> ## Content Index
> Fetch the complete content index at: https://longtalkshort.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Hard Takeoff Has a Supply Problem: Bob McGrew on Why Superintelligence Needs Factories
- URL: https://longtalkshort.com/the-hard-takeoff-has-a-supply-problem-bob-mcgrew-physical-world/
- Published: 2026-11-20T02:15:18.000Z
- Updated: 2026-11-20T02:15:18.000Z
- Description: Former OpenAI research chief Bob McGrew tells No Priors that a hard AI takeoff fails because GPUs, labs, and factories — not researchers — are the bottleneck. He runs Arda, which sells into that problem. Not investment advice.
- Author: Truman
- Tags: Science, Bob McGrew, No Priors, superintelligence, manufacturing

If you have been waiting for AI to remake the world you can walk through, not just the one you type into, Bob McGrew’s claim is that you are still waiting — and that superintelligence, if it arrives, has to wait on the same thing. The former OpenAI chief research officer tells the hosts of No Priors that you will not get there “without changes in the physical world.” Chatbots and software agents have moved fast. Factories, wet labs, and the supply of GPUs have not. The interesting part of the conversation is not a product pitch. It is why he thinks the popular story of a hard takeoff is “straightforwardly wrong.”

*The episode is “Superintelligence Needs the Physical World with CEO and Co-Founder of Arda Bob McGrew,” on No Priors, at* [*youtube.com/watch?v=\_wLzJINt9iQ*](https://www.youtube.com/watch?v=%5FwLzJINt9iQ&ref=longtalkshort.com)*. Sarah Guo and Elad Gil host. McGrew led product and engineering at Palantir through its first decade and led research at OpenAI, leaving as chief research officer in 2024\. This article stays with his argument about physical bottlenecks, not with a tour of the company. Auto-generated captions were used for wording; obvious names have been corrected from the official title and description.*

**Disclosure:** McGrew is the chief executive and co-founder of Arda, a venture-backed company that sells software for manufacturing and counts Anduril as its first design partner and customer ([Arda’s launch announcement](https://www.businesswire.com/news/home/20261020574393/en/Arda-Emerges-from-Stealth-to-Speed-Up-American-Manufacturing?ref=longtalkshort.com)). His argument that the next bottleneck is physical manufacturing is also the case for his company. That interest is a reason to hear him, and a conflict to keep in view. Arda comes up here only through the one example he gives.

*In short: McGrew argues that a “hard takeoff” fails because researchers are not the bottleneck — GPUs, wet labs, and factories are — so automated research is a large one-time gain, not a runaway loop. He thinks the leap that matters is AI acting in flexible, high-mix manufacturing, not a consumer humanoid. These are one founder’s views. This is not investment advice.*

His picture of the last few years is generous about software and blunt about matter. “We’ve seen a huge amount of improvement” from chatbots and agents, he says. A Waymo, as a host puts it, is only “halfway towards the physical world,” because “you didn’t change the car.” You are still “bounded by all of the physical barriers.” McGrew’s wager is that the next leap is systems that can act in factories, labs, and supply chains — and that superintelligence itself depends on that move.

![A steep digital-capability line rising inside a data-center box, and a nearly flat line on the factory floor and lab bench, with a gap labeled ‘the physical world.’](https://storage.ghost.io/c/d7/93/d793216c-cf13-441f-bdae-4f3aef998c0e/content/images/2026/11/cover-338.png)

His claim is not that the models are weak. It is that they have barely begun to change the world of objects, and that the world of objects is what bounds the next jump.

## 1\. Why he thinks the hard takeoff story fails

The argument he has been hearing “since like 2018,” he says, is a clean one. Automate 10% of a researcher’s time and you go a little faster. Automate 50% and you have a 2x. Automate 90% and you have a 10x; automate 99% and you have a 100x. Keep going and you get “a runaway acceleration,” a singularity, a hard takeoff. “I think this is wrong,” he says. “This argument is just straightforwardly wrong.”

The reason, in his telling, is that researchers are not the bottleneck. “The amount of time it takes in order for you to discover something is not primarily bounded by the number of researchers you have.” They are running experiments. Even in 2024, when he was chief research officer at OpenAI, “the actual limiting factor was not the number of researchers we could hire. It was the amount of GPUs that we had on hand” — whether the compute was in the right places, on the right networks. You can “completely automate researchers,” he says, and that would be “a big effect.” He is willing to compare it to the one-time jump from GPT-4o without reasoning to o3 with reasoning and tools. “That was a huge one-time acceleration, but it’s not RSI.”

His preferred image is not a flywheel. It is a one-time move up an S-curve. Imagine the GPUs are only “35% of the way to optimal” at delivering discoveries, he says. Automating more of the messy work — picking experiments, scheduling them, cutting engineering waste — might get them to 90%. The step from 90% to 99%, he argues, is “actually not a big difference.” So automated research is “a big one-time benefit, but it doesn’t kick in a self-improvement cycle.” Then “we have to go do something else.” The same logic, he says, applies wherever the physical world is “the primary limiter.” When hypotheses get cheap, “you’re just going to bang harder and harder against the limited number of wet labs that you have.”

