NVIDIA Just Showed a $1 Trillion Roadmap. GTC 2026 Changed Everything.

Jensen Huang walked onstage today and casually mentioned he sees at least one trillion dollars in purchase orders for Blackwell and Vera Rubin hardware through 2027. That is double the $500 billion projection he gave last year. Then he spent two hours showing why that number might be conservative.

GTC 2026 was not a product launch. It was a declaration that NVIDIA intends to own every layer of the Artificial Intelligence (AI) stack — silicon, networking, software, models, agents, simulation, and now literal orbital compute. If you work in security, AI infrastructure, or anything adjacent, this keynote just redrew the map.

Here is everything that happened, why it matters, and where the attack surface is expanding.

Vera Rubin: The Agentic AI Platform

The headliner is Vera Rubin, NVIDIA's next-generation platform built explicitly for agentic AI workloads. This is not a single chip. It is seven chips, five rack-scale systems, and one supercomputer designed to run autonomous AI agents at data center scale.

The numbers are absurd. The NVL72 configuration delivers 3.6 exaflops of compute with 260 terabytes per second of NVLink bandwidth. Rubin Ultra connects up to 144 Graphics Processing Units (GPUs) into a single coherent fabric. The entire platform ships in the second half of 2026.

For context, 3.6 exaflops is roughly the combined compute of every Top500 supercomputer on Earth a few years ago. NVIDIA is putting that in a single rack-scale system you can order from a cloud provider.

Vera CPU: The First CPU Built for Agentic AI

This is the announcement that should worry Intel and AMD the most. NVIDIA revealed Vera, which the company calls the world's first Central Processing Unit (CPU) purpose-built for agentic AI workloads. It runs on custom Olympus Arm cores with a neural branch predictor — the branch predictor literally uses a neural network to predict execution paths.

NVIDIA claims 3x the memory bandwidth and 1.5x the performance per core compared to current x86 processors. A 256-CPU liquid-cooled rack delivers over 22,500 cores and 400 terabytes of memory. This is a host processor designed to feed GPU clusters without becoming the bottleneck.

Alibaba, ByteDance, Meta, Oracle, and CoreWeave have already committed to deploying Vera. That is not a future roadmap slide. That is a customer list.

Groq 3 LPX: 35x Throughput Per Megawatt

NVIDIA is done pretending inference is a software problem. The Groq 3 LPX rack packs 256 Language Processing Units (LPUs) into a single rack that sits alongside Vera Rubin systems. Inference is disaggregated from training at the hardware level using NVIDIA Dynamo 1.0, an open source inference orchestration framework.

The result: 35x throughput per megawatt compared to current solutions. That is not an incremental improvement. That is a generational leap in inference economics.

If you have been watching the "inference will be cheap" narrative, this is the hardware that makes it real. The energy efficiency numbers alone change the calculus for every AI deployment at scale.

NemoClaw: Open Source AI Agent Security in One Command

This is the one security practitioners need to pay attention to.

NemoClaw is an open source reference stack for deploying enterprise AI agents based on OpenClaw, the agent framework Huang called "the most popular open source project in the history of humanity." Hyperbole aside, the pitch is compelling: a single command spins up a secure, observable, controllable AI agent with guardrails baked in.

Microsoft Security is already partnering on NemoClaw integration. The framework includes role-based access controls, audit logging, and policy enforcement layers designed to give security teams visibility into what autonomous agents are doing before they do it.

This matters because the agentic AI security problem is real and largely unsolved. Most organizations deploying AI agents today are bolting security on after the fact. NemoClaw is NVIDIA's bet that security should be a first-class citizen in the agent deployment stack, not an afterthought.

Whether NemoClaw actually delivers on that promise is a different question. But the fact that NVIDIA is positioning agent security as a core platform capability — not a third-party add-on — is significant.

Nemotron 3 Super: NVIDIA Wants the Model Layer Too

NVIDIA is no longer content to sell shovels. Nemotron 3 Super is a 120 billion parameter open model using Mixture of Experts (MoE) architecture, with only 12 billion parameters active at inference time. That is competitive with frontier models at a fraction of the compute cost.

The Nemotron Coalition includes Perplexity, Black Forest Labs, and LangChain. NVIDIA is building an ecosystem around its own models, which means the company now competes at every layer: hardware, networking, runtime, model weights, and agent frameworks.

For the industry, this is both exciting and concerning. NVIDIA building high-quality open models pushes the ecosystem forward. NVIDIA building a vertically integrated stack from silicon to software raises questions about lock-in that the industry has been debating for decades.

DLSS 5: Neural Rendering Arrives

Huang called DLSS 5 "the most significant breakthrough in computer graphics since 2018." That is a bold claim, but the demo footage backed it up.

DLSS 5 introduces neural rendering — generative AI integrated directly into the graphics render pipeline. Instead of traditional rasterization or ray tracing for every pixel, the GPU uses learned models to generate photorealistic frames with dramatically less compute. This ships fall 2026 and will roll out across GeForce hardware.

The security angle here is subtle but real. Neural rendering means the provenance of visual content becomes even harder to verify. When every game engine and creative tool can generate photorealistic imagery at the render pipeline level, the distinction between "captured" and "generated" visual content erodes further.

