I’ve been saying for over two years that the U.S. was missing the point on AI — or at least only seeing half of the story. And last week we saw the consequences: a data center-induced market correction that knocked 5% off the Nasdaq.
Stocks rebounded the next day, but to be clear: this is the beginning of an ongoing, long-term realignment of the entire North American AI market. There will be winners and losers — and most of the losers deserve what’s coming to them, so that’s entirely OK.
But it’s also important to understand what isn’t happening. We are not at the beginning of a catastrophic AI bubble implosion like the optical/internet ones in 2000 or the subprime mortgage catastrophe of 2008. Those were different, and not in a good way.
A correction in data center buildouts is not a correction in the AI market as a whole. It’s really important that you (yes, you) understand this.
The data center issue has come about for a few reasons, but the most immediate is debt — too much, too expensive, without a rational business case to justify it.
The numbers are absurd. According to Goldman Sachs, nearly $500 billion in AI-related debt had already been issued by early August 2026. The hyperscalers alone accounted for approximately $194 billion, with Goldman estimating that borrowing would finance a third of their capital expenditure this year.
And that's just the visible part. J.P. Morgan estimates that hyperscalers have approximately $1.4 trillion in data center lease commitments, around $1.1 trillion of which has yet to appear on their balance sheets, plus another $1.5 trillion in purchase commitments — creating enormous future financial obligations.
These companies are putting the debt horse before the revenue cart, borrowing hundreds of billions to build infrastructure before establishing whether customers will generate enough revenue to pay for it. That's not a business model. It's a bet that demand will eventually catch up with the debt. And the interest clock is already ticking.
It so happened that I was chatting to Jeetu Patel, president and chief product officer at Cisco, on the same day as the market correction, so I asked him if we were over-rotated on the data center buildout in the U.S.
“I don't agree that we are over-rotated,” he told me. “A country or an organization's ability to effectively and efficiently and securely generate tokens is directly proportional to its economic prosperity. The supply of tokens you can bring online right now is being consumed instantly. And we're in the very, very early stages of AI. The price of tokens will keep going down, but demand for tokens is going up at a faster rate than the price per token is falling. And so data center buildout is very, very important.”
I recently put the same question to Ericsson CEO Per Narvinger. “I believe there’s a huge value in data centers being built, whether they’re at the right pace, at the right valuation, time will tell,” he told me.
Brace! Brace! Brace!
Debt bad, data centers good. Got it? What data centers do is no less important today than it was at the beginning of last week. Giant probabilistic LLM frontier models sitting in data centers really are a critical component of the new digital industrial revolution.
That’s what passes for the good news.
The bad news is there is a much larger AI problem looming behind the current correction — one which goes far beyond last week’s drama to the level of a black swan event — something that, left unaddressed, could destabilize the entire U.S. economy through the midpoint of this century.
The larger problem, and this is one that has almost completely blindsided the U.S., is that data centers aren’t the only important part of the AI market.
There is a whole other type of AI — industrial, predictive machine learning — that sits outside of the data centers and is well on its way to transforming the world, except in the U.S. (and the U.K., the only developed nation in the world that is as ill-informed about this trend as America).
Machine learning isn’t an alternative to frontier models — they do different things. LLMs are probabilistic engines built to predict the next word in a sequence. ML is engineered to operate within strict physical rules and real-time operational guardrails. That makes machine learning the only type of AI that can be used safely and predictably in industry.
Arista CEO Jayshree Ullal sees the next architectural shift moving away from the center. “As you get enough centralized training, this will become a distributed inference and edge problem,” she told me.
That is a very different investment thesis from simply building ever-larger AI factories.
Below is a diagram that models the world in terms of market opportunity. The U.S. is all in on the little circle at the bottom — the core — where the data centers live, which it dominates. Much of the rest of the world is focusing on the bit out at the edge — that vast hinterland of industry which accounts for about 80% of global GDP.
Trailing the world
And I do mean the rest of the world. In addition to Finland, Sweden, Japan, Singapore and others in the vanguard of industrial digitalization, there is another giant swathe of countries that have the potential to leapfrog Industry 4.0 laggards in the West: the countries that buy products from Huawei using money borrowed from the Chinese government.
China began implementing its full-stack national digital strategy in 2006 (two decades later, and the U.S. still doesn’t have one). But that plan didn’t stop at its borders. The strategy has been exported by two vehicles: Huawei, China’s communications incumbent, and the Chinese government, which lends the money to developing countries to buy that technology — and build the other necessary utility infrastructure — to enable Industry 4.0.
Huawei has customers in more than 170 countries and regions. Not all of them are undergoing full-scale industrial digitalization, but the transformation is already well underway in many.
In China, autonomous trucks and cranes operate at Tianjin Port, while a fleet of 100 driverless electric mining trucks works at the Yimin mine in Inner Mongolia. In Thailand, Midea is using private 5G, edge computing and AI to automate manufacturing. In Hungary, Huawei's European Supply Center uses autonomous forklifts and AI-driven quality control. And in Botswana, Debswana has connected hundreds of pieces of heavy mining equipment through a private industrial network.
These aren't speculative applications waiting for a future generation of AI models. They're working industrial systems, operating in the physical economy today.
Missing the point
The spread of China’s Industry 4.0 diaspora is by far the biggest story that nobody in the U.S. communications industry seems particularly interested in discussing. And you have to wonder why.
