Micron's Print Has Nowhere to Hide: The Quarter That Audits the AI Memory Trade

Consensus has lapped Micron's own guidance in 90 days. When management guided Q4 to $50.0 billion (±$1.0B) on June 24, 2026, the LSEG consensus was $43.58 billion — meaning the print itself was already 14.8% above what institutional models expected. Three months later, that consensus has risen to $50.86 billion (Zacks, Sept. 25) and UBS now models $52.4 billion with $32.50 EPS. A Sept. 30 number inside the guide band no longer clears the bar — it disappoints. The market is paying $1,082.28 per share for $50 billion in expected revenue, an EV/sales of13.5x trailing, with Wall Street modeling EPS jumping 117% to $158.67 in fiscal 2027.

Burry opened his Micron short at $1,051.87 on July 1, 2026 — three days after the Q3 print — and labeled the stock a 'destroyer of capital' on his Substack (per Morningstar, July 3, 2026). He added to the position again on Tuesday Sept. 23, alongside fresh shorts in Palantir, Nebius, and the iShares Semiconductor ETF. The thesis is explicit: SOX is the most overextended versus its 200-day moving average since 2000, with a price-to-sales above 16 — a 'pure form of overvaluation in an index,' he wrote on June 30. VanEck's SMH just printed its worst July in 30 years (-5.26% for the month, -17.59% trailing 30 days through late July, per Business Insider). Burry rolled his puts to March 2027 expirations at SOXX strike prices in the low-to-mid 400s.

The macro layer matches. UBS now forecasts hyperscaler capital expenditure growth decelerating from 76% in 2026 (~$673 billion) to25% in 2027 and just 6% in 2028 (Reuters, July 16, 2026). Goldman Sachs still models $1.01 trillion of 2027 capex, up 32% from $765 billion in 2026, but the marginal year-over-year delta is the lowest since the cycle began. Even SK Hynix' own CEO has broken the price-discipline script, calling current memory prices 'abnormally high' and warning that expanding capacity and accepting slimmer margins may be necessary to avoid 'chipflation.' The same statement that bulls cite as evidence of shortage conviction is the statement bears use as the cycle's first concession.

Counterpoint's Q2 2026 HBM data is the under-appreciated shift. SK Hynix' HBM share has fallen to 50%, down from 64% a year earlier, while Samsung Electronics is closing the gap as its HBM4 yield approaches 80% — the same milestone Micron confirmed on Aug. 26, 2026. All three makers reportedly sold out calendar 2027 capacity in August, with no incremental supply currently planned (DigiTimes, via Reddit and Cybernews coverage). That is the opposite of a typical cycle peak: instead of inventory building, the industry is rationing against a constraint that the suppliers themselves say will persist into 2028.

The packaging bottleneck compounds the memory math. TSMC's CoWoS-L capacity is the gating step for every NVIDIA Rubin, every Broadcom TPU-variant, and every custom hyperscaler ASIC that pairs with HBM4. TSMC's trailing EV/sales is 13.97x with capex running at 35.2% of revenue — and roughly half its2026 capex is being absorbed by advanced packaging and leading-edge nodes earmarked for AI accelerators. When HBM, CoWoS, and Nvidia's allocation interact, the chain tightens or breaks together. A Q4 print that under-delivers on HBM4 ramp visibility will be read first as a TSMC demand signal, then as a Broadcom capex risk, before Micron's stock moves at all.

The fiscal Q3 print already locked in the inflection. Revenue of $41.456 billion, up 73.7% sequentially and 345.7% year-over-year, against the $9.30 billion base a year earlier. Non-GAAP gross margin of 84.9% — a company record, up from 39% in the prior-year quarter and above every other large-cap U.S. tech peer at the time of the release. Operating margin printed 81.2%; GAAP operating income was $33.318 billion on $1.738 billion of operating expenses (Micron 10-Q for the quarter ended May 28, 2026, filed June 25, 2026). The reason margins are the story isn't the level — it's the leverage. Every incremental dollar of HBM revenue at this point carries almost no marginal cost, which means any softness in conventional DRAM pricing hits the model harder than the HBM beats lift it.

The contracting structure has hardened into a moat. Micron now discloses 16 Strategic Customer Agreements — mostly five-year take-or-pay contracts covering roughly 20% of DRAM volume, with HBM4 already past $1 billion in cumulative revenue and ramping twice as fast as HBM3E did at the equivalent stage. The micro-shift that matters for Q4 commentary is HBM4 mix: management has said HBM3E and HBM4 are fully booked through calendar 2027, but HBM4 carries roughly 80% higher pricing per stack than HBM3E (~$550 vs ~$300 for 36GB 12-Hi, per silicon analyst pricing references). Every percentage point of mix toward HBM4 widens the gross-margin gap to SK Hynix' 76.2% TTM and Samsung Electronics' 56.9%.

Days-to-quarters: the print itself. A Q4 print inside the $50B guide with a Q1 FY27 guide above $46B and a reiteration of sold-out status through 2027 confirms the bull case and forces a re-rate toward the $1,295-$1,625 PT band. A print inside the guide with flat-to-down Q1 guidance — the Seeking Alpha framing the market is sitting with — is the capitulation scenario, because it implies Micron's own team now believes the deceleration UBS modeled for hyperscalers is showing up in the order book. Either outcome is binary; the gray zone is narrow.

One-to-three years: the structural test. If NVIDIA Rubin and Broadcom's custom-silicon roadmaps ship at the cadence implied by the NVIDIA Q3 FY26 revenue run-rate ($303B trailing) and the 76% 2026 capex ramp, HBM4 demand absorbs the entire 2027 capacity and the oligopoly keeps its pricing discipline. If hyperscaler capex growth does what UBS forecasts — 25% in 2027, 6% in 2028 — the cycle's second half looks like SK Hynix' 'abnormally high' warning coming home: pricing has to give, mix has to widen, and the marginal HBM stack starts competing with conventional DRAM for the same wafer. That is the scenario Burry's puts are written against, and it is the scenario a 'seeking peak on Sept. 30' tape fully prices in.

The data still sides with the bulls on facts and with the bears on framing. Facts: 84.9% gross margin, $41.5 billion quarter, 16 strategic contracts, 2027 sold out, NVIDIA Q3 FY26 revenue up105.9% YoY, hyperscaler capex +76% in 2026. Framing: consensus $50.86B already above Micron's $50B guide, SOX P/S >16x, Burry puts struck at $1,051.87, UBS forecasting the capex cliff, SK Hynix calling its own prices 'abnormally high.' The print is the first event that forces both sides to reconcile: either Q1 FY27 guidance extends the sold-out language into 2028 and re-anchors the multiple, or the cycle's pricing power starts to leak.

The honest call for the next 90 days is mixed, not bullish. Micron into the print is priced for perfection against a slowing backdrop, which means upside is capped by the guide and downside is open on any Q1 framing slip. The cleanest read is to own the suppliers — SK Hynix at 8.1x P/E and 6.99x P/S with 50% HBM share and 76.2% TTM gross margin, TSMC on the packaging bottleneck — and underweight the leveraged downstream (Broadcom at 44.9x trailing P/E versus 19.4x forward is the most exposed if the 2028 capex cliff arrives on schedule). NVIDIA sits in between: the largest single buyer of HBM and the most diversified monetizer of the demand it creates.