Capex Signal: Q1 2026 Earnings Sweep — Memory Supercycle, Custom Silicon, and the Tariff Overhang

The Q1 2026 earnings season for the AI infrastructure stack is largely in the books, and the numbers are extraordinary. Micron printed a 196% YoY revenue increase. Marvell guided FY2027 to $11 billion. ASML's Q4 2025 bookings came in at nearly double analyst estimates. Synopsys surged 66% YoYin its first quarter with Ansys. The demand signal across memory, custom silicon, EDA, and wafer fab equipment is as clean as I've seen it.

The backdrop: trade policy uncertainty has been a persistent overhang since the April 2025 reciprocal tariff episode, where a 10% universal baseline tariff triggered a sharp semiconductor selloff before electronics and chips were carved out via presidential memorandum. The January 2026 Section 232 action then imposed a 25% tariff on a narrow category of advanced AI chipsexported to China, alongside BIS rule changes shifting the review policy for advanced computing semiconductors destined for China from “presumption of denial” to case-by-case. That's the real policy environment we're operating in right now — not a market-crashing shock, but a slow-building reshoring and export-control pressure that's reshaping supply chain geography.

Here's the full earnings sweep and what I'm taking away.

Q1 2026 Earnings Scorecard

Micron (MU) — The Number That Broke My Model

A 196% year-over-year revenue jump to $23.9 billion is not a normal earnings report. Micron is now guiding Q3 FY2026 to $33.5 billion— roughly the entire quarterly revenue of Intel at its peak — from a memory company. The driver is HBM4.

Micron has its entire calendar 2026 HBM supply sold out, and the company can only meet 55–60% of total DRAM customer demand. They raised capex guidance from $20B to over $25 billion this fiscal year to chase the HBM4 ramp — with mass production scaling in 2H 2026. The HBM TAM projection: from $35B in 2025 to $100 billion by 2028, arriving two years earlier than prior estimates.

My read: the market's ~7% post-earnings pullback was a valuation argument, not a thesis argument. The underlying demand signal — HBM supply fully committed, HBM4 required for NVIDIA's Rubin platform, 40% CAGR TAM through 2028 — is as clean a constraint signal as I've seen. The Constraint is real. The Intensity is accelerating. The Durability is locked in by architecture: every future NVIDIA generation needs more HBM per chip, not less.

Marvell (MRVL) — Custom Silicon Is Now a $1.5B Business

Marvell's Q4 FY2026 was the cleanest execution story of the quarter. Data center revenue hit $1.65 billion — now 75% of total sales — and the custom AI ASIC business that was essentially zero two years ago generated $1.5 billion in annual revenue in FY2026. Management guided FY2027 revenue to $11 billion and as high as $15 billion by FY2028.

What I'm watching closely: Marvell's 1.6T optical interconnect solutions. As GPU cluster sizes grow from thousands to tens of thousands of accelerators, the interconnect layer becomes the next bottleneck. Marvell is positioned in both the custom ASIC layer (doing for networking what Broadcom does for compute) and the optical transceiver space. The stock jumped 16% on the Q4 print for good reason.

Synopsys (SNPS) — The Ansys Flywheel Starts Spinning

The first quarter with Ansys fully consolidated showed what the combined entity looks like: revenue of $2.41 billion, up 66% YoY, with Ansys contributing $886 million. EDA (design automation) revenue surged 96% YoY to $2.0 billion. Non-GAAP operating margin was 42.1%.

The stock sold off despite the beat — partly valuation, partly the $13.5B of acquisition debt. But the market is underweighting the synergy timeline. The first joint SNPS+Ansys solutions hit in 1H 2026, with monetization in FY2027. If you're building a custom AI chip — and everyone is — you need both EDA (Synopsys) and multi-physics simulation (Ansys) in the same workflow. This is becoming the design stack for AI silicon. The backlog of $11.3 billionsays the funnel isn't thinning.

Wafer Fab Equipment — The $139B Market

SEMI is projecting global semiconductor equipment sales to hit $139 billion in 2026, on the way to $156 billion in 2027. WFE specifically is expected to grow ~9–10%this year to roughly $128–135 billion, with memory as the key incremental driver (Morgan Stanley recently revised WFE from +5% to +10%).

ASML is the clearest signal in the equipment space. Q4 2025 bookings of €13.2 billion came in at nearly double analyst expectations. EUV bookings alone hit €7.4 billion. The year-end EUV backlog sits at €25.5 billion, total backlog €38.8 billion. EUV systems are effectively sold out through 2027. For 2026, ASML guides €34–39 billion total revenue at 51–53% gross margins.

Applied Materials ran the same playbook: Q1 FY2026 revenue of $7.01 billion, gross margins at 49.1% — a 25-year high — and implicit guidance for 20%+ semiconductor equipment business growth in calendar 2026.

Hyperscaler Capex — $700 Billion and Accelerating

The four-hyperscaler aggregate capex number for 2026 has crystallized around $635–665 billion, up roughly 67–74%year-over-year. Amazon (~$200B), Alphabet (~$175–185B), Microsoft (~$145B), Meta (~$115–135B). Add Oracle and you're at $700B+.

About 75% of this spend is AI-related infrastructure — roughly $450–500 billion in AI-specific capex this year alone. Bank of America estimates these five companies will spend approximately 90% of their operating cash flow on capex in 2026. Amazon is projected to go free-cash-flow negative this year for the first time in years.

