American tech giants are making the biggest business bet in modern history by putting hundreds of billions of dollars into artificial intelligence. They believe that whoever has the most computing power will rule the next economic era. But behind the huge capital commitments is a growing amount of debt, a growing derivatives market, and investors' growing worry that the AI revolution may be built on a financial foundation that is much weaker than Silicon Valley would like to admit.
The Cowork Shockwave from Anthropic: The Catalyst
A single product announcement from Anthropic, the AI company that makes the Claude language model, on February 3, 2026, wiped out about $285 billion in market capitalization from tech stocks in a single trading session. The release of industry-specific plug-ins for Anthropic's Claude Cowork tool set off the chain reaction. This AI platform is meant to automate complicated business processes in areas like sales, marketing, finance, and legal. Cowork can do multi-step business tasks like reviewing contracts, summarizing risks, making reports, and updating internal records with little help from people, unlike regular AI chatbots.
The market reacted quickly and strongly. The Goldman Sachs basket of US software stocks fell 6% in one day, the worst drop since the tariff-fueled selloff in April 2025. An index of companies that offer financial services fell by almost 7%. Thomson Reuters had its biggest one-day drop ever, losing almost 16%, and LegalZoom lost almost 20%. Gartner and S&P Global, two companies that do analysis and data, lost 21% and 11%, respectively. European businesses were not spared: RELX, the company that owns LexisNexis and is based in London, fell 14%, its worst daily performance in five years. Publicis, WPP, and Omnicom, three big advertising companies, all lost between 9% and 12%.
Days later, the panic grew when Anthropic released Claude Opus 4.6, an advanced model that could deploy autonomous teams of AI agents to work on different parts of bigger projects at the same time. This directly challenged well-known SaaS vendors like Salesforce, Microsoft's Copilot platform, and Workday. By the middle of February, the total loss in market value from the selloff of software and services stocks was close to $1 trillion.
The episode made a fear that had been growing quietly in boardrooms and trading floors for months crystal clear: AI is no longer just a tool that makes existing software better; it is quickly becoming a direct replacement for it. And the companies that are spending the most money to build AI infrastructure are also ruining the business models of the companies that were supposed to be their customers.
The $700 billion arms race
The amount of money being put into the AI race is unprecedented in history. Amazon, Alphabet (Google), Meta, Microsoft, and Oracle are the five biggest US cloud and AI infrastructure providers. They have all promised to spend between $660 billion and $690 billion on capital expenditures in 2026, which is almost twice as much as they spent in 2025. If they followed their advice to the letter, the four biggest companies would spend about $665 billion, which is a huge 74% more than the $381 billion they spent in 2025.
Amazon is leading the way with an estimated $200 billion in capital expenditures in 2026, most of which will go toward AWS data centers to handle the growing workloads of AI, which are expected to rise by almost 50% each year. Alphabet comes next with a prediction of $175 to $185 billion, which has been raised three times from an initial range of $71 to $73 billion for 2025. CEO Sundar Pichai said that the scale is big enough to worry people inside the company, but he also pointed out that the cloud backlog grew by 55% in a row to over $240 billion. Meta plans to spend between $115 and $135 billion, which includes building a one-gigawatt data center. Microsoft is on track to spend $120 billion or more, having already spent $37.5 billion in just one quarter. Oracle adds to the group with a goal of $50 billion.
In 2025, tech capital spending as a percentage of US GDP was about 1.9%, which was almost the same size as the biggest capital projects of the 20th century. The Apollo Moon Landing program, the Interstate Highway system, and the quick growth of the electricity grid in 1949 each made up about 0.6% of GDP. Goldman Sachs says that AI capital spending in 2026 will be more than 3% of US GDP.
