Agentic AI takes center stage

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

Nvidia’s yearly showcase event — dubbed the ‘Super Bowl of AI’ by some — kicked off at the start of the week to much fanfare across the tech sector. The event sees tens of thousands of attendees gather in California to get the latest on the world’s most valuable company’s plans for the future.

Didn’t manage to snag a ticket? No problem. I caught up with CNBC’s Katie Tarasov, who was on the ground at the event, to get a sense of what went down.

Katie: I’m always watching for the biggest hardware announcements because it’s Nvidia chips that are filling AI data centers and powering almost every major company’s AI ambitions. We saw two big new chip announcements during CEO Jensen Huang’s keynote on Monday.

First was an entirely new type of chip called a Language Processing Unit, or LPU. It’s the first chip Nvidia’s unveiling using technology it acquired from chip startup Groq in December. That $20 billion deal was Nvidia’s biggest purchase ever. While Nvidia’s star graphics processing units have thousands of cores that perform many operations simultaneously, the Groq 3 LPU is built with a single core optimized for speeding up those GPUs.

The other big chip announcement was the unveiling of a rack filled entirely with Nvidia’s newest Vera central processing units, or CPUs. I wrote a piece last week explaining how the CPU is having a renaissance as Nvidia sees it as a coming bottleneck for agentic AI, which requires more data transfer and general-purpose compute typically handled by the CPU.

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

Katie: The biggest theme this year was that the future is agentic. We’re seeing a lot of everyday AI use shift from call-and-answer chatbots to more task-oriented AI agents. That’s creating a big need for faster inference because agents spawn other agents, requiring a lot of orchestration and data transfer between them all to accomplish their tasks.

Huang said agentic AI has reached an “inflection point” and is driving a fundamental shift in computing needs. That’s why we saw less focus on the GPU and more focus on new compute configurations like full racks of LPUs and CPUs.

Overall, I was personally blown away by the crowds and hype on the showroom floor — it was unrecognizable from my first GTC in 2019. I last sat down with Huang for a lengthy interview in 2022, and it was a fascinating conversation that’s informed much of my chip reporting since.

But I never would have guessed that four years later, Huang would reach the full-blown celebrity status I witnessed when he was swarmed by selfie-seekers at every turn, even within the private press Q&A session.

Katie: This is a difficult question to answer. Not even Huang’s expectations of $1 trillion in Blackwell and Vera Rubin purchase orders by 2027 moved the needle. It seems like Wall Street’s expectations are so high, it’s almost impossible for Nvidia to impress.

One guess is that GTC may have underwhelmed the street because Nvidia chose to unveil Vera Rubin earlier, at January’s Consumer Electronic Show, instead of at its own conference. We also saw very little movement despite seeing the fruits of December’s $20 billion spend on Groq, perhaps because a first-of-its-kind chip like the LPU isn’t always a homerun.

Huang also signaled more revenue is forthcoming from China with sales of the H200 restarting there, but export controls have been such a moving target that investors may be waiting to see those sales confirmed in the next earnings statement.

Katie: Overall, I think we’re seeing Nvidia shift its strategy to align with changing compute needs as agentic AI takes off. Instead of putting all its eggs in the GPU basket, Nvidia’s taking a more “soup-to-nuts” strategy, as Creative Strategies analyst Ben Bajarin put it to me.

But Huang also showed a sneak peak of what’s next for its most buzz-worthy line: the Kyber rack-scale architecture. It will integrate 144 GPUs in compute trays that sit vertically instead of horizontally in order to boost density and lower latency. The Kyber design will be available in Vera Rubin Ultra, Nvidia’s next rack-scale system, expected to ship in 2027.

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

OpenAI will combine its web browser, ChatGPT app and Codex coding app into a singular desktop super app, CNBC confirmed on Thursday.

China’s tech giants are holding public events to encourage everyday people to get viral AI assistant OpenClaw downloaded on their laptops and phones.

Bain Capital has started reaching out to potential buyers for its stake in Bridge Data Centres, two people familiar with the matter told CNBC, as the private equity firm eyes exits amid rising demand for AI infrastructure.

Chinese tech giant Alibaba reported net income dropped 66% in its December quarter on Thursday, as it missed analyst revenue expectations.

Chipmaker Micron reported it had tripled revenue and sailed past analysts’ estimates in its latest quarterly results on Wednesday, but champagne and party poppers were not the order of the day.

The company, which specializes in memory chips — a key component in the AI boom, saw its stock fall despite blockbuster earnings, echoing investors’ unenthusiastic response to Nvidia’s better-than-expected results in February. 

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