Skip to content
MRVL
~7 min read · 1,653 words ·updated 2026-04-28 · confidence 63%

Industry Dynamics

Pluggable Optics vs. Co-Packaged Optics (CPO) Debate

Market Context: Why CPO?

Pluggable optics (e.g., 800G/1.6T PAM4 DSP modules in small form factors) have dominated hyperscaler deployments through 2025. However, Co-Packaged Optics (CPO) integrate switching silicon + optical I/O on a single package, offering:

  1. Power efficiency: 30–40% lower power per Tbps vs. pluggable (eliminates switch-to-optics PCB routing losses).
  2. Latency reduction: Integrated switch/optics shorten signal path.
  3. Form factor density: 102.4 Tbps switch + embedded optics in smaller footprint.

Tradeoff: CPO requires early design engagement (custom layouts), higher NRE, longer lead times (~3 years from design to volume).

Competitor Positioning

NVIDIA Spectrum-X / Quantum-X (announced GTC 2025, ✓):

  • Custom switch ASICs with integrated silicon photonics (SoIC / 3D hybrid bonding via TSMC).
  • Target deployment: 2027–2028 in hyperscale clusters.
  • Strategy: NVIDIA controls full stack (GPU + switch + optics), bypassing merchant vendors.

Broadcom Bailly CPO (2024–2025 development, ✓):

  • CPO platform with emphasis on ecosystem openness (partners: Lumentum, Coherent, Viavi for photonics modules).
  • Target deployment: 2026–2027 (earlier than NVIDIA).
  • Strategy: Leverages Broadcom’s switching dominance + partner photonics vendors.

Marvell Position (evolved 2025–2026, ✓):

  • Pluggable optics dominance (2025–2026): Ara, Ara X, Petra, Aquila PAM4 DSP platforms hold majority share in 800G/1.6T deployed modules.
  • CPO entry (December 2025 + April 2026): Celestial AI acquisition ($3.25B) + NVIDIA partnership ($2B) position Marvell to participate in CPO ecosystem for 2027–2028 deployments.
  • Hedged strategy: Maintain pluggable DSP share through 2026–2027, ramp CPO capability in parallel.

Market Implications

CPO adoption trajectory is uncertain (◐):

  • Bull case: CPO becomes standard by 2028 for all new hyperscale deployments; pluggable DSP market shrinks 50%.
  • Bear case: Pluggable optics remain dominant through 2029; CPO adds incremental capacity, does not replace.

Base case (Marvell guidance): CPO reaches 20–30% of new transceiver deployments by 2028–2029; pluggable optics remain 70%+ through 2028.

Confidence: ◐ (CPO ramp timing is key industry uncertainty for 2026–2027).

Linear Pluggable Optics (LPO) / Linear Receive Optics (LRO)

Technology Overview

Linear PAM4 (LPO/LRO) eliminates DSP and uses analog linear equalization to receive PAM4 signals directly:

  • LPO (linear-drive pluggable): Transmit-side simplification (laser driver + TIA instead of DSP).
  • LRO (linear receive optics): Receive-side simplification (receiver without DSP).
  • Power benefit: Sub-8W 800G modules (vs. 12–15W DSP-based modules).

Marvell’s Response

OFC 2025 (March 2025, ✓): Marvell demonstrated and announced 1.6T LPO chipset (Spica Gen2-T), a transmit-only version of its 5nm Spica DSP. Enabling sub-8W 800G and 1.6T modules with 40% power reduction vs. full-DSP equivalents.

Adoption trend: Industry consensus growing that LPO/LRO is required at 200G/lane (1.6T+) for sub-10W power envelopes (LightCounting, ✓).

Cannibalization Risk

Does LPO cannibalize DSP revenue? (⚠)

  • Short answer: Moderately. LPO/LRO use simpler transimpedance amplifiers (TIA) and laser drivers (Marvell Spica Gen2), not full DSPs.
  • Impact on Marvell: Spica Gen2-T (LPO transmitter) is higher-volume, lower-ASP vs. full Spica DSP. Estimated 10–15% revenue impact from pluggable optics if LPO adoption reaches 30% by 2027.
  • Offset: LPO enables sub-8W form factors, which expand addressable hyperscaler deployment (power-constrained clusters), partially offsetting ASP loss.

Confidence: ◐ (revenue impact depends on LPO ramp speed, which is uncertain).

