← Writing
Theme Deep-Dives

When Implied Volatility Sits Below Realised: Reading the Memory Names in September 2026

5 September 2026 · 13 min read


What this is: analytical working, published for education. Every figure carries the date it was measured. Nothing here is a recommendation, a forecast, or investment advice, and the tickers are worked examples of a method rather than a list to trade. Parts of this analysis were produced with AI assistance and checked against primary sources; if you spot an error, tell us.


Run a premium screen across large-cap US names in early September 2026 and something odd shows up. On a lot of them, 30-day implied volatility is printing below trailing 30-day realised volatility. The options market is quoting less future movement than the stock has just delivered.

The reflexive reading: premium sellers are underpaid. Sell a put while implied vol sits under realised and you take less compensation than the stock's own recent behaviour says you should demand.

That reading is right often enough to be a useful default. It also fails in one specific, identifiable situation, and September 2026 is full of that situation.

1. The ratio compares a forecast to a measurement

Implied volatility is what the options market thinks the next 30 days hold. Realised volatility is what the last 30 days contained. Divide one by the other and you are comparing a forecast to history.

So iv/rv below 1.0 means "underpaid" only if the last 30 days are a good guide to the next 30. When something one-off has just happened and finished, a forecast below trailing realised vol is the market being correct.

Both stories look identical in the ratio. A name that gapped 20% on an earnings miss three weeks ago and has been calm since, and a name whose volatility is accelerating, can print the same iv/rv.

Splitting the two, measured

The test that separates them: does trailing 30-day realised vol or trailing 10-day realised vol better predict the next 30 days? Measured on 19,485 name-days across 17 large-cap US names over five years of daily closes, to 2026-09-04. Lower mean absolute error wins.

RegimenMAE, 30-day windowMAE, 10-day windowBetter forecast
All observations19,4850.1600.18730-day
RV30/RV10 between 1.3 and 1.62,7140.1530.16730-day
RV30/RV10 between 1.6 and 2.01,3610.1920.18810-day, marginally
RV30/RV10 above 2.07150.2710.20410-day, clearly
RV30/RV10 below 0.83,0170.1750.32430-day, by a wide margin

Read the top row first. Most of the time the 30-day window wins, so the default reading of iv/rv holds and the recent window is mostly noise.

The fourth row is the exception. When the 30-day reading is more than double the 10-day reading, a volatility event is decaying out of the window and the recent measurement forecasts better. Implied vol below trailing realised is then the options market pricing the future correctly, with the ratio misleading you.

The last row is the guardrail, and it is the half that matters more. When volatility is expanding, the 10-day window is dramatically worse, with nearly double the error. A rule that simply switched to the shorter window would fail exactly where failure costs most: on names that are deteriorating.

Two honest limits. The windows overlap, so observations are heavily autocorrelated and the effective sample size is far smaller than the row counts suggest. And this measures one realised-vol window against another. It says nothing about whether implied vol beat either. A premium seller cares about that comparison, and answering it needs a depth of implied-vol history most retail data sources do not carry.

What it looks like on real names

Illustrative, computed on daily closes to the 2026-09-04 close. A snapshot that went stale immediately.

NameRV30RV10RV30/RV10Better windowiv/rv on that window
Name A (memory)0.880.491.7910-day1.29
Name B (memory)0.830.471.7610-day1.38
Name C (connectivity)1.161.270.9130-day0.60
Name D (exchange)0.810.890.9130-day0.80

The first two names read as badly underpaid on the 30-day window and adequately paid on the window the data says forecasts better. The last two read as underpaid on both, because their volatility is rising, not decaying. Same superficial signal, opposite situations.

One thing to take from this section: check whether volatility is decaying or expanding before you read the ratio at all. The direction of travel decides which number means anything. Same reason delta is not your probability of assignment: a number that looks like an answer can be measuring something adjacent to the question.

2. The memory names have been reporting well, and it removed some fear

Volatility across the memory complex spent the middle of 2026 elevated on a familiar worry: that memory demand was peaking. We looked at this group when fat premium was the story in fat premium and the memory bottleneck, and at the capex chain behind it in from AI capex to ROI capture. Memory is the most cyclical business in semiconductors, and fear of a cycle top has historically de-rated the group early and hard.

The reported quarters argue against a demand peak. All figures from company press releases and SEC filings, not transcripts.

