AI demand called unlimited, but chip stocks fall as spending questions mount across industry

Who pays for AI?

Good morning. A 77 per cent increase in second-quarter earnings from Taiwan Semiconductor Manufacturing Co was not enough to boost its shares, those of the chip industry generally or indeed the Nasdaq 100. All of those fell yesterday. TSMC is the world's most irreplaceable company and it is firing on all cylinders. However high its flying, though, expectations are flying higher. More ruminations on this theme below. Email us: [email protected]. AI: monetisation or bust The era of asking cutting-edge AI models for chicken soup recipes is coming to an end. That should worry everyone on the long side of the AI trade -- which, in 2026, means just about everyone, directly or indirectly. For three years, the AI investment story has gone something like this. Frontier AI model-makers such as Anthropic and OpenAI provided their products at well below cost, in the form of consumer-facing chatbots, allowing users to burn through the most precious commodity in the AI supply chain: compute capacity. AI users lived a subsidised existence while investors happily picked up the bill, as each successive fundraising round increased their paper gains (old heads will hear the echo of the eyeball- and click-based valuations at the turn of the millennium). Most of the value went straight to the chip industry -- the only companies positioned to reap an immediate return from a technology that consumes more capital than it generates and may do so for years to come. Again, investors didn't mind. They believed that once the technology achieved mass adoption, the model makers and data centre builders would capture epic profits and live happily ever after. That story may be running out of road, as the model makers come under pressure to capture some of the value from their applications. In the past few months, a growing number of AI-related services have switched from flat fees to usage-based pricing, starting with Microsoft-owned coding platform GitHub, in April. This "seismic shift" marks the "end of the token subsidy" that has underwritten AI's rapid adoption to date, argues Juan Correa at BCA Research. Companies faced with ballooning AI bills are moving from "tokenmaxxing" to token rationing, says Bank of America. Switching to consumption-based pricing has had two unintended effects. The most notable is the rapid increase in usage of Chinese AI models, which tend to be open source and lower cost: The second effect of this switch is that it appears to be pulling down spending per token, as lower-cost models account for a higher proportion of token usage: On the face of it, this raises serious questions about how the hundreds of billions in investment is expected to turn into profits, let alone revenue. But some, such as Torsten Sløk at Apollo, see this as early evidence of Jevons paradox applying to AI, the idea that an increase in efficiency leads to an increase in usage. On that view, falling unit costs will lead companies to increase their total spending on AI as they apply AI to more and more of their business. This might well be true, and the profitability of the AI companies depends on it. An even more alarming notion is that AI will become commoditised. This is the idea that an abundance of AI supply and broadly interchangeable models will make the technology more like electricity than software, writes Tyler Frawley at RBC Wealth Management. In that scenario, the real economic value accrues not to the producer but to the companies that build the most effective systems on top of it. If that's right, the bulk of the investment in the model makers has been misallocated and the returns will be terrible. If the model layer is being commoditised, the infrastructure layer isn't obviously safer. Investors in hyperscalers are increasingly questioning how the massive capital spending needed for AI infrastructure -- estimated by Morgan Stanley to rise to $1.2tn next year, at just five companies -- will be financed. Many have pointed to the collapse in free cash flows and the tsunami of hyperscaler bond issuance. The ratio of capex to revenue is also mounting: Summing up, the AI trade rests on two big assumptions. The first is that AI will be profitable. But as noted above, just because a technology leads to a huge increase in productivity doesn't mean it will generate strong returns. The second is that there will be widespread demand for AI, and soon. AI adoption has been fast, to be sure. But we're still a long way from peak adoption. Goldman Sachs estimates that it could take as long as 15 years, which would still be faster than the median adoption rate of 29 years for previous general-purpose technologies. Bloomberg reported earlier this month that Meta is looking to sell its excess compute capacity. It probably won't be the last, and it smells of malinvestment. But perhaps the clearest evidence that all is not well in the AI trade is hyperscaler stock price performance over the past three months, which has gone nowhere even as capex guidance keeps rising: Excitement about AI as a technology is different from exuberance about AI as a trade. Unhedged has lots of the former but, at the present juncture, not much of the latter. Feel differently? Email us.

