AI in 15 — March 17, 2026

One trillion dollars. That's how much Jensen Huang says is coming NVIDIA's way through 2027. Not revenue. Purchase orders. Let that satisfying number sink in for a moment.

Welcome to AI in 15 for Tuesday, March 17, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Marcus, we promised full GTC keynote coverage and today we're delivering. Jensen didn't disappoint. Seven new chips, five rack-scale systems, a supercomputer, and somehow he also found time to announce a desktop AI machine that fits under your desk. Beyond GTC, Mistral dropped two major releases at once, Karpathy's weekend project broke the internet, and a CMU study just put hard numbers on what developers have been feeling about AI coding tools. Let's get into it.

NVIDIA's Vera Rubin platform redefines AI computing with ten times the performance per watt.

Mistral releases a formal verification agent and a unified small model on the same day.

And a Carnegie Mellon study finds AI coding assistants boost speed but degrade code quality by forty-two percent.

So Marcus, as we previewed Sunday and Monday, GTC was the event to watch. We talked about Vera Rubin, the surprise chip, the CPU pivot. Now we have the full picture. And it's even bigger than expected.

Significantly bigger. Let's start with the headline number. Jensen said he expects purchase orders between Blackwell and Vera Rubin to reach one trillion dollars through 2027. That's double last year's five hundred billion projection. He opened by saying computing demand has increased by one million times over the last few years. Even accounting for Jensen's flair for dramatic numbers, the scale here is unprecedented.

Walk us through Vera Rubin itself. We speculated about five-x inference gains on Monday. What did we actually get?

Ten-x performance per watt over Grace Blackwell. The Vera CPU has eighty-eight custom Olympus cores with something called Spatial Multithreading, plus over a terabyte per second of memory bandwidth. A single rack of two hundred and fifty-six liquid-cooled Vera CPUs can sustain over twenty-two thousand concurrent CPU environments. Jensen's framing was telling. He said the CPU is no longer simply supporting the model, it's driving it. That's a philosophical shift. NVIDIA is saying the future of AI isn't just about GPUs anymore. It's about the full stack working together for agentic workloads.

And the surprise chip we were speculating about?

That would be the Groq 3 Language Processing Unit. Remember NVIDIA acquired the Groq startup for twenty billion dollars back in December. This is the first product from that acquisition. A Groq 3 rack holds two hundred and fifty-six LPUs and sits alongside Vera Rubin racks, together delivering up to thirty-five times higher inference throughput per megawatt. Jensen literally said "NVIDIA is the inference king." He's integrating the acquisition at remarkable speed.

There was also a deskside supercomputer announcement that caught my eye. The DGX Station.

Seven hundred and forty-eight gigabytes of memory, twenty petaflops, supports models with up to one trillion parameters. And it sits under your desk. Two years ago that kind of compute required a data center rack. Now it's a workstation. For AI researchers and enterprise teams that need to run frontier-scale models on-premises for security or latency reasons, this changes what's possible.

And then there's the roadmap. He previewed the Feynman architecture for 2028.

Which tells the market that NVIDIA has at least two more generations planned. Feynman will feature the LP40 next-gen LPU, the Rosa CPU, and BlueField-5 networking. When you lay out a public roadmap that far ahead, you're telling every competitor and every customer: we're not slowing down, plan accordingly. The cloud partnerships are equally telling. AWS is deploying over a million NVIDIA GPUs. Alibaba, ByteDance, Meta, Oracle, CoreWeave, Lambda are all in for Vera Rubin. Even Cursor confirmed they'll use Vera CPUs for improved throughput.

We said Monday that when Jensen speaks, roadmaps change. Did he deliver?

He delivered a roadmap that makes everyone else's obsolete. The integration of Groq for inference, Vera for agentic computing, NemoClaw for secure agent deployment, and even a Space-1 architecture to put AI in orbit — NVIDIA isn't just selling chips anymore. They're architecting the entire AI infrastructure layer from orbit to your desk.

From hardware to models. Mistral had quite a day on Sunday, dropping two major releases simultaneously. Marcus, let's start with the one I find more fascinating. Leanstral.

Leanstral is genuinely novel. It's the first open-source AI agent built specifically for Lean 4 formal verification. Instead of just generating code and hoping it works, Leanstral can mathematically prove that implementations meet strict specifications. It uses a sparse architecture, only six billion active parameters out of a hundred and twenty billion total, which makes it incredibly efficient.

And the benchmarks are striking. It outperforms models that are orders of magnitude larger?

On the FLTEval benchmark, Leanstral scores twenty-six point three at pass-at-two, beating models with seven hundred and forty-four billion parameters and even one trillion parameter models. And the cost difference is dramatic. Thirty-six dollars versus five hundred and forty-nine dollars for comparable tasks with Anthropic's Sonnet. Model weights are Apache 2.0, so anyone can run it. Mistral is calling this "trustworthy vibe coding" which is a clever rebrand of formal verification for the AI age.

For people who aren't familiar with formal verification, why does this matter?

Think of it this way. Right now when AI writes code, you test it and hope the tests catch bugs. With formal verification, you mathematically prove the code is correct. For safety-critical systems — medical devices, autonomous vehicles, financial infrastructure — that's the difference between "it probably works" and "we can prove it works." If Leanstral makes formal verification accessible to mainstream developers, that's potentially transformative.

