Special Episode: Here's Why AI Costs Still Worry Investors

Summary

---

Bloomberg Tech — Special Episode: Here’s Why AI Costs Still Worry Investors

Date: November 22, 2025
Host: Stephen Carroll (Bloomberg)
Guest: Tom MacKenzie (Bloomberg Tech Europe anchor)
Episode Theme:
A deep dive into the mounting costs of Artificial Intelligence (AI) infrastructure—particularly in data centers and AI chips—and why investors remain cautious despite the sector’s massive growth and optimism.

---

Episode Overview

In this special crossover episode from Bloomberg’s “Here’s Why,” Stephen Carroll is joined by Tom MacKenzie to address a critical, often underplayed topic: the soaring and rapidly depreciating costs associated with building and maintaining AI infrastructure. While the AI boom has sparked bullish forecasts and massive investments from tech giants, beneath the surface, questions linger about sustainability, asset depreciation, and whether this rapid expansion echoes past bubbles in tech history.

---

Key Discussion Points & Insights

1. The AI Boom and Exponential Growth

• The episode opens with an analogy: “It’s 10:30pm in this AI party. It started 9pm and that party goes to 4am…”—implying we’re still early, but the clock is ticking on current AI euphoria.
• AI is described as “a qualitative leap” (01:03), resulting in exponential adoption and use across sectors.

2. Investor Worries: The Hidden Costs

• Stephen Carroll sets up the central tension: despite booming results, investors worry about running costs—namely, the short lifespan and rapid depreciation of expensive AI chips used in data centers.
• Tom MacKenzie: “We are putting mostly chips, silicon into these data centers that have a lifespan of perhaps four years.” (01:36)
• The rapid cadence of innovation drives obsolescence: “Even Nvidia, there’s a new chip every 18 months and it’s 10 times as powerful as the earlier ones.” (01:44)
• The risk: Huge capital is poured into assets (chips, data center infrastructure) that may not hold value long enough to pay off—output and adoption may not keep pace.

3. Dot-Com Parallels and Depreciation Concerns

• Tom MacKenzie references Michael Burry (of “The Big Short” fame) who has taken large short positions against Nvidia and Palantir due to:
  • The potential for rapid depreciation of high-value chips (“AI accelerators”).
  • Not all tech giants “are properly accounting for how quickly these assets depreciate.” (02:47)
• Concerns draw a direct analogy to the late-1990s dot-com bubble:
  • Telecom infrastructure companies “spent huge amounts of money… and ended up losing a lot of money because the gains didn’t come as quickly, the technology didn’t evolve as rapidly as they had expected.” (03:28)
  • Some winners (like Amazon) did emerge—but for most, capital was burned, not paid back.

4. Industry Pushback: Longer Chip Lifespans?

• Nvidia, led by CEO Jensen Huang, contends that older chips (e.g., Hopper) are “very versatile” and have “a lifespan of about six years…fully utilized by most of the companies that own those.” (04:41)
  • Chips can be repurposed from training to inference, expanding their usefulness.
• Bain Capital’s analysis:
  • By 2030, AI hyperscalers need to generate $2 trillion/year in revenue to justify current investment rates (05:40).
  • There remains a “huge gap, hundreds of billions of dollars… between investments into AI infrastructure and the actual revenues” (05:56).

5. The “AI Agent” Bet and Sovereign AI Investments

• Tech giants predict the next phase: proliferation of “agentic AI”—AI-powered agents managing everything from holiday bookings to healthcare logistics, increasing integration into daily enterprise tasks (06:50).
• Major sovereign investments, with regions such as the Mideast, Europe, and Japan seeking their own AI infrastructure. Early days, but “all those kind of things are going to come together.” (07:14)
• The crucial question: Investors’ patience—“Will they continue to invest in the hyperscalers if they’re not seeing real material returns?” (07:32)
  • Bloomberg Intelligence projects real scrutiny by end of 2026, a potential inflection point for continued investment.

