Verdict: KEEP

Benchmark weight 2.52%. Not a close call. MU is the strongest fundamental story in the selectable set and the selloff creates a short-term dislocation against uninformed counterparties.

The Setup

Micron just printed the best quarter in semiconductor history. Revenue $23.86B (+196% YoY), EPS $12.20 vs $9.31 consensus (+31% beat), gross margin 75%, free cash flow $6.9B (record). Guided FQ3 to $33.5B revenue at 81% gross margin and $19.15 EPS — a single quarter that exceeds all of FY2024.

The stock dropped 20% in six days.

Three forces drove the selloff: profit-taking after a +550% run from April 2025 lows, CapEx guidance raised to $25B+ (from ≈$20B), and Google's TurboQuant paper claiming 6x LLM memory compression. The question for filtration: does any of this make MU a drag on the basket over 15 weeks?

What the Market Is Pricing

At $355, MU trades at 5.8x on a real FY2026 EPS build of ≈$61 (FQ1 actual $4.78 + FQ2 actual $12.20 + FQ3 guided $19.15 + FQ4 estimated ≈$25). The published consensus of $35.27 from 50 analysts is stale — most haven't updated post-Q2. Anyone screening on published multiples sees "10.6x forward P/E" when reality is 5.8x.

The 5.8x multiple is the market's answer to the cyclical question. In the 2018 memory peak, MU traded at 4-5x forward at $64.66 before declining 56%. In the 2022 peak, it traded at 11-12x at $98.45 before declining 50%. Today's multiple sits squarely in the 2018 zone. The market is pricing this as the same movie.

The historical pattern is consistent: stock peaks 1-2 quarters before earnings peak, then declines 50%+ while fundamentals are still strong. MU peaked at $471 in early March, earnings are still accelerating. The -24.5% decline so far matches the early innings of both prior cycles.

Where the Market Is Wrong (and Where It Might Be Right)

Mispricing #1: TurboQuant — $25-30/share overreaction (HIGH confidence)

The ≈7% selloff on TurboQuant ($28B market cap haircut) prices in a scenario that first-principles math doesn't support.

TurboQuant compresses LLM key-value cache memory by 6x. KV cache is not total memory. Working from Llama 3.1 70B architecture:

  • KV cache as share of inference HBM varies by context length: 3-5% at short context, ≈53% at 128K with 4 concurrent users, ≈87% at 1M tokens
  • Realistic 2026 workload-weighted average: 19.2% of inference HBM
  • Maximum theoretical total HBM demand reduction: inference share (60%) x KV cache share (19.2%) x compression savings (5/6) = 9.6% of total AI HBM demand

This is the ceiling. It assumes universal deployment (precedent: INT8 took 6-8 years, FP8 took 2-3 years), perfect compression at all model scales (only tested on 8B parameters), and zero Jevons effect (contradicted by every prior compression advance — DLSS increased GPU demand, H.264 increased bandwidth demand).

Meanwhile HBM supply fulfills 50-67% of customer demand. A 9.6% demand reduction, fully deployed instantly, still leaves supply deeply underwater. The market priced in the ceiling at 100% probability with zero offset. That's ≈$25-30/share of recoverable mispricing on a 2-4 week timeline as the headline fades (DeepSeek precedent: similar scare, 3-month recovery).

Morgan Stanley (Joseph Moore, March 26): "From our industry chain surveys, there is no indication whatsoever that memory or storage demand is declining." Maintained Overweight.

Mispricing #2: Cyclical multiple on structural evidence — magnitude uncertain (LOWER confidence)

This is the real debate. The market prices MU at the 2018 cyclical peak multiple. The question is whether the structural evidence justifies a higher multiple.

What's genuinely different this cycle:

Five-year Strategic Customer Agreements. First one signed in Q2. Prior cycles were all spot or 1-year contracts. This is not incremental — it changes the revenue visibility model from commodity to strategic. Multiple additional SCAs in negotiation.

Supply constrained by physics, not just capacity. HBM requires 4x wafer area and advanced packaging (CoWoS bottleneck). You can't convert commodity DRAM lines to HBM overnight. No new fab capacity until H2 2027. MU can only fulfill 50-67% of key customer demand.

Elastic demand driver. Prior cycles were driven by PC/smartphone (bounded endpoints). AI inference demand is elastic — cheaper memory enables more inference, which drives more memory demand. Jevons paradox applies structurally.

Industry confirmed independently. SanDisk guided $12-14 EPS vs street $3.63 consensus and disclosed the industry is transitioning from "quarterly auction" to multi-year supply agreements. Netlist confirmed OEM DRAM up 3-4x, spot up 7-8x, "unprecedented." Korean memory equities repriced: SK Hynix +455% 1Y, Samsung +296% 1Y.

What's the same:

81% gross margin is 34 percentage points above the highest prior cyclical peak (59.6% in 2018, 46.9% in 2022). Even if structurally higher due to HBM mix, margins this extreme invite reversion.

CapEx ramp at $25B+ FY2026, stepping up "meaningfully" in FY2027. In both prior cycles, maximum capex commitment preceded the earnings peak by 2-4 quarters.

Samsung entering NVIDIA's HBM supply chain. Three-player HBM competition will moderate pricing power over 12-24 months.

My assessment: The structural evidence is real but doesn't resolve within 15 weeks. P(margins compress from 81% within 15 weeks) ≈ 5%. P(within 12 months) ≈ 30-40%. P(within 24 months) ≈ 60-70%. The market prices ≈40-50% probability of near-term reversion. I price ≈30-40%. That's a ≈10pp edge on the cyclical question — meaningful but not huge.

What is NOT mispriced

The earnings quality, the supply constraints, the HBM supercycle thesis — 29 of 32 analysts rate Buy. This is consensus. The edge is not in knowing the bull case. Everyone knows the bull case.

