원문 보기
132 billion US dollars. That's how much Google's parent company, Alphabet, spent on capital investments over the past 12 months. And if you add to this the spending of Amazon, Microsoft, and Meta, it comes out to over $510 billion in concrete, chips, and electricity in one year. And one legendary investor , who deliberately avoided tech booms for 60 years, is now directly involved in it.
Warren Buffett recently gave an interview to CNBC, drawing parallels with the money invested in the 19th century railroad boom . Many believe that Buffett is hinting at a market peak, but in reality he means something else. After all, he said this when Berkshire Hathaway owns 106 million shares of Alphabet.
So what does Buffett really say about the AI boom? Today we'll look at what Buffett is actually warning about , why the money invested doesn't yet have a clear path to return, and why owning Alphabet shares is not a contradiction. My name is Guy, this is The Finance Bureau. Warren Buffett's entire career can be summed up in one long quote.
The key to investing is not assessing how much an industry will impact society or how much it will grow, but determining the company's competitive advantage, and, above all, the durability of that advantage. This may sound like a cliché until you consider what it leaves behind the parentheses. After all, it eliminates the very question of whether this technology works.
At a conference in Sun Valley, Buffett noted that the aviation industry changed the world, but in doing so, it destroyed a huge amount of investor capital. With the retort that Karl Marx did less harm to capitalists than the Wright brothers. And he completely ignored the internet on that basis. He wasn't saying the Internet was a fad, but he refused to pay for a story he couldn't appreciate with his usual approach.
He was seriously mistaken in his forecast for Alphabet, as he has repeatedly stated, including at Berkshire's annual meeting . But the principle remained unchanged. Technology can change the world, but investing in it can still be a disaster. So, what does AI development look like if you put it through this filter?
First of all, you must accept the physical nature of this boom. Because the story of artificial intelligence is often perceived as a story about software, but that's not the case at all. Software has the best cost structure ever invented. You write it once, copy it for free , and the gross margin is as high as 70-80 percent.
However, data centers work completely differently. Data centers are land, steel, transformers, network connections, cooling, backup power, and a warehouse full of chips with a short shelf life. And that was Buffett's main point in his interview with CNBC in July. He said, "This is the game they're playing now , and they didn't play it with software." So, the most profitable business model in history voluntarily transformed into a utility company.
And the scale of this transformation is something that has not yet been fully realized. Analysts estimate that the combined capital expenditures of the Big Four in 2026 will be between $720 billion and $800 billion, compared to approximately $410 billion in 2025. This is an 80% increase in one year.
Alphabet raised its forecast for the year to $195 billion to $205 billion, and its CFO Anat Ashkenazi told analysts that figure would rise significantly again in 2027. UBS forecasts aggregate hyperscaler capital expenditures of $4.1 trillion over 2026–2028, roughly three times more than was spent in the previous 6 years combined.
Morgan Stanley expects the annual figure to exceed $1 trillion next year. But here is a figure that gives some idea of the tension of the situation. In the second quarter of 2026, the four largest US hyperscalers spent 99% of their combined operating cash flow on capital investments. $170 billion in capital expenditures versus $171 billion in operating cash flow.
Alphabet went even further and reported negative free cash flow of $5.86 billion for the quarter. This is a company with a market capitalization of $4.1 trillion that generated less money than it spent in 3 months . Spending in this market is seen as proof of success, and the logic has begun to distort.
The higher the expenses, the more serious the company seems, and the less likely anyone is to ask what those expenses actually bring. Let's be fair: if a hyperscaler were to announce to the market that it was cutting capital spending, the stock would probably plummet, and every CEO understands that.
Buffett also touched on this dynamic, saying that these companies are effectively playing a game they do n't really want to play. This brings us to a point that makes the situation even worse, because the 99% figure only applies to the money we see. A Nikkei Asia study published last month estimated that the five largest AI developers—Alphabet, Microsoft, Amazon, Meta, and Oracle— have between $1.65 trillion and $2.13 trillion in off-balance sheet commitments, guarantees, and agreements.
