Celestial AI

1.1 Identity, Headquarters, and Stage

Celestial AI, Inc. was incorporated as a Delaware corporation and headquartered at 2962 Bunker Hill Lane, Suite 200, Santa Clara, CA 95054. Additional offices were maintained in Irvine, CA; Toronto, Ontario; and Hyderabad, India. As of the report date the company is a wholly-owned subsidiary of Marvell Technology, Inc., following the completion of Marvell's acquisition on February 2, 2026. The company was founded in 2020 by David (Dave) Lazovsky and Preet Virk, both with backgrounds in semiconductors and photonics. A third co-founder, Michelle Tomasko, is also cited in funding disclosures. The company positioned itself as the creator of Photonic Fabric, described as an optical scale-up interconnect platform that replaces copper-based electrical interconnects inside AI accelerator systems and data centers. Celestial AI's single product line is this Photonic Fabric platform, targeting hyperscalers, AI chip companies, and data-center operators. At the time of its last independent funding round (March 2025), Celestial AI had approximately 141–152 employees and had completed four TSMC 5/4 nm tape-outs of its electro-optical integrated circuits (EICs). Its business model combined chip licensing and custom silicon partnerships with Tier-1 hyperscalers and semiconductor companies. [CO001, CO002, CO003, CO003, CO003]

1.2 Founding Team and Leadership

David (Dave) Lazovsky serves as Co-Founder and CEO. Before Celestial AI, Lazovsky founded Intermolecular, Inc. in 2004 and led it from inception through its 2011 IPO, serving as President, CEO, and board member until October 2014. He subsequently served as Executive Chairman of POET Technologies and as board member of ESS, Inc., giving him deep experience commercializing complex semiconductor technologies at scale. Preet Virk is Co-Founder and COO. Phil Winterbottom serves as CTO; he began his career in the Computer Research group at Bell Labs and co-founded two subsequent startups — Entrisphere (acquired by Ericsson) and Gainspeed (acquired by Nokia). The CTO's architecture work underpins the Photonic Fabric's core design, including the electro-absorption modulator (EAM) approach and analog-only SerDes. Kelyn Brannon was appointed CFO in June 2024, adding financial leadership in anticipation of the company's commercialization phase. The executive team also includes Narayan Kaniyur (SVP Engineering), Steven Metz (VP Sales), Ankur Aggarwal (VP Advanced Packaging & Supply Chain), Subal Sahni (VP Photonics Engineering), and Victoria Sevilla (VP People). The Board of Directors is chaired by Lip-Bu Tan (Walden International founding managing partner; former Cadence Design Systems CEO; joined January 2025), who also invested personally in the Series C1. Other board members include Diane Bryant (independent director; appointed April 2024), Bethany Mayer (cybersecurity background), Peter Tague (investor representative), Alonso Galvan (IAG Capital Partners co-founder), and Isaac Sigron (Koch Disruptive Technologies managing partner). The CEO sits on the board, creating a single point of key-person concentration. [CO003, CO003, CO004, CO005, CO003, CO006]

1.3 Funding History and Capital Structure

Celestial AI raised approximately $515 million in venture funding across five disclosed rounds before the Marvell acquisition. The earliest documented round was a seed raise of approximately $2.87 million in January 2021. In February 2022, the company completed a $56 million Series A led by imec.xpand and including Koch Disruptive Technologies, M Ventures, The Engine, Tyche Partners, and Xora Innovation (Temasek subsidiary). A $100 million Series B closed in June 2023, led by IAG Capital Partners, Koch Disruptive Technologies, and Xora Innovation, with Samsung Catalyst Fund, Smart Global Holdings, Porsche Automobil Holding SE, The Engine, imec.xpand, M Ventures, and Tyche Partners also participating. The Series B valued the company at approximately $500 million. In March 2024, a $175 million Series C closed, led by U.S. Innovative Technology Fund (USIT), at a valuation of approximately $1.22 billion. AMD Ventures joined as a strategic new investor alongside existing backers. In October 2024, Celestial AI acquired the Rockley Photonics patent portfolio to strengthen its Photonic Fabric IP. The Series C1 — $250 million (some reports cite $255 million) closed in March 2025 — was led by Fidelity Management & Research and brought in BlackRock, Maverick Capital (Maverick Silicon), Tiger Global Management, and Lip-Bu Tan as new investors. TSMC's VentureTech Alliance and Samsung Catalyst Fund also participated. This round valued Celestial AI at $2.5 billion, more than double the Series C valuation. On December 2, 2025, Marvell Technology announced a definitive agreement to acquire Celestial AI at $3.25 billion upfront ($1 billion cash plus approximately 27.2 million Marvell shares), with earn-outs of up to approximately $2.25 billion in additional equity if Celestial AI meets cumulative revenue milestones. The acquisition was completed on February 2, 2026. [CO003, CO003, CO008, CO009, CO009, CO009]

1.4 Cover Metrics and Scale

The headline metrics available from public sources are: $2.5 billion last-round independent valuation (March 2025 Series C1); $3.25 billion upfront Marvell acquisition price (announced December 2, 2025; closed February 2, 2026); total venture capital raised of $515 million; and employee headcount of approximately 141–152 at time of acquisition. Revenue and ARR are private. GetLatka reported $19.5 million revenue for Celestial AI prior to acquisition, but this figure is unverified by primary sources and should be treated as an estimate. Gross margins, burn rate, and cash runway were not disclosed. Customer count is undisclosed; the company described "multiple Tier-1 hyperscaler engagements" without naming customers publicly. Named customer deployments, production contracts, and NRR are not publicly available. Marvell's acquisition announcement included forward-looking revenue targets: $500 million annualized run-rate by Q4 FY2028 (ending January 2028) and $1 billion annualized by Q4 FY2029, suggesting Celestial AI had minimal recognized revenue at closing. These targets are company-claimed forecasts. [CO012, CO003, CO012, CO013, CO009]