'Almost unlimited': Execs says AI demand remains strong even as enterprises move to 'valuemaxxing'

AI-related chip stocks have been volatile amid a debate over AI demand and spending. Chip stocks have had a blistering rally over the past year as investors bet on the semiconductor sector's central role in the global AI infrastructure buildout. But renewed volatility around chip stocks has sparked a debate if this is a sign of broader concern about AI demand. In interviews with CNBC this week, several AI executives poured cold water over the idea that demand is slowing, even as they acknowledged that businesses are being more cautious on the cost of using AI. "I somewhat think of AI demand as almost unlimited," Pat Gelsinger, the former Intel CEO and now general partner at Playground Global, told CNBC on Wednesday, adding that energy availability is "the only real limiter." "Because how much economic value do you get for increased intelligence? Almost infinite across every industry imaginable," Gelsinger added. A number of factors have stoked volatility in markets around chip and AI data center-related stocks. An announcement from Meta that it will sell its excess AI computing capacity was in part a contributor to the sell-off. While Meta's stock popped on the news, it raised questions over whether this was a sign that there was broader overcapacity of compute out there. Elon Musk's xAI also rented its excess capacity out this year. And this week, Samsung, one of the world's biggest memory chip companies, forecast a gigantic rise in profit, but its stock fell. After a more than 360% rally in its shares over the last 12 months, the market questioned how much further it could go. None of these moves appears to have dampened demand for compute and the infrastructure behind it. "What we're experiencing in terms of demand is extraordinary. There's much more demand than we're able to fulfil, and that's been our experience for some time now," Marc Boroditsky, chief revenue officer at Nebius, told CNBC on Thursday. Nebius is building data centers using Nvidia's GPUs. Andrew Feldman, CEO of Cerebras Systems, said the example of Meta and xAI selling its excess capacity is a "unique" case. "For the industry as a whole, the demand for compute far outstrips available capacity, and we're short on data centers. I think we're short on, as an industry, many of the inputs to compute," Feldman told CNBC on Wednesday. Cerebras, which went public earlier this year, is one of a slew of semiconductor startups attempting to become major players in the data center market and challenge Nvidia. Rebellions, another chip startup from South Korea, which is backed by Samsung and SK Hynix, reported seeing similar ample demand. "AI infrastructure momentum [is] still huge," Sungyun Park, CEO of Rebellions, told CNBC on Wednesday. "I personally believe it's not the signal saying that ... all the hyperscalers [are overinvesting] in the infrastructure," Park added in reference to the Meta and xAI news. Lumentum, which sells photonics and optical products for connectivity in the data center, said its products are sold out for the next five years. "We're trying to build up our capacity as much as we possibly can to fulfil a demand that we see out five years at this point," Michael Hurlston, CEO of Lumentum, told CNBC on Wednesday. Lumentum's stock is up around 600% over the last 12 months as investors pile into companies addressing key bottlenecks in the buildout of AI data centers. Another big debate around the AI trade is how much enterprises are willing to pay for the technology. There has been a period of so-called 'tokenmaxxing' at enterprises where companies would encourage employees to use as much AI as possible no matter the result. The tools often used were those from frontier labs like OpenAI and Anthropic. But companies are now focusing more on the return on investment from AI, especially as those frontier models remain expensive relative to open source offerings from companies like DeepSeek or Alibaba. Nebius' Boroditsky said that tokenmaxxing is only worthwhile if an organization is seeing a return on investment as a result. "The CFO bringing the hammer down and slowing spend should actually be looking for value or valuemaxxing," Boroditsky said, adding that AI should be applied to create value that justifies the spending. "We're seeing a shift now to more rationalization. We've seen it with every tech cycle, and that rationalization will definitely continue the demand," Nebius' Boroditsky said. While frontier AI models are seen as the most advanced, there are a plethora of open source models that are close in performance and some that are less advanced. Different models have different capabilities, which can be used for specific tasks. Cerebras' Feldman said that in the future, certain models will be used in specific situations. For example, frontier models can be used for more advanced problems, while some workloads will shift to others. "I think it's probably the case that you don't need a giant bus to go to the grocery store," Feldman said. "Certain workloads migrate to some type of compute and easier workloads to others, and I think as we learn and become more sophisticated in our deployment of AI, the same thing will happen." Choose CNBC as your preferred source on Google and never miss a moment from the most trusted name in business news.