For context: their combined balance sheet debt is about $1.35 trillion . Therefore, liabilities that are not reflected in the balance sheet are approximately 1.2–1.5 times greater than those that are on it. And if you're not familiar with it, they perform this kind of financial alchemy using what's called a special tool.
They are creating a separate company, financed by private loans and pension money. They keep their ownership stake low enough not to consolidate it on their balance sheet. And the debt remains with this instrument, not directly with them. Meta has about $420 billion in such liabilities compared to official debt on its balance sheet of about $83 billion according to its second-quarter report.
And their auditor, Ernst&Young, had flagged one of these data center structures as a critical audit issue back in February. Oracle's off-balance sheet liabilities reached $273 billion , increasing its leverage to 4.4 times EBITDA, largely to service its $ 300 billion partnership with OpenAI. Broadcom, meanwhile, is negotiating a structure worth up to $100 billion to buy and lease chips to its own customers.
S&P Global has counted approximately $675 billion in leases that have already been signed but have n't even started yet. And the credit market noticed this. Oracle's credit default swaps, which are simply the cost of insuring debt against default, have reached record highs— about $215,000 per year for $10 million in exposure .
Broadcom's performance reached a record 122 basis points immediately after the announcement of this financial structure. The Bank for International Settlements has begun calling this shadow borrowing and estimates that private lenders now hold about $200 billion in AI data center debt , which is about 8% of all private loans outstanding.
This, considering that private credit is a place where a lot of pension and insurance money is stored, is not entirely ideal. So, that's money coming out. Now let's see what should come back. Legendary "bear " Michael Bury has been claiming all year that hyperscalers amortize GPUs within five to six years, while the real economic term is closer to two or three.
He estimates this understates depreciation costs by $176 billion between 2026 and 2028 . Another " bear" Jim Chanos calls this the same thing a " depreciation time bomb ." Amazon cut the expected life of servers from six to five years, incurring a loss of $677 million in nine months, while Meta took a different path and extended it to five and a half years, saving $2.9 billion in depreciation.
So, this is actually the same equipment in the same industry, but with a diametrically opposite accounting result. Try to explain this. These discrepancies between two of the most advanced companies on how quickly their own machines wear out are telling. And then there's the income side. David Hahn of Sequoia tried to calculate how much revenue the AI industry actually needs to generate to justify all of this spending.
He reached $200 billion, then $600 billion, and his essays were written when the numbers were a fraction of what they are today. Analysts, applying the same approach to 2026 spending levels, arrive at a need for trillions of dollars in annual revenue. The combined annual revenue of OpenAI and Anthropic is estimated to be somewhere between $80 billion and $100 billion.
And on the consumer side, a study by MIT's Nanda project found that 95% of corporate generative AI pilots failed to deliver any measurable results for the P&L . So, despite all the hype around AI, the actual revenue leaves much to be desired , to say the least. These days, tracking history through credit spreads and 13F forms on your own is almost impossible, which is why you should let us do it for you at analyze corporate capital expenditures and capital flows that will inevitably impact your portfolio.
So is QR code on your screen to get started. Good. Now back to Warren Buffett, where we have to ask the obvious question. If he sees this infrastructure expansion as a “ black hole” for capital, then why is Berkshire one of Alphabet’s largest shareholders? The stake first appeared in Berkshire's third- quarter 2025 report: approximately 17.8 million shares worth about $4.3 billion.
dollars. By the first quarter of 2026, it had more than tripled, to approximately 58 million shares. Then, on June 1, Alphabet announced a capital raise to fund AI infrastructure worth just under $85 billion, and Berkshire acted as an anchor investor, investing about $10 billion. A Form 13F filed in August showed the position growing by another 83% compared to the previous quarter to about 106 million shares worth somewhere in the neighborhood of $35 billion .