1.5 Milestones Chronology

Celestial AI's milestone trajectory spans from founding through its 2026 acquisition by Marvell. The company progressed from seed-stage concept in 2020 through successive fundraises, IP acquisition, leadership additions, and product validation — culminating in a major semiconductor industry exit in early 2026. Notable adverse events include the Rockley Photonics bankruptcy (which created the IP acquisition opportunity) and the acquisition's regulatory process, which subjected the deal to antitrust review before closing. The Global Semiconductor Alliance recognized Celestial AI with its 2024 Start-Up to Watch Award, reflecting the company's technical credibility in the semiconductor ecosystem. In August 2025, CTO Phil Winterbottom presented at IEEE Hot Chips 2025 at Stanford — the first major public technical disclosure of Celestial AI's EAM-based photonic architecture, measured latencies, and OMIB packaging approach. This disclosure significantly raised the company's visibility among hyperscaler engineers and accelerated the Marvell acquisition timeline that closed just months later. [CO014, CO014, CO015]

1.6 Exhibits

2.1 Market Boundary and Included Spend

The relevant market for Celestial AI's Photonic Fabric is the AI data center scale-up optical interconnect market — specifically the segment addressing chip-to-chip, chip-to-memory, and package-to-package optical links within and across AI accelerator systems. This differs from the broader optical interconnect market in three important ways: (1) it targets the short-reach intra-system domain rather than long-haul or metro data center interconnect; (2) it requires co-packaged or interposer-based photonic integration directly with accelerator chips (XPUs), not pluggable transceiver modules; and (3) it competes directly with proprietary copper-based scale-up fabrics such as NVIDIA's NVLink and the UALink standard backed by AMD and Broadcom. Included spend: silicon photonics interposers, co-packaged optics (CPO), optical chiplets, and scale-up optical interconnect modules designed for XPU-to-XPU and XPU-to-memory connectivity. Excluded spend: traditional pluggable transceivers (800G/1.6T), scale-out Ethernet/InfiniBand switching, data center interconnect (DCI) between facilities, copper active/passive electrical cables for intra-rack short-reach links, and coherent DCI optical modules. Adjacent markets include the broader silicon photonics market ($2.16B in 2024, growing to $9.65B by 2030 per MarketsandMarkets), the full data center interconnect market ($15.38B in 2025 per MarketsandMarkets and Fortune Business Insights), and the AI cluster optics segment as tracked by LightCounting ($5B in 2024, doubling to $10B+ by 2026). Status-quo substitutes include NVIDIA's NVLink copper-based scale-up fabric and AMD/Broadcom's UALink copper solutions — which are the dominant architectures Celestial AI/Marvell is positioned to replace. [CM001, CM002, CM003, CM004, CM001]

2.2 Market Sizing — Multiple Lenses

No single analyst report cleanly isolates the Celestial AI SAM (scale-up optical chiplets for AI XPUs). The most useful sizing lenses are: **Lens 1 — Broadest TAM (all optical interconnect):** Fortune Business Insights values the global optical interconnect market at $15.38B in 2025, growing to $43.14B by 2034 at a 12.2% CAGR. This includes all reach classes, technologies, and applications and significantly overstates the addressable market for Celestial AI. **Lens 2 — Data Center Interconnect (DCI) market:** MarketsandMarkets sizes the DCI market at $15.38B in 2025 growing to $25.89B in 2030 at an 11% CAGR. This includes both optical and packet-switching DCI solutions; it better captures data center optical spend but still over-includes non-AI applications and cross-facility links. **Lens 3 — Silicon Photonics (most focused technology lens):** MarketsandMarkets sizes the silicon photonics market at $2.16B in 2024 growing to $9.65B by 2030 at a 29.5% CAGR. Data centers and HPC are the highest-CAGR end-user segment. This lens best captures the technology layer Celestial AI operates in. **Lens 4 — AI Cluster Optics (most commercially relevant):** LightCounting's January 2025 Optics for AI Clusters report sizes the AI cluster optics market at $5B in 2024, growing to $10B+ by 2026 at roughly 40%+ near-term CAGR. This includes compute-node transceivers as well as scale-up network optics, and moderates to double-digit growth in 2028–2030 as LPO and CPO ramp. **Lens 5 — Co-Packaged Optics (most direct product analog):** Analyst forecasts for the CPO market range widely ($95M–$2B in 2025 for CPO-specific), with 30–40% CAGR to $575M–$4.67B by 2030. CPO penetration in AI data centers was only ~0.5% in 2026 per AI2Work/Mordor Intelligence, with forecasted reach of 35% by 2030. The most credible SAM estimate for Celestial AI's scale-up optical chiplet product is approximately $2–5B by 2028–2029, consistent with Marvell's revenue target of $500M ARR by Q4 FY2028 if Celestial AI captures 10–25% of this emerging sub-segment. This is an evidence-constrained estimate with high uncertainty given the pre-revenue stage of the technology. [CM002, CM003, CM004, CM005, CM006, CM007]

2.3 Buyer and Segment Segmentation

There are three primary buyer archetypes for scale-up optical interconnect technology: **Segment 1 — Hyperscalers:** AWS, Google, Microsoft, Meta, Alibaba, Oracle, and other large cloud operators are the ultimate end-buyers of AI infrastructure incorporating optical interconnects. They collectively spent $315–320B on data center capex in 2025. Budget ownership sits with infrastructure engineering organizations (VP/SVP level) under the CTO; procurement is typically multi-year. Adoption trigger: scaling AI clusters beyond the effective reach of copper (~1m at 400G), and achieving AI cluster sizes of tens of thousands to millions of XPUs. **Segment 2 — AI Chip Companies / XPU Designers:** NVIDIA, AMD, Intel, custom ASIC design houses (including Marvell's own custom XPU customers). These customers purchase or license optical interconnect chiplets to integrate directly into their XPU silicon packages. Budget ownership is within chip architecture and packaging engineering teams. Adoption trigger: exceeding electrical I/O bandwidth density limits (~3 Tbps/mm² for copper vs. ~10 Tbps/mm for silicon photonics) and reducing per-bit power to enable higher TDP packages. **Segment 3 — System Integrators and ODMs:** Companies such as Super Micro, Quanta, Foxconn, and others building AI server systems for hyperscaler deployment. They design the boards and packages incorporating optical chiplets. This is a derivative demand segment dependent on Segment 1 and 2 adoption decisions. The buyer journey is long and technically complex. A hyperscaler typically requires 12–36 months of joint hardware qualification before production deployment. The chiplet integration model (co-packaging with XPUs) requires early engagement at the XPU design stage, 24–36 months before volume production — meaning sales cycles began for products not yet shipping. [CM009, CM010, CM011, CM001, CM002]