The reason for this spending spree is based on strategy: whoever owns the most powerful data centers can train and scale more advanced AI models than their competitors. Microsoft said that it has a $80 billion backlog of Azure orders that it can't fill because of power issues. This shows that demand is still outpacing even its aggressive build-out pace. All of the big hyperscalers say that their markets are limited by supply, not by demand. Sundar Pichai, the CEO of Google, has said very clearly that "the risk of under-investing is much greater than the risk of over-investing."
But the hyperscalers are stuck in what game theorists would call a "prisoner's dilemma": if one of them doesn't join the AI arms race, there is a chance that a rival will win. This dynamic pretty much guarantees that everyone will invest too much, which is what has happened before every major technology bubble in history.
When confidence meets the reality of cash flow, investors revolt.
Even though a lot of money is being spent, people are becoming more and more doubtful that the "largest data center" argument alone can keep market share and make money over the long term. The financial effects of the spending spree are becoming too big to ignore.
After Alphabet announced its latest increases in capital spending, the stock dropped 7.5% in after-hours trading, even though it reported record quarterly revenues for the second quarter in a row. Microsoft's market value dropped by $430 billion at one point after it released its earnings report. So far this year, the company has lost about 17% of its value, making it the worst performer among the major hyperscalers. Amazon's stock fell more than 8% after the company announced its $200 billion spending plan. This brought the company's annual decline to 9%. Since the start of the year, Nvidia, the company that sells the "pickaxes and shovels" of the AI gold rush, has lost about 4% of its value.
The main issue is cash flow. In 2025, the four biggest US internet companies made $200 billion in free cash flow, down from $237 billion in 2024. This drop happened before the planned increase in spending. The picture gets a lot worse for 2026. Pivotal Research says that Alphabet's free cash flow will drop by almost 90%, from $73.3 billion in 2025 to just $8.2 billion. Morgan Stanley and Bank of America analysts say that Amazon will have negative free cash flow of between $17 billion and $28 billion. Amazon told investors in a recent SEC filing that it might need to raise more equity and debt to pay for its AI build-out.
For Big Tech, capital intensity—the ratio of capital spending to revenue—has reached its highest level in more than ten years. This is a big change from the asset-light, software-driven business models that have kept valuations high for the past ten years. Tech companies that used to be praised for making a lot of money with little physical infrastructure are quickly changing into something like the capital-intensive telecommunications and utility sectors of the 20th century.
Ewan McIntyre, a vice president at Gartner, called the $650 billion commitment a "Intelligence Supercycle," which is a time when disruptive investments change the way markets work. "But the real problem isn't adoption," he said. "It's creating value that matters."
The End of the Fortress Balance Sheet: Investing with Borrowed Money
For twenty years, Alphabet, Microsoft, and Amazon were known for having a lot of cash and very little debt. That time is officially over.
The hyperscalers are using debt markets more and more to make up for the fact that their AI capital expenditure budgets are getting bigger and their internal free cash flow generation is getting smaller. In 2025, Amazon, Alphabet, Meta, Microsoft, and Oracle together sold $121 billion in US corporate bonds. That's more than four times the five-year average of $28 billion. Morgan Stanley thinks that hyperscaler borrowing will reach $400 billion in 2026, which is more than double the amount in 2025. UBS analysts think that total capital expenditures could reach $770 billion and that global companies could take on as much as $900 billion in new corporate debt this year. This rise is likely to push the total amount of investment-grade bonds issued in the US to a record $2.25 to $2.46 trillion in 2026.
On February 9, 2026, Alphabet priced a $20 billion multi-currency bond offering, which was the largest ever for the company. It included a 100-year sterling bond that would mature in 2126, which got a lot of attention. This was the first bond issued by a tech company in 100 years. The last one was sold by Motorola in 1997, when the dot-com boom was at its peak. The offering got more than $100 billion in orders, making it one of the biggest order books ever for a corporate bond. Oracle had just launched a huge $25 billion bond offering that got a record-breaking $129 billion in orders.
