SK Hynix’s record profit raises an awkward question: is AI revenue keeping up?

SK Hynix, the world’s biggest maker of the high-bandwidth memory (HBM) that sits next to Nvidia’s AI processors, reported record results on Wednesday. Revenue reached 104.6 trillion won, about $79 billion, up 328% from a year earlier. Operating profit was 83.3 trillion won, about $63 billion, up 632%. That is an operating margin of 80%, up from under 50% a year ago.

The company says its DRAM and HBM are effectively sold out through 2029 and expects the memory shortage to last until 2030. It has long-term agreements with 14 major customers, including the big cloud companies, covering more than half of its capacity for 2027 and 2028.

Trillions of won

Show the numbers

	Operating profit
Q3 2025	11.4tn won
Q4 2025	19.2tn won
Q1 2026	37.6tn won
Q2 2026	60.5tn won
Q3 2026	83.3tn won

Source: SK Hynix quarterly results

Those numbers answer one question clearly: the buyers of memory want all they can get. They don’t answer a second one. Is the money that pays for all this memory coming from real AI sales, or mostly from bets on future sales?

Follow the money

The AI money moves along a chain. Companies and consumers pay AI developers such as OpenAI and Anthropic, and use AI features from Microsoft, Google and others. The developers rent computing power from cloud providers. The cloud providers build data centers, and they buy chips from Nvidia, which in turn buys memory from SK Hynix, Samsung and Micron.

In a healthy chain, the money flowing in at the end would roughly match what is spent at the start. Right now, it doesn’t. In April to June, the four biggest US tech companies spent about $171 billion on capital projects. The two leading AI developers, OpenAI and Anthropic, brought in about $18 billion of revenue between them.

Billions of dollars in one quarter. Micron’s figure covers March to May.

Show the numbers

	One quarter
Big Tech capital spending	$171.2bn
SK Hynix + Micron revenue	$96.7bn
AWS + Google Cloud revenue	$67bn
OpenAI + Anthropic revenue	$18.3bn

Source: Company reports; OpenAI figure as reported by the Wall Street Journal

The comparison is not perfect. Cloud revenue and data-center spending include work that has nothing to do with AI. AI revenue does not come only from OpenAI and Anthropic: Microsoft sells Copilot, Google sells Gemini, Meta uses AI to sell ads, and thousands of companies run their own AI on rented servers. Capital spending also pays for buildings and power that will last for years. But the gap is large enough that it doesn’t disappear with those adjustments.

AI revenue is growing fast, but promises are growing faster

The AI developers are not standing still. Anthropic’s revenue rose from $2.0 billion in the last quarter of 2025 to $11.6 billion in the second quarter of 2026, and the company told investors it reached about $16.8 billion in the third quarter. OpenAI’s revenue rose 18% to $6.7 billion in the second quarter, the Wall Street Journal reported.

Their commitments are much larger. Anthropic’s IPO filing lists about $430 billion of computing contracts it cannot cancel. OpenAI expects to spend about $856 billion on computing and infrastructure through 2030, against forecast revenue of about $36 billion this year, according to a company presentation seen by the Financial Times. The same presentation shows OpenAI burning through about $278 billion of cash before then.

In other words, the orders reaching SK Hynix today are not paid for by today’s AI revenue. They are paid for by cloud companies’ profits, by borrowing, which reached about $489 billion of AI-related debt in 2026 by July according to Goldman Sachs, and by investors betting on forecasts like OpenAI’s, which has its revenue rising roughly tenfold to $350 billion by 2030.

So is memory spending running ahead of demand?

• The case that demand is real: memory is sold out, prices are still rising, customers are signing five-year contracts with deposits, and cloud revenue is growing fast. Google Cloud grew 82% and AWS 37% in the second quarter.
• The case that it is running ahead: the end customers of AI bring in a small fraction of what is being spent, the gap is covered by debt and investor money, and growth at the AI developers is not a straight line. Anthropic added fewer new dollars of revenue in the third quarter than in the second.
• What history says: memory has always been the most cyclical part of the chip business. When orders slow, its prices fall first and fastest.

SK Hynix’s finance chief argued that this time is different, saying “more investment or more fabs does not necessarily mean more supply,” because newer memory takes more factory space to make. That limits the risk of a glut from the supply side. It does nothing about the demand side, which depends on AI revenue catching up with the spending.

What to watch

• Wednesday, October 28: Microsoft and Alphabet report after the US market closes. Azure and Google Cloud growth show how much AI demand is real, and their 2027 spending plans show how much memory they will need.
• Thursday, October 29: Samsung’s full results, and Amazon’s results with AWS growth.
• Coming months: whether OpenAI and Anthropic keep adding more revenue each quarter than the one before.

For now, the memory makers are the clearest winners of the AI boom, because they are paid in cash today for chips that are already sold. The risk is that they are also the furthest link from the customers whose spending has to justify it all.