Big Tech Firms Like Oracle Turn to Bonds to Finance A.I. Data Centers

Jeff Sommer writes Strategies, a weekly column on markets, finance and the economy. Wealth from Oracle, the giant tech company founded by Larry Ellison, is enabling Larry and his son, David, to become media moguls. Thanks to backing from Larry's Oracle billions, David has taken control of Paramount and is now engaged in a hotly contested $111 billion bid to take over Warner Bros. Discovery, too. They are trying build a media behemoth containing two big movie studios, multiple streaming services and the news networks CNN and CBS News, all under one enormous corporate roof. The fight over the Oracle-financed empire has, understandably, captured plenty of headlines. But what hasn't received nearly as much attention is another important development, the downgrading of Oracle debt. It now stands just one notch above junk bond status. That happened on July 9, when S&P Global said that Oracle's finances had been deteriorating. Oracle has also been hit hard in the stock market, reducing the value of Larry Ellison's holdings since September by about $230 billion, according to my calculations based on FactSet data. What has damaged Oracle's debt rating and disturbed its finances is the elephant stomping throughout financial markets: colossal spending on artificial intelligence. Data centers and the other infrastructure for A.I. involve staggering sums of money. These cascades of A.I.-driven cash have enriched diverse segments of the stock market, from semiconductor makers to engineering companies to utilities to energy producers. A.I. money is bolstering the entire U.S. economy, contributing perhaps 1.1 percent to the nation's economic growth, JPMorgan Asset Management estimates. But where's that money coming from? At this point, a major source is firms like Oracle, which has gone on an immense spending spree on A.I. data centers, increasingly selling bonds to raise the money. Oracle is not alone. Alphabet, Microsoft, Amazon and Meta are giant investors in data centers, too. (The industry jargon is "hyperscaler.) But their underlying finances are stronger than Oracle's, and their expenditures have not landed them in the same level of trouble in the markets. Microsoft, for example, has a Triple-A credit rating -- better than the U.S. government's. Whether Microsoft manages to retain that rating after its splurges on A.I. data centers remains to be seen. "Microsoft is starting from a much better place, financially, than Oracle is," Mariya Entina, a portfolio manager for DoubleLine, the money management company, said in an interview. "It's important to have enough information to be able to differentiate." These five companies combined are pouring more than $800 billion into A.I. investments this year, and plan to add more than $1.2 trillion in 2027, according to Morgan Stanley. To put that in context, as Robert Armstrong of The Financial Times noted, the U.S. military budget request for 2027 is less than that: $961 billion, according to the Congressional Budget Office. Few people outside the markets have paid attention to what goes on behind the financial curtain for artificial intelligence. These big companies are able to categorize the money as an investment -- a capital expenditure -- and not as an expense. So under current accounting rules, the bulk of the spending has not yet counted against their gaudy earnings. That is helping to propel the stock market to new heights under rosy assumptions that A.I. will transform the world, and that the companies behind it will be making money. With the notable exception of Oracle, which has borrowed aggressively for the last couple of years, most of these companies generated so much cash from their main businesses that, until recently, their spending on A.I. data centers barely weighed on the performance of their stock or on the solidity of their underlying finances. But this year is turning out to be different. A.I. data centers are increasingly running on borrowed money. The problem goes way beyond Oracle. Hungry for Money The gigantic A.I. infrastructure expenditures are outpacing growth in profits. According to Bank of America, total capital expenditures for Oracle, Alphabet, Microsoft, Amazon and Meta are exceeding their free cash flow. That's the money their businesses generate beyond what they need to operate and invest in the future. The hunger for cash is likely to mount. Bank of America noted that these big tech companies, which formerly operated on relatively little invested capital, are now as capital-intensive as old-line fossil fuel companies like Exxon Mobil and Chevron. So the tech companies are going to the capital markets, mainly the bond market, which has begun to charge premiums for what it considers to be heightened risk. Oracle and Amazon bond prices have been hard hit. So have those issued by SpaceX, which is also building A.I. data centers. Its bonds are rated as investment grade but have been trading at fire-sale prices, like junk bonds. One problem is that the expected revenue for the data centers isn't rock solid. Much of it is linked to A.I. start-ups like OpenAI and Anthropic, which themselves rely on borrowed funds and speculative investments by venture capitalists and private equity funds. Oracle's heavy dependence on OpenAI makes it especially vulnerable, S&P Global said. In a presentation to reporters earlier this month, Savita Subramanian, Bank of America's chief equity strategist, drew parallels with the dot-com era of the late 1990s and early 2000s. The big "hyperscalers" have far more solid business models than many of the old internet companies did, she said, but their immense need for borrowing "is a little nerve-racking." If their returns from A.I. investments don't pan out, or if their borrowing costs become onerous because of rising rates on debt, these companies may not be in an enviable position. There will be questions about whether their share pricing is "appropriate," she said, given their "leverage and capital intensity." A Great Winnowing There are signs that the markets may have started to recoil from some of the more extravagant A.I. bets. Four of the five big, long-established data-center companies have underperformed the S&P 500 this year. Oracle has been leading the pack downward, with a fall of more than 35 percent through Thursday. Alphabet, on the other hand, has been ahead of the market, with a stock gain of more than 13 percent. Alphabet's bonds are faring better, too. It may not just be that its Gemini A.I. model is highly rated. The company's finances are more solid than Oracle's. It has plans to raise more money through bonds -- but also through additional equity sales, which would dilute the value of existing stock shares. The stock market has so far shrugged off that move. SpaceX became a publicly traded company on June 8 -- and is building big A.I. data centers with borrowed money. Its share price has been otherworldly, although the company has no earnings. The consensus estimate is that it will generate some next year -- but only enough to give it a price-to-earnings ratio of 182, based on its current share price, according to FactSet. That number, which measures a stock price against a company's earnings, is still off the charts: It's 6.5 times the valuation of the average company in the S&P 500. This week, SpaceX shares for the first time fell below their initial public offering price, a move that I've suggested is warranted. One day earlier, IBM's shares lost 25.2 percent. That was its steepest daily decline since the 1960s, and it was set off by an earnings shortfall that its chief executive attributed, in part, to the spending underway on A.I. data centers. "We did not anticipate the magnitude of the capex reprioritization," Arvind Krishna, the company's chief executive, wrote in a letter to investors. Other companies spent so much money to build A.I. foundations, he said, that there wasn't as much left as expected for software service companies like IBM. These are early days. I have no doubt that artificial intelligence is an important technology. Great fortunes are already being made. But I'm also certain that there will be many losers, as there were in two other episodes of mammoth infrastructure investments in budding technologies: the railroads in the 19th century and the various early internet companies of the dot-com era. Well-run, diversified and deep-pocketed companies have a better chance of survival in epochs like these than those that take on inordinate risk with their capital investments. Even so, the future champions may not be any of the early giants. A great winnowing is coming, and prudent investors will accept that they cannot know in advance who the winners and losers will be.

