Skip to content

Podcast

The Truth About the AI Bubble

The AI Daily Brief: Artificial Intelligence News and Analysis

Source ↗ ← All highlights
  • ChatGPT Met A Rate-Hiking Cycle
    • ChatGPT launched around the same time as Jerome Powell’s fastest rate-hiking cycle.
    • That timing forced AI enthusiasm to compete with a tightening macroeconomic environment. Transcript: Nathaniel Whittemore ChatGPT launched at almost the exact same time that Jerome Powell started the fastest rate hiking cycle in history. And for most of that hiking period, it was AI enthusiasm versus the world when it came to Wall Street performance. In the wake of the initial disappointment around GPT-5, there was an interview with Sam Altman where he was widely interpreted as having said that there was a bubble, even though he Couched his language a lot more than the headline suggested. And really, it’s just a question that’s never fully gone away. Now, functionally, does this matter to those of us who are just using these tools or figuring out how they are going to impact our businesses? The short answer is no. If you have used these tools for your work or your personal life, especially if you are here listening to this show, you will have undoubtedly come to the conclusion that they are immensely Powerful and are likely to reshape much of what you do. (Time 0:01:30)
  • Booms Versus Bubbles
    • Bubbles differ from booms by whether fundamentals catch up to hype over time.
    • A boom becomes durable if revenues and productivity eventually match expectations. Transcript: Nathaniel Whittemore Azeem starts historically. Bubbles, he writes, are among the oldest stories of capitalism. They’re parables of excess, belief, and collapse. But bubbles are not just financial phenomena. They are cultural artifacts. They return again and again as morality tales about greed and folly. Tulip mania, often misremembered as a frenzy of bankrupt weavers and drowning merchants, was less disastrous than legend suggests. It was confined to wealthy merchants and left the Dutch economy largely unscathed. But the myth has endured and that is the point. Bubbles become stories we tell ourselves about the dangers of optimism. Some bubbles are financial. The South Sea frenzy of the 1720s, the rowing stock market of the 1920s, Japan’s real estate boom of the 1980s, and the housing crash of 2008. Some are technological. In the 1840s, railways were hailed as the veins of a new industrial body, and they were. But a body needs only so many veins, and tracks were soon laid in places commerce could not sustain. Telecoms in the 1990s promised a wired utopia, only for 70 million miles of excess fiber to lie dark underground. The dot-com boom gave us a vision of a new economy, much of which did eventually materialize, but not before valuations evaporated in 2000. The funny thing is, there doesn’t seem to be an academic consensus on what an investment bubble is. Nobel laureate in economics Eugene Fama has gone so far as to say that they don’t exist. Azeem’s goal with this piece is to go beyond the I know a bubble when I see it kind of idea. He suggests that the two key pieces are when stock markets become overvalued and collapse, and then secondly, whether the quantity of productive capital, such as the money going into CapEx or VC, also collapses. To get specific, he writes, we see a bubble as being a 50% drawdown from the peak equity value that has sustained for at least five years. In the case of the U.S. Housing bubble and the dot-com, that trough was roughly five years long. Full recovery to pre-bubble peaks took 10 years for U.S. Housing and 15 for the dot-com. Alongside, we would expect a substantial decline in the rate of productive capital deployed once again 50% from peak. Ultimately, he says, a bubble means a phase marked by a rapid escalation in prices and investment, where valuations drift materially away from the underlying prospects and realistic Earnings power of the assets involved. Bubbles thrive on abundant capital and seductive narratives, and they tend to end in a sharp and sustained reversal that wipes out much of the paper wealth created on the way up. A boom, by contrast, can look very similar in its early stages with rising valuations and accelerating investment. But the crucial distinction is that in a boom, fundamentals eventually catch up. The underlying cash flows, productivity gains, or genuine demand growth rise to meet the optimism. Booms can still overshoot, but they consolidate into durable industries and lasting economic value. (Time 0:04:18)
  • Five Gauges To Judge AI
    • Use a multi-gauge framework to judge AI’s market state rather than a single indicator.
