This article is an excerpt from Barron's Tech's free weekly email newsletter: Sign up here to get it delivered straight to your inbox.
Blackwell LampEarlier this year, Tesla's Andrei Karpathy
of
The former AI director and co-founder of OpenAI posted on social media that “money doesn't buy happiness.” But Nvidia
of
AI chips? That's another story.
By now, most people know about Nvidia's H100 AI graphics processing unit, which has become the hottest commodity in the tech world: The GPU helped drive data center revenue for the company to $22.6 billion in the most recent quarter, up from $4.3 billion in the same period last year — and this came despite a sales slump for much of the past year.
While the H100 was a breakthrough, the latest indications and research from Wall Street suggest that the company's updated GPUs, based on Nvidia's new Blackwell architecture, will further satisfy AI scientists and provide an even bigger boost to Nvidia's sales.
In March, Nvidia CEO Jensen Huang sounded bright about the new chip, which will contain 208 billion transistors and cost $10 billion to develop: “Blackwell will be our most successful product launch in history,” he said during a keynote at the company's GTC developers conference in San Jose.
Advertisement – Scroll to continue
At the conference, he unveiled various Blackwell products, including the B100, B200, GB200 superchips, HGX B200 server boards and the GB200 NVL72 server rack system, which are expected to be released in stages this year.
But the star of the show is the GB200 NVL72 AI server system. The NVL72 integrates 36 GB200 superchips, and each GB200 connects two Blackwell GPUs to an Nvidia Grace CPU to allow them to work together more efficiently. This means that each NVL72 system has 72 Blackwell GPUs all linked together, delivering unprecedented density of computing power.
This also means prices for Nvidia's top-end AI server systems are rising: Last year, Nvidia priced a two-system package with 16 H100 GPUs for $400,000. KeyBanc says the GB200 NVL72 is expected to cost $3.8 million, a far cry from the $32 per video game graphics chip Nvidia received about 25 years ago.
Advertisement – Scroll to continue
Higher pricing seems unlikely to hurt demand: KeyBanc analyst John Vinh predicts that NVL72 will account for 60% to 70% of Nvidia's GB200 server rack volume compared to cheaper configurations. “Feedback this quarter indicates demand for GB200 next year will be greater than what we initially heard last quarter, potentially driving more than $200 billion in data center revenue by 2025,” he wrote.
It would be an astounding feat if Nvidia's data-center division could achieve roughly $200 billion in revenue next year, up from $48 billion last year and above the $140 billion Wall Street is currently expecting. That would also put it on par with Apple's iPhone revenue, which Wall Street estimates at $210 billion in calendar 2025.
Customers are clamoring for NVL72 as it is much more efficient than previous models, helping companies save on overall costs of training and querying AI models.
Advertisement – Scroll to continue
Nvidia claims that the GB200 NVL72 delivers up to 30 times the performance of an equivalent H100 GPU for large-scale language model inference (the process of generating answers from an AI model), while also consuming less power per unit of compute. The new system is four times faster at training AI models than Nvidia's previous version.
When Huang jokes during a presentation that “the more you buy, the more you save,” audience members usually laugh, assuming he's kidding, but Blackwell proves there's an element of truth to his catchphrase.
The question is how long this high demand will last. Late last month, AI startup Anthropic CEO Dario Amodei said AI models continue to expand at an incredible rate. Amodei noted that a cutting-edge model currently costs $100 million to build, with $1 billion models in development. He expects $10 billion to $100 billion models to be developed by 2027.
Advertisement – Scroll to continue
Nvidia AI data center revenues will likely continue to grow until the capabilities of larger AI models show diminishing returns. Until then, there's no end in sight.
Barons Tech This Week
Email Tae Kim at tae.kim@barrons.com or follow @firstadopter on X.