AI demand is 'unlimited'. So why are chip stocks falling?

AI executives insist demand is "almost unlimited", with Pat Gelsinger naming energy as the only real limiter and Lumentum reporting products sold out five years ahead. Yet chip and data-centre stocks keep lurching, because a ~60% year-to-date rally in the PHLX chip index prices in flawless execution: Samsung forecast a huge profit rise and still fell, and Meta's plan to sell excess compute cut both ways. The gap is about expectations, not demand. Executives building the AI boom are unwavering. Demand is effectively bottomless, they say, even as the stocks that ride on it wobble, CNBC reports. Pat Gelsinger, the former Intel chief now at Playground Global, put it plainly. He thinks of AI demand as almost unlimited, with energy availability "the only real limiter". The order books support him. Lumentum, which supplies optical components for data-centre connectivity, says its products are sold out for the next five years. So why are the stocks jumpy? Because the price already assumes all of it. The PHLX chip index has gained roughly 60% this year, which prices in years of flawless execution. At that level, good news stops being good enough. Samsung forecast an enormous profit rise and its shares still fell, after a 12-month rally of more than 360%. The same pattern hit elsewhere, with Cerebras doubling its revenue only to watch the stock drop. When expectations run this hot, a beat can read as a miss. Meta added to the nerves by saying it would sell off its excess AI computing capacity. Investors could read that either as smart monetisation or as an admission the company bought more compute than it needs. The bull case and the bear case The bulls have real numbers behind them. Unlike the dot-com era, the companies driving this rally are extraordinarily profitable, and the demand signals from suppliers are not fabricated. SoftBank's Masayoshi Son has gone further, saying calling AI a bubble is an insult. On this view the build-out is a generational infrastructure project, not a mania. The bears do not really dispute the demand. They dispute the price, noting market concentration now exceeds 2000 levels and the returns on hundreds of billions in capex remain unproven. Both can be true at once. Demand can be genuine and the stocks can still be priced beyond what that demand will pay back on any sensible timeline. The constraint nobody can buy their way out of Gelsinger's caveat is the one worth sitting with. If energy is the binding limit, then chips are no longer the bottleneck, and the sector's valuations rest on infrastructure it does not control. Capital is chasing that gap, with Nvidia-backed startups raising to solve data-centre power. Grids, turbines, and planning permission move on timescales that ignore quarterly earnings. That is the awkward truth under the volatility. The industry has convinced itself demand is infinite, and it may well be right, but the electricity is finite and the share prices are not.

