The Debt Tsunami That Could Signal the End of the AI Boom

The Most Expensive Group Project in History

Are we currently witnessing the beginning of the end of the AI boom? Let’s start with a number so large it loses all meaning: roughly US$700 billion. That’s the combined capital expenditure the four biggest US hyperscalers, Amazon, Alphabet, Microsoft and Meta, are guiding toward for 2026 alone. Google, Amazon, Microsoft, and Meta collectively plan to allocate $725 billion to capital expenditures in 2026, up a staggering 77% from last year’s already record-breaking $410 billion. Throw in Oracle and the pure-play AI labs and the five largest US cloud and AI infrastructure providers have collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026, nearly doubling 2025 levels.

To put that in perspective, that’s more than the GDP of Switzerland being shovelled into warehouses full of Nvidia chips, copper cabling and industrial air-conditioning. Amazon alone is reportedly eyeing $200 billion in capex, Alphabet $175–185 billion, Meta $115–135 billion, and Microsoft set its calendar-2026 capex at a jaw-loosening $190 billion, well above the $152 billion average analyst estimate.

And here’s the part that should make you sit up: nobody can actually build fast enough. Despite historic investments in GPUs and data centers, hyperscale data center operators say they are unable to keep pace with the demand for capacity for AI services. So they are turning the dial up to 11. When the richest companies on earth tell you they’re “capacity-constrained,” that’s either the greatest demand signal in corporate history or the sound of a very expensive party where nobody’s checking the bar tab.

When the Whole Sector Got Pulled Along for the Ride

This tidal wave of spending didn’t just lift Nvidia (now flirting with a US$5 trillion market cap). It dragged the entire semiconductor food chain up with it, and then some. The memory makers, Micron, SK Hynix and Samsung, are having the time of their lives. In the second quarter of 2026, DRAM contract prices rose 58% to 63% quarter-over-quarter, while NAND Flash contract prices surged 70% to 75%, marking the largest increases in a decade. Goldman Sachs reckons the market is on the cusp of the most severe memory chip supply shortage in 15 years.

Why? Because the memory giants have quietly pivoted away from making the chips that go in your phone and laptop, toward the high-bandwidth memory (HBM) that AI accelerators devour. The economics are brutally simple: a single HBM3E module sells for roughly $60 to $100, compared to approximately $5 to $10 for a comparable amount of conventional DDR5 DRAM. When you can charge ten times as much, you don’t agonise over which customer to serve.

The result has been a sector-wide melt-up that would make a dot com day trader blush. The share prices of Micron, SanDisk and the like have exploded in the last 12 months. TSMC fabs the chips, ASML sells the lithography machines that make the chips possible, AMD plays gallant runner-up to Nvidia, and poor old Intel watches from the sidelines wondering where it all went wrong. Everyone’s invited to the AI party and the punchbowl is currently overflowing. The trouble with overflowing punchbowls, as any veteran investor over 50 will tell you, is the cleanup the morning after.

Does the AI Maths Actually Work?

Here’s the question nobody at the party wants to ask out loud. You’re spending US$700 billion a year. What are you selling? AI subscriptions and tokens, i.e. people paying for ChatGPT, Claude, Gemini and Copilot, plus enterprises buying API access by the truckload.

The revenue is real and growing at an almost absurd clip. OpenAI is reportedly running at $2 billion every month … passed $20 billion in annualized revenue by the end of 2025. Anthropic, the maker of Claude, has had an even more vertical ride: Anthropic went from $87M ARR in January 2024 to $47B ARR by May 2026, a 540x increase in 28 months, with Claude’s coding tools doing much of the heavy lifting. Impressive. Genuinely.

But now do the arithmetic. You’re laying out roughly $700 billion in a single year, against industry-wide AI software revenue measured in the low tens of billions. Even the optimists admit the gap. The widely-cited Sequoia analyst David Cahn estimated the AI “revenue gap”, i.e. the difference between infrastructure spend and the revenue needed to justify it, has grown to $600 billion. Meanwhile the labs themselves are torching cash: OpenAI does not expect to turn a profit until around 2030. HSBC analysts estimate the company could face a $207 billion funding shortfall by that time.

In plain English: the people selling the AI can’t yet pay for the infrastructure the AI runs on, and the people building the infrastructure are betting that future token revenue eventually shows up to bail everyone out. It might! But that’s a very large IOU written against a very optimistic forecast. And it gets more uncomfortable when you notice the money is starting to run in circles with Nvidia committing up to $100 billion to OpenAI, which then spends much of it… buying Nvidia chips. As one CFO cheerfully admitted, “Most of the money will go back to Nvidia.” Critics call this circular financing; optimists call it a “virtuous circle.” History calls it Lucent, Nortel and Cisco circa 1999!

