Nvidia GTC: Thinking Machines Gets 1GW, DGX Spark Debuts
Jensen Huang kicks off GTC 2026 with a gigawatt-scale partnership with Mira Murati's startup and new DGX Spark systems for developers.
Jensen Huang delivered exactly what the market wanted to hear.
Nvidia's GTC 2026 keynote Monday featured the company's biggest partnership announcement to date: a multiyear agreement with Thinking Machines Lab to deploy at least one gigawatt of Vera Rubin systems. That's frontier-lab scale compute going to a startup that didn't exist 18 months ago.
The conference, running March 16-19 in San Jose with 30,000 attendees, is Nvidia's annual showcase for where AI is heading. This year's message was clear: the inference era has arrived, and Nvidia intends to own it the same way it dominated training.
Thinking Machines Gets Frontier Scale
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, just became one of the largest AI compute customers on the planet. A gigawatt of capacity—a threshold only OpenAI, Google, and a handful of others have approached—signals Murati's team is building frontier models to compete at the highest level.
The partnership includes a "significant investment" from Nvidia, though neither company disclosed the amount. Deployment on Vera Rubin is targeted for early 2027, the same timeline Meta's Nebius deal follows.
For Nvidia, it's a strategic bet on the next generation of AI labs. Murati has credibility from her OpenAI tenure, and backing her startup creates another dependency on Nvidia hardware—plus whatever model architectures Thinking Machines develops that might favor Nvidia's compute characteristics.
DGX Spark Brings AI Local
The hardware surprise was DGX Spark, a new accelerated computing system available for purchase at the Nvidia Gear Store and Micro Center. It's aimed at developers who want to run AI workloads locally rather than relying entirely on cloud APIs.
The local-first positioning matters. OpenClaw, Nvidia's agentic AI framework, got significant keynote time. Huang described it as "the fastest-growing open source project in history"—hyperbole, perhaps, but the message is serious. Nvidia wants developers building persistent AI agents that run continuously, interacting with files, apps, and workflows.
DGX Spark plus OpenClaw creates an entry point for enterprise developers to prototype agent systems before scaling to cloud infrastructure. It's the same playbook that made CUDA dominant: give developers tools that lock them into Nvidia's ecosystem.
Vera Rubin Timeline Confirmed
The keynote confirmed what analysts expected: Vera Rubin is on track for production deployments in early 2027. The pre-GTC preview covered the architecture details—288GB HBM4, massive bandwidth improvements, 5x Blackwell performance.
Multiple customer announcements this week suggest Nvidia has strong visibility into demand. Between Thinking Machines, Meta via Nebius, and others, significant Vera Rubin capacity is already committed before chips reach volume production.
Market Reaction
Nvidia shares rose 2% in pre-market trading, a muted response given the keynote's scale. But the stock already ran into GTC—it's up 12% in March despite broader market weakness from oil price volatility.
The more important signal is what didn't happen: no downside from the Groq licensing deal or competitive pressure on inference costs. Nvidia's response—dedicated inference silicon and OpenClaw—addressed the DeepSeek concerns that rattled the stock in January.
What Comes Next
GTC runs through Wednesday with sessions on physical AI, robotics, and AI factories. The conference has historically been a catalyst for semiconductor stocks broadly—analyst upgrades tend to follow as firms digest the product roadmap.
Key questions remaining: Rubin pricing, inference chip specifications, and enterprise adoption metrics for OpenClaw. The answers will filter out over the week, giving traders multiple opportunities to position around the news flow.