Marvell trades at 77.4 times earnings as SK hynix's first commercial memory runs on a startup's chip

SK hynix▼−0.81% — SK hynix, down 0.81 percent today, the world's second-largest DRAM maker, spent the Open Compute Project summit in San Jose this month showing CXL Hybrid Memory, its first named commercial product↗ on CXL, the interconnect that pools server memory. The controller inside came from XCENA, a startup founded in 2022 by Samsung and SK hynix veterans, not from Marvell▲+1.60% — Marvell, up 1.60 percent today ¹.

Marvell has priced that link for investors. It sizes the market for silicon connecting AI chips at $37 billion↗, its own estimate relayed through MarketBeat ², and LightCounting founder Vladimir Kozlov puts AI scale-up networking above $50 billion by 2030 ³.

Marvell papered the rack in one week

On October 12 it announced second-generation Structera X memory expanders↗ carrying up to 9 terabytes per controller ⁴. On October 13 came the Alaska E 3.2T retimer↗, 16 lanes of 224G signaling, and a 3nm Alaska A 1.6T cable DSP drawing 12.8 watts, both already sampling ³. On October 19 it co-founded an OCP security group↗ with Microsoft▼−0.34% — Microsoft, down 0.34 percent today and Oracle▲+2.82% — Oracle, up 2.82 percent today ⁵. At its October 6 investor day, management raised its fiscal 2028 revenue outlook to about $20 billion↗, against $9.45 billion of trailing revenue, and set a fiscal 2031 target of $70 billion to $90 billion ⁶.

Compute moved inside the memory

Structera X makes memory bigger; the host processor still does the thinking. XCENA's MX1 puts 3,072 RISC-V cores inside the controller, and at Hot Chips 2026 it measured up to 4.7 times a host CPU's throughput over CXL at 18.7 times the energy efficiency, on 40 watts ¹. Samsung starts MX1 mass production on its foundry lines by year-end, with revenue from 2027 ¹. Marvell's second-generation expander samples in the first quarter of 2027 ⁴. The rival chip reaches production before Marvell's answer takes its first sample.

The DRAM makers hedged: Samsung and SK hynix both supplied endorsement quotes for the Structera launch even as they backed the startup ⁴. And on October 14 two accounts posted that Marvell's OCP booth showed "significant levels of cheating": waveform averaging used for an eye measurement, a TDECQ reading flagged with a question mark, an RLM of 0.9 against a 0.95 spec ⁸⁹. The claims are unverified, and no Marvell response appears in the coverage gathered.

The bill arrives in 2028

A KeyBanc note relayed on X on October 19↗ has Marvell likely to win Microsoft's Maia, the in-house AI accelerator, over MediaTek and Broadcom▲+1.36% — Broadcom, up 1.36 percent today for the Maia 400 generation ¹⁰. Maia 300 enters production in the first half of 2027 at roughly 500,000 units; Maia 400 ramps in the first half of 2028 ¹⁰. A sector table posted the same day↗ puts Marvell at 40 times forward earnings, against 17 for Nvidia▲+0.32% — Nvidia, up 0.32 percent today and 5 for Micron▲+1.26% — Micron, up 1.26 percent today ¹¹. The shares fell 17.2% across the five sessions to October 16 and still closed at 77.4 times trailing earnings ⁶.

Three dates settle it: XCENA on Samsung's production lines by year-end ¹, Marvell's Structera sampling in the first quarter of 2027 ⁴, and the report against a $3.15 billion quarterly guide due around November 30 ⁶. Nothing in the Maia thesis collects revenue before 2028 ¹⁰.