AI Bubble Fears Are Starting to Spill Over

Can't-miss innovations from the bleeding edge of science and tech Yet another domino appears to be falling as part of the Ruth Goldberg machine that will eventually pop the AI investment bubble. Earlier this month, economic forecasters were sounding the alarm that overspending on AI was at a level far more severe than Black Tuesday, the day that jump-started the worst economic catastrophe in the history of the industrial economy. Now, investors seem to be coming to that same conclusion all on their own. This week, Taiwanese semiconductor giant Taiwan Semiconductor Manufacturing (TSMC) posted its second-quarter earnings results, revealing a staggering revenue of over $40 billion -- a record-breaking sum for the company. While that should come as welcome news to investors, the results had the exact opposite effect, sending shares of TSMC stocks tumbling by four percent. That in turn led the tech-heavy Nasdaq 100 index to fall by 1.4 percent Thursday, compounding losses from Wednesday, Bloomberg reported. The issue seems to be TSMC's revision to its capital expenditure. As the key manufacturer for chip design company Nvidia -- arguably one of the most pivotal players in the AI boom -- TSMC is a major bellwether for investor confidence around the buzzy tech. The trouble is that, in addition to posting record revenue, the Taiwanese chip firm raised its 2026 spending forecast to a range of $60-64 billion, up from $52-56 billion. That shift will test how much more spending investors are willing to stomach on AI, a technology that has yet to justify the nearly $1.6 trillion spent developing it over the past decade. Zooming out, the market's response seems to indicate a few things. For starters, good news is clearly not enough to maintain confidence in the tech industry's ability to make AI a financial success. It likewise signals that belief in the AI bubble is no longer confined to a few brave contrarians, but is becoming a mainstream narrative, as investors grow wary of an industry which promises the Moon, yet only seems to deliver hot air. More on the AI bubble: Tech Billionaires Are Quietly Rooting for AI Bubble to Collapse

The AI boom is increasingly built on debt, but investor demand is plunging just as hyperscalers ramp up their bond blitz | Fortune