The Debt Tsunami Nobody Saw Coming

For years, the comforting story was that all this spending came straight out of the hyperscalers’ colossal free cash flow, equity risk, not credit risk, nothing for bondholders to fret about. That contract has now been torn up. The cash flow simply isn’t enough anymore.

In 2025, five of the largest hyperscalers, Amazon, Alphabet, Meta, Microsoft and Oracle, alone issued $121 billion in US corporate bonds versus an average of $28 billion per year between 2020 and 2024. That’s more than four times the historical run-rate. And 2026 is on track to blow past it. The five largest cloud and AI companies have collectively sold $159 billion in bonds during the first five months of 2026, with Amazon alone executing a near-record ~$54 billion deal in March. Alphabet even issued a 100-year bond, because apparently 30 years wasn’t a long enough leash. Over at SpaceX, after merging with Elon Musk’s xAI, the group took out a $20 billion bridge loan at a significantly cheaper rate and used the proceeds to pay off xAI’s debt stack.

This changes the maths fundamentally. A bondholder doesn’t care about your moonshot, they care about getting paid. As one fixed-income manager put it, “For years, we’ve been told this AI spend would be funded by generated cash flow. That it is equity risk, it is speculative, and not to worry about it from a credit point of view.” Now that the debt is piling onto balance sheets, AI now surpasses US banks as the largest sector in the JP Morgan US Liquid index. Free cash flow is getting crushed in the process with Amazon projected to turn negative this year. When you fund decades-long, fast-depreciating assets with borrowed money and the revenue is still hypothetical, you’ve introduced a new and very unforgiving referee into the game: the credit market.

Who Blinks First? The Anatomy of Dropping Out

Not everyone can win the AI race. There’s room for perhaps three or four hyperscale clouds and a handful of frontier labs. The rest are running an extraordinarily expensive marathon they will eventually need to quit. So, what triggers the first surrender?

History rhymes here, and the rhymes aren’t pretty. The shale revolution that prompted US oil and gas companies to issue $350 billion in debt to fund drilling led to hundreds of bankruptcies after oil prices swooned in 2014 and 2015. Go back further and the widespread adoption of electric power led to a buildout that saw roughly half of the 3,000 small utilities and power companies that existed either disappear or get sold during a brutal decade of consolidation. The technology was world-changing in both cases. The early financiers still got slaughtered.

The mechanism for dropping out is almost always the same: the moment debt-funded capex meets a revenue disappointment and a refinancing window. A player keeps spending because stopping looks like surrender; then a quarter comes where token revenue growth slows, GPUs depreciate faster than expected (these things are obsolete in 2–3 years, not 20), and the bond market suddenly demands a higher coupon. The weakest balance sheet, most likely a debt-heavy pure-play lab burning billions with no profit in sight, or a second-tier cloud without an ads-and-software cash cow to lean on, finds the door slammed shut. They don’t go bankrupt overnight; they quietly announce a “strategic partnership” (read: rescue), slash capex, and tell investors they’re “rationalising compute.” That’s how defeat is announced in polite company.

Of course, right now, it is near-impossible to say who that first AI drop out will be.

What It Means for Your Portfolio: 6 and 12 Months

Next 6 months: Expect the music to keep playing, but louder and more nervously. The memory supercycle has further to run, Micron’s order books stretching into 2027 and HBM sold out for the year mean Micron, SK Hynix, Samsung, TSMC and ASML will still be very busy in the next little while. But watch the warning lights flashing in the bond market and in free cash flow. The single most useful early-warning indicator costs you nothing: two consecutive months of declining DDR5 contract prices historically precede drawdowns of 40 to 60 percent in memory stocks. Bookmark TrendForce.

And note the spate of mega-IPOs — SpaceX, OpenAI and Anthropic, together targeting valuations near US$3.6 trillion, which will suck enormous liquidity out of the market and serve as the ultimate sentiment thermometer. If SpaceX starts to stumble further, that’s your canary.

Next 12 months: This is where we’d get genuinely cautious. The debt issued in 2025–26 starts to matter once growth merely decelerates, it doesn’t need to reverse. Sometime in the next year, we may see at least one high-profile capex “rationalisation,” a frontier lab, such as OpenAI or Anthropic, forced into a rescue raise on humbling terms, and a sharp, sentiment-driven repricing across the AI complex. The picks-and-shovels names (Nvidia, the memory trio, ASML, TSMC) tend to crack hardest because they’re priced for perpetual hypergrowth.

None of this means AI is a fad, quite the opposite. The internet survived the dot-com bust spectacularly. It means the financing of AI has quietly shifted from “what’s the upside?” to “who gets paid back?” And in every prior infrastructure mania, that question has eventually been answered with a tsunami. The trick isn’t to leave the party. It’s to dance near the exit.