3.2T Transition & Modulation Material Requirements

Performance Requirements at 3.2T

3.2T operating point: 8 lanes × 400 Gbaud PAM4 = 3.2 Tbps per port.

Challenges:

  1. Modulation bandwidth: 400 Gbaud requires modulator cutoff frequency >150 GHz (vs. ~30–50 GHz for 200G).
  2. Drive voltage: Higher bandwidth typically requires lower drive voltage (<1V) to meet power budgets.
  3. Linearity: PAM4 requires excellent modulation linearity; 400 Gbaud compounds error.

Modulator Materials Trade-offs

Traditional silicon Mach-Zehnder modulators (Si-MZM):

  • Bandwidth: ~50 GHz cutoff (insufficient for 400 Gbaud PAM4).
  • Marvell legacy platform (pre-2026).

Plasmonic-Organic Hybrid (POH) modulators:

  • Bandwidth: Up to 500 GHz cutoff; 0.07 V·mm drive voltage metric (excellent).
  • Marvell strategy via Polariton acquisition (April 2026, ✓): Polariton specializes in plasmonics-based modulation enabling 3.2T scaling.
  • Limitation: Not CMOS-compatible (requires manual micro-fabrication of metal contacts), complicating foundry adoption.

Silicon-Organic Hybrid (SOH) modulators (emerging, ✓):

  • Compromise: Silicon waveguides + organic electro-optic polymer core.
  • Bandwidth: 200+ GHz cutoff.
  • CMOS-compatible pathways (e.g., NLM Photonics SOH tapeout at GlobalFoundries, Tower Semiconductor).
  • Emerging as standard for 1.6T→3.2T scaling (industry consensus, 2025–2026).

Implications for Marvell

Polariton acquisition (April 2026) is strategic response (✓):

  • Secures POH-based modulation IP for 3.2T+ designs.
  • Enables Marvell to participate in next-generation DCI and coherent optical modules (2027–2029 timeframe).

Alternative risk: If industry standardizes on SOH rather than POH, Polariton’s POH focus could become niche (⚠). However, Polariton’s plasmonic approach offers superior power efficiency, likely preserving relevance.

Confidence: ✓ (acquisition announced); ◐ (modulation material standardization uncertain).

AI DSP Booking Trends & Deferred Revenue

Fiscal 2026 Bookings Momentum

Marvell Q3 FY2026 earnings (✓):

  • Record design wins in FY 2026 (specific number not disclosed, but described as “all-time record”).
  • Bookings continuing at record pace, implying strong DSP+DAD attach (digital/analog designs) for current customers and new wins.

Interpretation: DSP bookings in FY 2026 are tracking for significant revenue realization in FY 2027–2028, consistent with 18 design wins + 50+ socket pipeline ($75B lifetime value per Barclays).

Standard accounting: DSP and custom silicon revenue is recognized upon delivery (not upon booking). Large design wins and customer commitments create deferred revenue (customer payments received but products not yet shipped).

Marvell disclosure (FY 2025 10-K, ✓): Deferred revenue not specifically broken out by segment in public filings. However, management commentary in Q3 FY2026 earnings suggests strong backlog conversion (deferred → recognized revenue) in H2 FY2026.

Implied trend:

  • High deferred revenue balance entering FY 2026 (from FY 2025 design win ramp).
  • Rapid conversion in FY 2026 (ships as products ramp).
  • Further backlog accumulation entering FY 2027 (new customer wins, 2nd major XPU customer ramping).

Confidence: ◐ (deferred revenue not separately disclosed; inference from earnings commentary).

HBM Supply Constraints & AI Silicon Ramp Timing

2025–2026 HBM Supply Situation

Critical bottleneck (⚠):

  • Micron, SK Hynix, Samsung HBM3E capacity: Fully allocated through end of 2026 (✓).
  • 2026 HBM demand growth: ~70% YoY (driven by NVIDIA Blackwell ramp, next-gen accelerators).
  • Manufacturing complexity: Producing HBM requires ~300% more wafer capacity per bit vs. standard DDR5 (⚠).

Supply shock effects (Q2 2025 early warning, worsened through 2026):

  • NVIDIA restricts next-gen accelerator shipments (e.g., RTX 50-series cut 30–40% due to GDDR7 availability).
  • Custom ASIC ramps delayed: Marvell’s XPU customers require integrated HBM; if HBM is unavailable, XPU ramps slip.
  • Hyperscaler rationing: Large customers (Google, Microsoft, Meta) have priority access; smaller customers face delays.

