Phase Space AI

Research

Seagate Technology [STX]

Seagate Technology Holdings plc [STX] — Research

Date: 2026-07-29 · Spot: $764.20 (Alpaca last trade, 2026-07-29) · Framework: v1.5.1 Fiscal year ends early July. FY2026 ended 2026-07-03; results reported today, 2026-07-29.

---

0. The one-paragraph mechanism

Seagate is one of three remaining nearline HDD suppliers. Demand is set by hyperscaler cold/object-storage capacity; supply is set by heads-and-media capacity that the industry has explicitly stopped adding in unit terms. In FY2026 Seagate shipped exabytes 30%+ higher while revenue per terabyte rose, which is the inverse of the historical HDD price curve. The named mechanism is Mozaic/HAMR areal density (3TB/disk qualified at all major US CSPs; Mozaic 4 at 44TB/drive now in revenue at two of the largest CSPs) sold under build-to-order and exabyte-scale supply agreements that allocate nearline capacity through calendar 2027. That combination — volume growth, positive price, falling unit cost, and contracted volume — produced a FY2026 GAAP gross margin of 45.6% against a business whose prior-decade range was roughly 23–34%.

---

1. Reported results — FY2026, filed today (8-K 0001137789-26-000153, EX-99.1)

Revenue +34.1% y/y. FQ1 FY2027 guidance: revenue $4.1bn ±$100m (+56% y/y vs $2,629m), non-GAAP diluted EPS $7.30 ±$0.20. Dividend raised to $0.74/qtr.

This is unprecedented and it is verified — say so explicitly

A 52.3% quarterly gross margin has no precedent in Seagate's history; the pre-2025 decade sat at 23–34% and the FY2024 trough quarter printed 10.2% gross margin and a −8.9% operating margin. Per MEMO_BRIEF §"Do NOT reject data for exceeding historical precedent", the figure was checked for arithmetic possibility and against a second source, and it holds:

• gross profit + cost of revenue = revenue — 1,898 + 1,731 = 3,629 ✓
• D&A positive ($276m FY26 from the cash-flow statement), so non-GAAP EBITDA ($4,220m LTM) correctly exceeds EBIT ($4,094m GAAP) ✓
• net income / diluted shares = filed EPS — 1,294 / 232 = $5.58 ✓
• EDGAR XBRL (RevenueFromContractWithCustomerExcludingAssessedTax, OperatingIncomeLoss) agrees with Alpha Vantage on all seven quarters checked to within 1.9% ✓

No sanity ceiling was applied. A bound calibrated on 2016–2024 would have rejected every quarter of FY2026.

---

2. Exabytes shipped and revenue per exabyte — the real unit economics

Disclosed on the earnings calls by the CFO. Definitional warning: Seagate changed its exabyte disclosure basis in the December-2024 quarter ("With this change to all periods presented, Seagate's total nearline shipments were 126 exabytes"), and has since moved from "mass capacity" to "data center" as the reported segment. Growth rates must not be computed across that break.

The trend that matters: total revenue per exabyte rose 14.4 → 14.9 → 15.6 ($m/EB) over three consecutive quarters, +8.3%. Data-centre revenue per exabyte rose 13.3 → 14.3 (+7.2% q/q). Management states it directly: "this drove a mid-single-digit increase in year-over-year data center revenue per terabyte. We expect this trend to continue" (CFO Romano, FQ3'26 prepared remarks).

Why this is the whole thesis. In every prior HDD cycle, $/TB fell 10–20% per year — areal-density gains were handed to the customer, so exabyte growth had to exceed price decline just to hold revenue flat. Revenue per exabyte rising while exabytes grow 39% y/y means Seagate is now capturing the density gain rather than passing it through. That is a change in industry conduct, not a change in industry demand.

Cross-check: FQ3'26 exabytes +39% y/y and revenue +44% y/y → price/mix contributed ~+3.6pp. FQ4'26 revenue +48.5% y/y. Volume, not price, remains the larger term — this is not a memory-style price spike (compare SNDK/MU, where analysts on the WDC call described "over 100% sequential pricing increases").

---

3. Pricing structure and contract term — the exact wording

All from FQ3'26 (call for quarter ended 2026-04-03), prepared remarks, CEO Dave Mosley:

"Assurance of reliable supply is our customers' highest priority, particularly for nearline products, which accounted for close to 90% of total exabyte shipments in the March quarter. We have exabyte-scale supply agreements in place with nearly all major cloud and hyperscale customers, with nearline capacity almost fully allocated through calendar 2027. At the same time, we are finalizing build-to-order contracts with these customers through the end of fiscal 2027, which defines specific configuration and pricing. Our value-based pricing approach enables customers to plan with confidence… And we are actively engaged in strategic planning discussions now reaching into calendar 2028 and beyond."

