A couple days ago I wrote a post Nvidia GTC Starts Monday... about what I was watching for at Nvidia GTC as a dev. If you read it, cool. If not, the short version: I had five things I cared about Groq integration, NemoClaw, Vera Rubin chips, open models, and a speculative ARM CPU thing.
Jensen just finished the keynote. Two hours. He lifted server racks on stage, brought a walking Olaf robot from Frozen out, and closed the whole show with AI-generated robots singing country music around a campfire.
I need to talk about what actually happened.
The Groq thing landed harder than I expected
In my preview post I wrote that the main question was how Groq's tech would fit into Nvidia's ecosystem. Would it be a side feature? A rebrand? Something bolted on?
Turns out it's more like built in instead of bolted on than I imagined. They announced the Groq 3 LPU, shipping Q3 this year, sitting inside the Vera Rubin platform as a dedicated token accelerator alongside the GPUs. Jensen threw out a number: 35x more throughput per megawatt when you pair them together.
I don't know about you, but when I think about whether my side projects can afford real inference costs, a 35x efficiency jump changes that conversation entirely. I was hoping for "significantly cheaper." This is more like a different cost universe. Whether Nvidia actually passes those savings on to developers or just enjoys fatter margins is a different question but the hardware capability is real.
NemoClaw isn't what I thought it was (and that's fine)
I assumed NemoClaw would be another agent framework competing. Nope. It's an enterprise security layer for OpenClaw the agentic AI platform that's been blowing up lately.
If you are living under a rock or a in a cave you may not have heard of OpenClaw: it was built by an Austrian dev named Peter Steinberger, went massively viral, and the creator joined OpenAI last month. It lets you build AI agents that actually do things communicate via Telegram, execute tasks, operate on their own.
Jensen called it one of the most important open-source developments in 30 years. He compared it to an operating system. NemoClaw is Nvidia's way of making it safe for companies one command install, secure models, protections against agents leaking sensitive data.
Honestly, this is smarter than building a competing framework. Nvidia doesn't want to own the agent layer. They want to be the platform everything runs on. That's been their playbook since CUDA and it keeps working.
He also dropped a new acronym AGaaS. Agents as a Service. Said every company needs an agentic system strategy and that IT is shifting from SaaS to AGaaS. I'm filing that under "probably right but definitely too early to put on a slide deck at work without getting laughed at."
Vera Rubin is not what I was picturing
I was thinking about this as a next-gen GPU. Jensen presented it as a next-gen everything. Seven chips, five rack-scale systems, one supercomputer. A new CPU designed from scratch for agent workloads. A reinvented storage architecture. The Groq LPU plugged in as a token accelerator. The whole thing purpose-built as one vertically integrated system.
10x more performance per watt compared to the previous generation. Samples already shipping. Broader availability second half of this year.
The thing that stuck with me: Jensen specifically described how agents "pound on memory really hard" and need completely different storage systems than traditional workloads. He's not designing hardware for chatbots. He's designing hardware for autonomous systems that run for days, hammering databases and calling tools continuously. The entire platform is architected around that use case.
Also he announced Vera Rubin Ultra connecting 144 GPUs, and a Space-1 module for putting data centers in orbit. I wrote in my preview post that I might "look silly for bringing up" speculative stuff. I did not, however, anticipate needing to have an opinion about space data centers. I do not have one. Moving on.
The open models play was exactly what I expected
I wrote that Nvidia would position itself as the open ecosystem's best friend, especially with Meta's Avocado situation making the future of Llama uncertain. That's exactly what happened.
Jensen announced the Nemotron Coalition Perplexity, Mistral, Black Forest Labs, Cursor, and others working together on open frontier models. He said Nemotron 3 Ultra will be "the best base model in the world."
He didn't mention Meta once. Didn't need to. The whole announcement was "open models matter and here's who's making sure they keep existing." Nvidia sells more GPUs when more people train and run open models. Their incentives are perfectly aligned with this ecosystem. If you're worried about Meta pulling back on open-source AI (and you should be paying attention to that), the Nemotron Coalition is meaningful.
There's also an open models panel on Wednesday with the LangChain, Cursor, and A16Z folks. That might be more interesting for developers than today's keynote was.
The stuff I got wrong
I said gaming wouldn't show up at GTC. Then Jensen announced DLSS 5 what Nvidia is calling the biggest graphics breakthrough since the original DLSS in 2018. It combines traditional rendering with generative AI to produce photorealistic lighting and materials in real-time. Shipping this fall.
I'm not a game dev so I can't evaluate the claims, but the demo looked genuinely impressive and the underlying tech has implications beyond gaming for anyone working with real-time visualization.
I also speculated about ARM CPUs for consumer laptops and desktops. That didn't happen. The Vera CPU is ARM-based but purely for data centers. No consumer chip reveal. Speculation is speculation, and sometimes it's just wrong.
The one thing I keep coming back to
Jensen said he expects purchase orders for Blackwell and Vera Rubin to hit $1 trillion through 2027. Last year that number was $500 billion. It doubled in a year.
AI is now 60% of Nvidia's revenue. This isn't a company that does AI on the side. This is an AI company that also does other things.
Here's what I wrote in my preview: "The hardware dictates the economics. And the economics dictate what we can build." After today, the economics are shifting. Inference is getting radically cheaper. Agent workloads are getting purpose-built hardware. Open models are getting institutional backing.
Whether that translates into better tools for people like us or just bigger numbers on Nvidia's earnings call that's something we have to wait and watch.
For now, I need to process the fact that the keynote ended with robots singing country music about tokens and open-source software. That happened. In real life. In 2026. At a professional conference.
We live in interesting times.
This is a follow up to my previous preview post from Saturday.
Find me at @uncaughtex on X I catch the exceptions nobody handles.
Top comments (0)