Phase Space AI

SanDisk Trade Construction 2026-07-27

SanDisk [SNDK]

SanDisk Corporation [SNDK] — Criteria Analysis & Trade Construction

Task 5 · analysis 2026-07-27 · migrated to the Criteria framework 2026-07-29 · framework v1.5.1 Price $1,095.78 (scan, 2026-07-28). The 2026-07-27 build used $1,278.23.

---

2026-07-29 MIGRATION NOTICE — read this before anything below it

Migrated from the retired six-Gate / E[R]-vs-cash-hurdle framework onto the Criteria framework (references/criteria.md, references/valuation.md).

1. The position verdict is deleted. The prior version of this document opened with INVESTMENT DECISION: WATCHLIST (research direction: BEARISH). The memo no longer emits Long / Short / Watchlist / Avoid — it emits an analysis, and whether that analysis justifies a position is a question about a particular book.
2. E[R] versus the 4.7% cash hurdle is deleted, replaced by the reverse-DCF implied-path test and a 12-month target. Sensitivity now runs over the exit multiple, never over scenario probabilities.
3. Criteria are retired; every test is a named Criteria with a type, returning PASS / FAIL / INDETERMINATE. A missing input is INDETERMINATE, never FAIL.
4. A 76% revenue understatement has been corrected — see §2A. It is the most consequential change in this document.

Pre-computed inputs are taken from reports/scan/SNDK_analysis.json (as-of 2026-07-28) and are used rather than recomputed, except where explicitly reconciled below.

---

2A. THE DATA CORRECTION — last fiscal year → TTM

OLD revenue $7.355bn. NEW revenue $12.978bn. Last-FY understated SanDisk by 76%.

Where the error did and did not sit. SanDisk_Valuation_Analysis_2026-07-27.md states no EV/Sales figure, so its headline conclusions were not directly contaminated. The last-FY figure entered through the screen / scan / rescore path, where it set the demonstrated growth rate and the multiple that drove the scoring. That path is what rejected the name.

The correction runs in both directions. The scan's revenue_cagr_demonstrated of 9.9% is a CAGR on annual revenue ending at the same stale FY2025 endpoint. Measured to the same TTM endpoint as the revenue:

Resulting change in the valuation conclusion: the implied-path margin moves from −34.2pp (as recorded in the 2026-07-28 rescore, on a 20% hardcoded terminal margin) to −26.4pp on the scan's growth measure, to −4.8pp on the TTM-consistent growth measure. SanDisk still FAILs the Valuation Criteria — but by roughly 5 percentage points, not 34. The prior document's characterisation of the price as indefensible does not survive the correction; the price is demanding, and it is within one revision cycle of being defensible.

Note on SanDisk's history length. SanDisk was separated from Western Digital in February 2025. EDGAR carries three fiscal years and four quarterly TTM observations. Several inputs that would normally be well identified are not, and each is marked INDETERMINATE below rather than being filled with a guess or a peer median.

1. The four conclusions, kept separate

1. Fundamental conclusion. SanDisk is executing exceptionally. Revenue +251% y/y, gross margin 22.5% → 78.4%, Datacenter +645% y/y, $1,849m of debt repaid to zero, an accumulated deficit turned into $2,746m of retained earnings in nine months, and $41.6bn of contracted remaining performance obligations booked into the financial statements. Earnings quality is clean (accruals −0.13). This is not a deteriorating business and the memo does not claim it is.

2. Expectations conclusion. The Street carries FY2027 revenue of ~$41bn and EPS of $177–208, revised up twice in three months, with an average target of $2,188 and a Buy consensus. The house carries FY2027 revenue of $31.1bn and EPS of $103.04. ΔE = −24% on revenue, −42% to −50% on EPS. The disagreement reduces to one testable assumption: the direction of NAND ASPs through fiscal 2027. Consensus needs them up again from the Q4 FY2026 exit rate; the house has them down ~13%.

