Jensen Huang took the stage at SAP Center in San Jose yesterday in his leather jacket and delivered a two-hour keynote that redefines the enterprise AI infrastructure landscape for the next two years.
The headline number — $1 trillion in expected orders for Blackwell and Vera Rubin chips through 2027, double the $500 billion projection from last fall — tells you the scale. But the five specific announcements that matter for enterprises tell you the direction.
Here are the five things from GTC 2026 that will shape enterprise AI for the rest of this decade.
1. The Groq 3 LPU: Inference Gets Its Own Chip
The most technically significant announcement was the unveiling of the Groq 3 Language Processing Unit — Nvidia's first chip from the startup it acquired through a $20 billion licensing deal in December, its largest deal ever.
Groq was founded by the creators of Google's tensor processing unit and has built technology specifically optimised for AI inference — the process of running trained models to generate responses, process documents, handle customer interactions, and execute agent workflows. While Nvidia's GPUs have historically been optimised for training AI models, inference is where enterprise value is created. Every API call, every document processed, every voice interaction, every agent task — that is inference.
The Groq 3 LPU is designed to complement GPUs, with one core optimised for accelerating GPU inference performance. Nvidia introduced a full rack system — the Groq 3 LPX — housing 256 LPUs, designed to sit beside the Vera Rubin rack-scale system shipping later this year.
For enterprises, a dedicated inference chip means faster AI responses at lower cost per operation. When your document processing pipeline, your voice AI system, and your agent workflows all run on hardware purpose-built for inference rather than repurposed training hardware, every operation gets faster and cheaper.
Our Minnato orchestration platform routes enterprise AI tasks across the best available infrastructure. As inference-specific hardware like the Groq 3 LPU enters production, Minnato can route inference-heavy workloads — document extraction, voice processing, agent task execution — to infrastructure optimised for those operations, automatically capturing the performance and cost improvements without changing any enterprise workflow.
2. NemoClaw: The Enterprise AI Agent Operating System
Nvidia officially launched NemoClaw — the open-source enterprise AI agent platform we covered when it leaked two weeks ago. But the GTC presentation revealed something more significant than the platform itself.
Huang called NemoClaw an operating system for AI agents. “It's no different than how Windows allowed us to make personal computers,” he said. The platform provides a reference software stack for businesses to deploy OpenClaw-style autonomous agents with enterprise-grade security, privacy, and governance controls.
NemoClaw addresses the critical enterprise concern around AI agents: they can communicate externally and execute actions without intervention. For consumer users experimenting with OpenClaw, that is exciting. For enterprises handling sensitive data, that is a security risk. NemoClaw provides the guardrails that make agent deployment viable in production environments.
Huang urged every company to adopt what he called an “OpenClaw strategy” — deploying AI agents as core operational infrastructure. He noted that Nvidia itself has “a whole bunch of OpenClaw running in the company, continuously running, doing things for us, writing, developing tools, developing software.” The compute demand from these continuous agents, he said, has skyrocketed.
For enterprises, the message is direct: AI agents are not a future technology. They are present-tense infrastructure. And the platform for deploying them securely at enterprise scale is now open-source and available.
3. DGX Spark and DGX Station: AI Factories on Your Desk
One of the most practically significant announcements for enterprise deployment was the pairing of DGX Spark and DGX Station with NemoClaw to create what Nvidia described as the ultimate platform for locally developing and deploying autonomous, long-running agents.
These systems bring “AI-factory-class performance directly to where intelligence is created — at the desk and inside the enterprise.” In practical terms: enterprises can now run sophisticated AI agents locally — not in the cloud, not in a remote data centre, but on hardware sitting in their offices.
For enterprises operating under data sovereignty requirements, in regulated industries, or with sensitive information that cannot leave the premises, this is transformative. The same agentic AI capabilities that require cloud infrastructure for most organisations can now run entirely on-premises, with enterprise-grade security, on hardware designed specifically for the task.
Our Vult document intelligence platform is architected for exactly this deployment model. Sensitive financial documents, legal contracts, and regulatory filings can be processed by AI agents running on local DGX infrastructure — with no data leaving the enterprise premises. The AI factory sits inside your security perimeter. The processing happens where the data lives.
Our Dewply voice AI platform can deploy on local infrastructure for environments where voice data — customer conversations, support interactions, confidential communications — must remain on-premises. The DGX Spark and Station systems provide the compute power to run voice AI agents locally with the same performance as cloud-based alternatives.
4. Nemotron 4 Coalition: Open Models Go Enterprise
Nvidia announced that its Nemotron 3 Ultra will be “the best base model in the world” — and that to build its successor, Nemotron 4, the company is creating a coalition with Black Forest Labs, Perplexity, Mistral, and Cursor.