"The progress we have made gives us confidence to significantly increase our annual revenue growth target from the low- to mid-teens to a minimum of 20% over the next few years."

FQ2'26, CFO Romano, Q&A (conceded, therefore stronger evidence than the volunteered claim):

"if we can produce a bit more in a quarter, we will sell those additional exabytes in the open market at a good profit. However, I would say that the vast majority of the volume is already allocated."

Note what is NOT said. Seagate never uses the phrase "sold out" — zero occurrences across thirteen transcripts. The operative wording is "almost fully allocated through calendar 2027" and "the vast majority of the volume is already allocated". The commitments are volume-based (exabytes) with pricing defined per build-to-order contract, not a fixed-price take-or-pay. A residual open-market tranche remains, which is where price would break first.

---

4. HAMR / Mozaic qualification status — named customers, dated

Qualification at PMR-equivalent timelines is the single most important technical datapoint: HAMR's historical risk was that hyperscaler qualification would take 2–4 extra quarters. It did not.

---

5. Accounting quality — is the reported growth real?

Verdict: organic, recurring, no identified quality defect.

Nothing here resembles the 10x Genomics / AAOI / AIOT / Twist defect patterns. The growth is unit shipments to named hyperscalers, collected in cash within 39 days.

---

6. Capital structure — get net debt right

Debt reduced $1,442m in FY2026 and $302m in FQ4 alone. Equity turned positive for the first time in years. Cross-check: EDGAR LongTermDebtNoncurrent 3,465 + LongTermDebtCurrent 398 = $3,863m at 2026-04-03, and DebtLongtermAndShorttermCombinedAmount = $3,905m — consistent with $3,565m one quarter later after $302m of retirement plus discount amortisation. Do not use a stale net-cash field; net debt has fallen 63% in two years.

Interest expense fell $321m → $284m and is guided lower again. Exchangeable senior notes due 2028 add ~2–6m dilutive shares under if-converted; management explicitly includes "the estimated net dilutive impact from the Exchangeable Senior Notes due 2028" in FQ1'27 guidance.

---

7. Transcript mention-frequency (required core metric) — prepared / Q&A

Counts of regex matches per call. Fiscal-quarter labels; 2026Q3 = quarter ended 2026-04-03.

*Table runs to 26Q2 for the like-for-like split; 26Q3 figures quoted in text above.

Reading it. Two signals. (1) HAMR/Mozaic in prepared remarks hit an all-time high of 21 in 26Q2 — management has moved the product transition from a subsidiary topic to the headline. (2) pricing in Q&A went 13 → 32, and LTA/committed in Q&A went 4 → 12: the analyst community is interrogating the durability of price and the contract structure far harder than management volunteers it. That asymmetry (management volunteers HAMR, analysts extract pricing) is the correct read — the contract disclosure is being pulled out under pressure, which per the brief makes it stronger evidence than the prepared-remarks version.

"sold out" never appears. Any memo asserting Seagate is "sold out" is asserting something management has not said.

---

8. Structural or cyclical? — the position, the evidence, the falsifier

Position: the mid-cycle level has genuinely reset upward, and the reset is only partly durable. Call it "structurally higher trough, still-cyclical peak."

Evidence FOR structural (all from filings/transcripts, not narrative):

1. Supply is capped by an explicit industry decision, not by lag. WDC, FQ3'26 Q&A: "we are not making any investments in adding unit capacity… We will definitely look at heads and media capacity investments if it makes economic sense, but not unit capacity investments." Seagate's capex is 4.7% of revenue (FY26 $569m on $12,195m) against a 4–6% target. In prior cycles, margin recovery triggered capacity investment which then broke price. Here the capacity response is areal density (bounded by physics and roadmap cadence, ~1.3x/generation), not units.
2. Demand is contracted, not ordered. Nearline capacity "almost fully allocated through calendar 2027"; build-to-order contracts defining configuration and price through FY2027; LTAs at WDC extending into calendar 2029; manufacturing lead times "about a year", so most POs are placed a year ahead. Backlog visibility of 4–6 quarters is not a spot business.
3. Revenue per terabyte is rising, not falling. Confirmed by both duopolists independently (STX "mid-single-digit y/y increase in data center revenue per terabyte"; WDC "pricing was up 9% year over year"). This is a break in a 25-year price curve.
4. The demand driver is stock, not flow. Inference and agentic workloads generate data that is retained, so the installed base compounds. Top-3 CSP RPO ~$1.1tn (Seagate's own proxy).
5. Only three suppliers, and the third (Toshiba) is not HAMR-competitive at scale.

