In 2026, global stock markets have climbed hard, and one industry sits at the center of the move: artificial intelligence. In late May, Goldman Sachs raised its year-end target for the S&P 500 to 8,000 and argued that the year's rally has been "powered entirely by corporate profit growth rather than rising stock valuations" [source: Goldman Sachs, 2026]. A month later, on June 28, the Bank for International Settlements (BIS) — the central bank for central banks — devoted part of its flagship annual report to a warning that the AI spending boom had become a risk to financial stability [source: BIS, 2026]. The same phenomenon, read two ways.
That gap is exactly why this subject is worth a careful look right now. "The market is up because of AI" can mean two very different things: that companies are actually earning more (earnings), or that investors are simply willing to pay more for the same dollar of profit (valuation, or "multiple expansion"). These are different layers, and mixing them is how both hype and panic get manufactured. This article separates them in order — what is actually rising, whether it is earnings or valuation, where the money is going, and where the real risks lie — using primary research from institutions on both the optimistic and the skeptical side.
Table of Contents
- What is rising — the earnings-led rally of 2026
- Earnings or valuation — why the two layers must be separated
- Where the money goes — trillion-dollar capex and the power bottleneck
- Where the risk really is — concentration and financing
- So is it a bubble — measuring the distance from the dot-com era
- Conclusion — what to watch
What is rising — the earnings-led rally of 2026
The optimists' most specific claim
Start with the optimists' strongest claim, because it is more specific than "AI is hot." In late May, Goldman Sachs lifted its year-end target for the S&P 500 from 7,600 to 8,000, implying roughly 6% of further upside from where the index traded at the time [source: Goldman Sachs, 2026]. The number that matters underneath that target is earnings: the firm projects S&P 500 earnings per share of about $340 in 2026, a roughly 24% increase over the prior year, and about $385 in 2027, a further 13% gain [source: Goldman Sachs, 2026]. First-quarter profits were already up about 18% year over year, with the median company on track for its strongest quarterly earnings growth in a decade — not just a handful of giants dragging the average up [source: Goldman Sachs, 2026].
Crucially, the firm attributes about half of the index's entire earnings growth in both 2026 and 2027 to AI-infrastructure-related companies [source: Goldman Sachs, 2026]. That is the load-bearing claim: not that AI is exciting, but that a specific, measurable share of reported profit growth is coming from the businesses building and supplying AI. Ben Snider, the senior US equity strategist behind the call at Goldman Sachs Research, frames the rally as profits doing the work rather than sentiment [source: Goldman Sachs, 2026]. In this telling, the market is not floating on mood; it is being pushed up by earnings that are actually being reported.
There is a subtle but important point buried in that median-company figure. A rally can be broad or narrow, and the two mean very different things. If only a few giants are growing while the typical company stagnates, the "earnings" story is really just a handful of names in disguise. Goldman's claim is the opposite — that the middle of the index, not only its top, is on track for its best quarterly profit growth in a decade [source: Goldman Sachs, 2026]. Breadth of that kind is harder to dismiss as a single-theme mirage. Hold that thought, though: as the section on concentration will show, broad earnings growth and a price index dominated by a few names can coexist uncomfortably, and much of this article's tension lives in that coexistence.
A second institution frames it the same way
One firm making a bullish case could be talking its book. It is more telling when a second institution, looking at the whole world rather than one index, reaches the same diagnosis. In its July 2026 outlook, the asset manager Loomis Sayles states plainly that "earnings growth rather than multiple expansion has predominantly fueled recent equity market gains across the globe" [source: Loomis Sayles, 2026]. It expects that trend to continue, with S&P 500 earnings growth above 20% and small-cap earnings — the Russell 2000 index — growing more than 30% [source: Loomis Sayles, 2026].
Loomis adds that AI-related capital expenditure is "driving earnings with far-reaching multi-sector implications" — the effect reaches beyond technology into industrials and utilities, which supply the buildings, power, and equipment that AI needs [source: Loomis Sayles, 2026]. It is candid, too, about where the optimism comes from: "technology and AI-related industries are largely responsible for the upward revisions to 2026 consensus earnings growth" [source: Loomis Sayles, 2026]. That candor cuts both ways, and we will come back to it — a rally powered by one theme's upgrades is also a rally exposed to that one theme's disappointments.
