Listen to the audio version of this article (generated by AI).
Editor’s note: “Slower AI Models Could Mean Faster Growth Somewhere Else” was previously published in August 2026 with the title, “Nvidia Took Claude From 30% to 100% – Without Building a Better Model.” It has since been updated to include the most relevant information available.
On one of AI’s toughest interactive tests, Claude solved fewer than one in three challenges.
Then Nvidia (NVDA) surrounded the same model family with better memory, outside tools, a structured work loop, and a supervisor that stepped in whenever Claude got stuck.
The system solved every challenge – and its score jumped from 30.2% to 100%.
Nvidia changed other parts of the test setup, too, so this was not a perfect apples-to-apples comparison. Still, the result points toward a major shift in the AI market.
We first covered this result last month. It matters even more now.
With the industry’s safety push dominating headlines – and some labs signaling a more deliberate pace on their most capable releases – the central question in AI has changed: how much value can be unlocked around the models that already exist?
According to Nvidia, a lot.
Today’s best models may already contain more useful intelligence than their surrounding systems allow them to express. And the companies that turn that capability into dependable, cost-effective – and now, provably safe – workflows could become some of the next major AI winners.
Powerful AI Agents Still Need Guardrails
Today’s AI agents are pretty good at sprints. Marathons are where they lose the plot.
Give Claude or GPT a short assignment. It can browse the web, write some code, pull information from another app, and come back with an answer.
But long projects are a different story. Imagine asking an agent to migrate a company’s financial software, redesign a supply chain, or optimize thousands of lines of GPU code. The work may require hundreds of decisions made over hours or days.
Frontier models still struggle with that kind of sustained work. They lose context. Repeat old mistakes. Chase dead ends. Occasionally become so confused that they damage the project they were supposed to complete – and we’re left to try to clean up the mess.
Nvidia’s Agentic Variation Operators system, or AVO, was designed to keep that from happening.
AVO gives the model persistent memory and its own repeated cycle of planning, acting, testing, and revising. Nvidia also added a supervising agent that watches the work and nudges the main agent when it gets stuck or starts exploring an unproductive path.
Think of it as a talented employee with a good project-management system and an experienced boss nearby.
The intelligence was already there.
Nvidia helped it stay organized long enough to finish the job.
AI Safety Is Becoming Its Own Infrastructure Layer
For the first few years of this boom, the model leaderboard commanded almost all the attention.
Which lab had the best reasoning score? Which model had the largest context window or wrote the cleanest code?
Those are still important questions; better models still retain a huge competitive advantage.
But the new safety push is expanding the market around the model.
Independent evaluators can’t do their jobs from the outside. They need access – model checkpoints to probe, environments to test in, and the logs that show how a system actually behaved. Companies deploying agents need something different: control. An audit trail behind every action, a clear point where a human takes over, and software that can stop or redirect an AI the moment something goes wrong.
Those safeguards have to remain active after the model launches, too.
A company will want to know:
- What information the agent accessed
- Which tools it used
- Why it took a particular action
- When a human should intervene
- If the system can recover safely after a mistake
That requires real infrastructure: software that watches the agent, systems that secure it, memory that records what it did – plus simulation platforms and an enormous amount of testing.
It also requires more compute.
Each independent evaluation runs the model again. Each monitor adds another layer of processing. Simulations, safety checks, recovery loops… all of it consumes more infrastructure.
The labs may take more time before releasing their most capable systems.
But the work surrounding those systems – proving them safe before launch, watching them after – keeps expanding.
AI Safety Gets More Serious When Machines Start Moving
AI safety becomes a much bigger issue when intelligence leaves the screen.
A coding assistant can generate a bad line of software. A human can review it, reject it, and run the task again.
A robot operating on a factory floor deals with physical consequences.
It may be carrying a heavy part through a crowded warehouse, inches from expensive equipment – and people. A bad decision can cause serious damage.
That raises the bar dramatically.
A useful robot needs far more than a capable model.
It needs cameras and sensors to understand its surroundings. Control software has to translate a decision into precise movement. Simulation tools must expose the system to unusual situations before it encounters them in the real world.
The robot also needs a plan for the moments when things go sideways. If it loses track of an object, it has to know how to reacquire it. If a person steps into its path, it has to stop or reroute instantly. It has to recognize when it’s out of its depth and call for help. And afterward, it has to be able to prove it acted safely at every step.
Those requirements are becoming central to commercial robotics.
Factories and warehouses are not waiting for a robot that can do everything a person can do. They need machines that can perform a handful of useful tasks reliably, repeatedly, and safely.
Stronger AI guardrails can help unlock those deployments. That makes safety more than a regulatory cost.
Better AI Infrastructure Changes Both Safety and Economics
Performance is only half the equation, though. Companies also care what it costs to finish the job.
As Databricks CEO Ali Ghodsi explained to TechCrunch, two agent systems built around the same model can produce dramatically different bills. Choose the wrong setup, and the same task may cost roughly twice as much to complete.
A clumsy workflow sends a routine job to an expensive frontier model when a smaller one would do. Poor memory leads the system to reread huge amounts of old information again and again.
A better setup preserves what matters, sends each task to the right model, and eliminates unnecessary loops.
Safety systems add work of their own. But they can also unlock far more valuable tasks.
A company may happily accept a little more processing overhead if the result is an agent – or robot – it can trust with meaningful work.
Better systems can lower the cost of each completed task and reduce expensive failures, making more jobs worth automating.
In another Nvidia experiment, AVO tested more than 500 approaches to improving a piece of GPU software and saved 40 separate versions before beating a leading implementation by as much as 10.5%.
The agent kept experimenting, checking, and revising until it found a better answer.
Every loop consumed compute. Every successful result made that compute more valuable.
Frontier releases may slow at the margin. The market around them won’t.
Physical AI Startups Are Building the Layer Between Models and Machines
The public market remains focused on the largest model labs and the companies supplying their chips.
But much of the work required to bring AI safely into the physical world is happening inside smaller, private companies.
They are collecting robot-training data.
Building simulation software.
Developing machine vision, control systems, and safety layers.
Creating the tools that let robots learn new tasks and operate around people.
This is the “second layer” of the AI boom: companies taking raw intelligence and applying it to specific industries, factory floors, warehouses, and machines.
The safety turn could make that layer even more valuable.
A more deliberate frontier race gives robotics companies time to improve reliability, integrate monitoring, and turn today’s models into systems that businesses can actually deploy.
That brings me to one private company I have called the “Nvidia of Robotics.”
For a limited time, everyday investors can claim a stake with as little as $500. But the opportunity is set to close to new investors on Monday, Sept. 21.
No paywall. No credit card required.
