Andrew Bailey has made several statement about the dangers of an AI debt fuelled investment boom. It’s not happened yet, but he’s still warning about it.
Andrew Bailey, the Governor of the Bank of England, has warned of a severe downturn in global stock markets if the AI bubble bursts.
In a letter to G20 finance ministers, Mr Bailey raised concerns over the economic fallout if the debt-fuelled AI investment boom begins to unravel.
He warned that the soaring levels of debt that have been used to fund AI “could amplify a future market correction”.
I view it like any other tech bubble. There will end up being a few real long term winners. The market will overestimate the importance of the tech in the short run causing bubbles and underestimate the impact of the technology in the long run. (Amara’s Law)
Lots of players compete for leadership in AI at many levels. Chips, data centers, software, applications. In the end two or three leaders will emerge. Others will specialize or fold. Fringe players are risky. I wonder how many penny stocks claim to be AI players.
I still think AI is the future and it will drive increasing productivity. But investors must be selective. Some will fall by the wayside. Major crash when AI investing slows is possible but buildout may continue for decades.
Time will tell. The intuitive play to me is to continue to invest in what are already the best businesses with definable moats and producing high rates of ROI where AI will provide further efficiencies and improve margins. My preference is to act like an investor rather than a speculator.
I see this as a dot-com bust, and hopefully not a Great Recession. AI is a bubble. AI is useful. Both can be true. Just like it was for the Internet. But I think picking winners will be very hard. Just look at Intel - it took 26 years for it to reach a new all-time-high after the dot-com crash. Intel!
I would put it differently, intelligence, by itself, is useless. Applied intelligence is the future. Where does AI create value? How does AI create value? Any takers?
As I see it, Intel fell victim to the Innovator’s Dilemma, too worried about defending the x86 architecture. They had multiple opportunities to adopt ARM’s architecture and each time they decided against it.
GoogleAI:
You hit the nail on the head. Intel’s multi-decade trajectory is the textbook definition of Clayton Christensen’s Innovator’s Dilemma. For a long time, the company chosen to prioritize its incredibly high-margin x86 server and desktop cash cows over lower-margin, highly efficient architectures. [1, 2]
It literally took 26 years for Intel to break its dot-com bubble record, finally hitting a new all-time high of $140.94 in June 2026. [1, 2]
How the Dilemma Played Out
Intel’s historical reluctance to move past x86 created massive blind spots across consecutive tech eras: [1, 2]
The Mobile Miss: Famously, Intel turned down Apple’s request to build the chip for the original iPhone because Intel’s leadership didn’t believe the volume would offset the low margins. Apple went with an ARM architecture instead. Intel later tried to forcefully scale down x86 for mobile devices (Atom processors), but it was simply too power-hungry compared to native ARM alternatives. [1]
The AI Shift: By over-indexing on central processors (CPUs), Intel was slow to invest in the heavily parallelized processing capabilities of GPUs and AI accelerators. This gave NVIDIA a multi-year head start to capture the AI enterprise space. [1, 2]
The 2026 Resurgence
Intel’s dramatic breakout in 2026 wasn’t driven by x86 dominance, but by a massive structural pivot under new leadership and a revived foundry roadmap. The stock rocketed thanks to an unexpected chipmaking agreement with Apple, surging demand for server-side AI execution (inference) systems, and massive support from the U.S. CHIPS Act to establish a localized foundry ecosystem. [1, 2, 3, 4]
Also..
GoogleAI:
Intel partnered with Elon Musk’s companies—Tesla, SpaceX, and xAI—in April 2026 to join the ambitious Terafab Project . [1, 2]
Key Details of the Partnership
The Project: Terafab is an advanced semiconductor manufacturing complex located in Austin, Texas.[1, 2]
The Goal: The project aims to refactor silicon fab technology to produce up to one terawatt of AI compute per year for robotics, autonomous vehicles, and space-based data centers. [1, 2, 3]
Intel’s Role: Intel will help design, fabricate, and package ultra-high-performance chips at scale, utilizing its manufacturing expertise and advanced nodes like the 14A process. [1, 2]
Significance: This collaboration is a major milestone for Intel as it works to expand its foundry services and turn around its manufacturing operations. [1]
In defense of Intel there, they did realize their own x86 architecture was awful and tried to roll-out IA64, Itanium. It was a novel approach. It was an EPIC architecture, Explicitly Parallel Instruction Computing, which offers big savings in hardware, which ripples into power efficiency and cost. But it was a horrible processor, and frankly EPIC is an architecture that looks good on paper but is horrible in the real world. Turns out that while using lots of logic to determine instruction parallelism is costly, doing so in the compiler is near impossible.
RISC and CISC are the two classic design styles for computer processors, while Itanium is a distinct, specialized 64-bit architecture that tried a different approach called EPIC. [1, 2, 3]
CISC (Complex Instruction Set Computer)
Core idea: Uses a large set of complex, variable-length instructions where a single command can perform multiple low-level tasks, like loading data, adding, and storing.