    • Azeem proposes five gauges: economic strain, industry strain, revenue growth, valuation heat, and funding quality. Transcript: Nathaniel Whittemore There are five. Gauge one, economic strain. Is investment now large enough to bend the economy? Gauge two, industry strain. Are industry revenues commensurate with deployed capex? Gauge three, revenue growth. Is revenue rising or broadening fast enough to catch up? Gauge four, valuation heat. How hard are valuations? Compared to history, are stocks excessively overpriced? And gauge five, funding quality. What kind of money is funding this? Is it strong balance sheets or fragile flighty capital? What What Azeem does with the rest of the essay is to look at each of these five areas and put them on a green-yellow kind of scale, where ultimately he argues that two reds equals trouble When it comes to bubble analysis. (Time 0:07:53)
  • CapEx Scale And Depreciation Risk
    • AI CapEx is massive but not yet as economy-bending as past manias when measured as GDP share.
    • Shorter GPU lifespans make AI CapEx more demanding and may force quicker market corrections. Transcript: Nathaniel Whittemore Azeem writes, the investment underway is vast, with Morgan Stanley expecting $3 trillion in AI infrastructure spend by 2029. But it has not yet reached the runway extremes of history’s great blowouts. But as Azeem points out, it’s not just the sheer magnitude that is the question. The other big factor is what he calls dependence. He points out that in the US, a third of GDP growth right now can be traced to data center construction. You might remember this chart I shared last month that showed that the pace of data center construction was about to overtake the pace of general office construction. Now, when it comes to the economy’s overall dependence on this particular area, Azeem writes, it’s not inherently bad, but it may be dangerous if the momentum falters. An economy leaning this heavily on one sector for growth can find the ground falling away faster than expected. And by the way, I think that this is actually why the bubble narrative is so ever-present, and part of why people can’t just be excited about what’s happening. There is actually an interesting argument that people are so concerned about their analysis that this is a bubble, that they’re not recognizing that AI is one of very few technology Changes that actually has the chance to create jobs in the short term, even as it’s going through the phases of creative destruction. Creative destruction is, of course, the famous idea from Joseph Schumpter that new technologies, while inevitably creating new things, wreak a path of destruction in terms of old Processes and old jobs and old roles and old industries even that get replaced and competed away. In the process of creative destruction, we usually see the destruction before the creative. One of the reasons that there’s anxiety around AI and agent-related job loss is that it’s much easier to see the one-to replacement effects for certain highly capable agents on jobs That exist today than it is to imagine five years out what new industries are going to be created by the new capabilities that these technologies represent. This is sort of the normal pattern of new technology shifts. However, with all of this CapEx spend, with data center construction, there are big sectors of the economy, things like construction, that are going to be forced to hire new people, Upskill them in new ways, in ways that could be value accretive in the short term. But again, people are so concerned that it’s over-exuberant and it’s all going to go away that they’re not really letting themselves get excited about those short-term gains. Anyway, coming back to Azeem, he writes, The surge of CapEx poured into the physical infrastructure that AI demands is an act of optimism. This is what CapEx is. Money spent today in the belief that it will become a funnel of revenue tomorrow. If it’s well-placed today, it will eventually lead to productivity gains and economic expansions. AI data centers are not just factories for a single product, they are infrastructure. Microsoft, OpenAI, and the U.S. Government all see it this way. They see compute as a foundational utility of the 21st century, no less critical than highways, railways, power grids, or telecom networks were in earlier eras. To build such infrastructure inevitably requires historic sums, on a par with the railway or electricity build-outs of the past. Azeem points out that in sheer terms, AI data centers are likely to be among the largest infrastructure build-outs in modern history. However, he writes, useful though infrastructure is, especially when private capital gets involved, things can get divorced from reality. The financing structure matters as much as the technology itself. He then points out the difference between the way that railways were funded versus electricity and road systems. Railways were largely private and had a number of different investment bubbles, whereas