Evidence FOR cyclical:

1. 199 exabytes at 39% growth is a shipment rate, not a contract. The contracts are volume commitments with periodic pricing adjustment (WDC: "depending on the duration, there may be periods of pricing adjustment"). Volume commitments protect units, not price.
2. Hyperscaler capex is itself the cycle. Substituting "AI capex" for "PC cycle" changes the driver, not the existence of one.
3. Above-base volume prices at spot. WDC: "The LTA volume… does not meet their full requirement, and anything we deliver above and beyond the base volume requirement… is subject to a different pricing regime." The marginal exabyte is still a spot exabyte.
4. The customer has an obvious counter-move: at 50%+ supplier gross margins, hyperscalers will fund alternative capacity, second-source aggressively, or accelerate QLC NAND / tape substitution. Nothing in the filings prevents that; the LTAs merely delay it.
5. No trough test yet. The reset has been observed only through an up-leg. FY2024 printed a 10.2% gross margin quarter. Nothing in the disclosed contract structure would have prevented it.

Synthesis, stated as a position: the floor has structurally reset (contracted volume + capped unit capacity + rising installed base ⇒ a 2024-style 10% gross-margin quarter is now unlikely), and the ceiling has not (50%+ gross margin invites customer-funded supply and substitution). The correct mid-cycle assumption is therefore materially above the old cycle average and materially below the current exit rate — not a mean reversion to 10–15% operating margins, and not an extrapolation of 44.6%.

Falsifier (what would prove this wrong, either way):

• Structural thesis falsified if: revenue per exabyte turns negative y/y for two consecutive quarters while exabyte growth continues; OR any of the three suppliers announces unit-capacity (not areal-density) investment; OR nearline capacity ceases to be described as allocated more than two quarters forward; OR a hyperscaler announces a QLC/tape architecture displacing nearline for object storage.
• Cyclical thesis falsified if: fixed-price LTAs (not volume LTAs) extend past CY2028 at disclosed prices; OR gross margin holds above 45% through a quarter of negative sequential exabyte growth; OR the above-base spot tranche disappears entirely.

Both are checkable each quarter from the disclosures that already exist. That is the point.

---

9. Comparator set — and why it carries no exit-multiple information

Shared set for this cluster: STX, WDC, MU, SNDK. All four run on the same driver (AI-era storage demand), all four reported unprecedented margins in 2026.

A 7.4x dispersion in forward P/E across four names with the same demand driver, the same direction of margin surprise, and overlapping customers. There is no growth ordering that explains it — MU and SNDK are growing faster than STX and WDC. The market is pricing durability, not growth: it believes HDD earnings persist and memory earnings do not.

Per valuation.md rule 6, a reference class can be valid and still carry zero information. This one does: it cannot discriminate an exit multiple. The comparator-derived exit multiple is UNIDENTIFIED; the exit multiple used in valuation is derived from the identity EV_T/EBIT_T = (1−t)(1−g/ROIC)/(WACC−g) and the traded set is reported only as a sanity check that fails to discriminate.

MU as the quantitative read-across baseline — and how far HDD is from it

MU's figures, independently verified against EDGAR primary filings: quarter ended 2026-05-28, revenue $41,456m matching EDGAR to the dollar, gross margin 84.6%, TTM revenue $90.27bn, TTM operating margin 65.8% against a 20-year maximum annual operating margin of 49.3% — i.e. MU is currently earning 16.5pp above its own best year ever. That is real, verified, and not evidence of a data error. It is the benchmark for how far above precedent a verified figure can sit.

The comparison that decides the structural question for HDD:

All three are far above their own historical maxima. The market pays 4.7x for MU's and 26–35x for the HDD names'. That gap is the market's entire judgement, and it is a judgement about mechanism, not magnitude: memory capacity can be added with capital in ~18 months, and MU is spending to add it; HDD capacity cannot be added in units at all under the industry's stated policy, and both HDD suppliers have said so on the record. The evidence in this memo supports that distinction being real. It does not support it being worth a 7.4x multiple spread, which is why the implied-path test still fails.

