Here’s the thing about markets: sometimes they get it spectacularly wrong. Right now, AI stocks are getting hammered by geopolitical jitters and macro anxiety. Meanwhile, the actual demand for AI infrastructure is going absolutely bonkers.
Let me break down what’s actually happening behind the scenes.
The numbers are wild. Marvell revised its fiscal 2027 revenue forecast from $9.5 billion to $11 billion in just six months—a 30% jump. Broadcom is now seeing $100 billion in AI chip revenue visibility by 2027. And Jensen Huang? He’s gone from seeing $500 billion in AI demand a year ago to at least $1 trillion through 2027. Oh, and he casually mentioned: “We are going to be short.”
Translation: demand is outpacing supply, and it’s not even close.
The real story is the shift from training to inference. For the first couple years of generative AI, most compute went to training models. Now? Reasoning models like OpenAI’s o1 and agentic systems that actually do stuff are driving the real workload. Every action requires tokens. Every token requires compute. That’s a structural change—we’ve gone from a one-time training cost to a perpetual inference tax on everything AI does.
Think about the scale: computing demand has increased roughly 1 million times in two years. That’s the product of two multipliers: compute per inference session went up 10,000x as AI got smarter, and usage itself grew 100x. Multiply those together and you get a million-fold increase. That’s not hype. That’s math.
The bottlenecks are shifting too. GPUs were the constraint, then interconnects, and now it’s memory. Micron can only meet 50-66% of customer demand. Their gross margins jumped from 75% to 81% in a single quarter—that’s how tight supply has become. Their CFO was explicit: memory isn’t a commodity anymore. It’s “a defining strategic asset in the AI era.”
Here’s where it gets interesting: customers are locking in demand early. Oracle’s $553 billion remaining performance obligation (basically a signed backlog) is probably the most underappreciated number in tech right now. Three years ago, Oracle was fighting for relevance. Today, it’s the infrastructure of choice for large-scale AI workloads. Nvidia confirmed this at GTC, naming OpenAI, Cohere, and others as Oracle tenants.
And then there’s the custom silicon play. Broadcom now serves six XPU customers—Google, Meta, ByteDance, OpenAI, and others. These aren’t one-off deals. They’re multi-year partnerships tied to long-term AI roadmaps. OpenAI alone signed a 10-gigawatt agreement through 2029.
The geopolitical noise is real, but it’s temporary. The underlying incentives don’t support sustained escalation—the economic costs are too high for everyone involved. When that risk premium fades, the market’s focus snaps back to fundamentals.
And the fundamentals? They’ve been getting stronger while everyone was distracted.
The AI infrastructure buildout is accelerating. Demand is locked in. Supply is constrained. The companies positioned to capitalize on this—the ones turning raw compute into products and recurring revenue—are about to have their moment.
The data’s already on the table. The market’s just not reading it yet.
