Memory shortage 2028: the gun you can only fire once

Ben Thompson says the memory oligopoly overplayed its hand. Micron's CFO just pulled a $100 billion market forward by a year. Both timelines land in 2028.

Micron beat its own quarter by a mile last week, told investors that the market for its most important product is arriving a full year early, and the stock has fallen about 30 percent since. The gap between what the company says and what the price does is the most interesting argument in semiconductors right now, and it comes with a date attached: 2028.

A blowout quarter, then a 30 percent drawdown

Micron (MU) reported fiscal third quarter 2026 results on July 12, 2026. Adjusted earnings came in at $25.11 a share against the $20.49 analysts expected. Revenue was $41.46 billion against $35.69 billion expected. The stock rose 14.55 percent after the close, to $1,199.52.

On July 16, 2026, Micron closed at $832.41, down 7.95 percent on the day. That is roughly 30 percent below the post-earnings print, in five sessions. Its 14 day relative strength reading, a common gauge of whether a stock has been sold too hard, sits at 20.9. Anything under 30 is usually called oversold.

It was not alone. SanDisk (SNDK) fell 15.47 percent the same day to $1,365.16. Intel (INTC) fell 7.58 percent to $95.18.

So the market spent the week voting against memory, in the same week the memory industry told it the opposite. Two credible voices put the argument on the record. They disagree about nearly everything except one thing: when it gets settled.

What Ben Thompson actually said

Thompson made the sharpest case against the memory makers on the July 6, 2026 episode of Sharp Tech, titled "Did Memory Makers Overplay Their Hand?".

His argument starts with structure. There are three companies left that matter: Micron, SK Hynix and Samsung. In his words, "they're not explicitly colluding, I don't think. But everyone involved has been around the block." The game theory does the rest. If nobody adds capacity, everybody makes money.

That discipline is exactly what he thinks went too far:

"Guess what makes you endure pain? Paying 9x the price for memory."

And then the core of it:

"the memory guys not investing sooner, not moving faster to address this shortage, having this little three three company club of scratching each other's backs and not competing. They are creating the conditions for everyone else in the world to endure pain to reduce their dependency on those big three because the dependency on the big three hurts even more."

His image for it is the one worth remembering:

"there's a bit where if you have a gun you can only fire it once and then everyone's gonna know the gun could be fired and then you lose sort of the leverage of the gun"

Nine times pricing is enormously profitable today. It is also paying every customer on earth to go find another answer. Thompson called this "not investment advice" on the show, and framed it as "a long-term concern like over multiple years." That matters, because he is not calling a top.

The three threats, ranked by how fast they bite

Thompson numbered them himself. In his order:

1. Algorithmic efficiency. "every single per frontier lab on the planet, every hyperscaler is obsessed right now with figuring out how can we use less memory." His proof is Apple, which ships 8 gigabytes of RAM in a MacBook because it controls the chip and the operating system together. His conclusion is blunt: "Smartphone memory demand is going to be permanently lower in the long run even if memory prices go down."
2. Disaggregation of inference. High bandwidth memory, or HBM, is stacked DRAM built specially for Nvidia's chips. Thompson does not think it owns the future alone. On Nvidia's Rubin platform: "they're not just going to be shipping standalone CPU racks, but also standalone memory racks that are using things like flash memory uh to because to offload context because HBM like isn't even enough." Cheaper tiers of memory grow alongside it.
3. China. "this is probably the one that's going to bite first and hardest." The mechanism is unglamorous. HP is rumored to be considering different parts lists by region, approved memory in machines sold in the United States, Chinese memory in machines sold elsewhere. As Thompson puts it, "remember what the key thing about memory, it's fungible."

Here is the part most summaries of that episode missed. Thompson is not bearish on these businesses:

"I think they are structurally worth a lot more money than they were before... I think people are still underestimating how large AI is going to be. But are we gonna eternally be in a world of what we want is HBM, what we need is massive amounts of memory with no improvements and only three providers... No way."

Structurally bullish, skeptical on how long nine times pricing lasts. That is a more useful position than either camp shouting at the other.

Micron's answer, a $100 billion market pulled forward a year

Micron's response, on its own earnings call four days later, was to make the demand bigger and sooner. CFO Mark Murphy:

"Today we indicated that we expected market tightness to continue beyond 2027. Part of that reason is we did see the HBM TAM increase. We saw that. We had said previously that it would cross $100 billion in 2028. We see that now, the HBM TAM easily crossing $100 billion in 2027."

Read that again. A $100 billion market, arriving twelve months earlier than the company guided three months ago. Tightness continuing past 2027.

The spending backs it up. Micron guided fiscal 2027 capital spending "higher than the mid-forties" in billions of dollars. But the detail that matters most for this argument came from EVP of Global Operations Manish Bhatia, who noted that the majority of that fiscal 2027 spending is for construction, and that "those construction dollars are not going to be producing bits in that time horizon."

Micron expects its greenfield capacity in Idaho and China to start contributing meaningfully to output in calendar 2028.

Why both cases land in 2028

That last line is the whole story, and almost nobody has connected it.

Thompson's three threats and Micron's own new supply arrive in the same window. Not because anyone planned it, but because that is how long these things take. Building a fab takes about two years. Re-engineering your product to need less memory takes about the same.

So the honest read is this. 2027 belongs to the memory bulls, and Micron's CFO made that year bigger last week, not smaller. 2028 is genuinely contested. Anyone who tells you they know how that year resolves is guessing.

What the market is pricing right now

The market is not waiting for 2028. It is pricing the bear case today.

Micron trades at 5.7 times next year's expected earnings. Its PEG ratio, which compares the multiple to the growth rate, is 0.13. A reading under 1 is normally considered cheap. At 0.13 the market is saying the growth is temporary and will not repeat.

For scale, Nvidia trades at 16.2 times forward earnings and Intel at 60.1 times. The company at the center of the AI memory shortage is the cheapest large semiconductor name on the board, at a third of Nvidia's multiple and a tenth of Intel's.

That is the debate expressed as a number. The market has already decided the cycle ends. Micron's CFO says the tightness runs past 2027 and the new supply does not land until 2028.

One of them is wrong.

What to watch

• The parts lists, not the headlines. Thompson's fastest threat is China, and the mechanism is boring: computer makers splitting approved memory by region. If more manufacturers follow the HP rumor, that thesis is live.
• Whether efficiency work shows up in orders. Every lab says it wants to use less memory. Watch whether that turns into smaller memory content per server, or stays a research goal.
• Calendar 2028, not fiscal 2028. Micron's greenfield output and the bear case arrive together. That is the year the argument settles.
• Micron's own capital spending. If fiscal 2027 spending goes higher than the mid-forties, the company is telling you it believes its own demand forecast.
• The gap itself. A company at 5.7 times forward earnings, with a PEG of 0.13, is priced for the cycle to break. Everything management said last week points the other way. That distance is the story.

Bargo tracks how tight compute and memory supply actually is, live, on the Compute Tightness Index, and what AI workload demand is doing on the token demand dashboard.

This is not investment advice.

Sources

• Sharp Tech, "Did Memory Makers Overplay Their Hand?", July 6, 2026 — all Ben Thompson quotes, verified against the episode transcript (episode)
• Micron Q3 fiscal 2026 earnings call transcript, July 12, 2026 — Mark Murphy and Manish Bhatia quotes, results and capital spending guidance (transcript)
• Micron investor relations — company filings and results (investors.micron.com)
• Bargo live quote and valuation data, July 17, 2026 — prices, forward multiples, PEG and relative strength readings (bargo.ai)

More research at bargo.ai/research.