2.4 Growth Drivers and Adoption Constraints

**Primary growth drivers:** (1) AI cluster scaling: GPU/XPU cluster sizes are expanding from hundreds to millions of units; each node requires hundreds of gigabits per second of bandwidth to adjacent nodes. Jensen Huang noted at GTC 2025 that each GPU requires six pluggable transceivers consuming 30W each; scaling to a 1M GPU cluster creates a 180MW transceiver power problem that only optical co-packaging can solve. (2) Copper physical limits: at 400G SerDes speeds, passive copper collapses to under 1m effective reach; at 1.6T targets, copper is simply impractical. (3) Power efficiency: electrical interconnects were consuming ~30% of total cluster power at major hyperscalers as of early 2025; CPO reduces per-bit energy ~3x vs. pluggable optics (~5 pJ/bit vs. ~15 pJ/bit). (4) Data center capex growth: global data center capex grew 59% YoY in Q3 2025, creating enormous infrastructure investment that will include optical upgrade cycles. (5) NVLink/UALink transition: NVIDIA's NVLink Fusion partnership with Marvell signals the industry is moving to optical scale-up fabrics; UALink is similarly architected with optical interfaces in mind. **Adoption constraints:** (1) Capital intensity and manufacturing complexity: TSMC CoWoS advanced packaging is expensive, with limited capacity and long lead times — creating supply-side bottlenecks regardless of demand. (2) Integration complexity: chip designs must be redesigned to accommodate optical I/O; hyperscalers and chip companies require 2–3 year design cycles. (3) Technology maturity: CPO penetration was ~0.5% in 2026; high-volume deployment is not expected until 2028–2030 (LightCounting consensus). (4) Switching cost and ecosystem risk: hyperscalers operating copper NVLink/NVSwitch systems would need to redesign clusters to adopt optical scale-up; the transition is technically risky and capital intensive. (5) Competing proprietary solutions: NVIDIA owns NVLink and has strong hyperscaler lock-in; transitioning to open optical standards requires coordinated ecosystem effort. [CM001, CM003, CM004, CM001, CM002, CM003]

2.5 Sizing Diligence Gaps and Contradictory Estimates

Market size estimates for Celestial AI's specific segment — scale-up optical chiplets (silicon photonics interposers and CPO for XPU packages) — are highly variable and not directly comparable. The CPO market alone has estimates ranging from $95M to $2B for 2025 depending on what is included; the silicon photonics market uses different boundary definitions; and AI cluster optics figures from LightCounting include pluggable modules that are not Celestial AI's product. The most important sizing gap is that no independent analyst has published a dedicated forecast for the scale-up optical chiplet sub-segment that Celestial AI occupied. Marvell's $500M ARR target for Q4 FY2028 implies a specific company-level revenue projection but is a company-claimed forward projection, not an independent market size. Reconciling these figures requires a bottom-up build from hyperscaler cluster size assumptions (number of XPUs × chiplets per XPU × ASP per chiplet), which is not publicly available. Contradictory estimates between Mordor Intelligence (CPO CAGR 37%), Precedence Research (30.66%), and general silicon photonics forecasts (29.5%) reflect genuine uncertainty about timing and adoption rates. [CM004, CM005]

2.6 Exhibits

3.1 Competitive Landscape Overview

The AI data center optical interconnect competitive landscape contains four distinct tiers relevant to Celestial AI's Photonic Fabric platform. The first tier consists of direct optical chiplet peers — companies pursuing in-package or co-packaged silicon photonics chiplets for XPU-to-XPU and XPU-to-memory scale-up links — most importantly Ayar Labs and, to a lesser degree, Ranovus. The second tier comprises platform incumbents that have launched or are actively developing silicon photonics optical chiplets from a position of existing semiconductor or systems dominance: Intel Silicon Photonics, Broadcom (CPO), and NVIDIA (NVLink ecosystem). The third tier includes established optical component manufacturers — primarily Coherent Corp — that compete in adjacent CPO/LPO product lines at scale. The fourth tier features LightMatter, a photonic computing and interconnect startup whose Passage photonic interposer overlaps with Celestial AI's interconnect use case despite a different architectural approach. The status-quo technology that the entire optical chiplet category must displace is NVIDIA's copper-based NVLink scale-up fabric, which in its fifth generation (Blackwell) supports 1.8 TB/s total bandwidth per GPU and enables pods of up to 576 GPUs. The UALink consortium — spanning Broadcom, AMD, Intel, Google, Microsoft, Cisco, and Meta — published the open-standard UALink 1.0 specification in April 2025 targeting up to 1,024 accelerators per AI pod, but hardware is expected no earlier than late 2026, leaving NVLink as the sole shipping high-bandwidth scale-up fabric at the report date. Competitor funding has accelerated sharply: Ayar Labs raised $500M in a Series E in early 2026 (total $870M), LightMatter raised $400M in October 2024 (total $850M, valuation $4.4B), and Ranovus committed $100M+ to Ottawa manufacturing in August 2025. The competitive intensity is high and differentiators on manufacturing readiness, ecosystem access, and architectural specificity will determine design wins. [CP001, CP002, CP003, CP007, CP008, CP011]