The stock market selloff has raised fears the AI boom is running on borrowed time. But it's also running on borrowed money, and Wall Street is less eager to provide a seemingly endless stream of debt. As so-called hyperscalers plow hundreds of billions of dollars a year into AI infrastructure, they have increasingly tapped bond markets to raise capital. That's in addition to drawing on cash flows and issuing new equity. In fact, since the start of 2025, Alphabet, Meta, Amazon and Oracle alone have issued more than $300 billion in bonds, according to Bloomberg calculations. AI chip leader Nvidia also issued $25 billion in bonds last month, marking its first such sale in five years. And SpaceX, which is now an AI player after acquiring xAI, sold $25 billion in bonds just days after its record IPO that raised $86 billion in stock. AI's insatiable demand for cash means the debt binge isn't expected to slow down anytime soon. The top five hyperscalers are expected to issue $300 billion annually in the coming years, up from $175 billion in 2026. That doesn't include SpaceX. JPMorgan estimated it will see $375 billion in debt proceeds from 2026 to 2030. But just as AI debt supply soars, investors appear to be losing their appetite. Amazon had to sweeten a "surprise" $25 billion bond sale earlier this month by offering 18 to 21 basis points of extra yield on its longest-date debt. That's because demand dipped, with orders at just 2.5 times the bonds on offer, down from 3.2 times in March. "Investors are pushing back," Bank of America wrote in a note. "The deal should also inject even more uncertainty into the hyperscaler/AI supply outlook." Torsten Slok, chief economists at Apollo Global, pointed out in a note on Wednesday hyperscalers' cover ratio -- investor orders per every dollar of bonds -- has plunged. It was nearly 5x in February 2026 but tumbled to below 2x in July, "suggesting investors may need wider spreads to absorb additional hyperscaler supply," he warned. By contrast, the ratio for investment grade bonds overall only slipped by about half a point in that span. The dollar bond market, the world's largest, has become so saturated tech giants have been issuing debt in other currencies. As a result, issuers will likely have to provide more attractive terms, meaning their borrowing costs will rise. AI-related debt must also compete against the flood of debt coming from the Treasury Department as the federal deficit continues to deepen and is on track to hit $2 trillion this fiscal year. "The current hyperscaler widening is a byproduct of the high-grade investor community trying to rationally price in an accelerating pace of issuance," JPMorgan strategists wrote in a note on Tuesday. Bearishness over bond issuance is spilling over to the secondary market. For example, SpaceX's debt has sold off, sending yields higher, and is now trading at levels comparable to junk bonds. This adds to the carnage stock investors have been suffering. SpaceX stock has plunged below the IPO price of $135 a share and is now 45% below its high, which briefly put its market cap above Microsoft's. But the bulk of the selloff has been concentrated in once-highflying chip stocks. Even Nvidia hasn't been spared and was overtaken by Apple on Friday as the world's most valuable company. The latest trigger was the release of the new Kimi K3 model from Chinese AI startup Moonshot, which claimed it outperformed models from OpenAI and Anthropic. While Kimi K3 is costlier than other Chinese rivals, it's still much cheaper than top U.S. models, and the surprise performance data raised new concerns that the AI boom's spending orgy is becoming harder to sustain. If users shift to lower-cost Chinese AI models, then U.S. AI companies may generate less revenue and cut back on their capital expenditures. The effects would ripple across the U.S. economy. AI-related investment accounted for more than half of real GDP growth in recent quarters. If AI investment declines, it could generate a mild recession, Citi Research warned on Friday. "Consumer spending has also been supported by the run-up in equity prices. Spending has increased faster than incomes, meaning the savings rate has fallen to historically low levels," Citi added. "A significant decline in equity prices would push the savings rate higher and spending lower, further contributing to a slowdown in economic growth."

AI chipmakers: Among AI crowd, some investors position for slower hyperscaler spending growth

