Jensen Huang LIVE: Nvidia's Future, Physical AI, Rise of the Agent, Inference Explosion, AI PR Crisis

The episode opens with the hosts explaining that the regular weekly show was preempted for this live conversation. They joke that the show is only preempted for "president Trump, Jesus, and Jensen," and describe the event as one where "Every industry is here. Every tech company is here. Every AI company is here." The banter then turns to the Grok purchase, with a joke about how a fellow host would behave after it and a back-and-forth over whether the deal close took six weeks or two.

The first substantive answer frames the Grok deal as part of a strategy that was announced publicly in advance. The speaker says many of the company's strategies are "presented in in broad daylight at GTC years in advance," and points to Dynamo, introduced two and a half years earlier as "the operating system of the AI factory." The name refers to the Siemens dynamo that turned water into electricity and powered the factories of the last industrial revolution. The speaker presents it as a fitting name for the operating system of the next one.

The core technology inside Dynamo is described as disaggregated inference. Inference is called "the most complicated computing problem today," involving "Incredible scale, lots of mathematics of different shapes and sizes." The idea is to split the processing so that some parts run on some GPUs and the rest on different GPUs. According to the speaker, this led to the view that heterogeneous, disaggregated computing could make sense more broadly, and the same thinking is linked to an earlier acquisition (transcribed as "melonize").

You know, we just really evolved from a GPU company to an AI factory company.

The speaker lists the components that now make up the company's computing: GPUs, CPUs, scale-up and scale-out switches, and networking processors, with Grok to be added so that "the right workload" runs "on the right chips." A questioner calls this disaggregation the biggest takeaway from the event and points to a recommendation from the stage: high-value inference users should devote about 25% of data center space to the Grok LPU and GPU combination. The answer restates this as "Grok to about 25% of the Vera Rubins" in the data center.

The reason given for the change is the shift from large language model processing to agentic processing. An agent accesses working memory and long-term memory, uses tools, puts heavy load on storage, and works with other agents. Those agents run on a mix of very large and smaller models, diffusion and autoregressive models. Vera Rubin was built to run this "extraordinarily diverse workload."

The speaker also puts this in market-size terms. What "used to be a one rack company" has added four more racks, and the speaker estimates the company's TAM grows by "call it, you know, 33%, 50% higher." Much of that increase is expected to come from BlueField storage processors, Grok processors ("I'm hoping"), CPUs and networking processors. Together they run what the speaker calls the computer of the AI revolution: agents.

Asked about embedded uses, starting with a daughter's talking teddy bear, and whether those need custom ASICs or a broader set of edge tools, the speaker describes three computers. The first trains and develops the AI model. The second evaluates it. For robots and cars, evaluation takes place in "a virtual gym that represents the physical world," using software that "obeys the laws of physics"; the company calls this Omniverse. The third is the edge or robotics computer, which could be in a self-driving car, a robot, or "Little tiny one for a teddy bear."

The speaker singles out one edge project as among the most important: turning telecommunications base stations into part of the AI infrastructure. Telecom is described as "a $2,000,000,000,000 industry" that, "in time," will become an extension of AI infrastructure, with radios acting as edge devices. Factories and warehouses are named as other edge settings. The closing line of the answer is garbled in the transcript ("aren't gonna be necessary"), so it is unclear whether all three computers are meant to be necessary.

A questioner recalls the speaker predicting last year that inference would grow a thousand times, a claim that seemed hyperbolic while the world was focused on pre-training. The answer raises the forecast: "Is 1,000,000 x? It's gonna 1,000,000,000 x." The questioner says inference "has exploded" and that "We're inference constrained."

The questioner then presents the bear case they hear. The new inference factory is said to be about 10x better in throughput, but it would cost "40 or 50,000,000,000," compared with "25 to 30,000,000,000" for alternatives such as custom ASICs and AMD. The worry is that NVIDIA will lose share. The question is whether a roughly two-times premium makes sense.

The big takeaway, the big idea is that you should not equate the price of the factory and the price of the tokens, the cost of the tokens.

The speaker says the $50 billion factory will produce the lowest-cost tokens and claims "I can prove it," though no proof is given in this exchange. The argument is about cost structure. About $20 billion of the factory is "just LAN power and shell," and storage, networking, CPUs, servers and cooling are needed regardless of the chip. So a GPU at full price versus half price does not produce a gap of $50 billion versus $30 billion. Using the speaker's illustrative figures ("Pick your favorite number"), the gap is closer to $50 billion versus $40 billion. That difference "is not a large percentage" when the more expensive data center has 10 times the throughput.

if you can't keep up with the state of the technology and the pace that we're running, even when the chips are free, it's not cheap enough.

A broader strategy question notes the company's scale, citing a projection of "350 plus billion of revenue next year" and "200,000,000,000 of free cash flow," and asks how the CEO decides where to double down, pull back, or enter new markets. The answer says that "In a final analysis, that's the job of the CEO": defining the vision and strategy, informed by computer scientists, technologists and people across the company.