3.2 Direct Optical Chiplet Competitors

Ayar Labs is Celestial AI's most direct competitor. Founded in 2015 as a UC Berkeley spin-out, Ayar Labs has built the TeraPHY optical I/O chiplet, which uses micro-ring modulators (MRMs) and the UCIe chiplet interconnect standard for integration into AI accelerator packages. TeraPHY achieves up to 8 Tbps bandwidth per chiplet; the companion SuperNova multi-wavelength light source supports 16 wavelengths as an external laser input. Ayar Labs targets approximately 5 pJ/bit energy efficiency with a roadmap to under 1 pJ/bit. The company raised $155M in a Series D in December 2024 — with AMD Ventures, Intel Capital, and NVIDIA as investors alongside Advent Global Opportunities — and followed with a $500M Series E in early 2026 led by Neuberger Berman with AMD, NVIDIA, MediaTek, and Alchip participating, bringing total raised to $870M and valuation to $3.75–3.8B. The key architectural contrast: Celestial AI uses electro-absorption modulators (EAMs) and 3D EIC-on-PIC stacking, while Ayar Labs uses MRMs and co-packaging within existing CMOS process flows. EAMs offer advantages in thermal tolerance, bandwidth density, and extinction ratio; MRMs offer lower insertion loss but require tighter temperature management. Celestial AI's 3D stacking architecture uniquely targets disaggregated memory access at scale — a use case Ayar Labs does not explicitly address — potentially giving Celestial AI a differentiated niche. NVIDIA's simultaneous direct investment in Ayar Labs (both Series D and Series E) while partnering with Marvell on NVLink Fusion represents a structural conflict of interest that requires monitoring. If NVIDIA ultimately favors Ayar Labs' TeraPHY-based approach within its platforms, Marvell/Celestial AI's access to NVIDIA's scale-up ecosystem could be compromised. Ranovus (Ottawa, founded 2012) is a secondary direct competitor. Its Odin direct-drive CPO 3.0 platform supports 100 Gbps per optical channel and targets AI/ML ASIC integration. In March 2024 Ranovus announced a 6.4 Tbps co-packaged optics solution jointly developed with MediaTek. Ranovus received $36M from Canada's Strategic Innovation Fund in 2023 and announced a $100M+ manufacturing expansion in Ottawa in August 2025. While Ranovus is active in the CPO space, it lacks the deep hyperscaler investor backing and broad ecosystem relationships of Ayar Labs. [CP004, CP005, CP006, CP019, CP020, CP021]

3.3 Platform and Incumbent Competitors

Intel Silicon Photonics has the longest institutional history in commercial silicon photonics, having operated a transceiver business since 2012. Intel's Optical Connectivity for AI (OCI) initiative represents its optical I/O chiplet program. At OFC 2024, Intel demonstrated the first fully integrated bidirectional OCI optical I/O chiplet co-packaged with an Intel CPU. The OCI chiplet supports 64 channels at 32 Gbps each way (4 Tbps bidirectional), achieves less than 3–5 pJ/bit power efficiency, and supports reach up to 100 meters over standard fiber. Intel's advantage lies in vertical integration: internal silicon photonics R&D, packaging, and foundry capabilities. Intel has acqui-hired optical teams including from Glofast to bolster the program. The competitive risk from Intel is primarily on Intel CPU platform customers, where Intel may prefer an internal optical chiplet, narrowing Celestial AI/Marvell's addressable market among Intel-platform AI deployments. Broadcom's CPO program focuses on scale-out Ethernet switching rather than the scale-up XPU-to-XPU interconnect domain where Celestial AI competes. In March 2024 Broadcom launched Bailly, the industry's first 51.2 Tbps CPO Ethernet switch, integrating eight 6.4 Tbps silicon photonics optical engines with its Tomahawk5 chip — achieving a 70% power reduction versus pluggable transceivers — with field validation at Meta. The roadmap extends to Davisson at 102.4 Tbps. Broadcom also supplies NVLink Switch chips to NVIDIA and is a foundational member of the UALink consortium. At Hot Chips 2024, Broadcom presented early work on AI compute ASICs with optical attach, signaling potential expansion into scale-up adjacency. If Broadcom pivots CPO into the XPU scale-up space, it would represent the highest- resource competitive threat given Broadcom's hyperscaler relationships and manufacturing scale. NVIDIA controls the scale-up ecosystem through NVLink. NVLink v5 (Blackwell) achieves 1.8 TB/s per GPU; NVSwitch enables up to 576 GPU pods. At Computex 2025, NVIDIA unveiled NVLink Fusion — a chiplet interface enabling custom XPU silicon, including from Marvell, to natively connect to NVIDIA's NVLink ecosystem. Marvell partnered with NVIDIA on NVLink Fusion supporting up to 1.8 TB/s chiplet bidirectional bandwidth. NVLink Fusion creates a cooperative channel but also ensures NVIDIA controls optical adoption timing within its ecosystem. [CP007, CP008, CP009, CP010, CP011, CP012]

3.4 Adjacent Optical Ecosystem Competitors

Coherent Corp (NYSE: COHR), formerly II-VI Incorporated, is the world's largest optical transceiver supplier. Coherent reported FY2025 revenue of $5.81B, up 23% year-over-year, with AI datacenter segment revenue growing 61% in the same period. Coherent is primarily a transceiver and optical component supplier, not a direct chiplet competitor, but its expanding CPO and LPO (linear-drive pluggable optics) portfolio puts it in buyers' decision sets when hyperscalers and switch ASIC vendors evaluate optical integration options. Coherent has deepened partnerships with NVIDIA for CPO integration in NVIDIA's next-generation networking platforms. The competitive risk from Coherent is indirect: if hyperscalers prefer Coherent's module-based CPO over a chiplet approach, Celestial AI/Marvell's addressable integration tier shrinks. Coherent's $5.81B revenue and public company status represent formidable ecosystem scale. LightMatter is a photonic computing and interconnect startup that raised $400M in a Series D in October 2024 (total $850M, valuation $4.4B). LightMatter's Passage photonic interposer uses light to connect multiple chips placed on a single photonic substrate. The Passage M1000, announced March 2025, delivers over 114 Tbps total optical bandwidth. Manufacturing partnerships with GlobalFoundries and Amkor provide supply chain backing. LightMatter's architectural approach requires all co-located chips to share the same photonic substrate, limiting deployment flexibility compared to Celestial AI's discrete chiplet approach; however, LightMatter's higher valuation ($4.4B) and funding track record signal strong market confidence. LightMatter competes for the same hyperscaler attention and custom AI cluster design wins. Rockley Photonics filed for Chapter 11 bankruptcy in early 2023 and emerged from restructuring later that year. Celestial AI acquired Rockley's entire silicon photonics IP portfolio for $20M in October 2024, covering optoelectronic systems- in-package, electro-absorption modulators, and optical switch technology, bringing Celestial AI's total patent count to 200+. Rockley's trajectory illustrates the capital intensity and execution risk in silicon photonics commercialization; the acquisition converted a potential IP liability into a competitive asset for Celestial AI's EAM technology position. [CP017, CP018, CP019, CP020, CP021, CP022]