3. Valuation conclusion. Under the house forecast, six independent anchors produce $360–787, averaging $566.60. Reverse-engineering the current price produces no feasible solution in a reverting model — it requires a through-cycle gross margin of 111–183%, or terminal revenue of ~$128bn against a whole-industry NAND market of $70–90bn. The price is only reachable if reversion does not occur at all.

4. Portfolio conclusion. No position. The research view is bearish and well-evidenced; the trade fails on implementation, and separately on portfolio correlation. Details below.

---

2. THE CRITERIA BLOCK (framework v1.5.1)

Every Criteria is scored. The memo blocks on none of them.

The old "Liquidity Criteria FAIL — this is the binding gate" is retired. Under the current framework the Liquidity Criteria asks whether the name can be owned and exited at the intended size; SanDisk's equity can be. What actually failed was a vehicle, priced on a real chain — and a vehicle that cannot be filled or cannot be justified at 134.6% implied vol is a vehicle problem, not an admission test on the underlying.

3. VALUATION — TWO OUTPUTS, BOTH MANDATORY

The old §Valuation Criteria computed E[R] = 0.35(+80.7%) + 0.45(+71.8%) + 0.20(+55.8%) = +71.7% for a short and compared it to a 4.7% cash hurdle. Both the arithmetic and the hurdle are retired. Running a range across scenario probabilities while point-estimating the exit multiple is the defect this replaces.

3A. Implied-path test (reverse DCF) — the Valuation Criteria

RESULT — what the price requires: a 36.4% five-year revenue CAGR. Margin = demonstrated − required = 31.6% − 36.4% = −4.8pp. → FAIL, narrowly. On the scan's annual-FY growth measure of 9.9%: −26.4pp. On the 2026-07-28 rescore's hardcoded 20% terminal margin: −34.2pp. The correction moves this name from "rejected by 34 points" to "missed by 5".

Implied compression, as a number. EV/EBIT compression is INDETERMINATE — FY2025 EBIT is negative, so today's EV/EBIT does not exist. On a sales basis the exit multiple is 22.6 × 18.45% = 4.17x EV/Sales against today's 12.22x = −65.9%. That is the compression the price is asking the business to absorb, and it is the honest way to state it when the earnings multiple is undefined.

Exit-multiple identification. The scan anchor is GROWTH_MATCHED with n=30, and the constituents do bracket SanDisk's growth — but only 5 of the 30 are semiconductor or optical companies (KLAC 30x, FN 38x, MU 39x, MRVL 113x, AMD 198x); the remainder are healthcare and software. Restricting the same growth-matched screen to the semis/optical cohort leaves exactly n = 5, the minimum, at a median of 39.2x — which would put the required CAGR at 22.2% and the margin at +9.4pp, a PASS. Both numbers are reported; neither is claimed to be the answer. This spread is the identification problem, and it is why the sensitivity below matters more than the point estimate.

Sensitivity — over the EXIT MULTIPLE, never over probabilities:

The flip point sits between the anchor set's median and its 75th percentile — i.e. within the ordinary dispersion of the anchor set itself. That is the correct place to state the uncertainty, and it is precisely where the old probability-weighted method could not look.

Terminal value is far above 60% of EV (SanDisk earns nothing on the filed FY2025), so the reverse DCF is mandatory as the primary long-horizon output and the two-tranche DCF in SanDisk_Valuation_Analysis_2026-07-27.md §2 is supporting evidence.

3B. The 12-month target

• Revenue base: FY2027 Street revenue ~$41bn. SanDisk's FY2027 ends ~July 2027, so this is very nearly the revenue that will be trailing at the target date. Sourcing caveat, carried forward from the original build and not softened: this is an aggregator figure, not a primary feed.
• Named product-cycle events inside 12 months (dated in the catalyst calendar, none invented): Q4 FY2026 results, 5 Aug 2026; Investor Day, 13 Aug 2026, at which management has said it will publish a new long-term model; the first sequential ASP decline, expected at Q1 FY2027 results ~November 2026.
• Multiple anchor — and here the honest answer is that there isn't one.