How it works: The processor hardware handles the heavy lifting of decoding these varied instructions at runtime.
Examples: Intel and AMD x86 processors used in most desktop PCs and laptops.
Pros & Cons: Great for code density and software compatibility, but leads to complex chip designs and higher power use. [1, 2, 3, 4, 5]
RISC (Reduced Instruction Set Computer)
Core idea: Uses a small, simple, and fixed-size set of instructions where each command executes in a single clock cycle.
How it works: Relies on a load/store architecture (separating memory access from math) and lets software compilers handle complex tasks by stringing simple commands together.
Examples: ARM chips used in most smartphones, tablets, and modern Apple Silicon.
Pros & Cons: Highly energy-efficient and fast, but requires more instructions to complete complex jobs.[1, 2, 3]
Itanium (IA-64)
Core idea: Developed by HP and Intel and launched in 2001, Itanium is neither traditional RISC nor CISC. [1, 2, 3]
EPIC Architecture: It uses Explicitly Parallel Instruction Computing (EPIC), meaning the compiler explicitly decides which instructions run in parallel instead of letting the processor figure it out dynamically. [1, 2]
Market: Built for high-end enterprise servers and heavy computing. [1]
Outcome: It largely failed in the market due to poor performance on legacy x86 software, high costs, and the rapid advancement of traditional multi-core x86 chips, leading to its official discontinuation in 2020. [1]
Google misses the key difference between EPIC and the other two. And the other two both do instruction level parallelism and re-ordering in hardware. Instruction level parallelism and re-ordering requires a LOT of hardware. Lots of logic goes into deciding if any two instructions can be executed in parallel down separate hardware paths, or if they can be re-ordered. EPIC moves that from a run-time determination using transistors and logic gates (which burns power, introduces latency, and takes more gates/area which means money) and puts it into the compiler. Big savings in terms of logic. But there are multiple real problems with this approach and is why nobody uses it.
EPIC has nothing to do with being RISC or CISC.
Tangent. A surprisingly small amount of the logic in a typical CPU (RISC or CISC) is devoted to actually executing an instruction. Most of the logic revolves around how and when to execute the instruction. Instruction level parallelism, stalling, reordering, branch prediction and speculative execution all take large amounts of complex logic, but we do it because the payoffs in execution speed are very real. Modern software would be notoriously slow without these features. Also GHz-level clock frequencies would not have happened. Machine code does NOT execute in program order and has not in a very long time. It simply appears to.
Second tangent: GPUs don’t necessarily do much of that, but is also why they are not “general purpose” compute cores. They typically execute a very narrow type of program, one that is inherently massively parallel from the beginning. But if you tried to execute more general purpose code on them, they’d be slow.
(Me: 30+ years in processor design and verification)
By the time I retired Intel had not gone EPIC but there were all sorts of projects trying to make hardware more like software, ASICs that morphed. I was very much in the ARM RISC camp because it seemed simpler. I posted quite a bit about it at my website, softwaretimes.com
Also, being a Mac developer I was not much in favor of the x86, Intel, Windoze camp. Might as well confess my biases.
It’s impossible to know, there are so many potential exit points. Perhaps financial realities will set in and the builders will realize you can’t keep selling a product for a dollar that costs $10 to make. Perhaps companies will awaken (as many have) that it isn’t providing the productivity boost that’s been so widely announced, particularly once they have to start paying for it.
Maybe cities and power companies will say “Well, it’s time to face reality, and the truth is there isn’t enough water and/or there aren’t enough electrons” or maybe “even if you bring you own power, the air pollution you will create is ouitside the boundaries of the law.”
Or possibly, and this one I hold as the least likely, people will wise up and understand that it isn’t Artificial Intelligence, and it certainly isn’t Artificial Wisdom, it’s just a collection of facts that silicon chips accumulate, sift, and offer up a “down the middle” answer, even when “down the middle” is substantially wrong. (Where, I note, would Intel be if researchers hadn’t gone outside the mainstream to build silicon wafers? Thomas Edison would still be carrying kerosene to oil lamps. And Henry Ford might have the best horse-buggy company in the world, had they not broken all convention and gone another way.)
It’s entirely possible that this dies not with bang but with a whimper, but with this much investment and stock market capital riding, and so many itchy fingers on the trigger (except for those on the inside ring of the merry-go-round) I’d be betting more on a rush for the exits if/when one of the above conditions comes true - if it happens in some sort of dramatic fashion.
Are there uses for it? Sure. I can’t wait to be on hold with AI chatbots more than I am already. Are there dangers? Aren’t there always?
Meanwhile, here’s one of my favorite new videos, from Ed Zitron, (I can only find it on Facebook at the moment) titled “AI is a financial scam.” Perilously good viewing:
