The AI Bubble Question Isn't New. What Grantham Gets Right (and Wrong) About Robotics Valuations

Most of the coverage of Jeremy Grantham's latest bubble warnings has focused on the stock market angle. Fair enough, that's his lane. But if you're trying to understand what his analysis actually means for robotics and industrial automation investment, you're mostly on your own. Let me try to fill that gap.

Grantham, co-founder and long-term investment strategist at GMO, has been making the rounds to promote his new memoir, The Making of a Permabear: The Perils of Long-Term Investing in a Short-Term World. In a recent appearance on Bloomberg's Odd Lots podcast, he drew the comparison a lot of people in tech have been quietly dreading: today's AI excitement looks, to him, a lot like the dot-com bubble of the early 2000s. He's watching the Magnificent 7's multibillion-dollar AI commitments with visible skepticism.

He called it correctly before. That matters. But the industrial automation sector is not the consumer internet circa 1999, and treating it as a monolith is where most of the hot takes go sideways.

What Does Grantham Actually Argue?

The core of Grantham's position, as laid out in his Odd Lots interview, is that several classic bubble indicators are flashing. He points to frothy valuations, speculative enthusiasm disconnected from near-term revenue, and a watershed shift in how Big Tech stocks are being priced relative to earnings. He's been here before, having spotted the dot-com collapse and the 2008 housing bubble before either fully unraveled.

His concern isn't that AI technology is fake. It's that the market is pricing in outcomes that are, at minimum, a decade away from generating the returns being baked into current valuations. That's a meaningful distinction, and one that gets lost when coverage reduces his argument to "Grantham says AI is a bubble."

The dot-com analogy is imperfect but not useless. The internet was real. It did transform the economy. It also wiped out a generation of investor wealth in the short term because the timeline and the winners were mispriced. Grantham's implicit argument is that something similar is happening now.

Why Industrial Robotics Is a Different Animal

Here's where I think the financial press coverage falls short. Industrial automation and the AI software layer are not the same investment category, and they don't carry the same risk profile.

I've seen enough spec sheets to know that capital equipment in manufacturing has a fundamentally different sales cycle than enterprise software. A robot arm installed on a production line in 2024 doesn't get repriced to zero because a language model underperforms expectations. The hardware is already deployed. It's already generating throughput numbers. The ROI is measurable in cycle times and defect rates, not in projected ad revenue.

Consider the numbers that are actually on the table right now. Global industrial robot installations hit approximately 590,000 units in 2023, according to the International Federation of Robotics. That's not speculative. Those robots are running shifts. The market for industrial automation hardware was valued at roughly $175 billion in 2023 and is projected to grow at a compound annual rate somewhere between 9% and 11% through 2030, depending on which analyst you ask. Those projections could be optimistic. But they're grounded in order books, not vibes.

The AI layer sitting on top of that hardware is where Grantham's concerns become more relevant. Computer vision systems, predictive maintenance software, AI-driven quality control, these are areas where the valuations can absolutely get detached from reality. A startup claiming its AI vision system reduces defect rates by 40% across all manufacturing verticals is making a very different claim than a company reporting actual cycle time improvements at a named customer facility. The former is a press release. The latter is a data point.

Look, I'm not saying the bubble concern is wrong for AI software. Some of those valuations are genuinely hard to justify with current revenue. That's an ambitious number for almost any AI software company pitching enterprise manufacturing clients right now.

What Are the Real Risk Signals to Watch?

If you're trying to apply Grantham's framework specifically to robotics and automation investment, the indicators worth tracking are different from the ones he's focused on at the macro level.

The signals that actually matter for this sector:

• Production volume versus pilot announcements. A company announcing a pilot program with a single automotive OEM is not the same as a company shipping 10,000 units a quarter. The real test is production volume, full stop.
• Integration complexity disclosures. Vendors who are upfront about integration timelines and failure rates are more credible than those who aren't. If a company can't tell you how long average deployment takes, that's a flag.
• Revenue recognition timing. Hardware revenue is recognized at delivery. Software subscription revenue is recognized over time. A company mixing both can present very flattering near-term numbers that don't reflect actual customer adoption.
• Customer concentration. If 60% of revenue comes from two customers, the TAM story is not yet proven.
• Capex commitments from end users. This one cuts both ways. The Magnificent 7 pouring money into AI infrastructure is exactly what Grantham is flagging as potentially overheated. But manufacturers committing capital to automation equipment are making multi-year bets that are harder to reverse than a software contract.

From my time in hardware, the metric that always told the real story was reorder rate. Customers who bought once and didn't come back were telling you something. Customers who expanded deployments were telling you something else entirely.

Is the Dot-Com Comparison Actually Useful Here?

In a way, yes. But the lessons are more specific than the analogy suggests.

The dot-com collapse didn't kill the internet. It killed the companies that had no path to profitability and were burning cash on customer acquisition in markets that weren't ready. The infrastructure companies, the ones laying fiber and building server capacity, many of them also collapsed in the short term. But the infrastructure turned out to be necessary. It just took longer to monetize than the market priced in 1999.

That's the more precise parallel for industrial AI. The hardware infrastructure, the robot arms, the sensors, the vision systems, is being installed now, and it has real utility now. The AI software layer on top of it is where the timeline and winner-selection problems are most acute. Which companies will own the operating system for factory AI in 2030? It remains genuinely unclear. I only found limited data on long-term software retention rates in industrial AI deployments, which is itself a signal that this market is still early.

Grantham is probably right that some of the current valuations in AI broadly are going to look embarrassing in five years. He has a better track record than most on this. But the coverage treating his argument as a blanket warning on all things AI and robotics is doing readers a disservice. The question isn't whether there's a bubble somewhere in this ecosystem. The question is where, exactly, and how exposed any given investment is to the part that's overheated versus the part that's just quietly installing robots on factory floors and counting the throughput.

The latter is a lot less exciting to write about. It's also a lot less likely to end badly.