Pension funds and insurance companies were the main institutional investors in the century bond because they wanted to match their long-term liabilities. But it also made people think about what it means when a company with $126 billion in cash and $73 billion in free cash flow each year feels like it has to lock in 100 years of debt to pay for investments this year. According to one fixed-income portfolio manager at Mirabaud Asset Management, this shift toward the bond market is changing the relationship between hyperscalers and their investors in a big way. He said it was breaking what he called a "unspoken contract" that had kept speculative AI spending mostly separate from debt markets.
BlackRock, the biggest asset manager in the world, called the mega-cap tech borrowing spree a "bonanza" meant to make up for the difference between current investments and future revenues. However, they also said that more corporate borrowing puts more pressure on bond markets that are already having trouble dealing with large public deficits.
Alphabet's long-term debt grew to $46.5 billion in 2025, four times what it was before. Meta's spending on AI has gone from almost $30 billion in cash in 2023 to negative $7 billion. Oracle has the highest debt-to-equity ratio of any company its size, at about 500%. Amazon's is about 50%, and Microsoft's is even lower.
The CDS Market Returns: Like 2008
The credit default swap market, which became famous during the 2008 global financial crisis, is growing at an alarming rate. This is one of the most worrying things happening in the AI financing world right now.
Credit default swaps are like insurance policies that pay out if a borrower doesn't pay back their loan. During the subprime mortgage crisis, CDSs made systemic risk worse by linking banks, insurers, and hedge funds in huge networks of exposure. Now, they are back to check the creditworthiness of the biggest AI companies.
The CDSs for Alphabet, Amazon, Meta, Microsoft, Nvidia, and Oracle add up to a market worth about €10 billion. The speed of this change is what makes it so interesting. Less than a year ago, companies like Alphabet and Meta didn't even have an active CDS market because they didn't have much debt. According to the Depository Trust & Clearing Corporation, these instruments are now one of the most actively traded parts of the credit derivatives market.
The most worrying thing is Oracle's CDS spreads. Oracle's CDS cost went from a low of about 40 basis points in September 2025 to 140–160 basis points by early 2026. This meant that the cost of insuring against Oracle's default went up by four or five times. Barclays lowered Oracle's debt rating to underweight in November 2025, saying that the company could drop to BBB-, the lowest investment-grade rating before becoming junk. Oracle's bonds still have official ratings of Baa2 from Moody's and BBB from Standard & Poor's, but they now trade in secondary markets at spreads that are more typical of speculative-grade debt. Barclays analysts have said that Oracle could run out of money by November 2026 if it keeps spending the way it is now.
Since mid-2025, the larger CDS market for hyperscalers has also grown. This is because credit risk has gone up and investors are more aware of the risks involved in these huge capital expenditure programs. The cost of insuring all types of hyperscaler debt has gone up a lot, and a growing derivatives market is being built around the idea that AI investments might not make enough money to pay off the debt that was used to finance them.
Critics have pointed out worrying structural similarities to 2008. Before the last crisis, Deutsche Bank was one of the main players that made money by selling CDSs linked to subprime mortgages. Now, they are reportedly giving multibillion-dollar loans for AI infrastructure and using Synthetic Risk Transfer (SRT) mechanisms to protect against default risks. Critics say these structures are very similar to the Collateralized Debt Obligations that made the financial crisis worse. Michael Burry, a famous hedge fund manager and the main character in "The Big Short," is said to have taken short positions against AI stocks by buying put options on one million shares of Nvidia and five million shares of Palantir.
The Revenue Problem: What's the Point?
A basic question that the tech industry hasn't yet answered well is where the money will come from to pay for all of these investments.
In 2025, the total yearly income from AI goods and services was thought to be less than $50 billion, even though more than a trillion dollars had been put into them. OpenAI, which is probably the most well-known AI company in the world, made about $12 billion in sales in 2025 but lost $8 billion in operations. According to the company's own internal forecasts, losses will double to $17 billion in 2026 and then double again to $35 billion in 2027.