CompanyQuarterReportedRevenueVersus own guidance
MicronFQ3 FY262026-06-24$41.46Bguided $33.5B ±0.75B
SK hynix2Q262026-07-29KRW 79.32tnrecord, 76% operating margin
Samsung (Device Solutions)Q2 20262026-07-30KRW 127.5tnrecord, sales +56% QoQ
SeagateFQ4 FY262026-07-29$3.629Bguided $3.45B ±0.10B
Western DigitalFQ4 FY262026-08-05$3.747Bguided $3.65B ±0.10B

Micron beat its own guidance midpoint by 24%, grew revenue roughly 346% year over year, took non-GAAP gross margin from 39.0% to 84.9% in four quarters, and guided fiscal Q4 to $50.0B at about 86% gross margin. From its 10-Q:

"AI-driven memory and storage growth is outpacing industry supply... demand for memory and storage exceeds overall industry supply."

The split between price and volume matters more than the headline. Quarter over quarter, Micron's DRAM sales rose 67% on a low-60% ASP increase and only a low-single-digit increase in bits shipped. NAND rose 99% on a mid-80% ASP increase and mid-single-digit bits. Nearly all of the growth is price, because bits cannot move much when supply is constrained.

Two structural signals point the same way. Micron disclosed 16 agreements representing roughly $100B of minimum contracted revenue, backed by about $22B of customer deposits including $18B in cash. SK hynix finalised long-term agreements with around 10 customers. Customers do not wire $18B in advance for capacity they expect to be cheap later.

The fear was specific, and the evidence removed it. Volatility came down with it. Back to section 1: several memory names show implied vol below trailing 30-day realised vol because the 30-day window still contains the fear that has since been answered. Their RV30/RV10 ratios cluster around 1.76-1.79. Four names, one shape, decaying on one schedule. Sector-wide worries clear like that.

Three markets, one label

These are not all the same business. DRAM and HBM are the memory the AI shortage is about. NAND flash is a separate market. Hard-disk drives are separate again: Western Digital completed the separation of its flash business into Sandisk on 2025-02-21, so WDC today is a pure hard-drive company, as Seagate always was. Nearline HDD demand for AI-generated data is real and durable, and it runs on a different cycle with a different competitive structure. Treating all five as one memory trade is a correlation assumption. Make it deliberately.

3. The bear cases that survive, and how rates could trigger them

Good quarters do not close the argument. Several bear cases survive the evidence above, and they are stronger than the ones that did not.

The evidence for a turn

Customer inventory is building while customers keep ordering. TrendForce reported on 2026-07-09 that US cloud providers accumulated DRAM inventory through Q2 2026, because server CPU shortages slowed system assembly, and that they kept procuring aggressively anyway in anticipation of a 2027 shortage. Inventory piling up at the customer while the customer orders harder out of scarcity fear is double-ordering, and it is how memory cycles have historically topped. Buyer resistance is already visible elsewhere: TrendForce noted on 2026-07-03 that "elevated OEM inventories have significantly reduced buyers' willingness to accept another round of price increases" in client SSDs.

Price momentum has gone flat at a record level. Contract prices rose 13-18% quarter over quarter for conventional DRAM in 3Q26 and 10-15% for NAND, described as a "noticeably slower pace than previous quarters." Weekly spot on 2026-09-04: DDR5 16Gb $54.07, +0.12%; DDR4 16Gb $93.50, +0.63%; DDR3 4Gb -0.15%. A market that ran up several hundred percent in a year is printing sub-1% weekly moves, with the oldest node already negative. Spot typically leads contract by about a quarter.

The largest customer is designing memory out. NVIDIA halved the SOCAMM capacity on next-generation Vera Rubin modules and is evaluating lower HBM stack configurations for Rubin Ultra than its original baseline. Cloud providers cut RDIMM capacities in the first half of 2026. Architectural workarounds like CXL switching and KV-cache compression are shipping. Demand destruction by engineering is slower than demand destruction by recession, and harder to reverse.

The affordability ceiling has a number. TrendForce projects memory will consume 47% of major cloud-provider capital expenditure in 2026, rising to 68% in 2027, with some providers already at negative free cash flow. Sixty-eight percent of a capital budget going to one input line does not hold for long.

Earnings revision breadth is rolling over. Morgan Stanley's Shawn Kim published on 2026-07-21 that the net EPS upgrade rate across the group fell from a peak of 92% to 77%, with contract prices peaking in Q4 2026. Read that note carefully. Its conclusion is that the cycle "will lengthen rather than collapse," and the stance is neutral. It calls a peak in the rate of change. A downcycle is a different claim.

Supply arrives eventually. Chinese producer CXMT is estimated at roughly 350,000 wafers per month in 2026, close to Micron's projected 375,000, with China's total DRAM capacity approaching 600,000 wafers per month. The gating constraint is immersion DUV lithography, and no domestic Chinese supplier has yet delivered it commercially, so meaningful domestically-tooled capacity is an early-2030s event. New capacity from the incumbents does not produce material bits until 2H27 at the earliest.