This is the demand foundation that makes the rest of the watchlist make sense. When hyperscalers are willing to burn through all of their operating cash flow on AI infrastructure, the supply chain — NVDA, MU, TSM, AVGO, MRVL, AMAT, LRCX, KLAC, ASML — has essentially locked-in demand for years. The constraint isn't demand. The constraint is supply.

NVIDIA — Blackwell Sold Out, Rubin Next

NVIDIA doesn't report until later this spring, but the signals are unambiguous. Blackwell B200 and GB200 NVL72 systems are effectively sold out through mid-2026 with a backlog estimated at 3.6 million units from cloud providers alone. Volume production of the Blackwell architecture ramped in February 2026. Roughly 35,000 NVL72 racks are expected to ship this year — at ~$3M per rack, that's ~$105 billion in NVDA rack revenue before discrete GPU sales.

The power reality: a single GB200 NVL72 rack draws up to 120 kW. At 35,000 racks that's 4.2 GW of new power demand from Blackwell alone. Power infrastructure is becoming the binding variable on deployment velocity — which is why Wolfspeed's 10kV SiC MOSFET (30% lower system power cost, 300% higher power density) and 300mm SiC wafer platform matter more than they get credit for.

TSMC — CoWoS Bottleneck, N2 Ramping, $52B Capex

TSMC reports Q1 2026 results on April 16, with guidance of $34.6–35.8 billion — up to 40% YoY growth. The company is investing $52–56 billion in capex this year, up 30% from 2025, with more than 10% going to advanced packaging.

CoWoS capacity is scaling from ~35,000 wafers per month in late 2024 to a projected 130,000 wafers/month by end of 2026 — nearly 4x. Even so, NVIDIA alone will consume over half of total global CoWoS demand (~685,000 units vs. projected supply of ~590,000). The gap is still ~20% shortof NVIDIA's needs. This packaging bottleneck is why advanced packaging — CoWoS, SoIC, InFO — belongs in every AI infrastructure investor's mental model.

N2 is ramping. Mass production kicked off in late 2025 and scales through 2026, with A16 (HPC-optimized with Super Power Rail) in volume by 2H 2026. This is the process node that will define the next generation of AI accelerators — NVIDIA's Rubin, custom XPUs at the hyperscalers.

Broadcom — Six XPU Customers, $100B AI Revenue Target for '27

Covered AVGO Q1 FY2026 in depth in the March 4 research note. The short version: AI revenue of $8.4 billion (+106% YoY), Q2 guidance of $22 billion ($1.44B above consensus), and a confirmed sixth XPU customer in OpenAI with 1 GW deployment targeted for 2027. Management gave line-of-sight to $100 billion in AI chip revenue in 2027from named customers with concrete deployment timelines. That's not a blue-sky projection — it's a bottoms-up construction from six named customers with multi-gigawatt targets.

Trade Policy — The Real Overhang in 2026

The April 2025 reciprocal tariff episode was a market shock that resolved quickly for semiconductors — chips and equipment were carved out. What matters more in 2026 is the policy architecture that emerged from it:

My read on the policy environment: the reshoring pressure is real and structural, but it plays out over three to five years, not quarters. The companies with U.S. manufacturing exposure — AMAT, LRCX, KLAC (equipment), Micron (Idaho and New York fabs), Intel (Ohio and Arizona) — have a long-duration structural tailwind from this. TSMC's Arizona N3 fab is the crown jewel of reshoring; if it ramps on schedule, it changes the geopolitical risk profile of the entire advanced logic supply chain.

Constraint → Intensity → Durability: Where We Stand

What to Watch Next

• ›TSMC Q1 2026 earnings (April 16): The most important semiconductor print of Q1. Guide is $34.6–35.8B. If they print above the top of that range and hold CoWoS pricing, the equipment names catch a bid immediately.
• ›ASML Q1 2026 earnings (April 16): Revenue guide €8.2–8.9B. The bookings number matters most — can they sustain the record €13.2B pace from Q4?
• ›LRCX Q3 FY2026 (late April): Lam guided 10–15% WFE growth for 2026. Watch for any update on China exposure post-BIS rule changes.
• ›Hyperscaler Q1 earnings (late April/early May): AMZN, MSFT, GOOG, META all report. Capex guidance updates will be the most watched data point in the sector.
• ›Sector-specific semiconductor tariff details: The Commerce Department Section 232 investigation into semiconductors has been signaled; the specific rate structure and exemption criteria will determine the actual supply chain impact.

Bottom Line

The Q1 2026 earnings data is uniformly strong. Micron at $23.9B with a $33.5B Q3 guide. Marvell at $2.22B with $11B in FY2027 sights. Broadcom $8.4B in AI chips and a credible path to $100B. ASML with €38.8B in backlog. SNPS with $11.3B in backlog. Trade policy is the real overhang, but the 2025 tariff episode showed that chips and equipment have enough strategic importance that they get carved out when it matters.

The numbers say the AI infrastructure capex supercycle is not plateauing. Micron can only meet 55–60% of DRAM demand. CoWoS is 20% short of NVIDIA's needs. HBM4 supply for 2026 is fully committed. That's the actual story. Everything else is noise.

— J