THE 12-MONTH MULTIPLE ANCHOR IS UNIDENTIFIED. SanDisk has traded as a separate company since February 2025. A trailing EV/Sales series requires four tagged quarters, which first exist on 2026-01-30 — giving 123 daily observations, all of them inside a six-month window during which trailing revenue rose from $8.9bn to $13.0bn. The observed range (11.11x – 27.43x, median 16.46x) describes a market pricing a ramp, not a multiple regime. Applying it to $41bn of revenue produces targets of +251% to +483%, which is not a valuation — it is the same multiple counted twice. Per criteria.md, an input that cannot be identified is INDETERMINATE and must be declared, not replaced with a peer median. It is declared.

What can be stated instead, and it is falsifiable:

The multiple required merely to HOLD spot on Street FY2027 revenue is 3.87x trailing EV/Sales — a 68.4% de-rating from today's 12.22x. On the house's own FY2027 revenue of $31.1bn it is 5.10x, a 58% de-rating. Today's price does not require multiple expansion; it requires the revenue to arrive while the multiple compresses by roughly two-thirds.

12-month target, stated as a range with its basis declared:

12-MONTH TARGET: INDETERMINATE on the own-multiple rule. The house point estimate of $566.60 (−48.3%) is retained from the original build as the best available figure and is flagged as resting on a DCF-and-scenario construction the current framework demotes, not on the name's own multiple history. This is reported as INDETERMINATE rather than manufactured, which is the whole point of the rule.

Sanity band. The Street average of $2,188 is +99.7% to spot and 3.9x the house figure. That gap is not resolved here and is not tuned away. What can be said precisely: the Street target implies roughly 7.8x trailing EV/Sales on its own FY2027 revenue — inside SanDisk's observed 11–27x range only if the ramp completes and the multiple holds, which is the same double-count identified above. The external target is a check on the output, never a calibration target for the model.

4. Portfolio context — the portfolio-book contract

Book state (portfolio_book.json, as_of 2026-07-27): 100% cash. Zero positions. Watchlist: MU, ISRG, NET, SMR, NBIS. Limits: 5% max single name, 25% max sector concentration, 0.60 max pairwise correlation without disclosure, 4.7% cash hurdle, −2% of book hard stop per position.

4.1 The correlation finding — stated explicitly, as required

Daily-return correlations, common window 2025-12-04 to 2026-07-27 (n=159, bounded by SanDisk's listing history):

An SNDK position plus an MU position is one factor bet in two tickers, not two positions. At ρ = +0.78 the diversification benefit of holding both is close to nil: a two-name equal-weight basket has ~94% of the variance of a single doubled position. The book's 0.60 threshold is breached by a wide margin and the two names must be sized as a single semiconductor/AI-hardware cluster exposure, capped in aggregate, not sized independently.

The same applies to WDC (+0.72) and STX (+0.71) — the entire memory complex is one trade. The screen said this in its thematic read ("one factor bet wearing twenty tickers"); the correlation matrix confirms it numerically.

CIEN at +0.55 is meaningfully less correlated to SNDK than MU is — a genuine distinction between the two names underwritten in this batch, and relevant to which one deserves the cluster's capital.

Only NET (−0.00) is genuinely orthogonal.

4.2 Capital competition — RETIRED as a test

The 4.7% cash-hurdle test is retired and is not applied. Under the current framework, capital competition is resolved by slot competition inside the strategy, not by a return threshold inside the memo, and the memo emits no direction to test. The prior text here concluded that "a long returns −55.7% against a 4.7% hurdle… SanDisk does not survive underwriting as a long." That conclusion is deleted.

What survives, and is genuinely useful to the screener: SanDisk was the screen's #1-ranked long of 47, and the underwriting found the implied-path margin negative. On the corrected TTM inputs that margin is −4.8pp, not the −34.2pp the last-FY data produced. The screen was far closer to right than the underwriting said it was, and the gap between them was a data artifact.