This is Berkshire's third- largest holding of public stock , behind only Apple and American Express. And Buffett ended speculation about who was responsible for it with three words in that CNBC interview. Quote: “I initiated this. Not Greg Abel, who took over as CEO this year, nor Todd Combs, nor Ted Weschler.
So either the 96-year-old Buffett is a little confused, or he's drawing a line that the rest of the market doesn't see . Well, it looks like the latter, because it's worth taking a closer look at what he actually bought. Alphabet trades at a price-to- earnings multiple of more than 16, which is lower than the S&P 500 as a whole for a company that generated $120 billion in revenue in a single quarter.
Buffett says she's more likely to be a winner, given her track record, compared to 95% of what's sold on Wall Street, because Wall Street only cares if they can sell something. And notice that he didn't speak too enthusiastically about her either. He also said, and I quote, "I don't like it as much as at least four or five other companies that we own." So Buffett clearly hasn't given in to the hype surrounding artificial intelligence.
He simply buys an established business with an “ economic moat” and cash flow, which happens to spend some of that money on AI. So the line he draws is not AI versus no AI. Buffett prefers businesses that fund it with profits over those that fund it with faith, and his favorite businesses could be wrong about AI and still survive in the future.
And if you want to know what happens to the types of businesses Buffett does n't like here, well , history offers some clear lessons. In the late 1990s, telecommunications operators raised more than $500 billion in debt and equity to lay fiber- optic cable across America and under the oceans. Demand forecasts were impressive.
The technology worked exactly as promised. And then about 95% of the fiber optic cable laid was never turned on . Global telecommunications stocks lost more than $2 trillion in market value between 2000 and 2002. WorldCom filed for bankruptcy with $107 billion in assets against $41 billion in debt, with an $11 billion accounting fraud as the icing on the cake.
Holders of unsecured bonds returned about 36 cents on the dollar. Global Crossing went first in January 2002 with $12.4 billion in debt after its founders sold more than $700 million worth of their stock . Its network was eventually acquired by Level 3 in 2011 for $3 billion . Nortel, which at one time accounted for more than a third of the entire value of the Toronto Stock Exchange, has effectively depreciated to zero.
What about fiber optics? Well, it was bought out of bankruptcy for next to nothing and later became the foundation of the streaming and cloud era . The infrastructure survived and became extremely valuable, but the people who funded it certainly did not. Looking even further back, the same structure appears in Britain in the 1840s.
At its peak, railway construction in Britain consumed over 6.5% of national income and about half of all gross capital accumulation in the country. In 1846, Parliament passed over 260 acts authorizing the creation of new railway companies to build some 9,500 miles of proposed routes. Approximately a third of the railways for which permits were issued in those years were never built.
All of this led to the Panic of 1847 , so severe that the government suspended the Banking Charter Act to prevent the system from collapsing. However, 6,220 miles of track from that period were still built and served Britain for the next century. The tracks survived, but the investors did not. In July, Buffett used this very comparison, noting that today's numbers far exceed those invested in railroads.
But how will the modern version end then? It depends on whether cash flow appears before debt starts to strangle these companies. And that's exactly what Buffett is trying to say. He doesn't need to raise funds or sell shares, and there's no reason to be diplomatic about it . Buffett argues that people have never been in such a gambling mood, and it is difficult to find value when everyone prefers to gamble.
However, his warning is not that artificial intelligence is a complete bubble filled with fraud. Buffett says that hundreds of billions have already been invested, while the returns needed to justify those expenditures remain highly uncertain. This money has already gone to data centers, chips, and energy infrastructure.
Now these investments should pay off. And Buffett says that if he had to pick one company to come out of this ordeal intact, it would be Google. Although, of course, this is just one person's opinion. What do you think? Is Google the safest bet on AI, or is Buffett going against everything he's been warning about for decades?
Let us know what you think how this off-balance sheet financial alchemy works, , see you in the next video. Guy was with you, see you soon .
Comments 0