3.5 Competitive Moat, Positioning, and Risk Assessment

Celestial AI/Marvell's competitive positioning rests on five pillars: (1) TSMC 5/4nm- qualified electro-optical integrated circuit (EIC) process with multiple tape-outs completed; (2) EAM modulator technology offering thermal tolerance and bandwidth density advantages over MRM-based competitors such as Ayar Labs; (3) 3D EIC-on-PIC stacking architecture uniquely enabling disaggregated memory access — a use case not addressed by Ayar Labs TeraPHY or Broadcom CPO; (4) a 200+ patent portfolio reinforced by the Rockley Photonics IP acquisition; and (5) Marvell's hyperscaler custom silicon distribution relationships and NVLink Fusion partnership with NVIDIA. The primary competitive risks are: (A) Ayar Labs enters 2026 better-funded ($870M raised vs Celestial AI's ~$515M pre-acquisition) with a broader strategic investor base spanning NVIDIA, AMD, and Intel Capital simultaneously — suggesting it may reach production scale faster. (B) NVIDIA's direct investment in Ayar Labs while also partnering with Marvell on NVLink Fusion creates a dual-track strategy where NVIDIA may favor Ayar Labs within its ecosystem. (C) Broadcom, with $350B+ market cap and all major hyperscaler relationships, could pivot its CPO capabilities toward the scale-up adjacency — representing the highest-resource potential entrant. (D) Intel's vertically integrated OCI program could produce a preferred internal chiplet for Intel-platform AI deployments, narrowing Celestial AI's addressable market on Intel's installed base. The UALink consortium's April 2025 specification publication signals a future open-standard scale-up fabric that could reduce NVIDIA's NVLink lock-in and create a more level playing field. UALink hardware not shipping until late 2026 leaves near-term design wins concentrated in the NVLink ecosystem, where Celestial AI/Marvell's NVLink Fusion partnership is a tactical advantage. The memory disaggregation architectural differentiation remains the most defensible long-term niche for Celestial AI/Marvell if it can demonstrate customer validation at scale. [CP021, CP016, CP021, CP016, CP028, CP030]

3.6 Exhibits

4.1 Funding History and Capital Structure

Celestial AI raised approximately $515 million across five external financing rounds between January 2021 and March 2025. The seed round of $2.87 million in January 2021 provided initial capital to advance the Photonic Fabric concept. A $56 million Series A led by Lux Capital followed in February 2022, enabling early silicon design work and initial tape-outs. Tiger Global Management led a $100 million Series B in June 2023 as the company advanced its first commercial prototypes and expanded its engineering team. The March 2024 Series C of $175 million, led by U.S. Innovative Technology Fund (USIT), was a pivotal strategic milestone: it introduced AMD Ventures, TSMC Ventures (VentureTech Alliance), Samsung Catalyst, and multiple hyperscaler-adjacent investors, validating the photonics-for-AI thesis at the supply-chain level. The round valued Celestial AI at $1.22 billion post-money, making it a unicorn. Thirteen months later, in March 2025, Celestial AI closed a $250 million Series C1 led by Fidelity Management and Research, with new participation from BlackRock, Maverick Silicon, and personal investment from Lip-Bu Tan (who simultaneously joined the board). This round valued the company at $2.5 billion post-money. Between the Series C and Series C1, Celestial AI acquired the Rockley Photonics IP portfolio in October 2024 for approximately $20 million — a balance-sheet use deploying cash at distressed prices to add over 200 patents covering electro-absorption modulators, optoelectronic systems-in-package, and optical switch technologies. Rockley had filed for Chapter 11 in early 2023, making this a strategic opportunistic purchase that reinforced Celestial AI's IP moat ahead of its largest financing round and eventual exit. [CI001, CI002, CI003, CI004, CI005, CI006]

4.2 Acquisition Financial Terms

Marvell Technology announced the definitive acquisition agreement on December 2, 2025, alongside its fiscal Q3 FY2026 earnings, and closed the transaction on February 2, 2026. The deal structure comprises two components. The upfront consideration is approximately $1.0 billion in cash plus approximately 27.2 million shares of MRVL common stock valued at approximately $2.25 billion based on the 10-day volume-weighted average price prior to announcement, yielding approximately $3.25 billion in total upfront value. The contingent earn-out consideration is up to approximately 27.2 million additional MRVL shares (~$2.25 billion) tied to cumulative revenue milestones by the end of Marvell's fiscal year 2029. The earn-out is structured in tranches: one-third unlocks if cumulative Celestial AI revenue reaches at least $500 million by FY2029 year-end, and the full earn-out vests only if cumulative revenue exceeds $2.0 billion. The cash portion was funded from Marvell's existing cash on hand, reducing its cash balance by $1 billion at closing and lowering expected annual interest income by approximately $38 million. The stock consideration required issuance of approximately 27 million new MRVL shares, increasing the diluted weighted-average share count by that amount. As part of the broader strategic partnership, Marvell simultaneously issued Amazon a warrant to purchase approximately 1,045,171 MRVL shares at $87.0029 per share, with vesting based on Amazon's cumulative purchases of Photonic Fabric products through the end of 2030. The acquisition also adds approximately $50 million in annual non-GAAP operating expenses to Marvell's run rate, reflecting Celestial AI's engineering and development workforce. [CI008, CI009, CI010, CI011, CI012, CI013]

4.3 Revenue Model and Marvell Projections

Celestial AI had de minimis recognized revenue at acquisition, consistent with its position as an early-stage semiconductor design company completing development-phase tape-outs and transitioning to commercialization. Marvell's official closing press release projects that revenue contributions will begin in the second half of fiscal year 2028 (Marvell's fiscal year ends in late January), with revenue ramping meaningfully in the fourth quarter to a $500 million annualized run rate. Revenue is then expected to double to a $1 billion annualized run rate by the fourth quarter of fiscal year 2029. These projections are Marvell management's company- claimed forward guidance carrying standard forward-looking statement caveats. By Q4 FY2026 earnings (released March 2026), Marvell updated its combined near-term outlook: Celestial AI and XConn Technologies (another acquisition closed after fiscal year-end) are expected to contribute approximately $250 million in aggregate to FY2028 revenue. This figure implies the longer-term $500 million ARR trajectory remains intact but is stated as a combined contribution, not broken out by entity. Marvell's CEO Matt Murphy cited a total addressable market of approximately $10 billion for scale-up connectivity, positioning Celestial AI as the primary vehicle for capturing that market. Marvell reported FY2026 actual revenue of $8.195 billion (+42% year-over-year), with FY2027 guidance of approximately $11 billion and FY2028 guidance of approximately $15 billion — a trajectory that implicitly requires Celestial AI and other photonics programs to contribute material incremental revenue within two fiscal years. The revenue model within Marvell follows a custom-silicon and licensing structure consistent with Marvell's Data Center Group. Hyperscaler customers — including Amazon, which holds a warrant — are expected to be primary revenue sources. No specific pricing, volume commitments, or contract structures have been publicly disclosed. [CI003, CI003, CI017, CI018, CI010, CI019]

4.4 Unit Economics and Cash Consumption

Celestial AI never disclosed standalone revenue, gross margin, or operating cost structure prior to acquisition. As a private semiconductor startup completing development-stage tape-outs and transitioning to commercialization, the company had de minimis revenues but incurred material operating expenses including chip design costs, packaging R&D, manufacturing tape-outs, and IP maintenance — including the $20 million Rockley Photonics IP purchase. No quarterly or annual income statement data is available from any public source. Implied cash burn can be estimated from the funding timeline. The $250 million Series C1 closed in March 2025, followed by the acquisition announcement nine months later in December 2025. A 141 to 152 employee semiconductor design house with advanced packaging programs and multiple tape-out cycles typically operates at $15 to $25 million per month in cash consumption. At that range, the $250 million Series C1 would have provided approximately 10 to 17 months of runway, placing the acquisition announcement at or near the end of funded runway. This is a secondary inference, not a disclosed figure, and should be treated with low confidence. At the $3.25 billion upfront acquisition price and approximately 141 to 152 employees, the implied acquisition price per employee is approximately $21 to $23 million — a metric reflecting the premium paid for the IP portfolio, engineering talent, and Marvell's projected future revenue generation rather than any current financial performance. The venture return for Series C1 investors who invested at a $2.5 billion pre-money valuation is approximately 1.3 times on the upfront consideration alone, rising to potentially 2.2 times if full earn-outs are achieved. [CI022, CI023, CI024, CI025, CI011, CI022]

4.5 Capital Adequacy and Financing

As an acquired subsidiary, Celestial AI no longer carries independent capital requirements. The relevant analysis shifts to Marvell's capacity to fund continued development and ramp to revenue. Marvell funded the $1 billion cash consideration from existing cash on hand, which is manageable given Marvell's free cash flow of approximately $1.39 billion in FY2025 and investment-grade credit rating. The incremental $50 million per year in non-GAAP operating expenses is modest relative to Marvell's overall operating expense base. The earn-out structure — up to 27.2 million additional MRVL shares — represents a contingent liability that does not consume cash. Instead, it creates future dilution risk contingent on successful revenue ramp. One-third of the earn-out (approximately 9.1 million shares) unlocks at $500 million in cumulative Celestial AI revenue by FY2029 year-end; the full 27.2 million shares vest only if cumulative revenue exceeds $2.0 billion — implying sustained and significant revenue generation well ahead of Marvell's Q4 FY2028 run-rate target. Standalone Celestial AI capital adequacy pre-acquisition was entirely dependent on continued external equity financing. The Series C1 of $250 million in March 2025 was the final financing event before the acquisition. No bridge financing, debt facilities, or revenue-backed credit has been disclosed. The company had no material revenue to generate operating cash flow, making it entirely equity-financed through the close of the Marvell transaction. [CI011, CI012, CI013, CI016, CI026, CI027]

4.6 Financial Risks and Verdict

The primary financial risk to the investment thesis is earn-out non-achievement. If Celestial AI fails to generate $500 million in cumulative revenue by Marvell's FY2029 year-end, approximately one-third of the earn-out will not vest, and if cumulative revenue remains below $2.0 billion, the full earn-out is forfeited — capping total transaction value at $3.25 billion. Given that meaningful revenue is not expected until H2 FY2028, Marvell has approximately a two-year window (H2 FY2028 through FY2029) to accumulate $500 million to $2.0 billion in cumulative revenue from a technology that had no commercial deployments at acquisition. Additional financial risks include: (1) integration execution — photonics-for-AI is technically demanding and requires customer qualification cycles typically exceeding 12 to 18 months; (2) revenue recognition timing — semiconductor custom silicon revenue is lumpy and milestone- driven, making quarterly run-rate targets sensitive to a handful of design-win deployments; (3) competitive displacement — Broadcom's and NVIDIA's internal photonics programs could erode the available market before Celestial AI reaches scale; (4) cost overruns — the $50 million annual non-GAAP opex figure may expand as manufacturing scale-up requires additional headcount; and (5) valuation re-rating risk — Marvell's stock price at announcement was approximately $82 to $87 per share, and a sustained decline would reduce the effective value of both the upfront stock consideration and earn-out shares. Financial verdict: Celestial AI's profile at acquisition was that of a pre-revenue, high-burn semiconductor design startup with strong IP and strategic positioning but no independent path to profitability. The investment rationale is entirely forward-looking, dependent on Marvell's ability to commercialize Photonic Fabric technology for hyperscaler customers within a 24 to 36 month window. Gross margin, unit economics, and pricing have not been disclosed and represent blocking diligence gaps for assessing earn-out achievability. [CI027, CI007, CI007, CI001, CI011, CI024]

4.7 Exhibits

5.1 Product Overview and Positioning

Celestial AI's Photonic Fabric is a silicon photonics optical interconnect platform designed to replace copper-based scale-up fabrics in AI data center infrastructure. In the customer workflow, AI system operators deploying clusters of XPUs — GPUs, TPUs, or custom AI accelerators — face a fundamental bandwidth and latency constraint: copper interconnects such as NVLink and UALink impose limits on aggregate bandwidth, power per bit, and the number of XPUs that can be tightly coupled in a single scale-up domain. Photonic Fabric addresses this by co-packaging an optical chiplet directly with the XPU, enabling chip-to-chip optical links that deliver greater than 10 Tbps bandwidth per module while consuming approximately 3.1 pJ per bit end-to-end — compared to 15 pJ/bit or more for pluggable optical modules and significantly more for copper at the same scale. The primary use cases are scale-up (connecting XPU-to-XPU within a server or across racks) and memory disaggregation (connecting XPU-to-remote memory pools via the Photonic Fabric Appliance, which supports up to 32 TB of unified HBM3e-backed DDR5 memory accessible at under 200 ns latency). These use cases map directly to the hyperscaler requirement to scale LLM inference and training workloads beyond the physical constraints of a single server node, allowing batch sizes and sequence lengths previously impossible with conventional memory architectures. The Photonic Fabric Module is delivered as a chiplet for direct co-packaging with XPUs. It integrates a TSMC 5nm custom control ASIC, HBM3e and DDR5 memory controllers, a silicon photonic interposer, and a multi-fiber array in a 2.5D system-in-package. The Photonic Fabric Appliance is a rack-form-factor product (2U) delivering up to 115 Tbps aggregate optical bandwidth for memory disaggregation scenarios, using the same underlying chiplet technology with a 16-port switching fabric. These two form factors cover both scale-up (chiplet, co-packaged) and scale-out (appliance, rack-attached) connectivity needs. Following Marvell's acquisition in February 2026, Photonic Fabric positions Marvell as the only custom XPU supplier offering fully optical scale-up connectivity as a co-packaged component. The NVLink Fusion partnership with NVIDIA, backed by a $2 billion NVIDIA investment in Marvell (March 2026), expands the total addressable market to include NVIDIA-ecosystem AI clusters, not only UALink-based architectures. [CE001, CE002, CE003, CE004, CE009, CE020]

5.2 Core Technology Architecture

The Photonic Fabric's technological foundation rests on three interrelated innovations: use of electro-absorption modulators (EAMs) as the primary optical modulation element; placement of optical I/O anywhere inside the die rather than only at the chip perimeter; and the Optical Multi-Chip Interconnect Bridge (OMIB) that integrates photonics into package bridges for multi-chiplet connectivity. Electro-Absorption Modulators operate by shifting the absorption edge of a semiconductor material — typically Ge/Si quantum wells on TSMC's silicon photonics process — in response to an applied voltage, modulating optical intensity without requiring resonant cavities. This makes EAMs intrinsically tolerant to temperature variations exceeding 85°C and transient thermal ramps of 200°C per second — conditions that defeat microring resonators, which require active thermal tuning heaters adding power overhead per ring. The analog-only SerDes design consumes approximately 2.4 pJ/bit for the electronics plus 0.7 pJ/bit for the laser, yielding approximately 3.1 pJ/bit total, compared to 5–15 pJ/bit for competing digital signal processor (DSP)-based approaches, including pluggable optical modules. The in-die optical I/O architecture eliminates the traditional beachfront constraint by embedding photonic structures via a silicon photonic interposer layer within the TSMC 5nm CoWoS package. This allows optical transceivers to be placed anywhere on the die surface, freeing the die perimeter for HBM and DDR I/O, increasing aggregate optical bandwidth density, and enabling placement of optical engines under 500W–900W GPU dies without causing thermal hotspots in adjacent circuitry. The OMIB (Optical Multi-Chip Interconnect Bridge) extends optical connectivity across multiple chiplets in a package and across packages in a multi-chip module. OMIB-enabled links achieve chip-to-memory latencies below 200 ns end-to-end, comparable to local HBM access for latency-tolerant operations. Wavelength division multiplexing (WDM) is applied to maximize bandwidth density per fiber, with pre-FEC BER below 1e-8 at -10 dBm OMA input. Four TSMC 5nm/4nm tapeouts had been completed as of the OFC 2025 and Hot Chips 2025 presentations, validating the TSMC silicon photonics process integration. Manufacturing relies on TSMC's SiPh process combined with CoWoS 2.5D interposer packaging, which TSMC was actively maturing toward CPO production capability — targeting 1.6 Tbps optical module mass production in H2 2025 for leading-edge AI customers. [CE001, CE002, CE004, CE005, CE006, CE007]

5.3 IP and Technology Stack

Celestial AI's intellectual property portfolio combines organic patents filed from the company's founding in 2020 with the October 2024 distressed acquisition of Rockley Photonics' patent portfolio. The combined portfolio exceeds 200 patents worldwide, covering worldwide issued and pending applications across three technology pillars: optoelectronic systems-in-package, electro-absorption modulators, and optical switch technology. Phil Winterbottom (CTO) brings deep telecom photonics expertise from Bell Labs' 1127 Computer Science Research Center and a track record as a serial founder (Entrisphere and Gainspeed). His architectural contributions — particularly in-die EAM placement and the OMIB bridge concept — are widely cited as central to Photonic Fabric's differentiation from ring-resonator-based competitors. David Lazovsky (CEO) and Preet Virk (COO) bring semiconductor commercialization backgrounds that complement the CTO's technical depth. The Rockley Photonics IP was acquired in October 2024 for approximately $20 million following Rockley's Chapter 11 bankruptcy emergence and subsequent restructuring. Rockley patents date from as early as 2014, predating most silicon photonics startups, and include foundational EAM waveguide structures and optoelectronic packaging claims directly applicable to Celestial AI's in-die architecture. This acquisition reinforced the IP moat ahead of the Series C1 financing round and Marvell acquisition. The technology stack's manufacturing layer depends entirely on TSMC's SiPh process combined with CoWoS 2.5D advanced packaging, creating a single-foundry dependency that is both a strength (TSMC advanced node quality and TSMC Ventures strategic alignment) and a risk (CoWoS capacity constraints under heavy NVIDIA and Broadcom demand). The software layer provides fabric management, telemetry, reliability/availability/serviceability (RAS), and AI-driven analytics for hyperscale fleet integration. The software stack's maturity and completeness for production hyperscaler deployment has not been independently verified. [CE008, CE009, CE010, CE011, CE012, CE013]