4.3 Sizing arithmetic (offered to the strategy, not decided here)

Conviction composite (weights per references/trade-construction.md):

Volatility tier: HIGH (109.1% realised, far above the >45% threshold). Grid output at Medium × High = 1.0% of book, hard-capped at 1.5%, and counted inside a combined semiconductor/AI-hardware cluster cap of 3% alongside any MU exposure. (The prior build carried this arithmetic to the conclusion "below the minimum that survives a 109%-vol name → WATCHLIST". That step is deleted: sizing is a constraint, not a verdict, and inverse-volatility sizing is the operative control whatever view a book takes.)

Consensus positioning note: this would be a position against a Buy consensus that is actively strengthening (FY2027 EPS revised up twice in three months). The disagreement is well-decomposed — it traces to one falsifiable assumption about ASP direction — so it is not penalised merely for being contrarian. But going against strengthening revisions in a 109%-vol name is real crowding risk and is named here next to the grid output rather than absorbed into the volatility tier.

---

5. Vehicle analysis — priced, and rejected

4.1 Outright short equity

Borrow is cheap and available (5.22% of float, 1.5 days to cover). Max loss unbounded. Rejected: a 109%-vol name with an active $6bn buyback and a bullish Investor Day nine sessions away is not a responsible outright short at any size that would matter.

4.2 Defined-risk put spread (the default structure) — priced and rejected

SNDK 18-Sep-2026 $1,110 / $830 put spread (real quotes, Alpaca, 2026-07-27):

Rejected. The structure needs a 21.4% decline in 53 days merely to break even, and the market already prices ±45%. Paying 124.5% implied volatility against 109.1% realised, to express a twelve-month terminal-value view, through a window containing a bullish Investor Day, is a poor trade even if the fundamental view is right. This is the variance-risk-premium rule in references/alpaca-options.md operating as intended.

4.3 Naked long puts

Rejected outright. ATM implied volatility is 119–153% across the curve and skew is inverted. Buying premium here would require the house expected move to materially exceed the option-implied move; it does not.

4.4 Pre-authorised conversion structure (documented so conversion is executable)

Not to be used before the conversion triggers fire. If the triggers in §5 are met and implied volatility has normalised (ATM below ~90%), the structure of record is:

SNDK 15-Jan-2027 $1,110 / $830 put spread — the longest listed expiry with real open interest, chosen so the Q1 FY2027 print (~November 2026, the first quarter that must show a sequential ASP decline) falls inside it. Indicative quotes today: buy SNDK270115P01110000 at $291.35 ask (IV 114.6%, delta −0.288, vega 2.91, theta −0.90, OI 208); sell SNDK270115P00830000 at $140.28 bid (IV 116.8%, delta −0.177, vega 2.21, theta −0.71, OI 59). Net debit $151.07, max value $280, reward:risk 0.85:1 — which is why it is not being bought today. The structure only becomes attractive after implied volatility compresses or the underlying rallies.

---

6. Monitorable triggers — what would make the impairment case live

These are no longer conversion triggers for a position verdict; no verdict is emitted. They are the dated, falsifiable observations that would move the Criteria scores, listed so a book can act on pre-committed evidence.

All of A, and at least two of B, and C:

A — the variant must survive the near catalysts (binding): - 5 Aug 2026: FY2027 revenue guidance or commentary implying below ~$36bn (i.e. materially under the ~$41bn consensus), or a Q1 FY2027 gross-margin guide below 70%.

B — evidence (two or more): 1. Reported quarterly gross margin declines sequentially for the first time (first checkable: Q1 FY2027, ~November 2026). 2. TrendForce or an equivalent reports a negative q/q NAND contract-price change. 3. The 13 Aug Investor Day publishes a long-term gross-margin target below 55%, or declines again to publish one. 4. Reported RPO fails to grow from $41.6bn in the FY2026 10-K despite two additional NBMs having been signed in Q4. 5. FY2027 consensus EPS is revised down by more than 10% over any 30-day window. 6. Datacenter revenue growth decelerates below 40% y/y.

