Holding massive AI infrastructure-related orders is no longer enough to "protect" a company.
Oracle holds $500 billion in orders, and its stock price has plummeted 40% from its peak in September. Broadcom currently has a backlog of AI product orders of about $73 billion, and its stock price turned from gains to losses after the latest earnings report.
CoreWeave, known as "Nvidia's favorite son," has quarterly revenues in the billions of dollars but managed to win over $36 billion in orders from OpenAI and Meta in just one week. Over the past month, the company's stock has fallen 17%.
While there are concerns about whether they have enough capacity (money) to meet customer demands, there is also worry about whether the customers themselves are truly "reliable."
At the core of the AI infrastructure onion are a few key players: Meta, Google parent Alphabet, Microsoft, Amazon, Apple, Nvidia, and other giants, plus star AI startups like OpenAI and Anthropic.
The star startups are still immature; building infrastructure almost entirely relies on external financing, which carries obvious risks.
The giants should be like anchors—financially stable, cash-rich, and filling the next few years with trillions of dollars in疯狂 infrastructure plans.
But the returns from AI, which is at the center of spending, remain minimal. Whether using "old money" to fuel new dreams will drag down the giants depends entirely on whether those dreams materialize quickly enough.
Success brings universal joy; failure could mean losing everything.
01
Holding the "future" card, Oracle has experienced great joy and great sorrow in just a few months.
When the great joy arrived, Oracle's stock price soared 40% in a single day, and founder and CEO Larry Ellison briefly surpassed Musk to become the world's richest person.
At that time, Ellison exclaimed: "Artificial intelligence is everything!"
Artificial intelligence is indeed everything. For Oracle, it was the entire reason for this wave of great joy—OpenAI had then reached a five-year, $300 billion computing power procurement agreement with Oracle, which was the match that ignited Oracle's stock price.
However, just three months later, Oracle holds even more orders, but the "magic" has disappeared.
Oracle recently released its fiscal 2026 second-quarter (corresponding to September to November 2025) earnings report, with revenue increasing 14% year-over-year. The company stated that its backlog has reached $523 billion.
<极 style="font-size:16px">This number is $68 billion higher than the previous fiscal quarter.
Once the earnings were released, the stock price fell 11% that day, marking the company's largest single-day drop since January. From its peak in September, Oracle's stock price has fallen 40%.
Future orders, amid the current skepticism about an "AI bubble," have transformed from a beautiful hope into a heavy burden.
Oracle appears to be struggling somewhat—the earnings report showed Oracle's cash flow was negative $10 billion, and quarterly capital expenditure (CapEx) reached $12 billion, nearly $3.7 billion higher than analysts' predictions.
And Oracle's CFO revealed that the company's fiscal year expenditure has also been raised by $15 billion, reaching a level of $50 billion.
The market's greatest fear is: Does Oracle even have the ability to raise that much money to support such a massive scale of AI infrastructure?
Some analysts predict Oracle will need to borrow $100 billion to complete the construction. In the second fiscal quarter, the company raised $18 billion in debt, one of the largest debt issuances on record for a tech company.
On the conference call, Oracle vigorously defended itself, explicitly opposing the prediction of "needing to borrow $100 billion," stating that the actual financing amount would be significantly lower. The secret lies in Oracle's adoption of a "customer brings their own chips" cooperation model.
In other words, it's not Oracle buying chips and then renting them to customers; it's customers bringing their own chips, which is virtually unprecedented in the cloud services industry.
Additionally, Oracle emphasized that some suppliers are willing to lease rather than sell chips to them, allowing Oracle to synchronize payments and receipts.
If Oracle's claims are true, it could significantly reduce its upfront investment and greatly increase its return rate.
But for the market, the risk hasn't disappeared; it has shifted: from Oracle itself to Oracle's customers. Meta or OpenAI and others purchase expensive GPUs themselves and install them in Oracle's data centers.
Whether Oracle's hundreds of billions of future dollars materialize depends not only on its ability to "deliver" but also on the customers' ability to "pay." Of Oracle's nearly $500 billion in undelivered orders, about two-thirds come from the yet-unprofitable OpenAI, and another $20 billion is known to come from a new agreement with Meta.
