Analysis · Memory · Outlook

Why did memory stocks bounce back, and can it continue?

16 August 2026 · 16 min read

The short answer: memory stocks bounced back because three things landed inside three trading days. On Wednesday 12 August Nebius reported $37.49bn of remaining performance obligations, which is contracted AI compute that cannot be delivered without memory attached to it. On Thursday 13 August SanDisk used its investor day to publish a long-term margin model built on volume commitments rather than spot pricing. On Friday 14 August SanDisk added a $14bn buyback and disclosed $93.9bn of multi-year customer contracts, and Korea's Kospi — home to the world's two largest DRAM makers — crossed into a bull market. The bounce is real, it is documented, and it is narrower than it looks: essentially all of it happened in the final three sessions of the week.

The more interesting part is that this board called memory the weakest segment on the market five days before it became the strongest, and published the exact conditions that would flip it. Both conditions were met, on all four names. That is worth walking through, because it says something about what a price-based score can and cannot do.

All board figures below are the close on Friday 14 August 2026, the last complete session. Everything else is sourced at the foot of the page.

How far did memory stocks actually move?

Over the five sessions from the 7 August close to the 14 August close, all four memory and storage names on this board rose, and all four finished in the top seven of 34 by five-day move.

Name	Close 14 Aug	5-day	Sentiment	Board rank	Below 52-wk high
SanDisk (SNDK)	$1,641.11	+35.38%	82 · Hot	3 of 34	30.3%
Seagate (STX)	$973.44	+19.77%	82 · Hot	2 of 34	15.0%
Western Digital (WDC)	$508.80	+17.15%	77 · Hot	5 of 34	36.4%
Micron (MU)	$971.66	+10.72%	75 · Hot	6 of 34	22.6%

At segment level, memory and storage finished the week on a cap-weighted sentiment score of 77, the highest of the six segments on this board, against a whole-complex reading of 56. All four names were advancing on the final session.

What this board said about memory a week earlier

On 9 August this site published Are memory stocks cooling off?. The answer was yes. Over the ten sessions to 7 August, memory was the only one of six segments whose sentiment did not rise. It finished that window at 33 against a complex reading of 59, with none of its four names advancing, and three of the five lowest scores on the entire 34-name board were memory names.

That article ended by naming what would change the reading, and it named it mechanically rather than editorially:

What it would take to flip the segment is mechanical: momentum turning positive, and the group outpacing rather than trailing the complex.

Five sessions later, both conditions were met on every one of the four names:

Name	Momentum	Relative strength	Both conditions met?
SanDisk	+11.9	+9.6 (max)	Yes
Seagate	+11.7	+9.6 (max)	Yes
Western Digital	+11.3	+9.6 (max)	Yes
Micron	+8.7	+9.3	Yes

Three of the four are pinned at the maximum the relative strength component can contribute, which is worth reading carefully: it means each of them outpaced the cap-weighted complex by more than the component can express, so the true gap between them is unmeasured. Two names both showing +9.6 are not necessarily equal.

The segment went from last of six to first of six in five sessions — a 44-point move in a score that is bounded at 98.

It is worth being precise about what that does and does not demonstrate. The score did not predict this. It is built entirely from price and volume, it has no view on contract pricing or capacity, and it was describing a real state of affairs on 7 August, when memory genuinely was the weakest thing on the board. What it did was state, in advance and in public, the specific observable conditions under which its own reading would reverse — and then register the reversal within days of the news arriving. A score that describes accurately and updates fast is a different and more honest instrument than one that claims to forecast.

What actually happened, day by day

The single most useful fact about last week is that it was not a week-long recovery. It was a three-session event. Daily closes, from this site's own bar data:

Session	MU	SNDK	STX	WDC
Mon 10 Aug	-1.89%	+2.12%	-1.45%	+0.93%
Tue 11 Aug	+0.87%	+2.68%	+2.44%	-0.09%
Wed 12 Aug	+4.92%	+5.76%	+7.03%	+3.69%
Thu 13 Aug	+4.23%	+13.67%	+4.91%	+7.31%
Fri 14 Aug	+2.30%	+7.39%	+5.65%	+4.41%

Monday and Tuesday contributed almost nothing. Across those two sessions Micron was down 1.0%, Seagate up 1.0% and Western Digital up 0.8%; only SanDisk managed a meaningful 4.9%. From Tuesday's close to Friday's close, the same four names put on 11.9%, 29.1%, 18.6% and 16.2% respectively.

Whatever explains this rally has to explain Wednesday. Anything that was already true on Monday does not.

Why did memory bounce back? The three catalysts

1. Wednesday: Nebius put a number on contracted AI compute

Nebius reported second-quarter revenue of $582.3m, up 454% year on year, and disclosed $37.49bn in remaining performance obligations — revenue under contract but not yet delivered — alongside first-half capital expenditure of $8.13bn. The stock rose 34% on the day.

The read-through to memory is mechanical rather than sentimental. Contracted GPU capacity is not a forecast of demand; it is a signed obligation to build. Every rack of accelerators that obligation implies carries a fixed ratio of high-bandwidth memory beside the GPU and enterprise SSD capacity behind it. An RPO number of that size is, indirectly, a memory order book. That is why a neocloud print moved four memory stocks on the same afternoon.