C — the tape (a Momentum Criteria observation — ENTRY TIMING ONLY; it cannot make or unmake the case for owning the name, and given the momentum-crash configuration it is the timing input that matters most): - A weekly close below the 200-day moving average (~$823), or - A failed rally producing a lower high below the 50-day moving average (~$1,727) — i.e. enter on strength, not into this drawdown.

And the vehicle test must independently clear (Liquidity Criteria, options limb): ATM implied volatility below ~90%, or an entry price high enough that a defined-risk structure produces at least 2:1 reward-to-risk. A vehicle that cannot be filled or justified is not a vehicle.

7. Evidence that would strengthen the implied path

• The next NBM tranche is struck at current (78%+ gross margin) economics rather than December-2025 floors, disclosed as a step-up in RPO per contracted bit; and
• The contracted share of FY2028 bits is disclosed above 50%; and
• Price below $700 (within the house valuation range, where the windfall cash alone provides support).

8. Analysis-invalidation triggers (for the bearish research direction)

• Price above $2,000 — the house valuation framework is wrong, or the market has decisively repriced the terminal. (The prior text read "Move to Avoid". No verdict is emitted; the observation stands, the instruction is deleted.)
• The 13 Aug Investor Day publishes a credible, contractually-supported long-term gross-margin target above 65% with the contract mechanics disclosed.
• NAND contract prices post a fourth consecutive quarterly increase above +20% q/q (i.e. TrendForce's rebalancing call is wrong).
• Reported bit shipments grow above 20% y/y while margin holds — this would break the "all price, no volume" finding, which is the spine of the whole argument.
• The FY2026 10-K discloses NBM pricing terms showing floors at or above current realised ASPs.

---

9. Answering the seven required questions

What is the market wrong about? The direction of NAND ASPs through fiscal 2027, and therefore which year's earnings deserve a multiple. Consensus applies a normal multiple to peak-year earnings and calls the result cheap.

What evidence supports the differentiated view? TrendForce's 21 July 2026 supply/demand reversal for CY2027; the contract-price second derivative already rolling from +72% to +12.5% q/q; ~$40bn of NAND conversion WFE pulled forward; flat y/y bit shipments; and the $41.6bn RPO's own shape — floors struck at ~51%-margin economics against a business running at 78–81%.

Why now rather than years from now? Because two dated events land in the next three weeks (5 and 13 August), the second of which forces management to attach a number to the terminal margin.

What prevents the company growing into the valuation? Nothing prevents it in FY2027 — the house explicitly concedes FY2027 will be strong. What prevents it thereafter is that half of Flash Ventures' fixed costs are payable regardless of output, bit growth is capped at mid-teens by the JV structure, and ~$40bn of industry conversion capacity is already funded.

What would invalidate it? §7 above, each item dated and checkable.

What is the adverse scenario? An Investor Day that publishes a credible structural margin model, a $6bn buyback executing into a 45% drawdown, inverted skew unwinding into a melt-up, and 5.22% of float covering in 1.5 days. The stock has more than doubled twice in twelve months.

Why is taking the position superior to not owning it? The memo does not answer this, and under the current framework it must not. That is a question about a particular book's slots, its existing factor exposure and its volatility budget. What the analysis supplies is the input to it: the implied-path margin is −4.8pp on TTM-consistent inputs (not the −34.2pp the uncorrected data produced), realised volatility is 111.9%, the option-implied move over the catalyst window exceeds any move the house would forecast, and two dated events land within three weeks. The prior version of this section concluded "It is not, today." That conclusion is deleted.

---

Options data: Alpaca options snapshots, 578 contracts, 2026-07-27. Correlations: Alpaca SIP daily bars. Book state: portfolio_book.json as_of 2026-07-27 — read, not modified; the write-back is staged in PENDING_book_and_ledger_updates.json.