Agent AI takes center stage

This report is from this week’s The Tech Download newsletter. Is it what you see? You can subscribe here.

Nvidia’s The annual showcase event, dubbed by some the “Super Bowl of AI,” opened earlier in the week amid uproar across the tech industry. The event brings tens of thousands of attendees to California to learn about the future plans of the world’s most valuable companies.

Couldn’t get tickets? no problem. I spoke to CNBC’s Katie Tarasoff, who was at the event, to find out what happened.

Katie: I’m always on the lookout for the biggest hardware announcements, as it’s the Nvidia chips that fill AI data centers and power the AI ​​ambitions of nearly every major company. CEO Jensen Huang’s keynote speech on Monday included two major new chip announcements.

The first is an entirely new type of chip called an LPU (Language Processing Unit). This is the first chip NVIDIA has announced using technology it acquired from chip startup Groq in December. The $20 billion deal was Nvidia’s largest acquisition in history. Nvidia’s Star Graphics Processing Units have thousands of cores that perform many operations simultaneously, but the Groq 3 LPU is built with a single core optimized to accelerate these GPUs.

The other big chip announcement was the reveal of a rack completely filled with Nvidia’s latest Vera central processing units (CPUs). I wrote an article last week explaining how CPUs are undergoing a renaissance. Nvidia believes that CPU will be the future bottleneck for agent AI. The CPU requires more data transfer and general purpose computing that is typically handled by the CPU.

And there was one software mention that stood out. Nvidia announced NemoClaw, an enterprise-level version of OpenClaw that layers Nvidia’s software stack on top of an autonomous AI agent platform.

Katie: The big theme this year was that the future is agentic. Much of the day-to-day use of AI is moving from call-and-response chatbots to more task-oriented AI agents. There is a huge need for faster inference, as agents spawn other agents, and large amounts of orchestration and data transfer are required between all agents to accomplish tasks.

Huang said agent AI has reached an “inflection point” and is fundamentally changing computing needs. Therefore, we focus less on GPUs and more on new compute configurations such as full-rack LPUs and CPUs.

Overall, I was personally shocked by the crowd and hype on the showroom floor. I didn’t recognize it when I first attended GTC in 2019. I last sat down for a lengthy interview with Mr. Huang in 2022, and it was a fascinating conversation that has inspired many of my tip reports since then.

But little did I know that four years later, as I witnessed, Mr. Hwang would reach the status of a full-fledged celebrity, swarmed by people asking for selfies everywhere, even in private press Q&As.

Katie: This is a difficult question to answer. Even Mr. Huang’s prediction that Blackwell and Vera Rubin’s purchase orders would reach $1 trillion by 2027 failed to move the needle. Wall Street’s expectations are so high that it seems nearly impossible for Nvidia to impress.

One speculation is that Nvidia chose to announce Vera Rubin at the Consumer Electronic Show in January rather than at its own conference, which may have caused GTC to get crushed on the streets. And despite seeing results from December’s $20 billion spending on Groq, there was little movement. This is probably because first-of-its-kind chips like LPUs aren’t always home runs.

Huang also hinted that the resumption of H200 sales in China will lead to more revenue from China, but export restrictions are such a big target that investors may be waiting for confirmation of that sale in the next earnings report.

Katie: Overall, I think Nvidia is pivoting its strategy to align with changing computing needs as agentic AI becomes more prevalent. Creative Strategies analyst Ben Bajarin told me that Nvidia is taking a more “soup-to-nuts” strategy rather than putting all its eggs in the GPU basket.

But Huang also outlined what’s next for the company’s most notable product line, the Kyber rack-scale architecture. To increase density and reduce latency, 144 GPUs are integrated into a vertically oriented compute tray instead of horizontally. Kyber’s design will be available in Nvidia’s next rack-scale system, the Vera Rubin Ultra, scheduled to ship in 2027.

The U.S. Attorney’s Office for the Southern District of New York has indicted officials at an unidentified U.S. server maker for illegally diverting billions of dollars in Nvidia-powered servers to China.

OpenAI plans to combine a web browser, ChatGPT app, and Codex coding app into a single desktop super app, CNBC confirmed Thursday.

The Chinese tech giant is holding a public event to encourage the public to download its viral AI assistant OpenClaw onto their laptops and mobile phones.

Bain Capital has begun reaching out to potential buyers for a stake in Bridge Data Centers, two people familiar with the matter told CNBC, as Bain Capital looks to exit amid growing demand for AI infrastructure.

Chinese technology giant alibaba announced on Thursday its December quarter net profit fell 66%, missing analysts’ revenue expectations.

chip manufacturer micron reported sales tripled in its latest quarterly results Wednesday, beating analysts’ expectations, but champagne and party poppers weren’t popular at the time.

The company, which specializes in memory chips, a key component of the AI ​​boom, saw its stock fall despite big profits, reflecting investors’ lackluster reaction to Nvidia’s better-than-expected financial results in February.

Correction: This article has been updated to note that Nvidia’s $20 billion Groq deal took place in December.