If the distinction between contracted and consumption-based capacity is unfamiliar, we cover it in hyperscalers vs neoclouds.

2. Thursday: SanDisk published a margin model, not a forecast

SanDisk's "In Focus" 2026 Investor Day was the largest single-day move of the week — the stock rose 13.67% — and the reason is more specific than "good guidance".

The company laid out a long-term financial model covering FY2028 to FY2030: revenue growing mid-to-high teens in line with bit growth, non-GAAP gross margin sustained at approximately 80%, non-GAAP operating margin around 75%, operating expenses near 5% of revenue, and adjusted free cash flow margin of roughly 50%.

Those are not numbers a memory company is historically able to promise four years out, because memory margins are normally a function of a spot price nobody controls. What SanDisk anchored them to is its New Business Model agreements — contracts carrying committed volumes and minimum financial guarantees rather than quarterly price discovery. Management said it had signed NBMs with eight customers, covering roughly 50% of bits in FY2027 and about two-thirds of bits in FY2028.

On Friday the company put a value on that book: $93.9bn of multi-year customer contracts across eight clients, including three US hyperscale data centre operators, and authorised a $14bn share buyback.

3. Friday: Korea, and the analysts

Korea's Kospi crossed into bull market territory, up more than 20% from its July low. That matters to a US memory investor for one reason: SK Hynix and Samsung Electronics are the two largest DRAM and HBM producers in the world, and neither is listed in the US. When Seoul re-rates the two largest suppliers, the US-listed memory complex tends to follow, because they are exposure to the same product.

Sell-side targets were reset on SanDisk the same week: JPMorgan resumed coverage at Overweight with a $2,250 target, Evercore ISI reiterated Outperform at $2,800, and Susquehanna raised its target to $3,250 — all above the $1,641.11 Friday close.

Coverage is not causation

A necessary caveat, and one this site applies to itself. Most financial headlines are written after a move and about it. A story published the same day as a rally is evidence of what was written, not proof of what caused it.

The reason to treat last week as better-than-usual evidence is timing and breadth, not the existence of the articles. The SanDisk investor day was a scheduled event with a published deck; the Nebius RPO figure was in a filing; the buyback was an authorisation. These are dated, primary disclosures rather than commentary. And they line up with the turn: nothing much happened Monday and Tuesday, and everything happened from Wednesday.

What cannot be established from any of it is attribution weight. Whether Wednesday was 70% Nebius and 30% positioning, or the reverse, is not knowable from price data, and this site will not pretend otherwise. The market sentiment score deliberately has no opinion here — it is built from price and volume only, which is exactly what makes it reproducible by anyone with the same bars.

Is this the same memory cycle as every other one?

This is the question that actually matters, and it deserves the honest version rather than the exciting one.

The bear case is the historical base rate, and it is strong. Memory is the most reliably cyclical business in semiconductors. Every previous upcycle ended the same way: high prices funded new capacity, the capacity arrived 18 to 24 months later, spot prices collapsed, and margins went with them. Nothing in physics or economics has repealed that.

There is also already a visible deceleration in the rate of price increases. TrendForce's quarterly contract price series shows conventional DRAM rising 58-63% quarter on quarter in 2Q26, then 13-18% in 3Q26. NAND went from 70-75% to 10-15% over the same two quarters. Prices are still rising — but the second derivative turned down two quarters ago, while the stocks made new highs last week. That combination is the classic late-cycle signature, and anyone buying memory here should be able to say why this time it is not.

The bull case is that the contract structure has genuinely changed, and this is the part that is new rather than rhetorical. In a normal cycle, a spot price fall passes into revenue almost immediately, because bits are sold close to spot. What SanDisk described on Thursday is a book where two-thirds of FY2028 bits sit under agreements with committed volumes and minimum financial guarantees. Kioxia said earlier this year that hyperscaler customers had proposed long-term agreements covering 2027 and 2028, against its normal 12-month cycle. If a majority of industry bits are pre-sold on enforceable multi-year terms, the transmission mechanism that produced every previous memory bust is not eliminated, but it is meaningfully dampened.

On the demand side, Deutsche Bank's published supply model has 2026 DRAM demand at 2,261k wafer starts per month against capacity of 2,051k — roughly a 10% shortfall — widening to 3,563k against 2,769k by 2028, a 29% gap. Even granting that supply forecasts two years out are among the least reliable numbers in this industry, the gap is large enough that it does not close on rounding.

Where I think this goes from here

The section below is opinion. Everything above is sourced fact or this board's own recorded data. What follows is a judgement, and judgements are wrong more often than data is.

My view is that the structural argument is real and the easy money is gone, and I do not think those two statements are in tension.

The contracted-bits argument is the strongest genuinely new thing in memory in a decade. It converts part of a commodity into something closer to a utility with take-or-pay terms, and if SanDisk's two-thirds-of-FY2028 figure is representative of where Micron and Kioxia also end up, then the depth of the next downcycle should be materially shallower than the last one. That is a structural change in the quality of the earnings stream, and it deserves a higher multiple than memory has historically carried. I think that part of the re-rating is justified.