![A rising staircase labeled 10%, 50%, 90%, and 99% automated research, marked as a hard-takeoff story, beside a flatter constraint labeled GPUs, wet labs, and factories.](https://storage.ghost.io/c/d7/93/d793216c-cf13-441f-bdae-4f3aef998c0e/content/images/2026/11/fig1-101.png)

He is not saying automated research would be small. He is saying it is one more S-curve, bounded by the equipment, not a loop that runs away on its own.

## 2\. Closed worlds, work cells, and a country that designs what it does not build

After leaving OpenAI, he says, he noticed a familiar American split. “We are very good in America at designing products.” The iPhone is “designed in California, manufactured in China. It says that on the back of your iPhone.” The country is “amazing at designing defense materials,” and “we don’t manufacture very much of it here.” He spent a year talking to manufacturing leaders about why. The useful problem, as he frames it, is not a single rigid assembly line. Innovative products change, so the line has to change with them. His preferred tool is not, first, a general-purpose humanoid. It is “reconfigurable robotic work cells that you can immediately flow a design change into.” A humanoid is the most obvious copy of a person, he says; a work cell is what you want if the station already exists.

He draws a sharp line between open-world and closed-world robotics. Cleaning a bedroom is open-world: you have not seen the items, and the intelligence required is “very similar to human intelligence.” An assembly line is closed-world: you control the environment, the motions repeat, and “there’s no novelty that’s being introduced.” Miss a sock in a bedroom and it may not matter. Miss a step in a closed-world build and “99 or 99.9% of the time, you got to go construct the device that you were constructing.” That is why he thinks traditional automation already works on lines, and why the hard remaining work is flexible, high-mix manufacturing rather than a robot that can do the dishes in a messy house. This is not a conversation about a consumer humanoid arriving on a date.

The one Arda story that belongs here, because it is how he thinks the bottleneck yields, is a first user at Anduril. A virtual build that “would have taken him a few days” was done “in 90 minutes using our software,” he says. He describes the aim, lightly, as building the missing “IDE for manufacturing engineers,” then putting an agent on it. It is a founder’s example, not an audit; Arda’s launch announcement reports Anduril timings of the same kind, such as basic work instructions going from 10 hours to 15 minutes, and those are company-reported too. The portable claim is that the next constraint is the physical loop: designs that can change, lines that can change with them, and intelligence that is allowed to act there.

![Two columns, open-world robotics as a bedroom you have not seen, and closed-world robotics as a line you control, with a note that his bet is the reconfigurable work cell.](https://storage.ghost.io/c/d7/93/d793216c-cf13-441f-bdae-4f3aef998c0e/content/images/2026/11/fig2-101.png)

The contrast is the point. He is not dating a consumer humanoid. He is arguing that the useful next work is flexible, high-mix manufacturing.

## Weighing it: what holds up

McGrew is arguing from the companies he has run research for and the company he is building. The GPU-and-lab bottleneck is a real constraint, and his rejection of a researcher-only hard takeoff is a coherent industry reading, not a proof that recursive self-improvement cannot occur some other way. The 35%-to-90% image is a teaching sketch, not a measured utilization study. The Anduril 90-minute build is his account of an early user, and Arda’s own figures are company claims. Because Arda sells into the problem he describes, weigh the argument with that in mind. Auto-generated captions are close, not word-perfect. This is not investment advice.

**Related reading:** [The Timeline Just Moved](https://longtalkshort.com/the-timeline-just-moved-patel-rsi/) · [Production Is Deterrence](https://longtalkshort.com/a16z-american-dynamism-50-production-is-deterrence/) · [A Race Is Not a Delivery Date](https://longtalkshort.com/a-race-is-not-a-delivery-date-brett-adcock-humanoids/)

## Long Talk Short

- McGrew says you will not reach superintelligence without changes in the physical world, because software gains are still bounded by GPUs, labs, and factories.
- He calls the hard-takeoff story — automate the researcher, then watch a runaway loop — “straightforwardly wrong,” because researchers are not the bottleneck.
- Automated research, in his telling, can be a large one-time S-curve, comparable to the jump into reasoning models, without becoming recursive self-improvement.
- He is more interested in reconfigurable work cells in a closed-world factory than in a general humanoid, and he points to America’s habit of designing what it does not build.

**The one thing to remember:** If McGrew is right, the pace of the next AI leap is set less by how many researchers or how clever the models are than by GPUs, labs, and factories — and the interesting work is getting intelligence to act there.

*Independent summary; not affiliated with No Priors, Arda, Anduril, or Bob McGrew. Not investment advice.*

Source: No Priors, “Superintelligence Needs the Physical World with CEO and Co-Founder of Arda Bob McGrew,” on YouTube (linked above); Arda’s launch announcement is linked in the article. Independent summary; not affiliated with the show or its guest.