as he puts it, electricity and road systems benefited from greater public investment and Coordination and were less prone to speculative excess. A boom becomes dangerous when the resources it demands start to bend the whole economy around it. Wages get sucked into one sector, supply chains reorient to serve it, and capital markets grow dependent on it. The snapback is vicious when expectations break. Consequently, he suggests that a good way to gauge the economic strain is to look at investment as a share of GDP. He calls this a crude but telling ratio, showing how heavily the economy leans on one technological bet. By this measure of past bubbles, the railway bubbles were the heaviest. In the US, railway spending peaked at about 4% of GDP in 1872, which was just before the first crash in that area. On the other end of the spectrum, the telecom boom of the late 1990s topped out at around 1% of GDP. Azim writes, the AI buildout sits in the middle zone. Around 370 billion is expected to flow into data centers globally in 2025, with perhaps 70% earmarked for the US or roughly 0.9% of American GDP. Goldman Sachs projects spending will climb by another 17% in 2026. My own forecasts are in line with this view, annual capex of 800 billion by 2030, perhaps 60% of the US, which would bring the American share to 1.6% of 2025 GDP. So for his economic strain gauge, he places the green segment as a technology representing up to 1% of GDP, the yellow as up to 2% of GDP, and the red as above 2%. Generative AI today at 0.9% is still in the green, although it could be heading into the yellow soon. The biggest caveat with the economic strain analysis here is that the depreciation of AI CapEx could be much faster than the comparative depreciation of railway track or telecom fiber. GPUs, he writes, by contrast, age in dog years. Their useful life for frontier applications such as model training is perhaps three years, after which they are relegated to lower-intensity tasks. Roughly a third of hyperscaler capex is going into such short-lived assets. They remain, in theory, monetizable in years five and six. The rest goes into shells, power, and cooling that last two or three decades. Adjusting for asset life makes the AI build-out look even more demanding. Unlike railroads or fiber, the system must earn its keep in a handful of years, not generations. Interestingly, he says, while the negative implications of that are clear, there is also an optimistic case. He writes, Shorter depreciation cycles may impose financial discipline on incoming investors. During the railway mania, decades-long asset lives masked the weakness of many business models. Companies could stagger on for years before insolvency. In AI, the flaws may surface quickly, forcing either rapid adaptation or rapid failure. Ultimately, he concludes, the strain is noticeable but not yet unbearable. (Time 0:08:35)
  • AI Build-Out Creates Short-Term Jobs
    • AI build-out can create near-term jobs in sectors like construction even amid creative destruction.
    • People often miss short-term employment gains because destruction effects are more visible than creation. Transcript: Nathaniel Whittemore And by the way, I think that this is actually why the bubble narrative is so ever-present, and part of why people can’t just be excited about what’s happening. There is actually an interesting argument that people are so concerned about their analysis that this is a bubble, that they’re not recognizing that AI is one of very few technology Changes that actually has the chance to create jobs in the short term, even as it’s going through the phases of creative destruction. Creative destruction is, of course, the famous idea from Joseph Schumpter that new technologies, while inevitably creating new things, wreak a path of destruction in terms of old Processes and old jobs and old roles and old industries even that get replaced and competed away. In the process of creative destruction, we usually see the destruction before the creative. One of the reasons that there’s anxiety around AI and agent-related job loss is that it’s much easier to see the one-to replacement effects for certain highly capable agents on jobs That exist today than it is to imagine five years out what new industries are going to be created by the new capabilities that these technologies represent. This is sort of the normal pattern of new technology shifts. However, with all of this CapEx spend, with data center construction, there are big sectors of the economy, things like construction, that are going to be forced to hire new people, Upskill them in new ways, in ways that could be value accretive in the short term. But again, people are so concerned that it’s over-exuberant and it’s all going to go away that they’re not really letting themselves get excited about those short-term gains. (Time 0:09:21)
  • Factor GPU Depreciation Into Valuations
    • Consider asset life when evaluating AI infrastructure investments because GPUs decay quickly.