Does MU's existing memo now look stale? — a finding about the corpus

Reviewed reports/companies/Micron Technology [MU]/Micron_Valuation_Analysis_2026-07-27.md (dated 2026-07-27, two days old).

Not stale on the boom's visibility. The memo already carries the 84.6% quarterly gross margin, explicitly labels it "25.7pp above the all-time annual peak", identifies CDBU/MCBU at 87% gross margin, and models the long-term supply agreements (SCAs) with their CY2030 expiry. It is not a pre-boom document.

Three specific respects in which it is now wrong or self-contradictory:

1. Spot is stale by −12.8%. The memo's reference spot is $820.51; MU closed today at $715.86 after a −12.8% single-day move. Every percentage-to-spot figure in it — including the headline "+52.3%" — is now wrong. The $1,250 target is +74.6% to today's price, not +52.3%.
2. It carries two contradictory outputs 2.6x apart. A 12-month target of $1,250 (+52.3%) and a "weighted price target" of ~$474 (−47%) appear in the same document. The memo says the 5-year DCF target is "retired", yet still prints it as a target with a percentage to spot. A retired output that is still printed with a sign on it is not retired.
3. The 48% normalised gross margin is a duration mismatch, and it flatters the wrong way. The memo justifies 48% as a "pure merchant" post-contract margin because the SCAs expire CY2030. But its own explicit forecast window is FY2027–FY2031, of which four of five years sit inside the contract period. Applying a post-contract merchant margin across the whole window prices the contracts as if they did not exist for four of the five years, and 48% sits 36.6pp below the trailing actual of 84.6% — squarely inside the defect class the brief names (21 of 84 names with a terminal margin below the trailing actual; AVGO −24.3pp, MSFT −23.7pp). The right construction is a duration-matched two-stage margin: contracted through CY2030, merchant thereafter.

By contrast, its 13x exit multiple is consistent with this cluster (12.6x for STX, 12.4x for WDC from the same identity), which is a point in its favour and suggests the multiple discipline in the corpus is now working even where the margin discipline is not.

Recommendation to the book: MU's memo needs a spot refresh and a duration-matched terminal margin. It does not need re-underwriting.

---

10. Downside case with a named cause

Named cause: a hyperscaler-funded second source, or CY2028 LTA renewal at flat-to-down $/TB.

Mechanics: nearline capacity is allocated through CY2027 and BTO pricing is fixed through FY2027. The exposure is the CY2028 renewal. If hyperscalers respond to 50% supplier gross margins by pre-funding capacity (the pattern already visible in NAND wafer prepayments) or by holding $/TB flat at renewal, Seagate's revenue growth reverts to exabyte growth alone (mid-20s%) with gross margin compressing toward the mid-30s as areal-density gains are again passed through.

Quantified: FY2028 revenue $19.0bn (mid-20s% growth, no price), gross margin 34%, opex $1.35bn → operating margin 26.9%, EBIT $5.11bn. At 12.6x warranted exit → EV $64bn, equity ~$62bn, ~$267/share (−65%). Type: MEASURED — logged and scored, does not reject the name.

---

11. Defects found (see final report for magnitudes)

1. AV commonStockSharesOutstanding is the diluted weighted-average, not period-end shares. AV reports 229.0m at 2026-03-31; the 10-Q reports 224.4m outstanding (dei cover 224.2m) and 229m diluted WASO. +2.1% on STX, +9.0% on WDC. Directly inflates market cap.
2. AV EARNINGS_ESTIMATES returns an empty estimates: [] array for both STX and WDC despite the premium tier. The brief's "revision direction is a real input" is unavailable on these names.
3. AV EARNINGS_CALL_TRANSCRIPT quarter parameter is FISCAL, not calendar, on June-FY names. 2026Q2 returns the December-2025 call. A calendar mapping is off by two quarters and would have presented eight-month-old commentary as current.
4. AV operating income differs from EDGAR by up to $55m (STX 2025-10-03: AV 707 vs EDGAR 694, +1.9%; WDC 2026-01-02: AV 963 vs EDGAR 908, +6.1%).
5. av_vs_edgar.py returned "no EDGAR overlap" and trust=NO for both names while EDGAR companyfacts in fact carries seven directly comparable quarters for each. The audit tool silently produced zero comparisons rather than an error — a D1 "existence is not validity" case.
6. Seagate's exabyte disclosure basis was restated in Dec-2024 and the reported segment renamed twice (nearline → mass capacity → data center). Any 8-quarter exabyte CAGR computed from a naive scrape of the transcripts is wrong.