🕸️

MRVL

💡
Thesis. The credible number two in hyperscaler custom silicon — data centre now 79% of revenue growing 46%, the outlook raised twice, and Google taking an equity warrant tied to purchases. The threat is that co-packaged optics undermines the DSP franchise that is currently its fastest-growing line.
Reviewed 11 September 2026, after Q2 FY2027 results. Next review after Q3 earnings on 2 December 2026 — with an Investor Day on 6 October 2026 before that.

What it does

Four businesses, and which one you think matters determines how you value the company.
Custom AI silicon (XPUs) — bespoke accelerators designed with hyperscalers, the same business Broadcom is in. Optical and coherent DSPs — the Inphi franchise, the chips that process signals inside optical transceivers. Ethernet switching and NICs. And legacy storage controllers.
CEO is Matt Murphy.
A DSP is the chip inside a transceiver that cleans up and encodes the signal. Marvell doesn't make the transceiver — it makes the part that makes the transceiver work, which is why it shows up as a supplier to companies it otherwise looks nothing like.

Bull case

  • Q2 FY2027, reported 27 August: record revenue $2.739B, up 37% and 13% sequentially, beating consensus. Non-GAAP EPS $0.94
  • Data centre is now 79% of revenue, growing 46% year over year. Operating margin improved to 16.8% from 14.5%
  • The outlook was raised for a second consecutive quarter. FY27 data centre growth now around 60%, up from 50%. FY28 guided above 60%
  • Interconnect is leading, not lagging. Optical interconnect guided to grow more than 70% in FY27, which means the DSP franchise is working while the XPU ramp is still ahead
  • Google took equity. An expanded custom silicon partnership includes a warrant for Google to acquire up to 7% of Marvell, tied to revenue milestones — the same structure Amazon used with AAOI, and one of the strongest forward-demand signals a supplier can get
  • It bought the bottleneck-removers. Celestial AI (photonic interconnect for moving data between compute and memory) and XConn (CXL and PCIe switching), both completed February 2026
  • Guidance is well ahead. Q3 guided to $3.15B, about 4% above the $3.03B consensus

Bear case

  • Co-packaged optics is the structural threat. TSMC puts CPO mass production in the second half of 2026. Over five to seven years it disintermediates the pluggable DSP business. Marvell is hedging via Celestial AI, but hedging a franchise is not extending one
  • Programme concentration. Losing a single XPU socket is material — and the Google warrant deepens the dependency at the same time as it secures it
  • Margins are thin by comparison. 16.8% operating margin against Broadcom's 67.9%
  • The XPU ramp is still a promise. Custom silicon growth above 20% is slower than the interconnect line, and the second-half ramp has to land
  • The market is hard to please. Shares fell on a beat-and-raise after a 161% year-to-date run, with options having priced a roughly 14% move

Major customers

  • Google — the anchor. Expanded custom silicon partnership with a warrant for up to 7% of Marvell shares tied to revenue milestones
  • Amazon — Trainium-related custom silicon and optical DSPs, historically the largest custom customer
  • Microsoft — custom accelerator programmes
  • Optical module makers — Coherent, Lumentum and Applied Optoelectronics all buy Marvell's PAM4 and coherent DSPs
  • Switch and NIC OEMs — Ethernet silicon
🔗
Read-through. Marvell and
🕸️
AVGO are the merchant custom-AI-silicon pair. Broadcom's September print, with AI revenue up 221%, is the bellwether for Marvell's own second-half XPU ramp — read one to anticipate the other.
Marvell's DSPs sit inside the transceivers that
🔆
COHR,
🔆
LITE and
🔆
AAOI build, so it is a supplier to that category rather than a competitor.
The shared exposure worth understanding: co-packaged optics would disintermediate Marvell's DSP and the pluggable transceiver business simultaneously.
🏛️
TSM naming mass production for the second half of 2026 is therefore a single event that reads across four pages at once.
🕸️
CRDO is the other name Marvell could absorb, by integrating the SerDes function into its own silicon.

What would change the view

  • Investor Day on 6 October — the AI infrastructure strategy, and any quantification of the Google agreement
  • The custom silicon second-half ramp landing as guided
  • Whether optical interconnect holds above 70% growth
  • Operating margin progression from 16.8%
  • CPO adoption pace — the structural threat to the DSP franchise
  • Any second hyperscaler warrant or equivalent lock-in

Update log

  • 19 Sep 2026 — Converted to the standard page template. Note: this page has no strengthen/weaken probabilities recorded, unlike every other entry, so its Constructive verdict is not currently derivable. Needs setting.
  • 7 Sep 2026 — Full review after Q2 FY2027 results: record revenue $2.739B (+37%), data centre 79% of revenue growing 46%, FY27 and FY28 outlook raised a second time. The Google 7% equity warrant and the Celestial AI and XConn acquisitions added; co-packaged optics flagged as the dominant structural risk.
Research and education only — not investment advice.