    • Short depreciation can force faster adaptation or rapid failure, so assess replacement cycles tightly. Transcript: Nathaniel Whittemore The financing structure matters as much as the technology itself. He then points out the difference between the way that railways were funded versus electricity and road systems. Railways were largely private and had a number of different investment bubbles, whereas as he puts it, electricity and road systems benefited from greater public investment and Coordination and were less prone to speculative excess. A boom becomes dangerous when the resources it demands start to bend the whole economy around it. Wages get sucked into one sector, supply chains reorient to serve it, and capital markets grow dependent on it. The snapback is vicious when expectations break. Consequently, he suggests that a good way to gauge the economic strain is to look at investment as a share of GDP. He calls this a crude but telling ratio, showing how heavily the economy leans on one technological bet. By this measure of past bubbles, the railway bubbles were the heaviest. In the US, railway spending peaked at about 4% of GDP in 1872, which was just before the first crash in that area. On the other end of the spectrum, the telecom boom of the late 1990s topped out at around 1% of GDP. Azim writes, the AI buildout sits in the middle zone. Around 370 billion is expected to flow into data centers globally in 2025, with perhaps 70% earmarked for the US or roughly 0.9% of American GDP. Goldman Sachs projects spending will climb by another 17% in 2026. My own forecasts are in line with this view, annual capex of 800 billion by 2030, perhaps 60% of the US, which would bring the American share to 1.6% of 2025 GDP. So for his economic strain gauge, he places the green segment as a technology representing up to 1% of GDP, the yellow as up to 2% of GDP, and the red as above 2%. Generative AI today at 0.9% is still in the green, although it could be heading into the yellow soon. The biggest caveat with the economic strain analysis here is that the depreciation of AI CapEx could be much faster than the comparative depreciation of railway track or telecom fiber. GPUs, he writes, by contrast, age in dog years. Their useful life for frontier applications such as model training is perhaps three years, after which they are relegated to lower-intensity tasks. Roughly a third of hyperscaler capex is going into such short-lived assets. They remain, in theory, monetizable in years five and six. The rest goes into shells, power, and cooling that last two or three decades. Adjusting for asset life makes the AI build-out look even more demanding. Unlike railroads or fiber, the system must earn its keep in a handful of years, not generations. Interestingly, he says, while the negative implications of that are clear, there is also an optimistic case. He writes, Shorter depreciation cycles may impose financial discipline on incoming investors. During the railway mania, decades-long asset lives masked the weakness of many business models. Companies could stagger on for years before insolvency. (Time 0:11:40)
  • CapEx-To-Revenue Tells A Strain Story
    • Industry strain measures CapEx relative to revenues and flags overbuilding risk.
    • Gen AI runs at an unusually high CapEx-to-revenue ratio compared to past booms, nearing warning levels. Transcript: Nathaniel Whittemore Every boom, he writes, needs to prove that the money poured into new equipment is starting to earn its keep. In any growth stage, it is unlikely that revenues will cover investment, but they should be non-zero. This gauge looks at the ratio of capex to revenues. Now, there have been a number of different estimates of how to look at Gen.ai revenues. Some of the most common that you see, especially when people are trying to say that there’s a bubble, is just adding up the revenue of Open.ai, Anthropic, and a handful of other startups, Usually pointing to a number that comes to between $15 and $18 billion, and saying something to the effect of how could that possibly justify the hundreds of billions being spent in Infrastructure? Azeem’s estimates are over $60 billion this year, and even that, he says, could undercount the He writes, value. He writes, And yet it’s undeniable that CapEx intensity is also increasing. He writes, In 2021, before ChachiBT, hyperscalers invested about 44% of their operating cash flow in CapEx. By 2024, that had risen to 68%, and in 2028, it will be higher still. But these firms can’t absorb this shift by replatforming, with structurally higher capital intensity driving growth and efficiency gains. This dynamic has been in place for a decade already. Between 2015 and 2018, he writes, Microsoft Azure’s CapEx represented between 70 and 90 percent of revenues. It was an investment in the future. He continues, This makes for an interesting comparison to earlier boom cycles. The railroads are particularly pertinent. The railroad’s direct revenue contribution was tiny compared to the value the railroads created in the U.S. Economy. Railway bubbles were always tethered to the reality of cash flow. The bonds issued to finance new track and rolling stock had to be serviced out of passenger fares and freight revenues. Whenever CapEx outpaced earnings, the strain showed. The manias of 1873, 1883, and 1887 all followed the same pattern, a sharp decline in the ratio of annual revenues to capital spending, and in some cases outright revenue contraction. At the height of the U.S. Railroad expansion in 1872, CapEx was around two times revenues. In the late 1990s telecom bubble, CapEx amounted to just under four times revenues. By contrast, today’s Gen AI boom runs on roughly $60 billion in revenues, against about $370 billion in global data center CapEx, a CapEx-to ratio of six times, the most stretched of The three. On the industry strain gauge then, railways sat healthily in the green, Gen.ai was in the yellow, but Gen.ai is in the yellow nearing red. (Time 0:17:42)
  • Revenue Momentum Supports A Boom
    • Revenue growth for Gen AI is rapid and accelerating, unlike pre-crash patterns in past bubbles.