What I am much more cautious about is the starting price. Micron was reported up 233% year to date at Friday's close - a figure from the coverage linked below rather than one this board computes. SanDisk closed Friday at $1,641 having been at $1,212 five sessions earlier. The published sell-side targets sit 37% to 98% above the Friday close. These are not prices that embed disappointment; they embed the good version of the story being delivered on schedule. The asymmetry that existed in early August — when this board was scoring the whole segment at 33 and three of its names sat among the five weakest on the market — has largely closed. Buying the same structural thesis is a different proposition at 35% higher.

Concretely, into 2027 I would expect: contract prices to keep rising but at a decelerating rate, consistent with what TrendForce is already showing; HBM to remain the tightest and most defensible part of the stack, which favours Micron on mix; and enterprise NAND to be where the surprise sits, in either direction, because that is where the contract restructuring is furthest along and where inference workloads are changing the shape of demand fastest. The name I find hardest to underwrite at these levels is the one that moved most: a 35% week on a long-term margin model for fiscal years that begin two years from now is the market pricing execution it has not yet seen.

The honest summary: I would be a buyer of the memory thesis on weakness and I would not chase it here. If that sounds unsatisfying, it is because the single best entry point of the last month was five sessions ago, and this board was publishing the reason it existed at the time.

What would change my mind

A capex announcement cycle from Samsung, SK Hynix or Micron that front-runs the 2028 gap. The supply deficit is the load-bearing assumption. Aggressive capacity additions announced in 2026 land in 2028, precisely when Deutsche Bank's model has the gap widest. That would be bearish, and it would arrive as good news.
The share of bits under long-term agreement stalling or reversing. SanDisk disclosed roughly two-thirds of FY2028 bits. If that figure does not keep climbing across the industry — or if Micron and Kioxia disclose materially lower coverage — then the structural argument is company-specific rather than industry-wide, and memory is still just memory.
A negative quarter-on-quarter contract price print. Deceleration is expected. Outright decline while capacity is still being added would mean demand, not supply, is the binding constraint, and that is a different and worse story.
Hyperscaler capex guidance being cut. Everything above rests on AI infrastructure spending continuing. The contracts are enforceable, but a customer base under pressure renegotiates.

On this board, the mechanical version of the same watch-list is simpler: momentum and relative strength are what carried the segment from 33 to 77, and they are what will carry it back. Both are visible daily on the memory and storage segment page and in today's movers.

The bottom line

Memory stocks bounced back because three dated, verifiable disclosures landed in three sessions — a contracted-compute number from Nebius, a contract-anchored margin model from SanDisk, and a buyback plus a Korean re-rating of the two largest DRAM producers on the planet. The segment went from the weakest on this board to the strongest in five sessions, and the two mechanical conditions this site published on 9 August as the test for that reversal were both met, on all four names.

Whether it continues is a question about capacity in 2028 and about how much of the industry's bit supply ends up under enforceable multi-year contract. The first is unknowable today. The second is disclosed quarterly, in filings, by companies that have started competing on it — which makes it the most useful thing to actually watch.

Frequently asked questions

Why did memory stocks go up last week?

Three catalysts landed in three sessions from 12 to 14 August 2026: Nebius disclosed $37.49bn of remaining performance obligations, implying contracted AI compute that requires memory; SanDisk published a long-term margin model at its investor day anchored on committed-volume contracts covering about two-thirds of FY2028 bits; and SanDisk added a $14bn buyback and $93.9bn of multi-year customer contracts while Korea's Kospi, home to the two largest DRAM producers, entered a bull market.

How much did memory stocks rise?

Over the five sessions to the 14 August 2026 close, SanDisk rose 35.38%, Seagate 19.77%, Western Digital 17.15% and Micron 10.72%. All four finished in the top seven of the 34 names on this board by five-day move. Almost all of the gain came in the final three sessions.

Is the memory shortage going to continue into 2027?

Contract prices are still rising but decelerating: TrendForce shows conventional DRAM up 58-63% quarter on quarter in 2Q26 and 13-18% in 3Q26, with NAND slowing from 70-75% to 10-15%. Deutsche Bank models a roughly 10% DRAM supply shortfall in 2026 widening to about 29% by 2028. The main risk to that view is a capacity response: additions announced in 2026 arrive around 2028.

Which memory stock has the strongest sentiment score right now?

At the 14 August 2026 close, SanDisk and Seagate both scored 82 out of 100, ranking third and second of the 34 names on this board. Western Digital scored 77 and Micron 75. The score is built from price and volume only and describes how a stock is behaving, not whether it is good value.

Does ChipSentiment's sentiment score predict memory stock moves?

No. The score is calculated entirely from price and volume, so it describes behaviour rather than forecasting it. What it did do on 9 August 2026 was publish the specific mechanical conditions under which its reading of memory would reverse - momentum turning positive and the group outpacing the complex - and both were met five sessions later.