Earnings are up 30%. So why are valuations falling?

Global share markets have had an unusual 2026. Analysts usually start the year too optimistic and cut their forecasts as the months go by. This year their forecasts weren't optimistic enough.

At a recent Fidelity International media briefing, portfolio manager Maroun Younes, who co-manages the global small-to-mid cap strategy, said that in January, analysts expected the MSCI World Index's earnings to grow by about 12–14% this year. Ten months in, that forecast is closer to 25–30%. Mid caps have also risen from about 13–14% to 17–18%.

However, valuations have moved the other way. Rosie Malcolm, who manages Fidelity's large cap global equities fund, said earnings expectations rose about 30% to the end of September, but "the actual multiple on the index came down by about half of that". The index has returned about 15%.

Who pays for the AI build-out?

Much of this year's earnings growth has been paid for by the balance sheets of the hyperscalers.

"We've seen a real transfer of cash," Malcolm said. "So a lot of the earnings that are coming into these companies have actually been the transfer of the balance sheets of the hyperscalers into the earnings of the memory companies and other parts of the AI chain."

That spending has been funded by cash, debt and new equity, and Malcolm doubts it can continue indefinitely.

"You can't just increase debt forever. That's not a sustainable platform to have that sort of earnings growth for a long period of time," she said. "I actually think that's something the market recognises and that's why we've seen that multiple come down."

The next phase needs consumers and businesses paying for AI on an ongoing basis. Malcolm is watching estimates from OpenAI and Anthropic, and the hyperscalers' cloud results, for signs that this is happening.

That uncertainty is a big reason her portfolio leans towards the hyperscalers. "We don't know right now where that profit of all that money that's being spent is actually which part of the chain it's going to land. They're in a lot of parts of the chain."

She also sees them as the more defensive way into AI. If the building stops, their cash flow should jump because spending would fall.

For now, suppliers are reporting firm commitments. Malcolm said memory chip buyers are signing five-year contracts and putting down large deposits that can only be used against the final year of orders. 

"That's actually a much more sustainable base that we're seeing in terms of very large companies being willing to step behind this build-out."

Not 2000 - yet

The dot-com comparison is never far away.

Younes pointed to a key difference. After the dotcom bust, only about 3% of US fibre was in use. Today, the hyperscalers say they are short of capacity. 

"If you listen to all the hyperscalers today, every single one of them will tell you their single biggest issue is they don't have enough." 

They say land, power and chips are all bottlenecks. Malcolm noted that Nvidia's A100 chips, about five years old, are being rented out again at higher prices than before.

Younes also said history suggests the biggest uses for AI may not exist yet. Social media, for example, emerged nearly a decade after the internet boom. "Some of the biggest use cases of AI likely, at least history would show us, haven't even been developed yet as a concept."

Malcolm added, "I don't think we've even begun to optimise the use of AI." Users have yet to settle on which models suit which tasks, or how much they are willing to pay.

James Abela, who co-manages the Fidelity Global Future Leaders Strategy with Younes, described the market as being in a "valley of productivity", as companies work out what AI actually delivers. Historically, he said, value has moved towards application software three to five years after a new technology arrives.

"I think we're within a year of that productivity valley."

Physical limits pushing change

Malcolm said the build-out is now running into physical limits. "One of the things about the AI build out is that it's about physics and we're seeing companies run into the physics of trying to build data centres big enough to run this model."

That is reshaping where the opportunities lie.

• Optical over copper. Copper can't carry enough data far enough, so Nvidia is moving its 2027–28 data centre designs to optical technology, which sends data as light.

• Cooling. Heat is one of the sector's biggest problems, and the fixes now reach down to the circuit board.

• CPUs over GPUs. On a June trip to Silicon Valley, companies told Malcolm that as AI moves from training models to everyday use, more of the work will shift from GPUs to CPUs.

Malcolm said that hyperscalers have flagged on recent earnings calls that their data centre spending comes in two parts. "The chips are a five-to-six year depreciation cycle in theory, but the land and the building is 20 years."

That raises a question for the industrial and electrical companies building the data centres. 

"How long will that trend go for them? Whereas it might actually go a lot longer for the chip companies."

On the other hand, a Meta analyst notes that land is becoming so scarce that hyperscalers may buy it just to lock it down. "They don't have to build on it. They're just going to make sure they've got the land in case they want it."

Younes said spending has moved well beyond chips and that hyperscaler spending is key for the revenue of the companies they look at for the small-to-mid cap fund. 

"Two years ago that spending was pretty much all on chips and semiconductors and now there's a lot being spent on things like power generation behind-the-metre power, alternative sources, cooling systems."

The risks

Younes is “constructive but cautious”. He’s watching three things:

• Rates. Most developed-market central banks are tightening. "It doesn't mean equities can't perform in a period of rising rates. It just means things become a little bit more volatile, things become a little bit more choppy."

• Narrow leadership. Bull markets tend to last when gains are broad. Long stretches where a few stocks lead have historically come before market tops.

• Credit. Stress is emerging in CCC-rated debt. Younes said it would be a bigger warning sign if that spread to stronger borrowers.

Malcolm added that the market is jumpy. “We’re also seeing the market take comments from different executives and get very upset very quickly over things.” Asked whether AI stocks could fall in 2027, she said: “Absolutely they can fall.”

Outside the AI trade

Abela argued that AI has made investors less willing to count on earnings far into the future, with any business seen as disruptable valued as if it has only a few years left.

A challenged value stock would normally trade on a free cash flow yield of about 7%, he said. Some stocks in his universe are on 30%. 

"I spend a lot of time trying to find things that are outside of that paradigm of positive or negative," he said. 

"If you find something that's independently grown, independently successful, and its duration is kind of not being questioned, then that's a nice place to be."

The team is also selectively buying back into stocks the market sold off as “AI losers” earlier this year, where it judges the business more resilient than the market believes.

They gave four examples:

• GTT (EPA: GTT): The French company holds about 90% of the market for membrane linings that insulate LNG tankers. "It's an energy company, but it's not capital intensive. It's not digging out wells. It's not involved in exploration. It's simply R&D, so it's human capital and it collects a royalty stream," Younes said. It earns returns of about 50–60% on the capital it invests, has net cash and yields about 4–5%. Demand for its membranes is rising as countries diversify their energy supply.

• Interactive Brokers (NASDAQ: IBKR) Abela said the broker has kept growing and held its valuation, because it operates in finance, an industry built on trust.

• Aritzia (TSE: ATZ): The Canadian women's fashion retailer is expanding into the US. Abela said the stock fell about 30% on consumer concerns, then rose 22% on results day. "The market is throwing a big blanket over fear generally, but when companies can actually deliver and show, no, we're actually growing and we're actually very comfortable, they are getting rewarded."

• Mobility Global (NYSE: MBGL) The S&P Global spinoff owns CARFAX, which sells US car history reports. Index funds that received the shares had no reason to keep them. "If you're passive, you get this stock, you don't hold it, it's not part of your index," Younes said. Fidelity took a small position after the selling pushed the price down.

ETF

Fidelity Global Future Leaders Fund and Active ETF (FCAP)

Global Shares

View

Managed Fund

Fidelity Global Equities Fund

Global Shares

View