    • Continued doubling and strong enterprise demand support the case this is a boom, not a bubble. Transcript: Nathaniel Whittemore He writes, the problem in the railroad and telecom booms was not sector strain per se, but that revenues ran out of momentum. Investment expects a return. After the railway bubble burst in 1873, revenue declined by 3% year over year. Telecoms did slightly better, declining 0.5%. Before the crashes, revenue growth was hardly explosive. Railways in 1873 expanded 22%, enough to double in three years. Telecom in the late 1990s managed only 16%, a doubling time of just over four years. By contrast, Gen AI revenues are still accelerating. By our estimates, Gen AI revenues will grow about twofold this year. And this is likely a conservative forecast. Citi estimates that model makers’ revenue will grow 483% in 2025. Open AI forecasts annualized growth of about 73% to 2030, while analysts like Morgan Stanley estimate this market could be as large as $1 trillion by 2028, equivalent to compound growth Of 122% a year over the period. This puts generative AI, in his estimation, very squarely in the green when it comes to revenue growth, given that basically every estimate has, at the very minimum, gen AI doubling Every year. Azeem continues, In my conversations with large companies, I get the strong sense that they can’t get enough of this technology right now, and this likely supports the strong growth Rates. IBM’s CEO survey shows that Gen.ai is already expanding IT budgets, with 62% of respondents indicating that they will increase their AI investments in 2025. If you listened to Friday’s show, you will have heard me talk about KPMG’s latest Pulse survey, where the anticipated investment among their 130 survey companies with a billion dollars In revenue each was $130 million over the next 12 months, which was up from a Azeem continues, million earlier in the year and $88 million in Q4 of last year. Ultimately, he points out that we are still at the, quote, foothills of enterprise use. For now, firms can barely secure enough tokens to meet their needs. He also notes that the consumer side tells a parallel story. U.S. Consumers, he writes, already spend around $1.4 trillion a year online. This could plausibly double to $3 trillion by 2030 if it grows at 15% to 17% a year, and it has grown at more than 14% a year since 2013. Against this backdrop, a Gen. AI app sector rising from today’s $10 billion to $500 billion within five years looks less far-fetched. Growth rates of 300 to 500% are already visible in mid-sized startups and the large model providers, suggesting that even a small reallocation of consumer digital spending could Drive revenues into the hundreds of billions. Taken together, he says, these signals point to an industry still in strong ascent, unlike the relatively meager revenue growth that preceded the railway and telecom busts. If Gen. AI revenues were to grow at even half the pace of last year, then on my conservative forecast, they would reach $100 billion by 2026, covering about 25% of that year’s capex. (Time 0:20:53)
  • Valuations Are Elevated But Cooler Than Dot-Com
    • Valuation multiples today are elevated but far below dot-com extremes.
    • Market PE levels suggest investors are betting on growth but not in the same manic way as 2000. Transcript: Nathaniel Whittemore He writes, this is often where bubbles reveal themselves most clearly, how exuberantly investors are pricing the sector regardless of fundamentals. And this is almost a default part of new technology cycles. He writes, as Carlotta Perez has argued for decades, financial markets tend to overshoot in the early installation phase of each technological revolution, pouring in capital far Beyond what near-term revenues justify. The frenzy looks irrational in the moment, but it is the mechanism by which society lays down the new infrastructure. The challenge is whether the frenzy can evolve into about this one is the dot-com bubble. Companies raising tens or hundreds of millions of dollars on absolutely no revenue and no business model to speak of ultimately led to the bubble popping and a long, slow rebuild where So many of the ideas that were initially present actually came to fruition backed by real revenue and real growth. Azeem importantly makes the point, what’s going on in Gen. AI does not compare to this. The key measure here is the price earnings ratio, or PE, a shorthand for how many years of current profits an investor is effectively paying for. A high PE means companies are betting on rapid future growth, but too high for too long, investors might be buying into a fantasy. This was the case in the dot-com era. At the peak, the Nasdaq traded at a P.E. Of about 72. One detailed study estimated that internet stocks alone carried an implied P.E. Of 605. In other words, investors were willing to pay for more than six centuries of current earnings. The issue wasn’t that the demand disappeared. Amazon’s revenues grew from $2.76 billion in 2000 to $3.12 billion in 2001, but that no company could grow fast enough to justify those sky-high expectations. In other words, he writes the fundamentals improved, but expectations collapsed. Today, he says the picture is much calmer. The Nasdaq PE is about 32, half of the dot-com era. The broader tech market is higher than the long-run average, but nowhere near dot-com territory. Ultimately, he puts this category in the green as well. (Time 0:23:35)
  • VC Hype Has Limited Market Impact
    • High venture valuations matter less to market stability because VC is small and expected to fail often.