6. The Cash Pile Paradox: Are the Tech Giants Really at Risk?

• Carroll asks why there’s concern given the giants’ huge cash reserves: “Why is there concern at all about how they’re going to pay for this?” (07:40)
• MacKenzie: “When we talk about hyperscalers, these are companies with massive balance sheets and huge cash reserves. These are incredibly profitable businesses… These are not non-profitable, major punts and risky parts of the market.” (07:56)
  • However, the risk comes for smaller players (“Neo clouds”—data center leasers, coreweave, nClouds) and even loss-making AI startups like OpenAI and Anthropic.
  • Some big names (even profitable ones) are now raising debt (“tapping the public markets”) to finance infrastructure, which raises questions about long-term leverage and risk tolerance (08:43).
  • Circular deals (e.g., Nvidia invests in OpenAI, which buys chips from Nvidia) add complexity: “All of these companies become increasingly enmeshed and intertwined… on what is a bet on the future.” (09:26)

---

Notable Quotes & Memorable Moments

• On the tech cycle’s late hour:

  “It’s 10:30pm in this AI party…almost every investor knows it’s all going to turn into pumpkins and mice at midnight. Only as Buffett would say, no one in the room has a clock.”
  —Stephen Carroll, 01:51

• On the core depreciation issue:

  “We are putting mostly chips, silicon into these data centers that have a lifespan of perhaps four years.”
  —Tom MacKenzie, 01:36

• The bull-bear split:

  “Michael Burry making the argument that companies, the hyperscalers…are not properly accounting for how quickly these assets depreciate.”
  —Tom MacKenzie, 02:47

• Historical perspective:

  “Comparisons…with what happened in the late 1990s…the dot com bubble, when…telecom equipment makers…spent huge amounts of money on building infrastructure…ended up losing a lot of money because the gains didn’t come as quickly…”
  —Tom MacKenzie, 03:28

• On chip lifespans and repurposing:

  “Their older AI chips, one of their older versions is called Hopper, has a lifespan of about six years and is very versatile…”
  —Tom MacKenzie quoting Nvidia, 04:41–05:00

• The coming crunch:

  “By 2030, the hyperscalers and other AI giants would have to be turning around revenues of about $2 trillion. And…right now, there’s a huge gap…between investments into the AI infrastructure and the actual revenues…”
  —Tom MacKenzie, 05:45–06:00

---

Timestamps for Key Segments

• 00:42–01:24 — Introduction to AI’s rapid growth and the “party” metaphor
• 01:36–02:27 — AI chip depreciation, Burry’s concerns, and dot-com parallels
• 04:30–05:40 — AI industry’s response; repurposable chips and longer lifecycles
• 05:40–06:55 — The revenue vs. investment gap; necessity of finding “go-to-market” fit; future AI use cases
• 07:40–09:26 — Who really faces risk? Breakdown of company types and the circular nature of AI partnerships

---

Conclusion & Takeaways

• Despite AI’s breathtaking ascent and the confidence of major tech players, cracks are appearing in the investment story—centered on soaring infrastructure costs, asset depreciation, and uncertain near-term business models.
• For the “hyperscalers,” balance sheets are strong, but even they’re beginning to borrow heavily and engage in circular financing.
• The real crunch is projected for end of 2026, when investors will expect hard proof that massive spending translates to sustained, significant revenue.
• For now, the AI party rages on, but “almost every investor knows it’s all going to turn into pumpkins and mice at midnight”—the only question is when.

---

For more explanations and deep dives, visit Bloomberg’s explainer archive at Bloomberg.com/explainers.

Transcript

[A]

Hi, this is Caroline Hyde from Bloomberg Tech. Today we're sharing something a little different in your feed, an episode from our colleagues at here's why, Bloomberg's weekly show that answers one big question. In under 10 minutes, host Stephen Carroll is joined by our Bloomberg Tech Europe anchor Tom McKenzie to dive into a story that's right at the heart of the tech world. The massive investments in AI data centers and the hidden costs that come with them. If you'd like to hear more episodes of here's why, you'll find a link to the podcast feed in the show Notes. Hope you enjoy Bloomberg Audio Studios Podcasts.

[B]

Radio News.

[C]

I'm Stephen Carroll and this is here's why, where we take one news story and explain it in just a few minutes with our experts here at Bloomberg. It's 10:30pm in this AI party. It started 9pm and that party goes to 4am and the reality is like, look, this is going to be a.

