#AI大跌后高盛仍看好数据中心 #WCTCS9 The AI data center train: every dip is an opportunity to get on board!
Wall Street has been arguing fiercely over the past two days.
Citadel’s Griffin said, “We can be almost certain that we’re in an AI bubble.”
But Giavarone, CIO of Federated Hermes, which manages hundreds of billions of dollars, said, “Fears of a bubble are premature—the current infrastructure buildout still has three or four years to go.”
State Street CEO O’Hanley shot back with a simple point: “These data centers aren’t being built blindly. They’re backed by real orders from real enterprise customers.”
Blackstone President Gray was even more direct: “Demand is growing parabolically. This is completely different from the 1990s telecom bubble, when companies ‘laid fiber first and waited for traffic.’”
For all the noise on Wall Street, there’s really just one question:
AI demand is real, but can growth keep up with the money being spent?
Let’s look at three sets of data that don’t lie.
First: orders have already been signed, and customers are lined up waiting for delivery. As of the end of the third quarter this year, Google Cloud had a backlog of $560.8 billion, AWS about $496 billion, and Microsoft commercial remaining performance obligations of $677 billion, up 73% year over year.
What does that mean? Customers have already paid; they’re just waiting for the data centers to be built and powered up. Microsoft even acknowledged that its capital spending this quarter was below plan—not because demand was weak, but because third-party capacity couldn’t keep up.
Second: if you want computing power, you’ll have to wait until 2027.
Rental prices for Nvidia’s B200 chips have climbed to $5/GPU·hour, up 25%-30% since tracking began last August. New orders are largely no longer being accepted, and delivery schedules generally extend into the first half of 2027.
Why is supply so tight?
Expanding chip foundry capacity takes 3-5 years, and getting power infrastructure for data centers from approval to grid connection takes 5-10 years—these are physical realities that can’t be sped up overnight just by throwing money at them. On one side, demand has multiplied by 1,000 in two years (daily token calls to Chinese large models have surged from 100 billion to 140 trillion); on the other, supply is like an old highway that can only be widened a little each year.
The cars are racing ahead while the road is being widened. What can be done? All we can do is build roads as fast as possible.
Third: the profit margins of those “money-losing” giants are actually rising.
On the market’s biggest concern, “negative free cash flow,” Nomura Securities hit the nail on the head: this is a classic J-curve effect—money goes out first, while revenue only starts coming in gradually once the assets are operational. Cash is negative on the financial statements, but the businesses themselves are making money. The evidence is right in front of us: AWS reported third-quarter revenue of $46 billion, its fastest growth since 2021; Google Cloud’s operating margin jumped from 23.7% to 36.2%; and the three cloud giants generated a combined $110 billion in revenue, with margins still rising even as capital spending surged.
The word “loss-making” simply doesn’t fit.
So what’s behind CoreWeave’s plunge?
It’s certainly alarming: bond yields have surged above 15%, the stock has fallen 30% in a month, and its market cap has shrunk from $90 billion to $31 billion. But look closely at its balance sheet and you’ll see a highly leveraged “neocloud” company that borrows to buy GPUs and then rents out computing power—its net debt is 7.5 times EBITDA, compared with just 0.1 times for large tech companies. China International Capital Corporation calls it a “subprime-like, weak-credit issuer in the AI financing system.”
In a word: CoreWeave’s cracks are real, but they’re a localized problem with its financing structure, not the end of AI demand. Its capital spending accounts for just 4%-5% of global cloud providers’ AI infrastructure investment—even if it shut down tomorrow, the giants’ excavators digging data center foundations wouldn’t stop. This is the truth most easily overlooked in this debate:
Stocks price narratives; bonds price cash flow.
When a small, highly leveraged boat rocks, it doesn’t mean the entire ship is sinking.
The story in China is even more compelling.
The central government has included computing-power networks alongside water networks and new-type power grids in the “six networks,” with about 4 trillion yuan in additional direct investment in computing-power networks planned during the 15th Five-Year Plan.
On China’s A-share market, the electronics sector’s net profit attributable to shareholders surged 183.9% in the first three quarters of 2026, semiconductor profits grew about 1.7 times, and semiconductor gross margins rose by a full 24 percentage points. Domestic AI accelerator cards’ market share leaped from 41.1% to over 60% in a single year—a decisive shift in market share.
Bill Ackman ran the numbers, and every investor should remember them: a 1GW AI data center costs $40-50 billion to build, can generate $15-20 billion in annual revenue, has an incremental operating margin of 50%, and delivers a pre-tax return on invested capital of 15%-25%.
This isn’t a money-burning black hole; it’s a money-printing machine.
The key isn’t whether it burns money, but whether it can keep printing money once it’s built.
The answer is already written in the order book: $677 billion in orders on hand, waiting to be fulfilled.
So how should we view this AI data center sell-off?
Dudley—the former president of the New York Fed—offered what may be the most level-headed assessment right now: global AI infrastructure investment could indeed reach $10 trillion, and revenue pressure is real, but a repeat of the 2008 systemic crisis is unlikely—because the market is broadly aware of the risks, financing terms largely match asset lifetimes, and insurance companies and long-term capital are standing behind it.
An $80 billion Wall Street hedge fund warned that the market could shift from FOMO to FOGO, but institutional money is voting the opposite way: during the A-share pullback in early November, the ChiNext computing-power ETF saw net inflows for several consecutive days, attracting more than 31 million yuan in a single day.
What is smart money doing? Quietly accumulating shares while others panic.
Back to the original question: can investment in AI data centers stop?
The answer is hidden in a statement from Microsoft: “Customers’ money is already committed, and they’re waiting for delivery.”
The answer is hidden in the symposium urgently convened by South Korea’s Ministry of Finance: “The medium- to long-term growth outlook for the AI and semiconductor industries remains valid, and no signs of a contraction in AI investment or demand have been detected.”
The answer is hidden in that ironclad investment rule: when demand grows parabolically while supply expands linearly, the gap in between is the industry opportunity of the next decade.
Bubbles deserve caution, but missing out on a long-term industry opportunity deserves even more caution.
Trends are long; volatility is fast.
What can truly weather the cycle isn’t predicting every pullback, but holding on to core assets with “orders in hand and capacity is king.”
#Gate股票畅聊季 
