For many years, artificial intelligence in business has been an advantage. Smarter search bar. Degree of fraud. A recommendation engine tucked inside a larger product.
What’s up now is different. AI agents receive instructions, break them down into steps, and complete the task. The finance team’s AI agent does not report an anomaly and waits for a human to investigate it. He investigates, pulls relevant records, drafts the memo and forwards it for review. The computing infrastructure required to do this is nothing like what supported the last generation of AI.
“Agent AI has arrived,” Nvidia CEO Jensen Huang said in the company’s fiscal first quarter. Earnings call. “AI can now do productive and valuable work. Tokens are now profitable.” The infrastructure that powers agents to complete real work is no longer a discretionary expenditure. He earns a return for every task he completes.
Two layers behind the AI workflow
An AI agent doing real work within a company works in two layers. One deals with logic. The other is in charge of implementation.
“All the thinking happens in the GPUs,” Huang said. “All the coordination basically runs on CPUs. If the AI is going to do the searching, use the browser, which will run on the CPU.”
The agent processes the reasons for chargebacks through evidence on one layer of devices. It then pulls transaction logs, response files, and updates the status history onto a second layer. The logic layer has been around for several years. The implementation layer is what’s new. No chip has been built for it yet.
Advertisement: Scroll to continue
Nvidia’s Vera processor is designed specifically for this implementation layer. Previous chips were designed to rent processing power to many users simultaneously, because that’s how cloud economics work. Agents do not rent capacity. They need to complete the task as quickly as possible and at the lowest cost per completed action.
“The economics of AI in the future are tokens per dollar,” Huang said. Nvidia told investors it expects nearly $20 billion in Vera chip revenue this year from a market it has never addressed before.
The Humanitarian Partnership puts a tangible organization name on this application. Cloud models are embedded in document review, financial analysis and compliance workflows in companies across financial, legal and business services. Nvidia had no infrastructure relationship with Anthropic before this year. Nvidia is now enhancing its capabilities across Amazon Web Services, Microsoft Azure, and additional cloud providers where enterprise buyers are already operating.
“The amount of capacity we will bring online to Anthropic this year and next will be very large,” Huang said.
High costs
Nvidia now divides its business into two groups. One covers a few of the hyperscale platforms that most organizations rent as of today. The other covers providers of cloud AI services, companies that run AI on their own premises, and governments and industrial operators that need computing within their own facilities.
The second group grew by 31% in one quarter. Its AI cloud revenue more than tripled year-over-year. The number of large data centers dedicated to AI has nearly doubled in 12 months.
Hwang told investors that spending on AI infrastructure could reach $3 trillion to $4 trillion annually by the end of the decade. High-speed capital spending on AI alone is expected to exceed $1 trillion in 2027. “Computing is revenue. Computing is profit,” he said.
What else stood out?
- Nvidia’s AI for Physical Operations, which covers logistics robotics, warehouse automation and autonomous vehicles, has generated more than $9 billion over the past 12 months. The partnership with Uber will deploy the technology across the robotaxi fleet in nearly 30 cities and four continents by 2028.
- Regarding China: US export licenses for some Nvidia chips have been approved for Chinese buyers, but Nvidia has not collected any revenue and is excluding all Chinese data center revenue from its future projections while import approvals remain unresolved.
- Nvidia’s networking business nearly tripled year over year. The company’s Spectrum
- Consumer AI devices, including AI-equipped laptops and workstations, generated $6.4 billion, an increase of 29% year over year. Demand for professional workstations has been strong. Demand for consumer devices declined as rising memory prices drove up system costs.
Top line results
Total revenue for the fiscal first quarter was $82 billion, up 85% year-over-year and 20% sequentially, and the 14th consecutive quarter of consecutive growth. Data center revenue reached $75 billion, up 92% year over year. In data centers, chip revenue reached $60 billion, up 77% year over year. Network revenue reached $15 billion, up nearly 3x. Consumer and edge devices revenue was $6.4 billion, up 29% year over year.
For all of our PYMNTS AI coverage, subscribe to our daily newsletter Amnesty International newsletter.