For most of the past two years, the opposite trade prevailed: investors piled into semiconductor and infrastructure companies on the assumption that Microsoft, Amazon, Alphabet and Meta would keep accelerating spending on the buildout of data centers. The parabolic rally in AI chipmakers has run into turbulence amid concern about valuations and the sustainability of their bumper revenues, with some investors quietly positioning for a slowdown in the near-trillion dollar spending boom that could provide a boon to the hyperscalers footing the bill. For most of the past two years, the opposite trade prevailed: investors piled into semiconductor and infrastructure companies on the assumption that Microsoft, Amazon, Alphabet and Meta would keep accelerating spending on the buildout of data centers. But that spending now looks set to slow, with UBS estimating hyperscalers' capex will rise 76% this year to $673 billion, but will increase by only 25% next year and just 6% in 2028. Some active managers have already cut their exposure to chip stocks and are adding shares of hyperscalers themselves, which have sharply lagged the rally in chipmakers. They are also buying into software stocks ⁠and sectors ⁠expected to benefit from AI adoption, such as financials and healthcare. "Once they stop increasing their capex, it will definitely be a relief for hyperscalers and a negative signal for the semi industry," said Alexis Bossard, global equity portfolio manager at Edmond de Rothschild Asset Management, who has already cut exposure to semiconductor stocks, which he believes have become too expensive relative to expectations. The Philadelphia Semiconductor Index, whose top holdings include Nvidia, Broadcom, Micron, ASML and TSMC, has more than doubled over the past year, even with a near-18% drop from its June peak, compared with an 11% rise in the equal-weighted S&P 500, or an 8% gain in Europe's AI-light STOXX 600. Bank of America's July fund manager survey found 82% viewed semiconductors as the most crowded trade and none reported being short the sector. The question arises over how to position if AI spending remains strong, but no longer accelerates fast enough to support the expectations embedded across the AI infrastructure trade. Bossard has increased exposure to ⁠Amazon and favours areas such as liquid cooling, cybersecurity and selected software firms. "We have a massive underexposure to semis right now." LFG+ZEST CIO Alberto Conca has sharply cut positions in memory-chip and equipment makers, while building positions in hyperscalers and healthcare stocks, and has backed that view by buying put options on selected semiconductor names. After funding the initial AI ⁠buildout through their own cash, hyperscalers are increasingly turning to external financing, prompting questions over whether capital-market pressures may eventually constrain spending growth. The corporate debt market has absorbed billions in Big Tech issuance this year and investors have, until recently, lapped it up. Apollo Chief Economist Torsten Slok notes that cover ratios, a measure of investor demand for the bonds on offer relative to supply, have fallen to below 2 times in July, from nearly 5 times in February. In June, the Basel-based Bank for International Settlements warned disappointment in returns could trigger a sudden pullback in financing and turn the capex boom into a protracted bust. "Cash flow is starting to be almost completely drained by capex," Conca said, arguing hyperscalers will become more disciplined on spending growth. Against that backdrop, Empirical Research highlights a growing mismatch between moderating capex growth and lofty revenue expectations for chipmakers and other suppliers of AI infrastructure, implying that something will have to give. "Either the capex trajectory of the hyperscalers will be upgraded again, or the revenue growth pencilled in for their suppliers will have to come from elsewhere," it said. Madeleine Ronner, senior portfolio manager at DWS, expects earnings-season commentary from hyperscalers to remain supportive of further investment. "The surprise would be if it's not like that," she said, also noting that buy-side forecasts for 2027 spending remain materially above analyst estimates. DWS has taken some profits in semiconductor stocks after their strong run but remains overweight the sector, and ⁠certain funds have added industrial and electrical equipment exposure following the pullback. Growing local opposition to U.S. data centers could also stall spending growth. Empirical estimates about 70% of projects face some degree of pushback. New York on Tuesday became the first U.S. state to halt construction of large new data centers, imposing a one-year moratorium as concerns grow that the facilities driving the AI boom are raising power costs, straining water supplies and burdening local communities. Still, investor appetite for AI infrastructure remains strong. Morningstar data show chip-focused funds attracted record net inflows of $10 billion through May. Fidelity Investments' Director of Global Macro Jurrien Timmer says demand for compute capacity is robust and recent volatility may prove to be just another shakeout. He compared recent pullbacks to the periodic corrections seen during previous tech booms, noting that leading stocks during the late-1990s internet rally suffered repeated 20-30% declines before resuming their advance. "The AI story is well known, it's ongoing, the earnings are still supporting the trend," Timmer said. Even so, he believes investors should diversify, noting that beneficiaries of AI adoption such as financials may increasingly matter alongside beneficiaries of AI construction. "I want to participate in the boom, but I also want to protect myself in case that boom is overdone," Timmer said.