Similarly, Broadcom also holds massive orders but received negative market feedback.
Broadcom also released new earnings. For the fourth quarter of fiscal year 2025 ended November 2, it achieved core performance with both revenue and profit exceeding expectations, and AI semiconductor-related revenue grew 74% year-over-year.
On the conference call, Broadcom CEO Hock Tan stated that the company's current backlog of AI product orders is approximately $73 billion, to be delivered over the next six quarters. He emphasized this is a "minimum value," and as new orders continue to pour in, the backlog size is expected to expand further.
However, Broadcom refused to provide clear guidance for full-year 2026 AI revenue, citing uncertainty in customer deployment节奏, which could lead to quarterly fluctuations.
<极 style="font-family:Arial,Helvetica,sans-serif">After the earnings release, Broadcom's stock initially rose about 3% but then turned negative, falling over 4% in after-hours trading.
Compared to Oracle's great joy and sorrow, Broadcom only encountered a small bump, but the underlying market sentiment is similar—people are no longer optimistic about that "future" of大兴AI infrastructure.
Broadcom's customers are also relatively concentrated, with its AI-related orders primarily coming from OpenAI, Anthropic, Google parent Alphabet, and Meta, among others.
02
Peeling back the layers of the AI infrastructure onion, you always find those familiar companies—the Magnificent Seven and OpenAI, Anthropic.
Similarly, AI cloud infrastructure startup CoreWeave, which has also received much attention this year, went public in March this year, marking the largest tech startup IPO since 2021. Its stock price subsequently more than doubled, even surpassing the "Big Seven tech giants."
Its customer concentration is also extremely high, basically surviving on orders from Microsoft, OpenAI, Nvidia, and Meta.
Just this Monday (December 9), CoreWeave issued another $2 billion in convertible bonds, while its total debt as of the end of September had already reached $14 billion. Market concerns intensified, and its stock price has fallen 17% over the past month.
Again, the market has deep-seated doubts about the AI industry as a whole, not only regarding whether these AI infrastructure-related companies can deliver services as planned but also whether the疯狂 making deals的大客户们 can actually foot the bills.
And the complex web of circular transactions between all related parties has formed a tight yet opaque net, making everything even more unclear.
If we look at the types of customers, startups like OpenAI and Anthropic were the first to raise concerns.
The reason is simple: neither has stable profitability yet, at least not enough for their膨胀的 infrastructure plans. They need external financing, and the uncertainty is obvious.
In contrast, the giants are more like风向标s and safety nets on the playing field.
The giants spend hundreds of billions of dollars annually on capital expenditure, a significant portion of which is用于扩建数据中心. Their combined capital expenditure in 2026 will be more than four times the total spending of the U.S. listed energy industry on drilling exploration wells, extracting oil and gas, transporting gasoline to gas stations, and operating large chemical plants. Amazon alone has capital expenditures exceeding the entire U.S. energy industry's total.
Compared to the稚嫩的 startups, the giants are obviously财大气粗. They are financially stable and have ample cash flow. At least for now, the spending hasn't exceeded their capacity to bear it.
For example, Microsoft, Google, and Amazon will have spent over $600 billion collectively from 2023 to this year, with expected revenue of $750 billion.
If you look at their recent performance reports, you'll find they are quite strong; "exceeding expectations" is almost standard practice. It seems there's no need to worry—in other words, they can afford this大兴AI基建.
But upon closer inspection, none have fundamentally changed their revenue structure. AI has indeed started to generate returns, but its share of total revenue often remains a supporting role, yet it takes center stage in spending.
For instance, regarding Microsoft's quarterly earnings at the end of July, TheCUBE Research estimated that AI services contributed about 19% to Azure cloud growth, over $3 billion, but this accounts for less than one-tenth of Microsoft's total revenue.
Over half of Google's revenue still comes from advertising and search, and Amazon's e-commerce and advertising still account for over 70% of its revenue.
In other words, the giants are using their mature businesses to nourish the future of AI.
The question is, how long can they keep nourishing it?
03
The giants have already begun a "borrowing frenzy."