[B]

Two to three year lap in this bull cycle for tech. The tech sector is very strong because artificial intelligence is really a qualitative leap in the kind of technology that we've had over the last several decades. You're seeing an exponential growth of adoption and use of AI. The number of applications that are going to be using these AI is also growing.

[C]

Everyone has an opinion on where the AI frenzy is going next. But while optimism is rampant about the technology's potential, more questions are now being asked about AI's running costs.

[B]

We are putting mostly chips, silicon into these data centers that have a lifespan of perhaps four years.

[C]

Those chips, they depreciate very quickly. Even Nvidia, there's a new chip every 18 months and it's 10 times as.

[B]

Powerful as the earlier ones. The thing with the rally this year is that almost every investor knows it's all going turn into pumpkins and mice at midnight. Only as Buffett would say, no one in the room has a clock.

[C]

Even with bumper results and bullish revenue forecasts, here's why costs still worry investors. Tom Mackenzie, who hosts Bloomberg Tech Europe on Bloomberg Television, joins me now for more. Tom the investor Michael Burry of Big Short fame is among those who's worried about these future costs of AI and data centers in particular. What's the concern?

[B]

Yeah, absolutely. Michael Burry putting on famously short positions, so shorting the stocks of Nvidia and Palantir before he wrapped up his fund. His concern does focus on the depreciation of some of these assets. By assets I'm talking about specifically these AI chips, very expensive AI accelerators, 90% of the market share is dominated by Nvidia. So across the sale of these chips, Nvidia has that significant market gain versus its rivals. And the concern is that as you get newer versions of these chips, the older ones essentially become less valuable. And Michael Burry making the argument that companies, the hyperscalers, so the Microsofts and alphabets and metas of the world are not properly accounting for how quickly these, these assets depreciate. The other part of the concern and kind of ties into this that you hear voice from the skeptics around the AI bubble is that there are comparisons, they say, with what happened in the late 1990et, 1999, early 2000, the dot com bubble, when it was the telecom equipment makers that leading up to all of the online expectations around how our digital economy was going to change, spent huge amounts of money on building the infrastructure to power the.com era and ended up losing a lot of money because the gains didn't come as quickly, the technology didn't evolve as rapidly as they had expected. Of course on the back of that you did get some very significant players like Amazon who came through the dot com bubble and of course now remain one of the most valuable companies on the planet. But there was a lot of capital, there was a lot of investment that was burnt in that process. And so that is another comparison that people are making. It's the depreciation around the assets and the chips that they're worried about, but also comparisons with what happened during the.com era and the pain that was felt by those telecom equipment makers that sunk so much money into which they accumulated huge losses.

[C]

So how are the big AI players thinking about these costs at the moment?

[B]

So pushback to the depreciation argument would come from Nvidia, and we've heard that recently from the CEO Jensen Huang, and he's made the case that in fact even their older AI chips, one of their older versions is called Hopper, has a lifespan of about six years and is very versatile. So you can use it not just for the training of these large language models, but for the post training and for the inference that's when they're actually being used by us, by consumers and by enterprise. And so you can move them around, they have different functions and therefore they actually have a longer lifespan than some of the skeptics are suggesting. And our own analysis suggests that those Hopper chips, those older varieties of chips, have a lifespan of about six years and are fully utilized by most of the companies that own those. So that does address some of that concern. The question going forward to what extent these companies are going to be able to find products that match the investments that they are sinking into the AI infrastructure story. Bain Capital came out with a report recently suggesting that by 2030, the hyperscalers and other AI giants would have to be turning around revenues of about $2 trillion. And that right now there's a huge gap, hundreds of billions of dollars in terms of the gap between the investments into the AI infrastructure and the actual revenues that are coming about as customers and as enterprises and companies use the end product. So the go to market, the product fit is going to be really, really important. And what the big AI players say, whether that is the hyperscalers again, the likes of Meta and Alphabet and Amazon say, or the likes of OpenAI and Anthropic, is we're going to be in this world of agentic AI. We're going to have AI agents booking our holidays, checking up on our health care, finding good schools and universities for our students. All those kind of things are going to come together. Enterprises are going to be embedding AI much more than they already are. We're only in the first opening stages of that would be the argument. And then there's the sovereign AI story, where different countries, and we're seeing that in the Middle east, but also in Europe as well, and Japan are investing heavily to ensure that they have their own AI infrastructure and AI class that were very early in that story as well. Those are all the cases that the big AI players would underscore in terms of why this is going to be driving momentum going forward at least through 2026. Our own team at Bloomberg Intelligence say the end of 2026 is going to be a question mark as to whether or not investors continue to have pat. Will they continue to invest in the hyperscalers if they're not seeing real material returns, if that product fit and that custom use isn't there in a really, really significant way? So I think the patience of investors and to what extent they can continue to lean into the hyperscalers as they spend these huge amounts, is going to be a key question mark. And our own team think that that's really going to come to the fore at the end of 2026. They'll need to answer that question. They've spent hyperscalers have spent about $300 billion on infrastructure this year. And the projection be, according to Nvidia, Nvidia sees the hyperscalers spending upwards of about $600 billion next year.

