NVIDIA CEO Jensen Huang’s keynote address at NVIDIA’s annual developer event on Monday included new product announcements and insight into the direction of revenue. Here are the highlights of Jensen’s two biggest updates. New inference chip Jensen announced a new inference-focused chip from Nvidia built on technology it licensed from AI chip startup Groq for $20 billion late last year. On Friday, we published an in-depth look at the origins of Groq and the growing competition Nvidia faces in inferential computing. This is the name for everyday use of the AI model after training. While Nvidia’s graphics processing units (GPUs) dominate training, AI computing is maturing and evolving, requiring more specialized inference chips. This is where Groq’s technology, which calls the chip a language processing unit, comes into play. They are best designed for specific inference tasks where speed is most important, typically referred to as low latency. Nvidia is naming the Groq-powered processors LPX, and it’s worth noting that they’ll be launched alongside the Vera Rubin generation of chips (Vera is the CPU, Rubin is the GPU) that will be released later this year as a successor to the Blackwell family. Jensen said LPX is currently in mass production at third-party manufacturer Samsung and is expected to be available during the “third quarter time frame.” Nvidia offers inference chips in server racks with 256 LPX processors. When we talk about racks, we are referring to cabinet-sized computers that contain both the “processor” engine and the network that connects the chips together. A data center has rows of server racks. The idea, Jensen said, is not that LPX racks will completely replace Nvidia’s combined GPU and CPU servers for inference, but that LPX racks will coexist within the data center and work together to improve performance. Jensen said not every type of task will require LPX. “If most of your workloads are high-throughput, I would 100% stick with Vera Rubin exclusively,” Jensen says. “If a lot of your workloads involve coding or generating highly valuable engineering tokens, you’ll add Groq to that,” he said. “We’ll add Groq to probably 25% of our data centers, and the rest will be Vera Rubin servers.” He added, “This shows how you can add Groq to Vera Rubin to extend its performance and further extend its value.” Jensen also suggested that new and improved versions of LPX will arrive in the coming years, cementing LPX’s place in Nvidia’s broader roadmap, featuring new central processing units (CPUs), GPUs, and networking technologies. As part of the licensing agreement, Nvidia hired key employees from Groq, including co-founder and now former CEO Jonathan Ross. We believe this inference chip is meaningful and will help Nvidia fend off competition in inference from internal chip initiatives such as Google’s tensor processing unit (TPU), which Nvidia co-designed with Broadcom, and other chip designers such as Advanced Micro Devices. Outlook to 2027 Jensen said NVIDIA expects orders for Blackwell and Vera Rubin generation chips to total $1 trillion by 2027. This is the latest outlook for demand over the next few years. To put this into context, we need to rewind a little. At the NVIDIA conference last fall, Jensen said NVIDIA had $500 billion worth of orders for Blackwell and Rubin chips and related networking equipment by 2026. Then, during Nvidia’s February earnings call, CFO Colette Kress suggested that Nvidia sees upside in the “$500 billion revenue opportunity that Blackwell and Rubin shared last year.” We believe we have inventory and supply commitments by calendar to meet future demand, including shipments. 2027. ” Now back to Monday afternoon. After Jensen presented the $1 trillion figure, Nvidia stock gained momentum, rising about 4.8% to $188.88. However, the pop faded, and the stock ultimately closed at $183.22, up 1.65% during the session. Of course, it is difficult to pinpoint exactly why the market sometimes moves the way it does. But in this case, it appears that after traders and investors took a closer look at how that $1 trillion disclosure stacked up against Wall Street consensus and their own models, they may have decided that it wasn’t as much ahead of consensus as it first appeared. Our take on the situation: The 2027 consensus for Nvidia’s data center revenue appears likely to be further revised upwards in light of Jensen’s disclosure and, by extension, the revenue forecast. However, the exact size is not clear at this time. Nvidia has scheduled a Q&A with financial analysts on Tuesday, which could provide more clarity. Jim Cramer is also scheduled to interview Jensen on CNBC on Tuesday, and we may learn more in that conversation. In any case, Jensen’s comments should lend confidence to the 2027 forecast. In other words, investors who fear that the momentum in AI investing will soon stall should feel better as Jensen discusses the company’s outlook through next year. The duration of the AI capital spending boom has now been debated over Nvidia stock for several years, and shows no signs of slowing down just yet. In a note to clients Monday morning ahead of Jensen’s keynote, Morgan Stanley analysts cited the uncertainty surrounding the 2027 forecast and highlighted why Jensen’s fresh comments will be helpful. Morgan Stanley writes: “When it comes to the duration debate, the company has generally been right, and we have seen confirmation through the ecosystem that investment is sustainable, including in its supersized comments just last week. Our view is that it is only a matter of time before investors begin to feel comfortable with the outlook for 2027. This will require capital market vibrancy, which is the main risk, but we believe that enthusiasm for AI remains high and that semiconductor constraints will lead to excessive levels of investment. (The Jim Cramer Charitable Trust is long NVDA and AVGO. See the full list of stocks here.) Subscribers to Jim Cramer’s CNBC Investment Club will now wait 45 minutes after Jim sends out a trade alert before buying or selling stocks in the Charitable Trust’s portfolio, Jensen said. If you talk about stocks on CNBC TV, we will wait 72 hours after issuing a trade alert before executing a trade. The above investment club information is subject to our Terms of Use and Privacy Policy, along with our disclaimer. Receipt of information provided in connection with the investment club does not guarantee any particular outcome or benefit.
