For the last few years, “AI chip” has been almost synonymous with one company: Nvidia. And for good reason. Running a generative model like ChatGPT takes enormous amounts of computation, and the GPU is the engine that does that heavy lifting. Naturally, that’s where the money, the headlines, and the investor attention went.

But the question the market is asking has quietly changed. It’s no longer just “who makes the fastest chip?” It’s “where is the real bottleneck — in the compute, or in the memory?” And increasingly, the answer is pointing toward memory.

Why memory suddenly matters more

Here’s the part that gets missed. In the AI era, being good at calculating isn’t enough. What’s becoming just as important — maybe more so — is the ability to move data in and out fast enough to keep the calculation fed.

Think of a GPU as a world-class chef. The chef’s hands can be lightning quick, but if the ingredients show up late, the food still comes out slow. In an AI system, the “ingredient delivery” job belongs to memory — specifically to high-bandwidth memory, or HBM, which feeds data to the processor far faster than ordinary memory can.

That’s why, in modern AI hardware, the GPU and HBM essentially operate as one unit. The stronger Nvidia’s chips get, the more critical the memory sitting right next to them becomes. You can’t separate the brain from the blood supply that keeps it running.

The number that reframes everything

The clearest sign of this shift is where the value in an AI system is concentrating. According to recent industry projections, memory’s share of the total value inside an AI system could climb from the mid-40% range last year to over 70% by 2027.

That’s a bigger deal than it sounds. Not long ago, the GPU was the expensive centerpiece and memory was treated as a supporting part bolted on beside it. If these forecasts hold, memory could account for the majority of an AI system’s cost. The center of gravity in the whole AI value chain is moving — from whoever makes the best compute chip to whoever can reliably supply enough HBM.

The price signals are already showing up

This isn’t just a story about future forecasts. It’s already visible in pricing.

Recent German retail data showed prices for high-performance server DDR5 DRAM jumping 22% in a single month in June. That’s month-over-month, not year-over-year — which annualizes into triple-digit territory. One retail data point can’t define an entire industry, but the direction is unmistakable: general-purpose DRAM is rising, server memory is rising, and HBM — the high-value product at the top of the stack — is under the strongest upward pressure of all.

Analysts have responded by sharply raising their forecasts. Projected average selling price growth for DRAM in the third quarter was revised from roughly 10–15% up to around 30%, with further gains expected in the fourth quarter. In plain terms: pricing power is returning to the memory makers.

Why HBM is so hard to flood the market with

There’s a structural reason supply can’t simply rush in and cap those prices. HBM is genuinely difficult to produce — the manufacturing process is demanding, yields are hard to push up, and customer qualification is strict. You can’t just spin up extra output the way you can with commodity DRAM.

So you get a classic squeeze: demand is exploding while supply struggles to keep pace. That gap is the engine behind HBM’s price climb. Some projections see average HBM prices roughly doubling by 2027.

And here’s why that matters so much for profits. Memory is a high-leverage business. When prices rise, revenue climbs — but production costs don’t rise at the same speed. So a large chunk of that extra revenue drops straight to operating profit. As HBM prices firm up, yields improve, and long-term supply contracts pile up, margins can improve far more dramatically than in past cycles.

Micron as the bellwether

The company showing this most vividly right now is Micron. Its stock has been on a tear, occasionally moving more than 10% in a day and repeatedly setting new highs (with the inevitable pullbacks when the sector takes profits).

But the daily swings aren’t the point. What’s interesting is what Wall Street is using Micron to see. It’s being treated not as a plain memory company but as a leading indicator for an AI-memory “super-cycle.” Some analysts have floated scenarios with Micron’s gross margins reaching as high as 80% — an unusual level for a memory company historically — and even earnings approaching $150 per share by 2027.

Those are optimistic scenarios, not guarantees. But the underlying thesis is consistent: the next wave of AI beneficiaries may not stop at the GPU makers. It’s spreading to the companies that supply the memory.

What this means for Korea

This is where SK Hynix and Samsung enter the picture. The HBM market is essentially a three-way race between SK Hynix, Samsung, and Micron. And the deciding factor isn’t simply who can make the most. It’s who can hit the performance and quality a customer demands, at high yield, delivered on schedule.

In this market, supply capability is negotiating power. When HBM is scarce, customers can’t easily push prices down — and may even accept higher prices to lock in steady volume. That’s why the leverage held by memory makers is stronger now than it has been in a long time.

The caveats worth keeping

If Nvidia’s chips are the brain of the AI era, HBM is the bloodstream that keeps that brain running. Block the vessels and even the best brain stalls. That’s why the market has started treating memory not as a commodity component but as the critical bottleneck of the whole AI system.

Still, a real, durable super-cycle isn’t automatic. Three things have to line up: Big Tech’s AI investment has to keep flowing, HBM prices have to hold at elevated levels, and manufacturers have to keep lifting yields and output reliably. When all three hold together, this stops being an ordinary semiconductor cycle and becomes something closer to a genuine shift in where AI’s value lives.

The likely winners in the end won’t be the companies that make one good chip. They’ll be the ones that solve the entire chain of bottlenecks — compute, memory, power, data centers, networking. And right now, memory has moved to the center of that conversation.

---

This post is for general information and is not investment advice. Forecasts and price targets mentioned here are projections, not certainties — do your own research before making any financial decision.