Welcome, humans.
We begin today with a segment we call, “Dispatches from Corey's first GTC”…
In case you don’t know, GTC is NVIDIA’s major tech conference; you’ll read all about it below. Corey’s there live and in person! And he has thoughts:
First off, he says Jensen Huang is basically Henry Ford. If you're building the future, you're probably using his machines. NVIDIA already dominates AI training; the new Groq chip deal means the big ‘Vid is pushing to dominate inference, too. Jensen sees $1 trillion in demand through 2027. No other tech leader is staring at a market that big. And the real advantage = they're already designing multiple generations ahead. They're not selling GPUs anymore. They're quietly becoming the electrical grid of the AI economy. For more, read this.
Highlight reel: Corey met Peter Steinberger, creator of OpenClaw, who was a full-blown celebrity in the NemoClaw tent.
People were lining up left and right to talk to the guy (who just flew in from Vienna, as far as we understand, and also said multiple times he stayed up all night working on the NemoClaw project). Clearly the guy is so popular he can’t sleep, and will remain so now that Jensen basically said OpenClaw is the new Agentic Operating System. What’s the equivalent of buying someone a coffee, but for bedtime?? If you ever see him IRL, do that.
Here's what happened in AI today:
😼 NVIDIA announced seven new chips, five rack types, and an agentic AI operating system at GTC 2026
📰 Meta signed a $27B AI infrastructure deal with Nebius, then started planning sweeping layoffs
📰 GPT-5.4 hit 5 trillion tokens per day within one week, generating $1B in net-new annualized revenue
🍪 ElevenLabs launched a creative studio that generates full audio/video campaigns in one browser
🎓 How to make your AI prompts literally improve themselves overnight (using autoresearch)
P.S: Want to reach 675,000 AI-hungry readers? Click here to advertise with us.
😼 Everything NVIDIA Just Announced at GTC 2026: Seven Chips, Five Racks, One Giant Bet on Agentic AI
NVIDIA’s GTC conference kicked off yesterday, and it’s basically the company’s Super Bowl. Actually, at this point, it might as well be the official Super Bowl of the whole AI industry. No other event has convened as many industries and companies and AI influencers and VCs and startups all in one place. We say all this not to give you FOMO, but to clarify for non-tech people: this event is a big deal.
Now, last year at GTC, Jensen said he saw $500B in high-confidence demand. In this year’s keynote, he declared “at least $1 trillion through 2027.” He also dropped a stat that puts the whole era in context: 40 million times more compute in just 10 years.
The centerpiece of all this is the Vera Rubin platform, seven new chips and five rack types designed to function as one massive AI supercomputer. It pairs Rubin GPUs and Vera CPUs with the new Groq 3 LPX inference accelerator (the chip that will actually run your AI queries), delivering up to 35x higher inference throughput per megawatt. Actually, make that 50x.
Dylan Patel from amazing AI chip analyst firm SemiAnalysis (who got an epic shout-out; see below) ran his own benchmarks and accused Jensen of "sandbagging" (deliberately lowballing the number so the real results look even more impressive). Jensen's response? "He's not wrong."
What does that make NVIDIA? Basically, it makes NVIDIA the Inference King…
Here's what else dropped:
NemoClaw: Jensen called OpenClaw "the operating system for personal AI" and compared it to Linux and HTML. NemoClaw wraps it with enterprise security (OpenShell) so agents can't leak sensitive data or execute code they shouldn't. More on this below.
Dynamo 1.0: Now in production as the "operating system" for AI factories, boosting Blackwell inference by up to 7x. AWS, Azure, Google Cloud, and Oracle all onboard.
Nemotron Coalition: A new alliance with Mistral, Cursor, Perplexity, Black Forest Labs, and others to build open frontier models. Nemotron 3 is already top-3 on the OpenClaw leaderboard.
Robotics: Robotaxis launching with Uber across 28 markets by 2028 (BYD, Hyundai, Nissan, Mercedes, Toyota, GM). Jensen called AVs "the first multitrillion-dollar robotics industry." Disney's Olaf robot, trained in NVIDIA's Newton physics simulator, walked out on stage and got roasted by Jensen (“I thought you’d be taller”).
DLSS 5: Neural rendering this fall. Hollywood-level visuals in real time.
Space-1: A Vera Rubin Module going to orbit. AI in space. Not a metaphor.
Why this matters to you (AI Economics 101): Jensen showed a chart he says every CEO will study. Here’s the TL;DR:
Your data center has a fixed amount of power. It will never get more.
So the only number that matters is tokens per watt (how much AI output you squeeze from each watt). More tokens per watt = more revenue.
Vera Rubin produces 5x more revenue per gigawatt than Blackwell. With Groq, 35x at the premium tier.
Jensen predicts tokens will be priced in tiers from free to $150/million, and every engineer will soon get an annual token budget alongside their salary.
Jensen's clearest line of the day: "Every single SaaS company will become an AGaaS company", or an agent-as-a-service company, and that everyone needs an OpenClaw strategy.
What you can do right now: NemoClaw is basically that strategy. OpenClaw but make it safe is huge for businesses. You can deploy your own instance on NVIDIA's Brev console for $0.13/hr ($3.12/day), or try it free via OpenShell. The NemoClaw GitHub repo also has everything you need; just give those two repos to your AI and ask it to help you set it up.
