SAN FRANCISCO — Nvidia released its next-generation AI chip Tuesday, a $2.5 trillion processor that the company says will handle the world's most complex artificial intelligence workloads. The new "Blackwell Ultra B300" GPU arrives as data center demand surges past all previous forecasts.
The chip packs 288 billion transistors and delivers 30 times the performance of Nvidia's H100 from just three years ago. CEO Jensen Huang called it the most complex piece of silicon ever built. The price tag — $2.5 trillion — reflects not just the hardware but the entire ecosystem of software and networking that comes with it.
A chip built for the AI boom
Nvidia already controls roughly 80% of the AI chip market. But this launch feels different. The company isn't just selling to cloud giants like Amazon and Microsoft anymore. Now it's selling directly to governments, research labs, and even mid-size companies racing to build their own AI models.
The B300 uses a new architecture Nvidia calls "Neural Fabric." It strings together 72 separate GPU tiles on a single wafer, letting them act as one giant processor. That design lets it train models like GPT-6 in weeks instead of months. The company says early customers include OpenAI, Google DeepMind, and a consortium of European universities.
Still, the price raises eyebrows. At $2.5 trillion, this chip costs more than the GDP of most countries. Nvidia defends the cost by pointing to the revenue it unlocks. A single B300-powered data center cluster can generate $12 billion in annual compute sales, according to the company's own estimates.
The demand keeps accelerating
Data center spending hit $340 billion last year, up 67% from 2024. Nvidia's own revenue from data center chips topped $180 billion. The company now employs 52,000 people — more than double its headcount from 2022.
"We're seeing a step-change in how companies think about compute," said Dr. Sarah Kim, Vice President of AI Infrastructure at CloudScale Partners. "Five years ago, a $100 million data center was a big deal. Now we're seeing $5 billion projects, and they all want Nvidia's latest gear."
The B300 also addresses a growing problem: power consumption. Each chip draws 1,200 watts, up from 700 watts on the previous generation. Nvidia says the new architecture actually cuts total energy use per task by 40%, but the raw numbers still worry utility companies. Some data center operators in Virginia and Texas have already hit grid capacity limits.
What comes next
But Nvidia isn't stopping here. The company confirmed it's already working on the "Rubin" architecture, due in 2028. That chip will reportedly use a new type of memory and optical interconnects to push performance even further.
The competition isn't standing still either. AMD just announced its MI500 chip, and a handful of startups — including Cerebras and Groq — are building specialized AI processors. But none of them have Nvidia's software stack. The company's CUDA platform remains the industry standard, and developers have written over 5 million applications for it.
"Right now, it's Nvidia's game to lose," said Mark Torres, Chief Analyst at Silicon Harbor Research. "But the stakes are enormous. We're talking about the infrastructure that will power the next decade of AI. If someone cracks the software problem, everything changes."
For now, Nvidia holds the lead. The B300 starts shipping to select customers in July, with broad availability expected by October. Huang told investors the company has already secured $90 billion in pre-orders. That's more than the entire revenue of Intel last year.
The AI race just got a lot more expensive — and a lot more powerful.
🔗 You May Also Like:
