Marvell just made one of its biggest AI bets yet: a $3.25 billion deal to acquire Celestial AI, the photonics startup behind the Photonic Fabric optical I/O platform. Announced on December 2, 2025, the move signals that the next leap in AI data centers won’t come only from bigger chips—but from faster, lower‑power links that bind them together. Marvell press release • Reuters
Why this deal matters
AI clusters are scaling from single racks to multi‑rack “pods” that connect hundreds of accelerators with any‑to‑any, memory‑semantic links. Copper is running out of road on reach, bandwidth, and power—the reasons hyperscalers are eyeing all‑optical connections for the short‑reach, intra‑rack links that keep models fed and synchronized. Marvell says Celestial’s Photonic Fabric was “purpose‑built” for this inflection, delivering an optical scale‑up fabric that can be co‑packaged with custom XPUs and scale‑up switches. Marvell press release
- Celestial’s first‑gen Photonic Fabric chiplet provides 16 Tbps of bandwidth in a single chiplet—designed to be placed next to XPUs or switches for dramatic bandwidth density.
- Marvell says the optical fabric offers more than 2× the power efficiency of comparable copper interconnects, with nanosecond‑class latency and thermal stability suitable for multi‑kilowatt packages. <<stat label="Power efficiency vs. copper" value=">2×" source="Marvell press release, Dec 2, 2025">> Marvell press release
The deal at a glance
Key terms and milestones
| Item | Detail |
|---|---|
| Purchase price | ~$3.25B upfront ($1B cash + ~27.2M MRVL shares, valued ~$2.25B on VWAP basis) |
| Earn‑out | Up to ~27.2M additional MRVL shares (valued up to ~$2.25B) tied to revenue milestones through FY2029 (one‑third at $500M cumulative; full at >$2.0B) |
| Close timing | Expected in Q1 2026, subject to approvals |
| Revenue ramp (guidance) | “Meaningful” in 2H FY2028; ~$500M annualized run‑rate in Q4 FY2028; ~$1B run‑rate by Q4 FY2029 |
| AWS tie‑in | Marvell issued Amazon a warrant for up to 1,045,171 shares, vesting with photonic‑product purchases through Dec 31, 2030; exercise price ~$87.00, expiring Dec 2, 2031 |
Sources: Marvell press release • StreetInsider/SEC 8‑K extract • Reuters
Where it fits in Marvell’s AI strategy
This isn’t Marvell’s first big optics swing. In 2021 it absorbed Inphi, a leader in high‑speed electro‑optics used for data‑center interconnects and Ethernet optics—fortifying Marvell’s “scale‑out” (rack‑to‑rack, data‑center‑to‑data‑center) portfolio. Celestial fills the “scale‑up” hole: the ultra‑short‑reach, ultra‑low‑latency links that tie accelerators together inside systems and across racks—where copper struggles and co‑packaged optics excel. Marvell–Inphi
The timing lines up with industry standards too. The UALink Consortium—formed by heavyweights including AMD, Intel, Google, Microsoft, and others—published its 200G 1.0 spec in April 2025, enabling up to 1,024 accelerators per AI pod with a low‑latency, memory‑semantic fabric. Marvell is building UALink IP and positioning Celestial’s optical scale‑up to ride that ecosystem. UALink spec • Marvell UALink announcement • Reuters background
Signals from hyperscalers
Two clues suggest real demand: Amazon’s public statement and its warrant. AWS’s Dave Brown said optical interconnects will be “an important role in the future of AI infrastructure,” and the warrant vests only if Amazon buys photonic‑fabric products by 2030—effectively a demand‑signal contract. Marvell also told investors it expects a $10B photonics‑related opportunity and forecasts data‑center revenue up ~25% next fiscal year, with custom chip sales up ~20%. Marvell press release • Reuters
What changes for AI builders
- Memory‑semantic scale‑up at rack scale: Optical links with nanosecond‑class latency and 16 Tbps per chiplet could make disaggregated HBM and pooled memory appliances practical—reducing model sharding pain and tail‑latency spikes. Marvell press release
- Open fabric options: As UALink hardware arrives, a vendor‑neutral alternative to proprietary NVLink/NVSwitch becomes real. Marvell aims to be the “optical engine room” for that ecosystem. UALink spec
- Power and density: If optical scale‑up truly delivers >2× power efficiency versus copper at these reaches, operators can pack more accelerators per rack within the same power envelope. Marvell press release
Competitive landscape and caveats
- Nvidia: NVLink/NVSwitch is the incumbent scale‑up fabric. UALink aims to open that market, but Nvidia’s tight hardware‑software stack remains a moat. Reuters background
- Broadcom and others: Competing silicon photonics, Ethernet switching, and CPO efforts are well funded. Marvell’s advantage is breadth—from Ethernet optics (via Inphi) to custom silicon and now optical scale‑up. Marvell–Inphi
- Execution risk: Co‑packaging photonics with multi‑kilowatt XPUs at volume is non‑trivial. Supply chain, yields, and standard maturity can slip schedules. Notably, Marvell guides first meaningful Celestial revenue to 2H FY2028—so this is a medium‑term bet. Marvell press release • Reuters
Bottom line
Marvell isn’t just buying a startup—it’s buying a seat at the table where the next AI data‑center fabric is being set. If Celestial’s Photonic Fabric ships on the timeline Marvell projects and UALink takes hold, optical scale‑up could become a standard building block of AI pods by 2027–2028. That would reduce power per bit, expand memory footprints available to each accelerator, and—most importantly—unlock higher utilization for real‑world AI workloads. For teams planning 2026–2029 build‑outs, this deal is a green light to prototype optical scale‑up now. Reuters • Marvell press release
Sources
- Marvell Technology, “Marvell to Acquire Celestial AI…” (Dec 2, 2025) — definitive terms, product roadmap, AWS quote, revenue milestones. Link
- Reuters, “Marvell to buy chip startup Celestial AI for $3.25 billion…” (Dec 2, 2025) — deal size, composition, close timing, growth outlook, AWS warrant context. Link
- SEC Form 8‑K (Dec 2, 2025) — Amazon warrant specifics (share count, vesting, strike, expiry). Link
- Business Wire, “UALink Consortium Releases the 200G 1.0 Specification” (Apr 8, 2025) — 200G/lane, up to 1,024 accelerators per pod. Link
- Marvell Newsroom, “Marvell Expands Custom Compute Platform with UALink…” (Jun 11, 2025) — Marvell’s UALink IP and positioning. Link
- Marvell Newsroom, “Marvell Completes Acquisition of Inphi” (Apr 20, 2021) — optics portfolio background. Link
- Reuters, “Marvell shares jump as chipmaker bolsters AI ambitions with Celestial deal” (Dec 3, 2025) — market reaction, data‑center growth commentary. Link