Where the bear case is weak

Conventional DRAM is not weak. Prices rose 13-18% quarter over quarter, and on 2026-09-04 DDR4 16Gb traded at $93.50 against DDR5 16Gb at $54.07, the older node at 1.7x the price of the newer one. An inversion like that is a shortage signature in the non-HBM business, and it undercuts the argument that HBM strength is masking a weak core. Add $18B of prepaid customer cash, a supply wave that does not arrive until 2027-2028, and balance sheets strong enough to avoid the forced-capitulation dynamic that ends most cycles.

Where rates come in, and where they do not

The intuitive chain: rates rise, the companies funding the AI buildout feel pain, capital budgets tighten, memory demand falls. A more specific version deserves attention. The largest platforms fund capital spending out of advertising revenue, advertising is bought substantially by small businesses, small businesses are rate-sensitive, and firms experimenting with AI cut discretionary spend first. That gives the platforms two reasons to slow down.

The first link measures strongly. Regressing daily returns on the market and on changes in the 10-year yield, 753 trading days to 2026-09-04:

NameRate coefficientt-statistic
Large advertising platform A+0.02251.76
Large advertising platform B+0.00480.45
Independent ad-tech name+0.01970.83
Large enterprise-software platform+0.02823.11
Small-cap index-0.0307-5.97

Small caps are hit hard and unambiguously by rising yields. The most statistically significant relationship in the set, and the first link of the chain.

The transmission from there is invisible in prices. The advertising-funded platforms cluster around zero, and the enterprise-software platform is significantly positive. Equity markets are not currently pricing a rate-to-advertising-to-capex channel.

The chain may still be true. It would run on quarterly fundamentals, at a resolution daily prices do not show, and equity prices reprice on expectation long before revenue confirms anything. It also runs slower than most option expiries. No plausible mechanism gets a policy decision into memory revenue within a month.

One subtlety makes this worse for anyone underwriting these names. So much revenue is now locked into multi-year take-or-pay agreements with cash prepaid that reported financials have become a lagging indicator by design. A break in spot prices would show up in the equities well before it showed up in an income statement. The next strong quarter is not an all-clear.

One more wrinkle sits in Micron's own disclosure, and it comes from the company, not a sceptic. Its largest agreements carry "a ceiling price for existing products that approximates the market price in the second calendar quarter of 2026, and a floor price through the term," with management noting that even at floor pricing the margins clear any previous cycle peak. Capping your own upside at roughly today's price is what a management team does when it wants durability more than the top of a rise.

What to actually watch

Three observable things, all public, none requiring a subscription:

  1. Quarterly contract-price releases. The 3Q26 baseline is conventional DRAM +13-18% and NAND +10-15% quarter over quarter. A 4Q26 print below +10%, or a flat-to-negative 1Q27 guide, would mark the deceleration turning into something else. Expect NAND to break before DRAM.
  2. Weekly DRAM spot. DDR5 16Gb sat at $54.07 on 2026-09-04, up 0.12% on the week. Two consecutive negative weeks, or DDR4 16Gb breaking below $90, is the earliest and cheapest inflection signal available.
  3. Inventory days and capex guidance in the makers' filings. Rising inventory days alongside rising capital expenditure while prices are still at records is the late-cycle pattern, and it lands in the filings before it lands in the price.

What the premium is paying for

Implied volatility below realised volatility is not automatically underpaid premium. It is a forecast disagreeing with history, and sometimes the forecast is right because the history contains something that has finished happening. Check whether volatility is decaying or expanding before reading the ratio at all.

Several memory names show that decay pattern in September 2026, and the reported quarters give a reason for it: a specific fear about peaking demand was tested against results and did not survive in the form it was held. What survives is narrower and slower. Customer inventory building, price momentum flattening at a record level, memory eating an unsustainable share of customer budgets, and a supply wave dated to 2027-2028.

None of that is a view on where any share price goes next. It describes what the premium is compensating for, and that is the only thing an option seller can assess. To see how a specific name screens on quality, value and the risk sitting alongside the credit, use the screener.


Sources and dates. Company results: Micron FQ3 FY26 8-K exhibit 99.1 and 10-Q, filed 2026-06-24/25, quarter ended 2026-05-28; SK hynix 2Q26 results release 2026-07-29; Samsung Electronics Q2 2026 results 2026-07-30; Seagate FQ4 FY26 8-K exhibit 99.1, 2026-07-29; Western Digital FQ4 FY26 8-K exhibit 99.1, 2026-08-05; Sandisk FQ4 FY26 release 2026-08-05; Western Digital/Sandisk separation completed 2025-02-21. Industry data: TrendForce press releases dated 2026-07-03, 2026-07-09, 2026-07-21, 2026-08-04 and 2026-08-25; DRAMeXchange spot prices as of 2026-09-04. Analyst research: Morgan Stanley, Shawn Kim, 2026-07-21, read via secondary reprint rather than the primary note. Volatility and return figures computed from daily closing prices to the 2026-09-04 close. CXMT capacity estimates are third-party modelled figures, not company disclosure. Two CXMT claims circulating in early September 2026 — a 10% global DRAM share and HBM3E shipments — could not be verified beyond headlines and are excluded from the argument above.