[C]

One of the things that occurs to me in this as well as we're talking about some of the world's most valuable companies, they have massive cash piles in a lot of cases. Why is there concern at all about how they're going to pay for this given their revenue streams and how much money they have?

[B]

You're absolutely right. So when we talk about the hyperscalers, these are companies with massive balance sheets and huge cash reserves. These are incredibly profitable businesses that come through with very strong earnings. These are not non profitable major punts and risky parts of the market. These are not companies that no one's heard of. They're making real product, they're selling it to customers and they've been doing that for decades. Microsoft, Alphabet, Meta and Amazon, they have that balance sheet strength, they have that cash on hand. The concern then is around other parts of this ecosystem. So you can think about it in different baskets. You have those big ticket blue chip names in one basket and then you have maybe Neo clouds in the other basket. These are the core weaves or the nClouds, companies that lease out data centers to some of these hyperscalers and some of the large language models who have business models that are less proven than the hyperscalers. Then another bucket would be maybe some of the key large language models themselves. The OpenAI's and the anthropics that are losing money on an annual basis, even as they're seeing a lot of growth and revenues increase year on year, they're still not profitable. So you can break it down into different categories in terms of the level of risk. But even amongst the big publicly listed companies with those strong balance sheets, you have seen examples of them tapping the public markets and raising debt on the public markets. And so far that's been well received by the markets. But how long is that going to continue and to what extent is the leverage that now these companies are, are starting to tap into going to be acceptable to investors? And again, I think you have to put a different framework over the different companies in terms of how you answer that question. Then there's the circularity of the financing. So OpenAI, for example, doing deals with Nvidia and Nvidia investing in OpenAI and in response to that, OpenAI committing to buying a certain number of chips from Nvidia. Those circular financing deals as they've been described by some, have also caused some concern as all of these companies become increasingly enmeshed and intertwined in terms of their deals and their investments on what is a bet on the future. And how the future evolves.

[C]

And an expensive one at that. Tom MacKenzie, thank you very much for joining us, host of Bloomberg Tech Europe on Bloomberg Television. For more explanations like this from our team of 3,000 journalists and analysts around the world, go to Bloomberg.com explainers. I'm Stephen Carroll. This is here's why. I'll be back next week with more. Thanks for listening.

[A]

Save over $200 when you book weekly Stays with VRBO this winter. If you haven't seen your college besties since, well, college. You need a week to catch up in a snowy cabin. Take a week long vacation and save over $200. Book now.

[D]

@Vrbo.Com we buy insurance for peace of mind. But every year millions of claims are denied. Not because people did anything wrong, but because their policies quietly excluded what happened. Insurers every detail. Policyholders rarely do. That's why my policy advocate exists. For just 27 cents a day, their platform reads your policies and explains where you are vulnerable. They don't sell insurance, they deliver transparency. Before you trust your policy to protect you, let my policy advocate tell you what it really says. Go to mypolicyadvocate.com in the heat of.

[A]

Battle, your squad relies on you. Don't let them down. Unlock elite gaming [e-mail] Dominate every match with next level speed, seamless streaming and performance that won't quit. Push your game play beyond performance with Intel Core Ultra processors for the next era of gaming. Upgrade to smooth high quality streaming with Intel WiFi 6e and maximize game performance with enhanced overclocking. Win the tech search Power up at lenovo.

[C]

Com Lenovo Lenovo.