In September, Meta issued $30 billion in bonds. Alphabet recently also announced plans to issue approximately $17.5 billion in bonds in the U.S. market and about $3.5 billion worth of bonds in the European market.
Data from Bank of America shows that in September and October alone, large tech companies focused on artificial intelligence issued $75 billion in U.S. investment-grade bonds, more than double the industry's average annual issuance of $32 billion between 2015 and 2024.
These companies' revenue growth should currently be able to support the spending, but to keep pace with the AI field, they will ultimately need more debt.
The Wall Street Journal pointed out sharply in an analysis: AI is making giants weaker.
As of the end of the third quarter of this year, Microsoft's cash and short-term investments accounted for about 16% of total assets, down from about 43% in 2020. Alphabet's and Amazon's cash reserves have also significantly decreased.
Alphabet and Amazon's free cash flow this year is expected to be lower than last year. Although Microsoft's free cash flow over the last four quarters appears to have grown compared to the same period last year, its disclosed capital expenditure does not include spending on long-term leases for data centers and computing equipment. If these expenditures were included, its free cash flow would also decline.
This trend seems destined to continue.
Analysts estimate that Microsoft, including lease expenditures, is expected to spend approximately $159 billion next year; Amazon is expected to spend about $145 billion; Alphabet is expected to invest $112 billion. If predictions hold true, these companies will have cumulatively投入$1 trillion within four years, most of which will be spent on the artificial intelligence field.
Overall, these changes—reduced cash balances, reduced cash flow, increased debt—are fundamentally changing the business models of tech companies.
The tech industry is beginning to resemble industries like semiconductor manufacturing, where tens of billions of dollars are invested in building cutting-edge factories that take years to construct but even longer to yield returns.
Deploying trillions of dollars across hundreds of massive data centers presents clear and enormous challenges仅仅从执行角度来看for AI infrastructure.
Data centers consume enormous amounts of electricity—GPUs require大量电力进行计算—and current power grids cannot handle the surge in demand. Secondly, cooling is also a problem. GPUs run very hot and require vast amounts of fresh water to keep the equipment running. Some communities have begun opposing the construction of data centers, worried about the impact on water supply.
Nvidia and OpenAI announced a massive new agreement worth up to $100 billion this year, with OpenAI planning to deploy 10 gigawatts of Nvidia systems. But recently, Nvidia's CFO admitted that this plan is still at the letter of intent stage and has not been formally signed.
This, on the one hand, casts a shadow over the "credibility" of the热闹的AI基建交易, and on the other hand, hints at future uncertainty.
The reason for the delay in signing the agreement has not been made public, but the "risk factors" section of Nvidia's filing with the SEC can serve as a reference.
In the filing, Nvidia warned that if customers scale back demand, delay financing, or change direction, the company could face risks of "excess inventory," "cancellation fees for orders," or "inventory write-downs and impairments."
Additionally, the availability of "data center capacity, power, and capital" is key to the deployment of AI systems. The filing stated that power infrastructure construction is a "process that will take years" and will face "regulatory, technical, and construction challenges."
Even if the AI infrastructure ultimately progresses smoothly, it is not the end of "success."
AI infrastructure ultimately serves AI demand. If the infrastructure is in place but market demand fails to materialize, then low utilization rates of the infrastructure will result in huge losses.
Of course, not everyone is frowning with worry. Supporters believe this is a豪赌 worth taking because AI demand will grow at an exponential rate, not linearly.
Analyst Azeem Azhar calculated that the direct revenue from AI services has grown nearly ninefold in the past two years.
In other words, if this growth rate continues, it's only a matter of time before AI companies start generating record-breaking profits.
"I think people who obsess over the specific financing methods of these investments have an old mindset. Everyone assumes this technology will develop at a linear speed. But AI is an exponentially growing technology. It's a completely different model," Azhar said.
But the question is, will the moment when AI starts to bring "profits" explosively arrive, and when will it arrive?
Ultimately, whether AI infrastructure will drag down the giants is essentially a race of AI market demand catching up to AI infrastructure. If it catches up, the AI infrastructure is "worth it." If it doesn't, the massive data centers will ultimately become like "ghost towns." That would be the best proof that the giants' AI bets were incorrect and would have disastrous consequences.
