NVIDIA's AI factories

Plus: AI agent capabilities are doubling every 7 months, OpenAI releases new voice agent models, and more

Welcome back to Building AI Agents, your biweekly guide to everything new in the AI agent field!

Speeding up your work with agents is great. Giving them work you have no idea how to do yourself and just trusting the outputs, however, can get you in trouble.

In today’s issue…

  • AI factories power the new Industrial Revolution

  • Voice agent models from OpenAI

  • Email for AI agents

  • Agent capabilities double every 7 months

…and more

🔍 SPOTLIGHT

Source: NVIDIA

A new industrial revolution is underway, but its factories look nothing like anything ever seen before.

At Nvidia’s semi-annual GPU Technology Conference (GTC) last week, its CEO Jensen Huang gave a marathon keynote presentation of over two hours, in which he laid out the company’s vision for the coming era of AI agents. Although the talk was sprinkled with other interesting projects, such as new “personal AI supercomputers”, a physics simulation engine, and a foundation model for humanoid robots, its one overarching focus was clear: agentic AI, and the massive data centers that will be necessary to power it.

Nvidia’s explosive revenue growth and the resulting run-up in its stock over the past several years—which have made Huang one of the richest men in the world—have been driven by the recognition that the new world of large language models will require enormous amounts of computational power, something Nvidia is in a unique position to provide. However, the true magnitude of the demand has only become clear as the focus of the AI industry has pivoted from un-augmented chatbots to truly agentic systems capable of reasoning, planning, and taking actions in the real world. In his speech, Huang noted that the amount of compute required to power this new age is a hundredfold greater that what was expected even a year ago.

Of course, it’s easy to dismiss this as marketing hype—after all, Nvidia has a vested interest in creating a narrative about ever expanding demand for AI compute. But the claim stands up to scrutiny. Since the release of OpenAI o1 last September, we’ve seen the rise of reasoning models, which natively think through their outputs before providing them. A new family of such models explicitly designed to power AI agents was, in fact, one of the items announced in the keynote. These systems must often generate far more tokens than they would if they were responding directly. Agentic AI, in particular, is incredibly computation-hungry, with long chains of actions requiring huge amounts of text to be iteratively produced and then fed back into the model for the next step.

To meet this demand, Huang conjured up an image evocative of Victorian Britain: colossal “AI factories” which churn out tokens by the trillion. To orchestrate them, he announced NVIDIA Dynamo, described as the operating system of an AI factory, and fittingly named after the electrical dynamo that helped drive the Industrial Revolution by converting motion into electrical power. Then, of course, there were two new generations of GPUs—the engine driving the computational factory. If AI agents are the machines it powers, then naturally an assembly line is needed to integrate them. Enter NVIDIA AgentIQ toolkit, an open-source software library created to connect, profile, and optimize teams of agents, including those built in different frameworks, such as LangChain, Microsoft’s Semantic Kernel and CrewAI.

While the machinery of AI’s new industrial revolution is clear, its great names are still being written. Huang will almost certainly be remembered someday with Ford, Carnegie, and Rockefeller. Many others who build AI agents will as well.

Subscribe to keep reading

This content is completely free, but you must be subscribed to Building AI Agents to continue reading.

I consent to receive newsletters via email. Terms of use and Privacy policy.

Already a subscriber?Sign in.Not now