The Options Tell a Specific Story

Pre-earnings (June 18, 83 days): P/C OI 1.58, heavy put positioning. ATM puts outnumber calls 4:1. Defensive.

Post-earnings (July 17, 112 days): P/C OI 0.80 — flips to neutral. Max pain $390, 10% above current. Unusual activity: 10,703 contracts at $165 puts (11.9x OI) — institutional tail hedge on a large long, not a directional bet.

Both expirations: Call IV elevated above put IV (+4.9% June, +6.3% July). This is unusual — normally puts carry higher IV. Upside vol is being bid. Someone is paying for upside exposure while headline positioning looks bearish.

Translation: institutional money is long, hedging cheaply, and paying up for upside through earnings. The put OI is insurance, not conviction.

FQ3 earnings report June 24. Inside the 15-week window. If MU beats again (FQ2 beat guidance by 45% on EPS — $12.20 vs $8.42 guide), the cyclical-peak narrative gets harder to sustain. But the bar is elevated: street consensus ($34.76B revenue, $19.84 EPS) already sits above the high end of management guidance ($34.25B, $19.55). A "meet" is a "miss" in this setup.

Consensus

29 of 32 analysts rate Buy/Strong Buy. Median price target $510 (+43% from $355). 19 PT raises the day after earnings. Goldman Sachs is the lone prominent bear — Neutral, $400 target. Summit Insights downgraded to Hold citing H2 deceleration.

The sell-side unanimity is itself a risk. When everyone agrees, the pain trade is the other direction. But for filtration, the question is narrower: is MU a weak name? At 5.8x real earnings, 81% GM, $14B guided quarterly FCF, net cash, deleveraging — the answer is unambiguously no.

Counterparty Analysis

Who sold at $355 after a +31% EPS beat with $19.15 FQ3 guide?

  1. Profit-takers after +550% run — not expressing a fundamental view. Edge: their selling creates a dislocation from fundamental value.

  2. Momentum algos responding to broken 50DMA — systematic, not discretionary. They'll buy it back when the trend re-establishes.

  3. TurboQuant sellers who saw "6x memory reduction" and sold without doing the math — uninformed. Didn't distinguish KV cache from total HBM, didn't calculate the 9.6% ceiling, didn't check that HBM is sold out.

  4. Peak-cycle fearers who see 81% gross margins and assume reversion — partially informed. Right about the destination, wrong about the timeline.

No informed sellers with superior information identified. The $400B market cap means Goldman, Citadel, and every quant fund has the same Q2 data. The selling is technical and sentiment-driven.

Factor Decomposition

Regression (trailing 250 days): MU = -1.27xSPY + 2.28xXLK + 0.41xMTUM + e. Idiosyncratic variance 45.7% — well below the 75% target. MU is predominantly a tech sector bet against {SPY, XLK, MTUM}.

The orthogonal Sharpe of 2.54 (110.4% alpha / 43.5% idio vol) is inflated. The "alpha" is substantially the HBM/memory supercycle latent factor that XLK doesn't capture. XLK is 60%+ AAPL/MSFT/NVDA — a proper regression against SMH would absorb much of this.

For filtration this is acceptable. The QQQ short neutralizes most of the sector beta. The residual exposure is to the memory supercycle factor — which is the thesis. Keeping MU = a bet that memory outperforms QQQ's average constituent. Given the fundamental gap, that's a bet worth taking.

Semiconductor cluster risk: 10 semi names total 11.57% of the selectable set (MU, AMD, LRCX, AMAT, KLAC, QCOM, ASML, WDC, STX, MRVL). Filtration decisions on these names are correlated. Each evaluated independently, but the cluster exists.

Balance Sheet

Net cash $6.5B (record). Two credit upgrades to solid BBB in the quarter. FCF $6.9B (FQ2), guided to roughly double in FQ3 (≈$14B). March 25 8-K: commenced cash tender offers on six series of senior notes (5.3%-6.05% coupons, 2031-2035 maturities). You don't retire high-coupon debt at premiums if you think the cycle is peaking. 30% dividend increase to $0.15/quarter.

No going concern language. No material weaknesses. Netlist patent litigation ($445M jury verdict, not accrued — PTAB invalidated both patents, Netlist appealing at CAFC) is immaterial vs $7.5B/quarter revenue.

Kill Conditions

FILTER if any of these materialize:

  • Any major hyperscaler cuts AI capex guidance
  • DRAM spot prices break below 2x early-2025 levels
  • FQ3 revenue misses below $32.75B (low end of guidance)
  • Management language shifts from "tight" to "easing" on supply
  • A second major compression breakthrough stacks on TurboQuant
  • Samsung HBM4 qualification at NVIDIA confirmed ahead of schedule

15-Week Expected Path

Weeks 1-4: TurboQuant sentiment dissipates. Technical digestion continues. Stock likely trades $330-380 range as momentum algos finish unwinding. This is the maximum pain window.

Weeks 5-10: Consensus estimates update. Stale FY2026 P/E of "10.6x" converges toward real 5.8x. Fundamental floor reasserts. DRAM pricing data for Q2 confirms supply constraints.

Weeks 11-15: FQ3 earnings (June 24). If the beat pattern continues (FQ2 beat guide by 45% on EPS), the stock reprices. If MU merely meets guidance while consensus is above the high end, the reaction is muted or negative — but the fundamental trajectory is still accelerating.

P(MU outperforms average QQQ constituent over 15 weeks): ≈60%. P(roughly matches): ≈25%. P(underperforms by >2%): ≈15%.

Expected excess return: ~+5%, with wide confidence interval driven by the unresolved cyclical question and FQ3 earnings outcome.