This coalition model for developing frontier AI is significant for enterprise architecture. Rather than a single company building a proprietary model behind closed doors, multiple companies with different specialisations — search (Perplexity), lightweight models (Mistral), development tools (Cursor), and generative media (Black Forest Labs) — are collaborating on an open model that enterprises can access, deploy, and fine-tune.
The coalition approach produces models that are more versatile across enterprise use cases because they incorporate expertise from multiple domains. And the open nature means enterprises are not locked into a single provider's roadmap. They can fine-tune Nemotron models on their own data, deploy them on their own infrastructure, and switch to alternative models without rebuilding their entire AI stack.
This validates the model-agnostic architecture that is core to everything we build at Lynt-X. Our Minnato platform does not commit to a single model. When Nemotron 4 ships with frontier capabilities, Minnato can route appropriate tasks to it — while continuing to use Qwen 3.5 for cost-efficient local operations, GPT-5.4 for computer-use tasks, and specialised models for domain-specific workflows. The orchestration layer captures value from every model improvement, from every provider, automatically.
5. The $1 Trillion Signal: Infrastructure Is Not Slowing Down
Huang's $1 trillion order projection through 2027 — double the $500 billion estimate from just months ago — tells enterprise leaders something critical about the AI cost curve.
When $1 trillion flows into AI infrastructure over two years, the production capacity for AI compute expands dramatically. More chips in more data centres means more supply. More supply means more competitive pricing. More competitive pricing means lower inference costs for every enterprise consuming AI services through cloud providers.
The infrastructure investment also accelerates the technology curve. Vera Rubin systems shipping later this year deliver order-of-magnitude improvements in inference performance and energy efficiency compared to current-generation hardware. The Groq 3 LPU adds inference-specific acceleration on top of that. Photonics interconnects (which Nvidia invested $4 billion in earlier this month) further reduce data centre operating costs.
Each improvement compounds with the others. Faster chips connected by photonic interconnects running inference on dedicated LPUs — the combined effect is an AI cost curve that falls faster than any single improvement would suggest. Enterprise AI operations planned at today's pricing will be running at dramatically lower costs within 12 to 18 months.
For enterprises planning AI budgets, this means the business case for AI deployment improves continuously. Deploy now at today's costs, and your operations get cheaper every quarter as infrastructure improvements flow through to provider pricing. The enterprises that wait for costs to drop before deploying are the ones that fall behind competitors who deployed early and captured returns while costs were still declining.
What to Do This Week
Track the GTC announcements. The conference runs through March 19 with hundreds of sessions. Pay attention to partnership announcements, model releases, and enterprise deployment frameworks — they will shape the competitive landscape for the rest of 2026.
Evaluate local AI deployment. DGX Spark and Station systems paired with NemoClaw make on-premises AI agent deployment viable for the first time at enterprise scale. If data sovereignty, regulatory compliance, or security requirements have been barriers to AI adoption, assess whether local deployment removes those barriers.
Plan for the inference cost curve. The Groq 3 LPU, Vera Rubin systems, and photonics interconnects all point to dramatically lower inference costs within 12–18 months. Build your AI business case with the declining cost curve factored in — not just today's pricing.
Monitor the Nemotron 4 coalition. An open frontier model built by a coalition of Nvidia, Perplexity, Mistral, Cursor, and Black Forest Labs will create new options for enterprise AI deployment. Ensure your architecture is model-agnostic enough to capture value from it when it ships.
Start your OpenClaw strategy. Huang told every company at GTC to adopt an AI agent strategy. NemoClaw provides the enterprise-secure platform. DGX systems provide the local hardware. The infrastructure is ready. The question is whether your organisation has identified the workflows where autonomous agents deliver the most value.
“Jensen Huang just told 30,000 attendees from 190 countries that AI agents are the next computing platform, inference is getting its own dedicated chip, and $1 trillion in orders are flowing through 2027. For enterprises, GTC 2026 was not a keynote. It was the infrastructure blueprint for the next two years. The companies that build on this blueprint first capture the advantage that compounds with every quarterly improvement.”
The Blueprint Is Set
GTC 2026 was the moment enterprise AI infrastructure went from evolving to planned. The roadmap is now visible: dedicated inference chips (Groq 3 LPU), enterprise agent platforms (NemoClaw), desktop AI factories (DGX Spark and Station), coalition-built open models (Nemotron 4), and $1 trillion in hardware flowing into data centres worldwide.
Every enterprise consuming AI services will benefit from this infrastructure build-out — through faster inference, lower costs, and more capable models. The enterprises that design their architecture to capture these improvements automatically — through model-agnostic orchestration, multi-provider deployment, and governance-ready infrastructure — will compound the advantage with every quarter.
The blueprint is set. The infrastructure is shipping. The question is what you build on it.