    • Startups show unusually rapid revenue growth, softening classic VC bubble concerns. Transcript: Nathaniel Whittemore You will sometimes see people point to the incredibly high valuations that startups are getting in the venture realm as evidence of bubble. I think there are a couple things that make that not all that concerning when you look at it in this overall context. The first and most obvious one is that venture dollars ultimately are a tiny, tiny portion of the financial markets. In 2024, U.S. Venture investment deals were worth $215.4 billion. Last week, after Oracle announced its projections for the coming years, the company added $244 billion to its market cap, about $30 billion more than all of VC in 2024. The point is that VC is just ultimately relatively small and has an outsized impact on our imagination as compared to what it does in our markets. Second, its impact is even lessened because VC is meant explicitly to take risky bets. Venture capitalists are well-prepared, as are LPs, for the idea that most of their investments will fail. That’s just built into the industry. And finally, while valuations are high and moving quickly, the rate of revenue growth among many of these companies is unlike anything we’ve ever seen in startups before. One may quibble about how durable that revenue is. There is a huge amount of, for example, what some have called curiosity revenue, where enterprises and consumers try interesting new products but then don’t stick around, that maybe Are warping some of those growth numbers. But still, taken as a whole, companies are making more and faster in this technology shift than just about any we’ve ever seen before. (Time 0:25:33)
  • Who Pays Determines Fragility
    • Funding quality matters: who pays and how patient capital is determines systemic fragility.
    • Big hyperscalers can self-fund much of the build-out, but a large debt gap would reintroduce classic bubble vulnerabilities. Transcript: Nathaniel Whittemore Funding quality, Azim writes, is not a standard metric, but a composite judgment. It asks who the money is coming from, how it is structured, and whether the capital is willing to wait years for returns or rather chase quarterly pops. Low quality capital, in short, is short-termist, undisciplined, and debt-laden. It rushes and flees quickly. High quality capital is more patient, better underwritten, and able to withstand volatility. Every bubble has its signature weakness, invariably rooted in how it was financed. Railways were fueled by speculative retail investors with little capital behind them. By the early 1870s, funded debt averaged 46% of total assets among American railroads. When overbuilding Metacredit squeeze, financing evaporated. Dotcom firms a century later were a little sturdier. Venture capital was a boutique business in 1995 with only $5.3 billion deployed. By 2001, more than $237 billion had been poured into startups, often by new and inexperienced managers. The frenzy spilled into public markets. IPO volume between 1999 and 2000 ran six times above historical averages. Companies went public with little revenue. One other aside note on the manager thing, by the way, not only do you not have new and inexperienced managers running around venture capital right now, you actually have the exact opposite. Alpins have been starved of liquidity for so long that the entire industry has been forced to start to use secondaries as a mechanism to get some liquidity. And many, many funds in the wake, especially of rising interest rates in the post ZERP era have shuttered their doors and not been able to raise again, meaning the crop of people that Are around now are a lot more battle-hardened, even relative to venture capital terms. Anyways, back to Azeem, he continues, Telecoms in the late 1990s leaned on mountains of cheap debt. US and European carriers doubled and quadrupled their leverage in just a few years. Deutsche Telekom and France Telekom together added $78 billion in net debt between 1998 and 2001. When revenues failed to keep pace, defaults rippled through the sector. In each case, the capital that fueled the boom proved ephemeral, but the degree of fragility differed. Railways and telecoms were most exposed to credit crunches, with debt ratios ballooning. Dot-coms were hostage to market mood with equity values evaporating. On this front, today’s AI boom looks sturdier. Microsoft, Amazon, Alphabet, Meta, and NVIDIA are minting extraordinary cash flows, easily enough to bankroll their own build out. For now. But investment needs are racing