Why this distinction matters
This matters because it changes what kind of story 2026 is. If a rally is driven by multiple expansion, it depends on mood: investors paying 25 times earnings instead of 20 for no new reason. If it is driven by earnings, it rests on cash that companies say they are actually collecting. The claim on the table, from more than one institution, is that 2026 belongs to the second category. But note the careful wording throughout this section: "the source of the gains is earnings" is not the same as "valuations are cheap." That distinction is the whole next section.
Earnings or valuation — why the two layers must be separated
First, what a "multiple" actually is
Because the rest of the argument turns on it, it is worth being concrete about the vocabulary. A stock's price can be thought of as two numbers multiplied together: the earnings a company produces, and the "multiple" — how many dollars investors will pay for each dollar of those earnings, usually expressed as a price-to-earnings (P/E) ratio. A price can rise for either reason. If earnings climb while the multiple holds steady, that is an earnings-led gain. If the multiple climbs while earnings hold steady, that is multiple expansion — investors simply revaluing the same profit stream upward. The two feel identical on a chart and are completely different underneath, which is exactly why the layers have to be kept apart.
The level: still above the historical average
Here is the part that both cheerleaders and doomsayers tend to skip. Saying the rally is earnings-led does not mean stocks are inexpensive. The S&P 500's forward price-to-earnings ratio in mid-2026 sits somewhere in the low-to-mid twenties, against a ten-year average closer to nineteen. By that measure, prices are still above their historical norm, whatever is driving them. An earnings-led rally and an expensive market are not contradictions; in 2026 they are simply both true.
The direction: multiples actually compressed in 2026
What is genuinely striking is the direction of the multiple this year. Rather than expanding, valuations have compressed: the index's forward multiple is down from where it began 2026, and the group of megacap technology leaders — often called the "Magnificent Seven" — has seen its valuation come down as a group, by something on the order of 15% by some measures, with its premium over the rest of the index shrinking toward its lowest level in years. In plain terms, earnings grew faster than prices, so the "how many dollars per dollar of profit" number actually fell even as the index rose. That is the strongest evidence for the earnings-led thesis: the rally happened despite multiples cooling, not because they inflated.
Holding both truths together
So both things are true at once, and a fair reading holds them together. The engine of 2026's gains is profit growth, not a re-rating of sentiment — that supports the optimists. And yet the absolute level of valuation remains above the long-run average, which leaves less cushion if those profits disappoint — that supports the skeptics. Anyone who tells you it is purely one or the other is collapsing two layers that need to stay apart.
The framing also hands the reader a concrete test for the months ahead. As long as earnings keep climbing faster than prices, the multiple stays flat or falls and the "earnings-led" description holds. The moment earnings stall while prices keep rising, the multiple re-expands — and the rally quietly changes character from one built on profit to one built on willingness to pay. That is the tipping point to watch, and it is why the capex and financing questions in the next two sections matter so much: they are, ultimately, about whether the earnings can keep coming.
Where the money goes — trillion-dollar capex and the power bottleneck
The sheer scale of the spend
The engine behind the AI earnings story is capital expenditure — capex, the money companies sink into long-lived assets like data centers and chips — and its scale is hard to overstate. For 2026, the four largest US cloud companies — Amazon, Microsoft, Alphabet, and Meta — have guided to combined capital spending in the neighborhood of $725 billion, up roughly 77% from the prior year's figure of around $410 billion; Goldman's broader hyperscaler tally reaches about $754 billion for 2026, an 83% jump, and about $905 billion for 2027 [source: Goldman Sachs, 2026]. "Hyperscaler" is simply the industry's term for the handful of companies that run cloud computing at global scale, and they are the ones writing these checks.
By individual guidance, Amazon points to around $200 billion, Microsoft and Alphabet to roughly $190 billion each, and Meta to $125–145 billion for the year; both Microsoft and Meta have flagged rising memory-chip prices as one reason the bills keep climbing. These are company projections, not audited outcomes, and should be read as guidance rather than settled fact. The distinction is not pedantic: guidance is a plan, and plans at this scale are exactly what the skeptics later in this article are questioning.