And the second release, Mistral Small 4?

A hundred and nineteen billion parameter Mixture-of-Experts model that unifies what were previously separate models for reasoning, vision, and coding into a single deployment. It has a two-hundred-and-fifty-six-thousand-token context window and a per-request reasoning effort parameter, so developers can dial up thinking time for hard problems and dial it down for simple ones. Forty percent faster, three times more requests per second, and also Apache 2.0. Mistral is basically daring you not to use their models by removing every barrier.

Andrej Karpathy, OpenAI co-founder, former Tesla AI director, published what he called a "Saturday morning two-hour vibe coded project" that promptly went viral. It's an interactive visualization scoring every US occupation for AI exposure.

Three hundred and forty-two occupations covering a hundred and forty-three million jobs, each scored on a one-to-ten scale. And the headline finding upends the traditional automation narrative. The highest-paying jobs are the most exposed. Professions earning over a hundred thousand dollars averaged a six point seven exposure score. Those earning under thirty-five thousand scored just three point four.

So it's the white-collar knowledge workers, not the blue-collar workers, who are most at risk?

Exactly the opposite of what people feared for decades. Software developers, data scientists, financial analysts, paralegals, writers, graphic designers — all scored nine out of ten. Construction laborers, roofers, janitors, ironworkers — scored one. The treemap lets you toggle between AI exposure, BLS growth projections, median pay, and education requirements. The patterns are stark.

Elon Musk chimed in saying "all jobs will be optional." But the Hacker News discussion was more nuanced than that.

Much more. People pointed out that the BLS growth projections don't account for AI displacement at all. Someone noted there are as many jobs in the "top executives" category as in "retail sales worker," which is a fascinating structural observation. And curiously, Karpathy's GitHub repo for the project was deleted within hours, though the site stayed up. Coming from someone who literally helped build the AI systems creating this disruption, the visualization carries a weight that no government report or think-tank paper could match. It makes the abstract viscerally real.

Now for a study that connects directly to the developer frustration we've been covering all week. Carnegie Mellon analyzed over eight hundred GitHub repositories that adopted Cursor and compared them against nearly fourteen hundred that didn't. Marcus, what did they find?

The speed boost is real but temporary. Cursor adoption produced a statistically significant increase in development velocity that faded over time. Meanwhile, static analysis warnings increased by thirty percent and code complexity increased by forty-one point six percent. And here's the critical part — those quality degradations were persistent. They didn't fade. So you get a short-term speed hit that evaporates, and a long-term quality problem that sticks around.

That's essentially the opposite of what you'd want.

It creates a vicious cycle. The initial speed gains are eroded by the technical debt the tool itself generates. The researchers used rigorous econometric methods to show that the increased warnings and complexity directly drive long-term velocity slowdown. One Hacker News commenter captured the counterargument nicely — AI simultaneously increases code complexity and reduces the cost of dealing with that complexity, since you can use AI to understand and refactor it. But another commenter made the sharper point: "The agent can introduce something that will make you rip your hair out in six months, but tests are green today."

This is the most rigorous study we've seen on this, and it validates what Rob Englander argued in that viral essay we covered Sunday — AI makes the easy part easier but doesn't help with the hard part.

And it quantifies what we discussed Monday about sloppypasta. The seventeen percent reduction in developer skill mastery, the forty-two percent increase in code complexity — these numbers tell a consistent story. AI coding tools are powerful accelerants that require expert supervision. Without that supervision, you're not building faster. You're accumulating debt faster.

Quick update on xAI. As we reported Saturday, the co-founder exodus continues. The count is now nine of eleven original co-founders gone.

And Musk publicly saying xAI was "not built right first time around" while scrambling to hire from Cursor tells you how desperate the situation is. The ambitious Macrohard project, their plan to build an AI that can do anything a white-collar worker can, has been paused after its leader departed. At a company valued north of fifty billion dollars, losing ninety percent of your founding team isn't normal turnover. It's a crisis.

Especially with GTC showing the rest of the industry accelerating.

That's the timing problem. Every day xAI spends rebuilding is a day Anthropic and OpenAI pull further ahead in exactly the coding and agentic capabilities Musk cares about most.

Tuesday big picture. NVIDIA just laid out a trillion-dollar roadmap. Mistral is making formal verification accessible. Karpathy showed us exactly which jobs AI is coming for. And CMU proved that AI coding tools create as many problems as they solve. Marcus, what's the thread?

The infrastructure is extraordinary and accelerating. NVIDIA's roadmap through Feynman in 2028 guarantees the hardware keeps getting better. But the human layer isn't keeping pace. Karpathy's visualization shows who's exposed. The CMU study shows the tools aren't a clean win. xAI shows that even billions of dollars can't substitute for getting the fundamentals right. The technology is racing ahead. The question, as it has been all week, is whether we build the discipline, the judgment, and the institutions to use it well.

Jensen promised ten times more performance per watt. Maybe we should demand ten times more thoughtfulness per deployment.

Now that would be a real breakthrough.

That's your AI in 15 for Tuesday, March 17, 2026. See you tomorrow.