Imagining the Unthinkable for AI Investors

The most dangerous investment mistakes often begin with an inability to imagine outcomes that differ from what the market expects. That does not mean investors should reflexively bet against strong trends, because momentum is real and great businesses can stay great for a long time, but when an investment thesis depends on outcomes that are inherently unknowable, imagination becomes a risk-management tool. One useful exercise is steel-manning, which is simply taking the other side of an argument or debate and building the strongest version of it. Perhaps the most uncomfortable version of that exercise is imagining a future in which today's seemingly unquenchable AI data center and compute demand falls dramatically. That sounds implausible in 2026. AI infrastructure remains one of the most powerful capital spending cycles in the world. A recent industry forecast from Gartner predicts global data center electricity consumption will rise 26% in 2026 to 565 TWh, with AI-optimized servers accounting for 31% of that consumption and exceeding conventional server power draw by 2027. Also from Gartner, worldwide AI spending is forecast at $2.59 trillion in 2026, up 47% year over year, with AI infrastructure representing more than 45% of spending. Source: Bloomberg 7/15/2026 This is exactly why the exercise matters. The more obvious a theme becomes, the more important it is to ask what could go wrong. Imagine the following: The year is 2035. GPU-filled "AI data centers" are running at 50% utilization, up slightly from 2030 but still miles below the 90%+ levels seen during the supply-constrained buildout years of 2026-27. Vast amounts of new capacity came online just as gains in model training plateaued, while inference costs kept climbing, thanks to a combination of monopolistic behavior among East Asian semiconductor fabs and nationalistic policies on both sides of the Pacific that kept capable, abundant, low-cost Chinese chips largely out of global markets. The result: the vast majority of the Fortune 500 realized minimal to no efficiency benefits relative to what they spent. Broad consumer large language models (LLMs), like ChatGPT, continue to bleed money, with free tiers still serving 95% of users. Advertising revenue has kept these "frontier model application" companies shambling along in a quasi-zombie state, trimming losses just enough each year to keep the dream alive. The one sector in which LLMs have undeniably proven their worth is fast food restaurants: it turns out the only place people genuinely prefer talking to a robot is the drive-thru. We have no idea whether this made-up, hypothetical scenario is at all likely, but it is possible enough to analyze. Especially the bit about the drive-thru. Fast food jokes aside, the capital-cycle risk is straightforward. Based on consensus estimates, the four largest hyperscalers will have spent over $4 trillion in capex from 2025 through 2030. Assuming half of that is related to AI and a roughly 20% annual depreciation rate on those assets, that implies about $400 billion of annual depreciation expense, greater than the group's combined profits in 2025. This depreciation math is reflected in consensus estimates. Total hyperscaler depreciation in 2030, inclusive of non-AI business lines, is expected to be over $550 billion. If capacity is built, accounting costs eventually arrive, even if revenue disappoints. Servers, GPUs, buildings, power equipment, and networking assets depreciate on different schedules, meaning today's income statements may not fully reflect tomorrow's cost burden. What about demand? OpenAI reportedly crossed 900 million weekly active users in 2026 and reached $25 billion in annualized revenue by February 2026. The monetization math is more uncertain, with approximately 50 million paying subscribers against 900 million weekly active users. The implication being that the overwhelming majority of usage remains free or lightly monetized. If consumer willingness to pay remains limited, and enterprise productivity benefits prove narrower or slower than expected, the industry may discover that usage is not the same thing as profitable demand. There is also the possibility that AI becomes more efficient rather than endlessly more compute-intensive. After DeepSeek's R1 release in January 2025, hyperscalers largely reaffirmed or increased spending plans, arguing that cheaper AI would expand use cases and therefore increase aggregate compute demand. This is the "Jevons Paradox" argument. That may be right, but steel-manning requires asking the opposite: what if a future model is not 20% more efficient, but 100x more efficient? What if the next architecture dramatically reduces inference costs and makes today's data center buildout look oversized? If something like the 2035 scenario materialized, the industry implications would be severe. GPU suppliers would face order digestion and margin compression. Data center developers would be left with underutilized capacity. Utilities and grid equipment suppliers might still benefit from a broader electrification cycle, but the AI-specific urgency embedded in valuations would fade. Hyperscalers would face a harder question. Were they investing to build a durable competitive advantage, or were they trapped in an arms race where the benefits ultimately accrued to customers? This outcome has played out time and time again. Most recently, the fiber laid during the dot-com boom ultimately helped create the modern internet, while the returns on those investments certainly disappointed. Is our imagination too pessimistic? Perhaps. But for long-term investors, we see a steel-man exercise like this as simply good discipline. When everyone is expressing crystal-ball level confidence in the future, a bit of contrarian imagination can point your analytical attention in a less-traveled direction. It also serves as a useful reminder that in capital-intensive booms, returns on investment typically disappoint. Much of the narrative around AI today centers around its impact. That narrative and debate isn't going anywhere: Is AI akin to the next industrial revolution? Is AI as important a technological development as electricity? Is AI simply the next evolution in computer capabilities, like the microprocessor or the graphical user interface? But the most important question for investors is not the magnitude of a technology's societal impact. It is who earns the economics, over what time frame, and after how much capital has been spent to find out. *** Important Disclosures This material is for general information only and is not intended to provide specific advice or recommendations for any individual. There is no assurance that the views or strategies discussed are suitable for all investors. To determine which investment(s) may be appropriate for you, please consult your financial professional prior to investing. Investing involves risks including possible loss of principal. No investment strategy or risk management technique can guarantee return or eliminate risk.