Marvell’s mixed print: record Q2, softer Q3 guide — what it says about AI hardware demand

Marvell Technology (MRVL) delivered a record quarter but paired it with a cautious near-term outlook, sending the shares lower after hours. For the July quarter (Q2 FY26), revenue came in at $2.006B and non-GAAP EPS $0.67, both essentially in line to slightly ahead of guidance, with data center now 74% of sales (company release; see the investor deck for the mix and margin detail, including a 59.4% non-GAAP gross margin and data-center revenue of $1.491B, +69% YoY (PDF)). Despite the beat, the stock fell ~7–12% after hours as investors focused on Q3 guidance that was a touch light versus consensus (Reuters wrap; WSJ live market note here).

The numbers that matter

• Top line: $2.006B, +58% YoY and +6% QoQ (deck page 8–9).
• Profitability: Non-GAAP EPS $0.67; non-GAAP op margin 34.8%; cash from ops $462M (release).
• Q3 guide: revenue $2.06B ±5%, non-GAAP EPS $0.74 ±$0.05; non-GAAP GM 59.5–60.0% (outlook slide). Coverage highlighted that the revenue midpoint sat below Street models, which drove the initial selloff (Reuters after-hours).

What’s working: AI interconnects + custom silicon

The AI build-out continues to be Marvell’s flywheel. In Q2 the company began volume shipments of its next-gen 200G-per-lane PAM4 DSPs aimed at 1.6T optical modules, and ramped 51.2 Tbps switches — core plumbing for training clusters and AI-heavy clouds (deck, data-center highlights). Management also won additional hyperscaler custom sockets, adding to 18 prior wins, with “lifetime revenue forecasted at multiple billions of dollars.” That custom roadmap is why investors increasingly put Marvell in the “AI ASIC” bucket alongside Broadcom — a structural tailwind as hyperscalers seek lower-cost, power-efficient silicon for specific workloads (see recent market framing via MarketWatch on custom AI chips here).

Outside AI, enterprise networking and carrier infrastructure kept recovering as inventories normalized. Enterprise networking rose 28% YoY / 9% QoQ, while carrier was +71% YoY / −6% QoQ, with management expecting the combined pair to grow ~30% QoQ in Q3 (end-market slides). Marvell also closed the $2.5B sale of its Automotive Ethernet unit to Infineon, simplifying the portfolio ahead of the heavier AI cycle (announcement recap in deck).

Why the guide felt “meh”

Two points weighed on sentiment:

1. Data-center mix: Management said Q3 data-center revenue should be “flat sequentially” — interconnect strength (optical DSPs, switches) offset by lower custom revenue as program timing ebbs and flows quarter to quarter (data-center slide note). For investors primed for straight-line acceleration, “flat” reads conservative.
2. Consensus vs. cadence: The formal outlook — $2.06B ±5% — sits just shy of typical sell-side models, and that gap is all the market needed after a strong year-to-date run in AI hardware names (WSJ and Reuters both flagged this divergence: WSJ, Reuters).

The bigger picture: can AI revenue compound from here?

Structurally, the AI plumbing Marvell sells — optical DSPs, switch ASICs, storage, security HSMs, and custom accelerators — scales with GPU count and cluster size. The company’s pipeline now spans scores of custom designs across 10+ large customers, and it is rolling out 2nm building blocks (e.g., custom SRAM and die-to-die IP) to stay on the leading edge (deck: recent 2nm/IP announcements). If hyperscalers keep tilting spend toward application-specific silicon and higher-throughput optics, Marvell’s share of wallet can rise even if GPU unit growth normalizes.

Key watch-items:

• Optics cycle: Adoption of 1.6T modules (200G/lane) into 2026 would extend interconnect growth — the PAM4 DSP ramp starting this quarter is an early tell.
• Custom cadence: Program phasing will stay lumpy; investors will want clarity on 2026 lifetime value as new sockets move from tape-out to revenue.
• Gross margin: AI hardware is margin-dilutive vs. software, but mix + procurement can keep non-GAAP GM ~59–60% if interconnect volume scales (per Q3 guide in the outlook).
• Multi-market recovery: Sustained enterprise/carrier normalization hedges any pause in custom AI shipments.

Bottom line

Marvell’s Q2 shows AI demand is real in both custom and interconnect, but the Q3 outlook reminds us this build-out isn’t linear. For long-only investors, the thesis hinges on 2026 run-rate power: if 1.6T optics and a wider custom base hit production, today’s wobble looks like noise in a multi-year uptrend. If program slips stack up or hyperscaler budgets rotate, the multiple can compress quickly.

Not investment advice.