Common questions

What does it mean when implied volatility is lower than realised volatility?
The options market expects the next 30 days to be calmer than the last 30 days were. Implied volatility is a forecast of future movement and realised volatility is a measurement of past movement, so a ratio below 1.0 compares a forecast to history and does not by itself prove a mispricing. It is only mispriced if the past turns out to be the better guide to the future. When a stock has just come through a one-off event that has since resolved, a forecast below trailing realised vol is the market being right, and the seller is not necessarily underpaid.
Is selling options when IV is below RV always a bad trade?
No, and running it as an automatic rule loses money in one identifiable situation. The better question is which measurement window forecasts the next month more accurately. Measured across 19,485 name-days on 17 large-cap US names over five years, the trailing 30-day realised vol window was the better forecast overall, with a mean absolute error of 0.160 against 0.187 for the 10-day window. When the 30-day reading sat more than twice the 10-day reading, meaning a vol event was visibly decaying out of the window, the 10-day window won clearly at 0.204 against 0.271. The relationship inverts when vol is expanding: there the 10-day window was far worse, 0.324 against 0.175. The window matters more than the ratio.
Why is option premium high on memory stocks?
Premium pays you for the width of the range the market expects, and memory has an unusually wide one. Memory is the most cyclical business in semiconductors, with a long history of sharp boom-and-bust cycles, and it currently sits inside an AI-driven shortage that has pushed contract prices up violently. The bull case and the bear case both involve large moves, so implied volatility stays elevated. High premium describes the width of the distribution and says nothing about which side of it the stock lands on.
Are memory companies still growing as of September 2026?
On the most recent reported quarters, yes, and sharply. Micron's fiscal Q3 2026, reported 2026-06-24, showed revenue of $41.46B against its own guidance midpoint of $33.5B, up roughly 346% year over year, with non-GAAP gross margin at 84.9% versus 39.0% four quarters earlier, and it guided fiscal Q4 to $50.0B. SK hynix reported a 76% operating margin for 2Q26, and Samsung's Device Solutions division reported record results. The detail that matters is the split between price and volume: Micron's DRAM sales rose 67% quarter over quarter on a low-60% ASP increase and only a low-single-digit increase in bits shipped. Almost all of the revenue growth is price. Constrained supply does that, and expanding demand would instead show up in bits.
What would signal that the memory cycle is turning?
Three observable things, none of them needing a subscription. First, the quarterly contract-price releases: conventional DRAM rose 13-18% quarter over quarter for 3Q26, so a 4Q26 print below +10%, or a flat-to-negative 1Q27 guide, would mark the deceleration. Second, weekly spot prices: DDR5 16Gb printed $54.07 on 2026-09-04, up 0.12% on the week, and DDR3 4Gb was already negative, so consecutive negative weeks would be an early inflection, since spot typically leads contract by about a quarter. Third, inventory days and capital-expenditure guidance in the makers' own filings, because rising inventory alongside rising capex at record prices is the late-cycle pattern.
How would interest rates affect memory demand?
Not directly, and the transmission takes longer than it looks. The large cloud providers that drive memory demand fund their capital spending largely from operating cash flow, so their borrowing costs move very little with policy rates. Measured over 753 trading days to 2026-09-04, regressing daily returns on the market and on changes in the 10-year yield, the large advertising-funded platforms showed rate coefficients statistically indistinguishable from zero, while the small-cap index showed a strongly negative coefficient with a t-statistic of -5.97. Rate moves visibly hit small companies and do not visibly hit the large platforms. Any chain running from rates to memory has to pass through second-order effects on customer budgets over several quarters, and that is slower than an options expiry.
Does levelbox.ai recommend any of the stocks in this article?
No. levelbox.ai makes no recommendations, predictions or price forecasts, and every ticker in this article is a worked example of a method, not a suggestion to trade it. The tooling screens candidates for quality and value, prices a cash-secured put at a target delta, and shows the credit next to the maximum loss, the breakeven and the modelled whole-book loss in a drawdown, so reward and risk sit side by side and the decision stays with you. This is analytical and educational tooling, not investment advice.

Read next