One contributing factor is cultural. The people predicting the future of AI have spent their entire lives in a white-collar bubble, insulated from anything with a broom handle or a 3,200-amp insulated-case circuit breaker. Traveling a straight line from elite universities to investment banks, venture capital firms, consulting houses and technology companies, they’ve mistaken the factory floor for an office with heavier machinery. They understand spreadsheets, software subscriptions and stock vesting. They don’t understand operational technology or industrial control systems. Their experience of physical labor extends little further than watching someone else perform it. And therein lies the problem: the people deciding where trillions of dollars should be invested in AI have barely set foot in the heavy industries it should be transforming. No wonder America is missing the point.
Hang out with your Huang out
But a much larger part of the problem is that America's technology strategy is largely shaped by the people who stand to profit from it: Big Tech executives, hyperscalers and chipmakers.
Among them, Jensen Huang, CEO of Nvidia, has achieved almost messianic status. His five-layer AI cake — energy, chips, infrastructure, models and applications — has become the de facto framework for explaining the AI economy.
It's a useful description of the business Nvidia wants to build. But it isn't a complete description of the economy AI will transform.
In Huang's world, factories, robots, vehicles and utilities sit at the application end of an AI computing stack. In the real world, they are enormous industrial biomes in their own right, with their own infrastructure, operational technology, machine learning, control systems, physical constraints and regulatory obligations.
Water, for example, isn't simply a cooling requirement for an AI factory. It's a critical industrial resource. Machine learning isn't something that arrived with ChatGPT. It's been running industrial processes for decades.
The problem isn't that Huang's stack is wrong. It's that an entire industry appears increasingly willing to mistake a chipmaker's view of the world for the world itself.
And that is a very expensive mistake to make.
My own stack, which I published seven months before Huang’s, accounts for the world as it is becoming, not as America’s Big Tech industry would like it to be. Huang’s is U.S.-centric. Mine is modeled on what is happening around the world now, and over the rest of this century.
Localized risk
As the global AI infrastructure story plays out, the current correction will also affect U.S. AI ecosystem players (or any event company that just bought a giant data center trade show).
The effect, if any, largely depends on two things: how exposed these companies are to the debt situation, and how far they have diversified their value out of the data center GPU compute gold rush.
Neoclouds (essentially, debt-fueled GPU landlords) like CoreWeave had been widely recognized to be in a parlous situation even before last week's market wake-up call. The company's competitive moat is an OS that increases the efficiency of GPU operation. But surely that's just a feature that any cloud platform should offer — not a USP for a $25 billion company?
At the other end of the spectrum, Vultr — the world's largest privately held independent cloud infrastructure provider — has been on a mission to diversify out of raw compute resale and into high-growth industrial, sovereign and enterprise applications where the real value lies. Better.
Among my core constituents — companies like Cisco, Ericsson, Nokia and Huawei, that provide large tranches of AI infrastructure — the news is better. Those that sell optical capacity to connect data centers or speed up their insides can continue to do so. The data center market is not going away. And Ericsson, which does not share in that optical bounty, is ready for the Industry 4.0 opportunity beyond it.
Surprise!
And finally, one surprise winner: the carriers, historical Rodney Dangerfields of the comms industry.
For years, telcos have been treated as dumb pipes while cloud giants collected the cash. That dynamic changes in industrial AI. Controlling physical processes across thousands of geographically dispersed locations demands proximity, resilience, predictable performance, security and operational sovereignty. You cannot deliver that simply by building another enormous centralized data center.
“Mobile networks are probably one of the most distributed architectures out there in any industry. Sometimes you will need compute closer to where you are,” Narvinger pointed out in our recent conversation.
Carriers already own and operate the distributed edge infrastructure, local spectrum and mission-critical reliability that industrial intelligence requires — and Nvidia is investing heavily in the technologies to bring AI directly onto their networks. Hyperscalers aren't going away, and they will continue to dominate centralized cloud AI. But the assumption that they will own the entire industrial AI stack is totally wrong. For the first time in years, carriers have the right to a genuine seat at the industry’s high table, should they choose to occupy it.
The end game
The biggest question remaining unanswered is this: who will own the Industry 4.0 world on the factory floor? This is a zone with the greatest moat of all — OT, not IT, operating systems, backed by security and reliability requirements that make carrier standards seem tame in comparison. Today, that is the remit of industrial incumbents like Siemens and Rockwell. Yet they do not have the networking, AI or automation portfolios that communications vendors possess.
This is the real prize, and it is a trillion-dollar market.
Learn more about industrial AI on Fierce Network
AI’s full stack is looking a bit wobbly
The world’s AI economy enters the fifth dimension
Top 25 global AI infra index debuts
America’s great AI hallucination
Introducing: The Unified Infrastructure Stack
AI extinction risk will become real when intelligence gets physical
Special report: Physical AI moves toward the factory floor, but full autonomy remains years away
Ericsson CEO Per Narvinger wants a network without borders
Arista CEO Jayshree Ullal: The AI network can pay for itself
Ciena’s Gary Smith: AI is forcing a global rebuild of network infrastructure
How big is Industry 4.0? Bigger than you think
5G’s true potential: Why analysts are missing the mark on industrial digitalization
The AI opportunity heads to the edge
Stephen M. Saunders MBE is a communications analyst and USPTO-registered inventor examining how digital infrastructure — 5G, cloud and AI — is reshaping industry, power and society, as well as underpinning the emerging, ubiquitous global digital economy. As anchor of FNTV and a longtime industry insider, he focuses less on growth narratives and more on execution, risk and how hyperscale technology is distorting markets, governance and society at scale.
Opinion pieces from industry experts, analysts or our editorial staff do not represent the opinions of Fierce Network.