Jensen Huang took the stage in San Jose on Monday and did what he does best — made the already-staggering look modest. At Nvidia's annual GTC developer conference, Huang announced the company now expects purchase orders across its Blackwell and Vera Rubin chip families to hit $1 trillion through 2027. Last year's projection was half that. The crowd barely blinked.
If you work in software, AI, or anything that touches compute infrastructure, this keynote deserved your full attention. Here's what actually mattered.
The $1 Trillion Order Book
Nvidia's revenue growth has been almost comically consistent — 11 straight quarters above 55% year-over-year. Last month the company forecast Q1 revenue of roughly $78 billion, a 77% jump from the same period last year. And yet the demand signal keeps pointing up.
Huang's explanation was simple: as AI workloads shift from chatbot-style inference to agentic applications — systems that spawn child agents to accomplish complex multi-step tasks — the number of tokens being generated has exploded. More tokens means more GPU cycles, which means more Nvidia. "If they could just get more capacity, they could generate more tokens, their revenues would go up," Huang said plainly.
This is the infrastructure reality underpinning every AI product being built right now. The constraint isn't intelligence — it's bandwidth and compute.
Vera Rubin: 10x Performance Per Watt
Scheduled to ship later this year, Vera Rubin is Nvidia's next rack-scale system. The headline stat: 10 times more performance per watt than Grace Blackwell, its predecessor. That's a significant jump when energy consumption is one of the most critical bottlenecks in the AI buildout. Data centres are already straining power grids globally, and efficiency gains at this scale have real downstream effects — both on operating costs and on the feasibility of continued expansion.
Vera Rubin is built from 1.3 million components. That complexity is itself a story about how far GPU architecture has evolved beyond the gaming-focused roots that originally made Nvidia famous.
Groq 3 LPU: The $20 Billion Bet Pays Its First Dividend
Perhaps the most technically interesting announcement: the Nvidia Groq 3 Language Processing Unit, or LPU. Nvidia acquired Groq — the startup founded by the creators of Google's in-house Tensor Processing Unit — in a $20 billion asset purchase last December. The Groq 3 is the first chip to emerge from that acquisition and is expected to ship in Q3.
The design is clever. Rather than replacing the GPU, the Groq 3 LPU acts as a companion chip — one core is optimised for raw throughput (the GPU), the other for ultra-low latency (the LPU). Together, they address different bottlenecks in inference workloads. Huang introduced the Groq 3 LPX rack, which holds 256 LPUs and is designed to sit alongside the Vera Rubin GPU rack. The claimed improvement: 35x better tokens per watt versus Rubin GPUs alone.
"We unified two processors of extreme differences — one for high throughput, one for low latency," Huang said.
Kyber: The Generation After Rubin
Already teasing what comes after Vera Rubin, Huang showed a prototype of Kyber — a new rack architecture that packs 144 GPUs into vertical compute trays to improve density and reduce latency. Kyber will arrive as Vera Rubin Ultra, slated for 2027. The vertical tray design is more than aesthetic — it addresses thermal and signal integrity constraints that become acute at this level of integration.
NemoClaw and the Agentic Pivot
Alongside the hardware announcements, Nvidia introduced NemoClaw, an AI agent platform aimed at enterprises building multi-agent workflows. This signals where Nvidia sees the next major wave of platform lock-in forming — not in foundation models, which are increasingly commoditised, but in the orchestration and memory layers that run on top of them.
The robotic demos rounded out the show — including an Olaf robot in partnership with Disney, running simulations on Nvidia GPUs. Equal parts PR spectacle and genuine proof-of-concept for physical AI.
What This Means for Developers and Builders
For those of us building software products on top of AI APIs, the GTC story translates into a few practical realities:
Inference is getting faster and cheaper. The LPU + GPU combo, combined with Vera Rubin's efficiency gains, should continue pushing inference costs down. That's great for anyone building products where token cost is a meaningful input to unit economics.
Agentic workloads are the next compute category. If Nvidia is orienting $1 trillion in order projections around agentic AI, the infrastructure bets are being placed. Now is the time to understand what multi-agent architectures actually look like in production.
The energy bottleneck is real, but being attacked. 10x performance per watt isn't a marketing number — it's an engineering response to a physical constraint that was threatening to limit the whole industry's growth.
Nvidia's dominance in AI infrastructure was supposed to be temporary — a lucky accident of CUDA and GPU parallelism that competitors would eventually route around. So far, the company keeps widening the moat. With Vera Rubin, the Groq LPU, and Kyber already mapped out through 2027, the pace shows no sign of slowing.
The question for every AI product builder remains the same: are you designing for the infrastructure that exists today, or the infrastructure that's being built for?
Top comments (0)