to be a modified RTX A400 workstation GPU
US to patch loopholes that allow China to buy banned AI GPUs from other countries — new regulations include national quotas on GPU exports and a global licensing system
-
ezst036
I guess that kind of answers the one question I was having.Admin said:Indeed, Huang stressed that the company would rather sell supercomputers or DGX B200 SuperPODS with plenty of hardware and software that command premium prices. Therefore, the company does not list B200 cards or modules on its website, only DGB B200 systems and DGX B200 SuperPODs. That said, take the information about the pricing of Nvidia's B200 GPU with a grain of salt despite the incredibly well-connected source.
Hopper didn't come to consumers at retail that I know of, and it looks like Blackwell won't either. -
thisisaname
May be Hooper will be the next retail card. New architecture for the Datacentre, then getting handed down to retail when it gets replaced with the new one?ezst036 said:I guess that kind of answers the one question I was having.
Hopper didn't come to consumers at retail that I know of, and it looks like Blackwell won't either. -
thestryker Hopper is very similar to Ada in design as they were done in parallel, so while we'll never see Hopper consumer cards there are Ada enterprise cards. All signs point towards Blackwell being a unified client/enterprise architecture like Ampere was.Reply
As for pricing I assume whatever they say can be taken with a grain of salt like crypto era pricing. There's the price it should cost, and then the price it actually costs to acquire. -
ekio 40k per gpu, when the entire wafer that contains multiple dozen of dies cost like 7k to produce…Reply
This is a prime example of what pure dominant position abuse is lol.
When will they stop? When margins exceed 95 percent ? -
JTWrenn
Pricing your products high is not abuse. That is just competition. If they told customers you can't buy our product unless you sign something saying you will never buy a competitors, that would be abuse. This is just pure pricing based on supply and demand.ekio said:40k per gpu, when the entire wafer that contains multiple dozen of dies cost like 7k to produce…
This is a prime example of what pure dominant position abuse is lol.
When will they stop? When margins exceed 95 percent ? -
Alvar "Miles" Udell It'll cost whatever they want because nobody can compete with them toe to toe, and just like in consumer cards if anyone can even come close they aren't going to compete with them on price, they're going to give a discount so slight as to not be a reason to use them.Reply
Just like in consumer GPUs, if AMD -wanted- to they could force a price war and return GPU prices to normalcy, but instead they price them high and profit as much as possible at the detriment of consumers who are forced to buy one. -
The Historical Fidelity
The cost per wafer is $20,000 for TSMC N3ekio said:40k per gpu, when the entire wafer that contains multiple dozen of dies cost like 7k to produce…
This is a prime example of what pure dominant position abuse is lol.
When will they stop? When margins exceed 95 percent ? -
Stomx
Let's estimate. If each wafer has 10x10 = 100 successful pieces the cost of each is $200. Each B200 has 2 cores so this is $400 total. Sold by $40,000 or factor of 100x more. This is the same sort of numbers like with AMD EPYC/Threadripper with its 8-core chiplets with 6B transistors each which cost $10 per chiplet to produce if estimate the same way but are charged $1000. Margin 99% !! You charge whatever you want if you are leader or monopolyThe Historical Fidelity said:The cost per wafer is $20,000 for TSMC N3