Forecasting the AI bubble: When scarcity turns to surplus
SiliconANGLE Dave Vellante ● Covered by 3 sources
AI’s still booming, but the bill may arrive before the money does. The shortage of chips, power and packaging is hiding how much capacity is already being built.
Based on reporting by SiliconANGLE, Dave Vellante — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Artificial intelligence can be transformative and still sit inside a capital bubble. That’s the core point here: AI doesn’t have to fail for the bubble to pop. It only needs supply and financing to race ahead of the revenue those systems can actually earn.
The numbers explain why people are still excited. Global semiconductor revenue approached $800 billion in 2025, and the WSTS forecast cited here puts 2026 at about $1.51 trillion. Nvidia, Broadcom and AMD all posted big AI-related data-center or semiconductor revenue figures in their latest reports. But more than half of that projected 2026 market comes from memory, which means scarcity is doing a lot of the work.
That’s why this cycle is tricky to read. A rise in memory revenue does not automatically mean demand is exploding in a healthy way; Micron’s own results show how pricing can inflate the headline while physical bit shipments move only a little. The source’s warning is simple: watch what happens when pricing normalizes. If prices ease while bit demand, deployment and monetization keep rising, that’s probably fine. If prices fall and physical demand stalls too, the market may be clearing into surplus.
The bottleneck also keeps moving. GPUs need high-bandwidth memory. HBM needs advanced packaging. Racks need network fabric. And all of it still needs power, a site and financing before it becomes revenue-producing capacity. Solving one shortage just pushes pressure to the next one, which delays the moment when buyers finally have real choice and the market discovers whether it overbuilt.
That timing gap is where the risk lives, and Oracle, OpenAI and Stargate are the clearest example. OpenAI reportedly committed $300 billion over five years to buy compute from Oracle starting in 2027. Oracle has already seen capex jump to $55.7 billion in fiscal 2026, with guidance as high as $95 billion gross for fiscal 2027, while remaining performance obligations hit $638 billion. Yet only about 12% of that is expected to convert into revenue in the next 12 months. The money is moving now; the cash is not.
And that’s the real test for investors. Not whether AI works. Whether all this committed capital turns into energized, useful, cash-generating capacity before the shortage disappears and the financing gets less forgiving.
My take — AI-written commentary, not fact-checked reporting
This is what happens when everyone confuses backlog with reality. AI spending is being treated like proof of destiny, but a pile of contracts and half-finished data centers is not the same thing as earnings. The industry’s favorite trick is calling it “infrastructure” when it’s really just a very expensive waiting room.
Read more about this at: SiliconANGLE