ahead. Morgan Stanley reckons total global data center CapEx will hit $2.9 trillion between 2025 and 2028. Hyperscalers can cover perhaps half of that from internal cash. The rest must come from private credit, securitized finance, and new operators. Governments have also pledged $1.6 trillion in sovereign AI investments, and Gulf Capital is seeking new opportunities. Here is where the risk creeps in. Morgan Stanley itself points to a $1.5 trillion gap that will need to be plugged by debt markets and asset-backed securities. The sums are enormous. $800 billion from private credit, $150 billion in data center ABS, and hundreds of billions more in OEM loans and vendor financing. That $150 billion alone would triple the size of the data center securitized markets almost overnight. And not every borrower looks like Microsoft. So the point for him is that right now, as much as you could say companies are spending too much on CapEx, they’re spending their own money on CapEx. They’re not going into debt or finding weird novel instruments to do so. The question is how long that can persist. As Azim sums up, the foundations are stronger than in past bubbles, but the superstructure is starting to resemble the old pattern. Esoteric debt structures, concentrated counterparties, and hardware that may not hold value are reappearing. If Gen.ai revenues grow tenfold, creditors will be fine. If not, they may discover that a warehouse full of obsolete GPUs is a different thing to secure. (Time 0:27:00)
  • Net Assessment: Boom Not Bubble
    • Applying the five gauges today places AI in ‘boom’ territory rather than ‘bubble’.
    • Only industry strain edges toward red, while other gauges remain green under current data. Transcript: Nathaniel Whittemore Simply put, we are in boom territory, not bubble. Of Azeem’s five gauges, four of them, economic strain, revenue growth, valuation heat, and funding quality are all in the green right now. The closest to red is industry strain, which is again a measure of CapEx investment divided by revenue, where he is purposely going with a relatively modest or conservative revenue Number. The inescapable conclusion is AI is not a bubble. But of course, that doesn’t mean that it won’t become one. The pressure points that Azeem considers worth watching include more and more of the economy relying on AI, basically if investment climbs towards 2% of GDP, a sustained fall in enterprise Or consumer spending, especially if it’s followed by a shrinking order backlog for companies like NVIDIA, valuations jumping from where they are at a PE ratio of 32 right now to up to 50 or 60, and lastly, if a greater and greater portion of CAPEX starts to be covered off of the balance sheet. He concludes, my current heuristic is that if two of the five gauges head into the red, you’re in bubble territory. Time to sell up, buy the VIX, and take some deep breaths. Gen AI isn’t there yet. Racing fast, the engine is whining, but not overheating. (Time 0:30:27)
  • Use Two Red Gauges As A Sell Signal
    • Watch for two gauges turning red as your bubble signal and adjust exposure accordingly.
    • Monitor GDP share of AI investment, revenue contraction, valuation spikes, and rising leverage closely. Transcript: Nathaniel Whittemore But of course, that doesn’t mean that it won’t become one. The pressure points that Azeem considers worth watching include more and more of the economy relying on AI, basically if investment climbs towards 2% of GDP, a sustained fall in enterprise Or consumer spending, especially if it’s followed by a shrinking order backlog for companies like NVIDIA, valuations jumping from where they are at a PE ratio of 32 right now to up to 50 or 60, and lastly, if a greater and greater portion of CAPEX starts to be covered off of the balance sheet. He concludes, my current heuristic is that if two of the five gauges head into the red, you’re in bubble territory. Time to sell up, buy the VIX, and take some deep breaths. Gen AI isn’t there yet. Racing fast, the engine is whining, but not overheating. How long would it take for two gauges to get into the red? I’ve toyed around with combinations and most scary scenarios take a couple of years to play out. And not all scenarios are scary. That said, so many macro factors, from a recession in the US to rising inflation, a challenging interest rate environment, and domestic or international politics could dampen spirits. While we might not be solidly in bubble land, it would be hubristic to assume the AI investment cycle is immune to those exuberant dynamics. (Time 0:30:52)