Open App →
Back to Blog
Original BAB News analysis November 22, 2026 By Steve

The Forecast That Can’t Be Wrong

A memory price collapse with no date isn’t a forecast. It’s a free option for the forecaster, paid for by everyone who trades on it. Here’s the bear case made testable, and why the supply side says it isn’t close.

What the story said

On November 22, The Korea Times ran a cover story warning that the AI spending binge puts Samsung and SK hynix at risk of a memory price collapse. It quoted one analyst, Jim Handy of Objective Analysis.

His argument runs like this. AI demand won’t justify what the hyperscalers are spending. At some point they slow down, memory demand falls, oversupply follows, and prices collapse. He calls the current price surge temporary and puts the risk on the demand side, not in production. If the cycle follows past supercycles, gross margins above 80% could fall to near zero within about two quarters.

The line that carries the story is his: “That growth is temporary, and sometime in the future prices will fall back to where they were in 2024.”

The article gives no date and no timeline for the trigger. It says the timing is hard to predict.

Why “sometime” isn’t a forecast

A forecast with no date can’t be wrong. If memory prices hold through 2028, the call wasn’t missed. It just hasn’t arrived yet. If prices hold through 2030, same answer.

Memory has collapsed before, more than once. So “sometime in the future” is close to a guaranteed hit over a long enough horizon. That’s what makes it useless. A call that’s right on every timeline tells you nothing about any of them.

The costs don’t fall evenly, either. The forecaster carries no risk: there’s no date to miss. The cost lands on the investors who sell on the headline, the companies whose cost of capital moves with their share price, and the buyers who delay purchases waiting for a collapse that may be years away.

Timing matters here. Samsung and SK hynix shares fell about 33% and 34% over the three months before the story ran, in the same year Samsung posted the world’s highest quarterly operating profit. The bear case was already in the price. A cover story with no date doesn’t inform that market. It feeds it.

The supply side the story waved off

Handy puts the risk on the demand side, not in production. That’s exactly where the argument is weakest, because this cycle is being set by supply.

The clearest statement of it came from Samsung’s own memory sales chief, Jaejune Kim, on the company’s latest earnings call. Samsung has raised capex, but it doesn’t expect that to drive bit growth. The reason is HBM. AI memory takes far more wafer capacity per bit than conventional DRAM, and every step from HBM3E to HBM4 to HBM4E raises that ratio. Kim put it plainly:

“Amid the growing shift in bit production from conventional DRAM to HBM, I think this will act as a physical constraint on supply.”

That’s the mechanism the bear case skips. Money doesn’t create bits on demand. Building a fab takes years, and a new fab doesn’t deliver at full volume on day one. Micron CEO Sanjay Mehrotra said on his company’s October earnings call that output from new DRAM and NAND fabs ramps gradually and becomes meaningful only a few quarters after first wafers.

His conclusion, from the same call:

“We do not have line of sight to when supply and demand will return to balance.”

He repeated the outlook on November 18: 2027 and 2028 will be even tighter than 2026.

The Korea Times story actually contains the supply argument without noticing it. It notes that production plans were built on demand projections made two years earlier. In a shortage, that cuts against the bear case. Fabs sized for a smaller forecast don’t produce a glut. They produce a queue. New capacity fills the day it arrives, and there’s displaced demand from 2026 and 2027 waiting behind it, including PC refreshes that buyers deferred because of high prices.

Sold before it’s made

The fair objection is that memory executives always talk their book. A CEO in the middle of a historic upturn isn’t going to forecast the end of it.

That’s why contracts matter more than commentary. On October 28, SK hynix said that, counting long-term agreements already signed plus demand under negotiation, its DRAM and HBM supply is effectively sold out for the next three years, through 2029. Signed long-term agreements alone cover more than half of its total production capacity for 2027 and 2028. The company said it expects the shortage to persist through 2030.

Micron said more than 75% of its planned fiscal 2027 output is already committed, and most new customer discussions have moved to 2028 volumes. Samsung said its 2027 HBM4 supply is booked.

A forecast is a talking point. A signed long-term supply agreement is a customer putting money behind a view. This also changes how a downturn would play out. In the old cycle, a demand wobble could send contract prices down within a quarter or two, which is where the bear case gets its two-quarter collapse. With more than half of the next two years’ capacity already under contract, that shock gets dampened. The supercycles Handy is comparing to didn’t have that structure.

The trigger, and where to watch it

Give the bear case its due: it names a real trigger. If hyperscalers cut capex, memory demand falls. That part is mechanical.

The advantage of a named trigger is that you can watch it. A capex cut shows up in guidance and analyst estimates long before it shows up in memory contract prices. Right now that signal is pointing the other way. Goldman Sachs reported that Wall Street’s estimate for 2027 hyperscaler capex rose from $929 billion to more than $1 trillion during the last earnings season. After the midterms, Morgan Stanley held its forecast for hyperscaler capex at $1.6 trillion in 2027 and $1.8 trillion in 2028.

Estimates for next year are still going up. The trigger hasn’t fired.

The bear case the story should have written

There is a stronger bear argument than “sometime,” and it does have a clock. It’s about financing, not demand.

According to the same Goldman research, hyperscaler capex is set to exceed cash flow from operations from 2026 through 2028. S&P Global is modeling only one of the six largest AI spenders with positive free cash flow in 2027. That means a growing share of the buildout is being paid for with debt and outside capital instead of internal cash.

That’s a real risk, and unlike “sometime,” it’s testable. If credit tightens, spreads on AI infrastructure debt widen, or a major refinancing goes badly, capex guidance would have to follow, and memory demand would follow capex. Those are dated, observable events. A story built on that would have given readers something to watch instead of something to fear.

My call, with a date

The standard I’d ask of any forecast is one I should meet myself.

My read: memory stays tight through at least 2029. Supply is physically constrained by the shift to HBM, new fabs ramp over quarters rather than weeks, and the next three years of output at the largest suppliers is largely under contract.

Here’s what would prove me wrong, in the order it would show up:

If none of those happen, the collapse isn’t coming on any timeline that matters to a decision being made this year.

You don’t remove a bottleneck. You relocate it. A glut needs this one to disappear, and nothing on the supply side says it’s going anywhere soon.

The test for any forecast is simple: what has to happen first, and when would we see it? “Sometime” doesn’t pass.