Marvell Exposure

Direct impact on custom XPU ramp:

  • Marvell’s 18 design wins include HBM-intensive designs (e.g., Microsoft’s Cobalt for inference, Google’s custom TPU-class).
  • If HBM supply remains constrained through H2 2026, XPU revenue recognition could shift 2–3 quarters into FY 2027–2028 (⚠).
  • Management guidance (FY 2027 custom XPU ≥20% growth) assumes HBM supply improves by mid-2026; if not, guidance could miss (◐).

Indirect impact on optical DSP:

  • Optical DSP is NOT memory-intensive; minimal HBM exposure.
  • However, if hyperscaler capex decelerates due to HBM-driven delays in AI ramps, overall network fabric capex could soften (⚠).

HBM4 & Relief Timeline

HBM4 expected 2026 ramp, reaching mass production mid-2026–2027:

  • Bandwidth: 2 TB/s (double HBM3E).
  • Manufacturing: ~20% less complex than HBM3E, easing capacity constraints.

Implication: HBM supply relief likely H2 2026 → 2027, enabling XPU ramp acceleration in FY 2027–2028.

Confidence: ✓ (HBM supply situation widely documented); ◐ (HBM4 mass production timing uncertain).

Industry TAM/Competitive Positioning Summary

MarketTAM 2028Marvell PositionKey Risk
PAM4 DSP (800G/1.6T pluggable)~$8–10BMarket leader (Ara, Spica family)LPO cannibalization, CPO competition
Custom AI XPU$40.8B18 wins, $1.5B FY2026 revenueHBM supply, hyperscaler ROI scrutiny
Switching ASIC (Teralynx)~$8–12B10% share, niche strengthBroadcom dominance, CPO integration threat
CPO ecosystem$2–3B (nascent)Celestial AI partner, NVIDIA allianceNVIDIA/Broadcom lead; Marvell entry timing risk
Scale-up interconnect (Celestial AI)$6B (by 2030)Majority via Celestial AI dealNiche market, execution risk

Key Takeaways

  1. CPO adoption is the macro industry pivot for 2026–2028. Marvell is not first-mover (NVIDIA/Broadcom ahead) but positioned to participate via partnerships (Celestial AI, NVIDIA $2B investment). Pluggable DSP remains dominant through 2027; CPO becomes material in 2028+.

  2. Linear optics (LPO/LRO) are emerging alternative to full DSP, offering power efficiency but risking 10–15% DSP revenue cannibalization. Marvell’s Spica Gen2-T (transmit-only LPO) is responsive but plays second fiddle to full DSPs through 2026.

  3. 3.2T modulation scaling hinges on materials innovation: Marvell’s Polariton acquisition (POH modulators) is strategic but risky—SOH may become industry standard. Mitigates but doesn’t eliminate risk.

  4. HBM supply is the critical near-term bottleneck limiting AI accelerator ramps. Marvell’s XPU revenue could slip 2–3Q into FY 2027–2028 if HBM constraints persist. Manage expectations vs. FY 2027 20%+ growth guidance.

  5. DSP booking momentum is strong, with record design wins and backlog conversion in FY 2026. Deferred revenue is rising, supporting FY 2027 optical DSP revenue growth if CPO competition does not intensify unexpectedly.

  6. Competitive positioning is mixed: Marvell dominates PAM4 DSP (pluggable, 2025–2027), but trails NVIDIA/Broadcom in CPO. Switching (Teralynx) remains niche vs. Broadcom Tomahawk. Custom XPU is highest-potential growth segment but faces HBM supply and hyperscaler ROI headwinds.

Sources

  • Marvell Q3 FY2026 earnings call; Investor Day 2024
  • OFC 2025 announcements (Marvell, Broadcom, NVIDIA)
  • LightCounting optical interconnect and DSP analysis (PAM4, LPO, CPO trends)
  • Celestial AI and Polariton acquisition announcements (December 2025, April 2026)
  • NVIDIA GTC 2025 (Spectrum-X, Quantum-X announcements)
  • TrendForce HBM supply chain analysis
  • Polariton Technologies company background (ETH Zurich, plasmonic photonics)

Cross-references