A study by the National Bureau of Economic Research that came out in February 2026 found that 90% of the companies surveyed said AI had no measurable effect on productivity at work. However, executives still thought AI would boost productivity by 1.4% and output by 0.8%. Researchers compared this gap between what executives thought would happen and what actually happened to the long-documented productivity paradox. An MIT report said that 95% of AI pilot programs did not produce useful results. According to BCG, only 5% of businesses that used AI saw real benefits from it. According to Forrester Research, only 15% of business survey respondents said that AI had helped them make more money in the past year.
Even among the hyperscalers that had the best early returns—Microsoft said Azure AI added 16 percentage points to its 33% cloud revenue growth, and Google Cloud revenue grew 48% year-on-year to $17.7 billion in Q4 2025—the question is whether these revenue streams can grow quickly enough to make the hundreds of billions being spent worth it. Microsoft wants to make $25 billion in AI-related sales by the end of fiscal year 2026, but that number is only a small part of the money being spent.
The rise of effective rivals makes things even more complicated. In January 2025, the Chinese startup DeepSeek showed that it was possible to make competitive AI models for a lot less money than Western hyperscalers thought. This caused Nvidia's market value to drop by $600 billion in just one day. The investment thesis could be hurt if AI becomes a commodity product, like broadband internet or mobile phone services, instead of a high-margin monopoly business.
Will the bubble pop?
It's getting harder to ignore the similarities between the current AI investment cycle and past technology bubbles.
The US equity market capitalization is now almost twice the GDP, which is a lot more than it was at the height of the dot-com bubble. The Shiller CAPE ratio is at its highest level since 2000, and the price-to-book ratio for S&P 500 companies is 5.3, which is higher than the 5.1 level reached during the dot-com boom. More than 1,300 AI startups are now worth more than $100 million, and almost 500 AI "unicorns" are worth $1 billion or more.
Oliver Wyman has come up with two possible scenarios for the AI market to crash. If investors suddenly changed their minds about what they expected, the value of the stock market could drop by about $33 trillion, which is more than the entire US GDP. In a hybrid situation made worse by debt, the contagion could spread through corporate bonds, private credit, asset-backed securities, and even the banking system. If half of the $6 trillion in AI capital spending that is expected to happen between now and 2030 is financed by debt, the consulting firm says that the resulting credit buildup would be more than all of the broadband infrastructure investment since the internet began. Private credit is expected to be responsible for more than $1 trillion in AI-related debt.
Sundar Pichai, the CEO of Google, himself warned of the risk in very clear terms, saying that "no company is going to be immune" if an AI bubble bursts.
But there are big differences between this bubble and past ones that make it wise to be careful when making direct comparisons. In the dot-com era, many companies didn't make any money and didn't have any good products. But today's hyperscalers are very profitable businesses that make real cash flow, even if that cash flow is being used to pay for capital expenses. The assets being financed this time are productive data centers and computing infrastructure, not bad mortgages like in 2008. The rules are stricter now, and the companies that are driving the AI boom have balance sheets that are getting worse but are still much stronger than the ones that went under in 2008.
Pat Gelsinger, who used to be the CEO of Intel, put it best when he said, "Of course we are in a bubble." We're excited, we're speeding things up, and we're putting a lot of power into the system. But this bubble will last for a few years.
The final decision will depend on whether the AI revolution can deliver on its promise to change things quickly enough to pay off the huge amount of debt being taken on to build it. History shows that new technologies that change things often end up being worth the money spent on them during their risky early stages. The railroads, the telephone network, and the internet all made a lot of long-term value, even though the people who paid for their building often lost everything. The question is not if AI will change the economy. It's whether the financial structures being put in place to build the AI infrastructure will last long enough for that change to happen.
The credit default swap market keeps growing, the bonds keep flowing, and the biggest tech companies in the world keep borrowing at a rate not seen since the worst days of the dot-com era. They are betting everything on a future they say is certain, but they can't yet fully prove it.