In sum, the enterprise IT industry is being rebuilt from tools-for-humans to agents-that-do-the-work, and NVIDIA wants to own every layer of that stack. So what, does NVIDIA just own everything now? Guess that makes Jensen not just the Inference King, but the AI King!
FROM OUR PARTNERS
Something BIG is coming to SmartBear on March 18
AI has drastically increased development speed, but maintaining quality is harder than ever. On March 18 at 11 AM ET, SmartBear shares their solution. Join the livestream reveal with DOOM co-creator John Romero, plus a live Q&A and exclusive first look at the future of software quality.
🎓 AI Skill of the Day: Your AI prompts can now improve themselves overnight
Nick Saraev showed how to combine Claude Code skills with Andrej Karpathy's autoresearch methodology to improve your AI Skills, and it’s pretty universally applicable.
The idea: write simple yes/no evaluation criteria for your prompt (an "eval suite"), have an agent generate outputs every few minutes, score them, mutate the prompt, keep the winner. Repeat until near-perfect.
His diagram generator skill went from 32/40 to 39/40 (97.5%) within a few runs at about $0.20 per test cycle. His website optimization test cut load times from 1,100ms to 67ms over 67 experiments.
The key is defining the right binary evals: avoid Likert scales (1-7 ratings create compounding probability noise) and keep criteria simple enough the model can't game them. Here's a prompt to set it up for any skill you have:
Read this repo: https://github.com/karpathy/autoresearch
I want you to use the autoresearch convention to build a self-improving system for my [SKILL NAME] skill.
Eval criteria (binary yes/no):
1. [YOUR CRITERION 1]
2. [YOUR CRITERION 2]
3. [YOUR CRITERION 3]
4. [YOUR CRITERION 4]
Every 2 minutes, generate 10 outputs, evaluate all 10 against these criteria, count the score out of [TOTAL], iterate the prompt, and keep the winner. Run until you hit [TARGET] or higher.
Here’s a good tip from Nick: The list of failed experiments is almost as valuable as the final prompt. You can hand it to the next, smarter model and it'll pick up where the last one left off. Your research compounds even when the models get replaced
Want more tips like this? Check out our AI Skill of the Day Digest for this month.
Have a specific skill you want to learn? Request it here.
🍪 Treats to Try
ElevenCreative generates, edits, and localizes studio-grade audio/video content in one browser by mixing voices, music, sound effects, and video models, turning a single prompt into campaigns dubbed in 70+ languages —no pricing details.
GLM-5-Turbo (NEW)— Z.AI's agent-optimized model, 200K context, 128K output, outperforms GLM-5 on tool use and long-chain tasks
Mistral Small 4 (NEW)— 119B open-weight MoE with fused instruct/reasoning/agentic skills, NVFP4 and speculative decoding checkpoints for faster throughput
Adaptive Computer runs always-on AI agents that connect your tools (Gmail, Slack, Sheets, Notion) via one-click OAuth, build workflows, and handle everything from spreadsheet uploads to daily sales reports with human approval on sensitive steps—free month to try.
Lossless Claw fixes OpenClaw's biggest annoyance: forgetting what you were working on mid-session; it replaces the default sliding-window compaction with a DAG-based system that persists every message so your agent can drill back into anything it summarized (OpenClaw creator Peter Steinberger apparently recommended it over his own built-in memory; Ray Fernando demos the install here —free (code).
OpenAI released Subagents in Codex so you can spawn specialized parallel agents (default/worker/explorer or custom TOML configs) that keep main context clean, tackle subtasks simultaneously, and merge results.
Google Labs built the Stitch SDK so you (or your agents) can programmatically generate, edit, and extract HTML + screenshots of UI screens from natural-language prompts with project management, variants, and Vercel AI SDK integration.
📰 Around the Horn
Meta signed a deal worth up to $27B over five years with Nebius for dedicated AI compute capacity (including first Nvidia Vera Rubin chips at scale) while simultaneously planning sweeping layoffs as AI costs mount.
Greg Brockman announced GPT-5.4 is processing 5 trillion tokens per day within one week of launch, exceeding all prior API volume and generating $1B in annualized net-new revenue.
Alibaba created a new Token Hub unit consolidating all AI products under CEO Eddie Wu and is launching an agentic AI service for enterprises this week.
FROM OUR PARTNERS
Your AI is resolving tickets. Is it keeping customers?
Resolution rates look great. But Gladly's 2026 Customer Expectations Report reveals the metric most CIOs are missing — and what the data says about where AI investments actually translate into retention, not just throughput.
🔧 Tuesday Tool Tip:
Free GPUs and AI APIs most people don't know exist. NVIDIA's build.nvidia.com is the most underused resource in AI right now. You can:
Run free inference on frontier models (GLM-5, Kimi K2.5, Nemotron 3, MiniMax) with zero setup, just an API call
Launch GPU instances (B300, B200, H200, RTX PRO 6000) for prototyping in a sandbox
Deploy NemoClaw with one click via OpenShell to test secure AI agents
Browse Blueprints with workflow templates and code samples to build full AI apps from scratch
Free to start, runs in your browser. If you're building anything with AI, bookmark this before everyone else discovers it.
A Cat’s Commentary
 | That’s all for now.
|
