The AI Momentum Trade Is Over. The AI Investment Boom Is Not. Japan's AI stocks are entering a new phase.

by Kenichi Ogura - Former Chief Editor of PRESIDENT (Japan’s No.1 business magazine)

In July, AI stocks sold off hard. AI spending did not.

A record unwind in Asia's hedge funds

According to Goldman Sachs prime-brokerage data reported by Reuters, Asia-focused long-short hedge funds fell an average 18.6% through July 28. It was their worst monthly drawdown on record. These funds had done very well earlier in the year, helped by large bets on AI hardware, including South Korean chipmakers. Their year-to-date return reached 40% on July 22. Less than a week later, 21 percentage points had disappeared.

Asian hedge funds cut their market exposure for eight straight trading days through July 27. Five-day de-grossing, meaning the reduction of both long and short positions, hit a record. Vikas Pershad, a portfolio manager for Asian equities at M&G Investments, told Reuters: "In market cap terms, this is the biggest unwind we've ever seen."

The same kind of selling appeared in the U.S. Goldman reported heavy selling in popular hedge-fund holdings and lower exposure to AI, semiconductors and mega-cap technology stocks.

But data-center demand did not fall 18.6%.

That difference matters.

Prices move faster than business conditions

Goldman's own momentum factor showed the same stress. Its realized volatility over the previous three months rose to one of the highest levels in 45 years outside recession periods. That tells investors something about the trade itself: recent winners had become unusually unstable. It does not tell them that cloud companies suddenly stopped ordering equipment.

Stock prices can change much faster than business conditions. A supplier can still have a full order book while its shares fall because investors had already expected even better numbers. That is exactly the kind of gap the market is now exposing.

A crowded trade can fall very quickly even when the business behind it is still growing. If many funds own the same stocks, a price decline can force them to cut risk. Their selling pushes prices down again, which can lead to more selling. July looked much more like this kind of market unwind than the end of the AI investment cycle.

For investors, that changes how AI stocks should be judged. A chip-equipment maker, an optical-fiber company and a cloud platform can all benefit from AI spending, but their profits depend on very different things. The market is starting to price those differences again.

The focus shifts to cash flow

The more important question is now moving to the companies spending the money.

For the past few years, a bigger AI investment budget was usually treated as good news. If Microsoft, Alphabet, Amazon, Meta Platforms or Oracle spent more on AI, suppliers of chips, optical equipment, power systems and cooling systems could expect more demand.

That part is still true. But investors are paying more attention to what happens after the money is spent.

A Reuters analysis based on LSEG consensus estimates shows why. Between 2025 and 2027, Microsoft, Alphabet, Amazon, Meta and Oracle are expected to increase annual operating cash flow by about $340 billion. Over the same period, their capital expenditure is expected to rise by about $534 billion.

Spending is rising faster than the cash the businesses are expected to generate.

Current-year capex estimates for the five companies also rose sharply. In January, the total was about $485 billion. By July, it had climbed to around $730 billion.

There is another reason cash flow is getting more attention. Large technology companies spent years being rewarded for businesses that could grow without building much physical infrastructure. AI changes that. A new model may attract users quickly, but serving those users requires very expensive chips and data centers. Revenue can rise while free cash flow comes under pressure.

That does not make the investment wrong. It does mean that a rising capex budget is no longer enough by itself to support a higher valuation.

Some of that spending is already producing growth. But AI infrastructure is expensive. Companies need servers, chips, network equipment, land, electricity and cooling. The more physical infrastructure they build, the more cash they have to put back into the business.

Microsoft reported $35.8 billion of operating cash flow in its fiscal second quarter. Capital spending, including finance leases, reached $37.5 billion. Oracle spent $55.7 billion on capex in fiscal 2026, equal to 174% of its $32 billion in operating cash flow.

The market is starting to care more about these numbers.

Japan offers a useful test

Orders at Lasertec Corporation <6920> more than doubled in the year ended June 2026. The stock still fell about 14% on the first trading day after the company reported its results.

Lasertec makes inspection equipment used in advanced semiconductor production. For the year ended June 2026, sales fell 8.3% to 230.5 billion yen. Operating profit fell 14.3% to 105.3 billion yen. Net income fell 8.5% to 77.5 billion yen.

Orders, however, jumped to 237.5 billion yen from 105.2 billion yen a year earlier.

The company also gave a strong forecast for the year ending June 2027. It expects sales of 290 billion yen, operating profit of 125 billion yen and net income of 90 billion yen. At about 159 yen to the dollar, the sales forecast is roughly $1.8 billion.

Those numbers do not look weak. The problem was that investors were expecting more.

In the latest quarter, Lasertec's operating margin fell to about 44% from more than 52% a year earlier. The company still expects record sales, but the market is now looking at margins, delivery timing and how quickly orders turn into revenue.

A good AI story is no longer enough.

Fujikura and Furukawa: demand you can already see

Fujikura Ltd. <5803> shows the other side.

In the April-June quarter, Fujikura reported sales of 402.0 billion yen, up 50.1% from a year earlier. Operating profit rose 155.1% to 104.8 billion yen. Its telecommunications business grew especially fast, helped by strong demand for optical components used in data centers.

The company raised its full-year operating profit forecast from 310 billion yen to 432 billion yen. That is about $2.7 billion at 159 yen to the dollar.

Here, investors can already see AI-related demand in sales and profit. The story is not based only on what might happen in the future.

Furukawa Electric Co., Ltd. <5801> is seeing a similar benefit. It raised its full-year operating profit forecast to 123 billion yen from 95 billion yen because data-center demand was stronger than expected. First-quarter operating profit nearly tripled to 25.4 billion yen. The company also said better pricing helped earnings.

Fujikura and Furukawa are benefiting from the same broad data-center boom, but not in exactly the same way.

Uneven winners and a less forgiving backdrop

Memory is harder. Strong HBM demand does not help every supplier equally. A company that is late to qualify a new generation of high-bandwidth memory can miss much of the cycle even while the market itself is booming.

Semiconductor-equipment makers have their own problem. Orders can move sharply from quarter to quarter because customers spend in waves. A company can have a strong long-term AI story and still report a weak quarter.

Japan also has something that was largely absent during the early AI rally: rising interest rates. The Bank of Japan kept its short-term policy rate at 1.0% on July 31 after raising it in June.

Higher rates give investors more alternatives to expensive growth stocks and can support the yen, making the backdrop less forgiving for exporters. That does not decide which AI stocks will win, but it makes a one-way trade harder to sustain.

For Japanese suppliers, this creates a strange but useful market. The end customer may still be spending aggressively, yet a supplier can disappoint because its own margins are weaker, deliveries slip or investors had already priced in a perfect result. That is why company results now matter more than the broad AI label.

The next phase of the AI trade

The next earnings season should show more of this separation.

For U.S. hyperscalers, operating cash flow deserves as much attention as AI spending. For Japanese suppliers, orders are only the first step. Investors also need to see what margins those orders carry and how much new investment is needed to fill them.

AI spending is still growing in the real economy. Data centers still need chips, memory, optical links, electricity and cooling.

What has ended is the easy part of the trade.