From chips to electricity: the constraint moves
What is that money buying, and what is holding it back? Increasingly, the binding constraint is not chips but electricity. Advanced processors remain tight but obtainable with enough advance commitment; in a growing number of markets, the harder limit is power — the grid capacity and generation needed to run and cool vast data centers. A data center is, from the electricity grid's point of view, a permanent new city's worth of demand that has to be wired, supplied, and cooled around the clock. That is why the AI buildout has spilled into industrials and utilities, and why energy featured alongside technology as a leader in 2026 earnings-growth forecasts: the picks-and-shovels of AI are increasingly transformers and turbines, not just silicon.
Even the optimists flag overheating
There is a reason to keep both eyes open here. Loomis Sayles, an optimist on the earnings trend, still warns in the same breath that "optimism surrounding the massive AI-driven capital expenditure cycle is running hot," and that "major setbacks would likely disappoint markets and slow future economic and earnings growth" [source: Loomis Sayles, 2026]. Spending at this scale is a bet that demand will show up to pay for it. If the payback is slower than promised, the same capex that is lifting earnings today becomes a weight tomorrow — depreciation and interest that must be earned back whether or not the AI revenue arrives on schedule.
It helps to be concrete about what "payback" actually requires. For this spending to earn its keep, the hyperscalers have to turn those data centers into cloud and AI revenue large enough to cover not only the electricity to run them, but the depreciation of the hardware inside — chips that lose value quickly as newer ones arrive — and, increasingly, the interest on the debt that funded them. That single equation is where the optimistic and skeptical cases actually meet: Loomis's "running hot" caution, Goldman's confidence that the profits are already showing up, and the financing worries in the next section are all arguments about the same sum, viewed from the revenue side or the cost-and-debt side.
Where the risk really is — concentration and financing
Concentration: a few names carry the whole index
If there is a bubble worth worrying about, the sharpest institutional critiques suggest it is less about frothy prices and more about structure. The first structural fact is concentration. The ten largest companies in the S&P 500 now account for roughly 40% of the entire index by weight — nearly double the 18–23% range that held for most of the 1990s and 2000s — and the megacap technology names dominate that group. These are the so-called Magnificent Seven: Nvidia, Apple, Microsoft, Alphabet, Amazon, Meta, and Tesla, with a single chipmaker at the very top whose weight alone approaches 7% of the index.
That concentration changes the nature of the risk for ordinary savers. Most people hold stocks through index funds and retirement accounts precisely to be diversified. But when so much of an index rides on a handful of names tied to one theme, that diversification is partly an illusion: a disappointment in a few AI-linked stocks can move entire passive portfolios and pension balances, not just a sector. The index has quietly become a concentrated bet that most of its owners never consciously placed.
Financing: the maturity-mismatch worry (IMF)
The second concern is how the buildout is financed, and here the International Monetary Fund (IMF) is pointed. In its April 2026 Global Financial Stability Report, Tobias Adrian, the fund's director of monetary and capital markets, argued that the biggest danger is not a classic valuation bubble but a maturity mismatch: large technology firms are funding rapidly depreciating AI hardware with medium- and long-dated debt [source: IMF, 2026]. A maturity mismatch means borrowing over a long horizon to buy something with a short useful life — and AI accelerators can be technologically stale in a few years while the bonds that paid for them run far longer. "Major tech firms are starting to leverage up themselves," Adrian noted, marking a shift for companies that historically funded themselves from their own cash [source: IMF, 2026].
The fund put numbers on the trend: large technology companies issued roughly $159 billion in corporate bonds over about five months to finance the buildout, even as a substantial share of planned data centers — on the order of 60% — had not yet broken ground [source: IMF, 2026]. Borrowing long to buy something short-lived, before most of the projects are even under construction, is a familiar way for booms to turn fragile.
The credit market's contradiction (BIS)
The BIS sharpens the point from the credit side. In its 2026 annual report it judged that "equity valuations are elevated, particularly for firms at the core of AI development," that "AI valuations are volatile and concentrated," and that the five largest hyperscalers' AI commitments of more than $1 trillion across 2025–2026 are outpacing their earnings [source: BIS, 2026]. Its most telling observation is a mismatch between markets: the interest-rate spreads that private lenders charge AI firms sit close to those charged to non-AI firms, even though equity investors treat AI companies as exceptional. Either lenders are underpricing the risk or equity markets are overpricing the future — one of the two, the BIS implies, is wrong [source: BIS, 2026].