Physical AI: Robots, Cars, and a Walking Snowman

NVIDIA's physical AI push went from concept to deployment commitments today.

Healthcare robotics: CMR Surgical and Johnson & Johnson MedTech are building surgical robotics on NVIDIA's platform. This is not simulation. This is hardware going into operating rooms.

Autonomous vehicles: Uber announced a robotaxi fleet in 28 cities across 4 continents by 2028, running on NVIDIA infrastructure. BYD, Hyundai, Nissan, Geely, and Isuzu are all building Level 4 autonomous vehicles on NVIDIA Drive Hyperion.

The Olaf moment: Disney brought Olaf from Frozen onstage. Not as a video. As a walking, reacting physical robot trained entirely in NVIDIA simulation. It was simultaneously the most ridiculous and most impressive demo of the keynote. The point was clear: NVIDIA's simulation-to-reality pipeline works for embodied AI, whether the body is a surgical instrument, a delivery vehicle, or a cartoon snowman.

T-Mobile and Nokia are deploying physical AI for network infrastructure — autonomous systems that manage and optimize telecommunications hardware.

Every one of these deployments introduces new attack surfaces. Surgical robots, autonomous vehicles, and telecom infrastructure all require security models that most organizations have not built yet. The gap between physical AI deployment speed and physical AI security maturity is widening.

Vera Rubin Space-1: AI Data Centers in Orbit

Yes, really. NVIDIA announced Vera Rubin Space-1, an initiative to put AI compute in orbit. The claim is up to 25x H100-equivalent compute in space, designed for Earth observation, communications, and edge AI processing where terrestrial data centers cannot reach.

The security implications are significant and underexplored. Orbital compute means AI inference happening in jurisdictions that do not technically exist yet. Data sovereignty, supply chain integrity, and physical security all look very different when the hardware is in low Earth orbit.

Feynman: The 2028 Roadmap

NVIDIA revealed the Feynman architecture as the successor to Vera Rubin, targeting 2028. It includes a new GPU, a new LPU, and a new CPU called Rosa — named for Rosalind Franklin. The networking stack moves to BlueField-5, ConnectX-10, and Kyber co-packaged optics.

Co-packaged optics is the key detail here. Moving optical transceivers onto the chip package eliminates a major bottleneck in data center networking. Kyber means NVIDIA is building the interconnect fabric of the future directly into the silicon, not relying on external networking hardware.

This is a two-year roadmap, but it signals that NVIDIA's cadence is accelerating. Annual architecture refreshes are the new normal.

The Partnerships Tell the Story

The partner announcements round out the picture:

Adobe is integrating NVIDIA acceleration into Firefly for real-time generative content creation.
Microsoft Security is building on Nemotron and NemoClaw for AI-powered security operations.
Roche has deployed over 3,500 Blackwell GPUs for drug discovery and genomics.
IBM is integrating NVIDIA hardware into WatsonX for enterprise AI workloads.
T-Mobile and Nokia are deploying physical AI for telecommunications infrastructure.

The thread connecting all of these is that NVIDIA is no longer a GPU company. It is an AI platform company, and every major enterprise partner is building on that platform.

Why Security Practitioners Should Care

If you work in security, GTC 2026 just expanded your threat model.

Agentic AI infrastructure is real now. Vera Rubin and NemoClaw together represent a full-stack platform for deploying autonomous AI agents. That means autonomous systems making decisions, accessing APIs (Application Programming Interfaces), and taking actions without human approval loops. Every agent is an identity. Every agent action is an authorization decision. Most Identity and Access Management (IAM) frameworks are not built for this.

NemoClaw is a security opportunity and a single point of failure. If the industry converges on NemoClaw as the standard agent security layer, a vulnerability in NemoClaw becomes a vulnerability in every agent deployment that depends on it. Open source helps with transparency. It does not automatically help with security.

Physical AI attack surfaces are multiplying. Surgical robots, autonomous vehicles, orbital compute, telecom infrastructure — each of these has safety-critical failure modes that traditional Application Security (AppSec) does not address. The security community needs to move fast on physical AI security frameworks before deployment outpaces defense.

The inference explosion changes the economics of abuse. Thirty-five times more throughput per megawatt means inference costs drop dramatically. That is great for legitimate use cases. It also means adversarial AI becomes cheaper — more phishing at scale, more deepfakes, more automated vulnerability discovery by attackers.

Orbital compute introduces novel jurisdictional and supply chain risks. There is no existing regulatory framework for AI data centers in space. Whoever gets there first writes the rules.

The Bottom Line

GTC 2026 was NVIDIA declaring that the agentic AI era is not coming — it is here, it is shipping in the second half of this year, and the company intends to own the infrastructure from silicon to orbit.

The $1 trillion demand forecast is not just a number. It is a signal that the world's largest technology companies believe this platform is the foundation of everything they are building for the next decade. Jensen Huang just showed them what that foundation looks like.

For security professionals, the message is clear: the infrastructure is moving faster than the security models. NemoClaw is a start, but we need independent security frameworks, red team methodologies, and governance models for agentic AI at this scale. The window to build those before deployment overtakes us is closing.

The future NVIDIA showed today is extraordinary. Making it safe is on us.

Eric Fleming is a senior security practitioner and researcher who writes about AI, cybersecurity, and emerging technology at ericfleming.ai.