Jensen Huang’s annual leather jacket showcase came with a $1 trillion proclamation this year. Speaking at GTC 2026 in San Jose yesterday, Nvidia’s CEO announced that combined purchase orders for Blackwell and the upcoming Vera Rubin chips are on track to reach $1 trillion through 2027 - double the company’s previous $500 billion projection.
The three-hour keynote packed in enough announcements to make competitors nervous: next-generation Rubin GPUs, an 88-core Vera CPU, SRAM-packed inference chips from the acquired Groq team, and a full-court press on what Nvidia is calling the “agentic AI” era.
Vera Rubin: The Next-Gen Platform
The Vera Rubin platform represents a complete architectural overhaul. The Rubin GPU features a dual-die design with two reticle-sized compute chiplets containing 336 billion transistors combined. Key specifications:
- 288GB HBM4 memory
- 50 PFLOPS inference performance
- 22 TB/s memory bandwidth
- Built on TSMC’s 3nm process
- 10x performance per watt over Grace Blackwell
- 5x inference throughput via new NVFP4 4-bit floating point format
The accompanying Vera CPU brings 88 custom Nvidia Olympus cores with “Spatial Multithreading” that runs two tasks per core. It uses LPDDR5X memory with up to 1.2 TB/s bandwidth. Partner availability begins in the second half of 2026.
The Groq Play: Nvidia’s Inference Land Grab
Perhaps the most significant announcement wasn’t a new chip at all - it was the unveiling of technology from Nvidia’s December acquisition of Groq.
The $20 billion deal - Nvidia’s largest ever - brought founder Jonathan Ross, president Sunny Madra, and Groq’s inference-optimized architecture in-house. The result is the Nvidia Groq 3 LPU, a Language Processing Unit designed specifically for high-volume, low-latency token generation.
The numbers are striking. Each Groq 3 LPU packs 500MB of SRAM with 150 TB/s bandwidth - nearly 7x faster than the 22 TB/s from Rubin’s 288GB HBM4. The tradeoff? SRAM doesn’t hold as much data, but it accesses what it has almost instantaneously.
Nvidia’s Groq 3 LPX platform stacks 128 LPUs into a single server rack. Paired with a Vera Rubin NVL72 rack, the company claims 35x higher throughput per megawatt and 10x more revenue opportunity for data center operators.
The strategy is clear: Nvidia wants to own both sides of the AI compute market. GPUs for training, LPUs for inference at scale.
Agentic AI: The Central Theme
Every major announcement at GTC 2026 reinforced Nvidia’s “agentic AI” thesis - the idea that AI systems will increasingly act autonomously rather than simply responding to prompts.
The DGX Station was explicitly positioned as a platform for building and running autonomous agents. The DGX Spark, powered by the GB10 Grace Blackwell Superchip, delivers 1 petaFLOP of FP4 AI performance with 128GB memory in a compact form factor.
More interesting is NemoClaw, a new open-source agent development platform. It integrates Nvidia’s Nemotron models and the OpenShell runtime into the OpenClaw agent framework with a single command, adding privacy and security guardrails to autonomous AI agents. The emphasis on security controls suggests Nvidia understands the risks of long-running, unsupervised AI systems - even if they’re betting heavily on building them.
DLSS 5: The “GPT Moment for Graphics”
Nvidia calls DLSS 5 the “GPT moment for graphics.” The system introduces a real-time neural rendering model that enhances lighting and materials using generative AI while staying anchored to the underlying 3D scene.
Unlike previous DLSS versions that focused on upscaling, DLSS 5 actually modifies how pixels look - adding photorealistic lighting effects that weren’t in the original render. It runs at up to 4K in real time and maintains frame-to-frame consistency, a critical requirement for games.
Publishers including Bethesda, Capcom, Ubisoft, and Warner Bros. Games have signed on. Launch titles include Starfield, Assassin’s Creed Shadows, Hogwarts Legacy, and The Elder Scrolls IV: Oblivion Remastered. The feature arrives this fall.
Autonomous Vehicles: Level 4 Takes Off
The automotive announcements were substantial. BYD, Geely, Nissan, and Isuzu are all building Level 4-ready vehicles on Nvidia’s DRIVE Hyperion platform. The system runs on DRIVE AGX Thor chips using the Blackwell architecture and integrates with Nvidia’s new Halos OS safety framework.
Uber announced that Nvidia-powered robotaxis will launch across 28 markets by 2028, starting with Los Angeles and San Francisco in early 2027. Bolt, Grab, and Lyft are also scaling robotaxi development on the Hyperion platform.
What This Means
Nvidia is executing a multi-front strategy that’s becoming increasingly difficult to counter:
Training dominance continues. Vera Rubin extends Nvidia’s lead in training large models. The 10x performance-per-watt improvement matters because data center power is becoming the binding constraint.
Inference is the new battleground. The Groq acquisition signals that Nvidia sees inference - running trained models at scale - as an equally large market. Rather than compete against specialized inference chips, Nvidia bought the best one.
Agentic AI is the bet. Every announcement assumed a future where AI systems run autonomously for extended periods. DGX Station, NemoClaw, and the emphasis on “long-running agents” all point the same direction.
The stack keeps growing. CPUs, GPUs, LPUs, networking, software, safety frameworks, and now gaming graphics - Nvidia is building a vertically integrated AI empire.
The Bottom Line
GTC 2026 wasn’t just a product announcement - it was Nvidia declaring that it intends to own the entire AI compute infrastructure, from model training to real-time inference to autonomous agents. With $1 trillion in projected orders and no clear competitor matching this breadth, the rest of the industry is playing catch-up.