That observation is worth unpacking, because it turns on a market most readers never see. "Private credit" refers to loans made not through public bond markets but by non-bank lenders such as specialized investment funds, frequently to finance exactly this kind of buildout; a "spread" is the extra interest a borrower pays above safe government debt to compensate the lender for the risk of not being repaid. When the spread charged to AI borrowers barely differs from the spread charged to everyone else, those lenders are effectively pricing AI as ordinary-risk credit — at the very moment equity investors are paying premium prices on the conviction that AI is extraordinary. The two markets are staring at the same companies and disagreeing about how risky they are. Whichever one turns out to be wrong, the repricing does not stay politely in its own lane.
The report does not soften the warning. It cautions that "optimism surrounding AI may not last, despite its promise of future productivity gains," and that "the current surge in capital expenditure could prove unsustainable if supply bottlenecks restrain production" [source: BIS, 2026]. It adds that AI financing is "increasingly leveraged," with "complex interactions within the AI supply chain" — the kind of tangled, debt-fed linkages that transmit trouble from one link to the next when a single assumption fails [source: BIS, 2026].
So is it a bubble — measuring the distance from the dot-com era
The resemblance
The comparison everyone reaches for is the dot-com boom of the late 1990s, and the IMF has made it explicitly: "it was the internet then; it is AI now" [source: IMF, 2025]. The fund has flagged stretched valuations, vulnerabilities among non-bank financial institutions, and the growing concentration of gains in a few megacap technology companies — a checklist that rhymes uncomfortably with 1999.
One item on that checklist deserves a plain-language gloss, because it is where a market wobble can turn into a financial one. "Non-bank financial institutions" — investment funds, insurers, private-credit vehicles and the like — now hold much of the risk that once sat on bank balance sheets, typically with thinner regulatory cushions and less public visibility. The IMF's worry is not only that AI valuations are stretched, but that if they fall, the losses may surface in these lightly-watched corners rather than at well-capitalized banks — making the fallout harder to see coming and harder to contain [source: IMF, 2025]. It is the difference between a fire in a room you are watching and a fire in the walls.
The measurement
But the same institution insists on measuring the distance, not just the resemblance, and the measurement cuts against the most alarmist reading. The IMF has estimated that US AI-related investment has risen by less than 0.4% of GDP since 2022, materially smaller than the roughly 1.2% of GDP that the dot-com buildout added between 1995 and 2000 [source: IMF, 2025]. A smaller macroeconomic footprint implies that even a sharp AI correction would spill over to the broader economy less violently than the tech crash did. That is not a claim that nothing can go wrong; it is a claim about proportion, and it is the single most useful counterweight to the loudest bubble talk.
Reframing "earnings or bubble"
So the honest answer to "earnings or bubble" is that the frame is slightly wrong. On the specific question of what drove 2026's gains, the evidence points to earnings, and multiples actually cooled. On the question of whether risk is building, it plainly is — but concentrated in financing structure, index concentration, and the assumption that enormous capex will be repaid, rather than in a simple story of prices detached from all profit. Both can be true, and the BIS and the optimists are, in a sense, describing the same animal from different ends: one is looking at the profits coming in the front door, the other at the debt and commitments piling up at the back.
Conclusion — what to watch
The most defensible position in mid-2026 is not "boom" or "bust" but a short list of things to watch. First, payback: whether the roughly three-quarters of a trillion dollars in annual hyperscaler capex converts into the cash flow that today's earnings forecasts assume [source: Goldman Sachs, 2026]. Second, the power bottleneck: whether electricity and grid capacity can keep pace, since that, more than chips, now sets the ceiling. Third, financing stress: whether the spreads on AI-related corporate debt and private credit reprice, which is where the IMF and BIS locate the real fragility [source: IMF, 2026][source: BIS, 2026]. Fourth, concentration: whether the market broadens beyond a handful of names or leans even harder on them.
For a general reader, the practical takeaway is a discipline rather than a forecast. When you hear "the market is up because of AI," ask which layer is meant — earnings or valuation — because the two carry very different risks. In 2026, the profits appear real, the valuations are full but not runaway, and the genuine questions are about leverage, power, and whether one theme is carrying too much of the whole. That is a more useful thing to know than whether to shout "bubble."