5.4 Technology Roadmap and Marvell Integration

Marvell's $3.25 billion acquisition of Celestial AI, closed on February 2, 2026, was explicitly motivated by accelerating Marvell's co-packaged optics roadmap for its custom AI silicon (XPU) business. The acquisition positions Photonic Fabric as a standard offering co-packaged with Marvell-designed XPUs sold to hyperscaler customers, extending Marvell's data center optics revenue beyond its existing pluggable DSP portfolio. The revenue contribution roadmap targets initial Celestial AI revenue in H2 FY2028 (approximately H2 calendar 2027), scaling to a $500 million annualized run rate by Q4 FY2028 and $1 billion by Q4 FY2029. These milestones drive the earn-out structure tied to Celestial AI shareholders. To meet the H2 FY2028 timeline, Photonic Fabric must achieve TSMC CoWoS production-ready status and pass customer qualification by late calendar 2026 to mid-2027 — a compressed window given the complexity of photonic interposer co-packaging in volume production. The NVIDIA NVLink Fusion partnership, announced on March 31, 2026, accompanied by a $2 billion NVIDIA investment in Marvell, opens a second pathway for Photonic Fabric deployment. NVLink Fusion allows Marvell's XPUs to interconnect directly with NVIDIA GPUs at up to 1.8 TB/s, and Photonic Fabric will be critical for enabling rack-scale and multi-rack NVLink Fusion topologies. NVIDIA's involvement substantially de-risks the customer pipeline for Marvell's custom XPU plus Photonic Fabric combination, and accelerates hyperscaler qualification timelines. The technical roadmap phases include: (1) first-generation in-die optical I/O chiplet for chip-to-chip scale-up, demonstrated at OFC 2025; (2) Photonic Fabric Appliance for memory disaggregation at 32 TB and 115 Tbps, demonstrated at Hot Chips 2025; (3) next-generation chiplet with higher bandwidth density targeting Marvell custom ASIC integration (2026–2027); and (4) volume production qualification at TSMC CoWoS line (estimated 2027). The four completed 5nm/4nm tapeouts validate the TSMC process integration, but production-scale yield and hyperscaler customer acceptance testing remain ahead of the revenue milestones. [CE012, CE013, CE016, CE017, CE018, CE019]

5.5 Technology Risks and Mitigation Gaps

Five categories of technology risk are material to the Photonic Fabric investment thesis following the Marvell acquisition. EAM insertion loss: While EAMs outperform MRMs on thermal stability and bandwidth linearity, EAMs can exhibit higher optical insertion loss under specific waveguide geometry and bias configurations. Higher insertion loss demands more powerful laser sources, increasing system power and cost, and can reduce link budget margin at scale. No independent laboratory has published a head-to-head silicon comparison of EAM versus MRM insertion loss under identical production-equivalent conditions. TSMC CoWoS capacity concentration: Photonic Fabric depends exclusively on TSMC's SiPh and CoWoS processes. TSMC's CoWoS capacity is heavily allocated to NVIDIA's HBM stacking and Broadcom's custom ASICs. Adding Celestial AI's photonic interposer layers introduces scheduling complexity. Any TSMC capacity constraint or yield shortfall could push volume production beyond the FY2028 revenue targets, directly threatening the Marvell earn-out structure. Thermal management at co-packaged XPU scale: Optical engines co-packaged with 500W–900W XPUs generate heat that must be dissipated without thermally coupling into adjacent photonic structures. While EAMs tolerate 85°C swings, sustained high-power operation creates heat flux management challenges at the package level that are not yet resolved at production scale. CPO standardization risk: The OIF CPO Implementation Agreement (3.2 Tbps module) and IEEE 802.3 Ethernet working group standards for embedded optics are still evolving. Lack of interoperability standards means each hyperscaler must qualify Photonic Fabric independently, extending qualification cycles and increasing per-customer integration costs. Technology readiness uncertainty: Despite four TSMC tapeouts, no external production qualification or independent third-party silicon characterization has been publicly disclosed. All performance claims derive from company presentations and an arXiv preprint — not peer-reviewed publications with independent silicon measurements. This gap increases customer hesitancy before committing to production deployments. [CE005, CE013, CE021, CE022, CE006]

5.6 Exhibits

6.1 Customer Overview and Market Position

Celestial AI was a pre-revenue company at the time of its February 2, 2026 acquisition by Marvell Technology. The company operated in the semiconductor hardware category where revenue recognition requires production-volume chip deployment at hyperscaler data centers—a process that typically spans 24 to 36 months from initial design-in to first revenue. No publicly confirmed production customers existed at acquisition. Marvell explicitly guided to de minimis revenue contributions from Celestial AI for fiscal years 2026 and 2027. Despite the absence of production revenue, Celestial AI was not without documented customer traction. In March 2024, at the Optical Fiber Communication (OFC) conference, the company announced that a leading (unnamed) hyperscaler had formally adopted Photonic Fabric for next-generation AI processors—described as one of the earliest large-scale commitments to replace copper interconnects with optical alternatives in AI system design. CEO Dave Lazovsky stated in multiple interviews that additional hyperscalers were engaged in NDA-based technical evaluations. The volume implied by the design win—hundreds of thousands of chiplet units per quarter at production ramp—signals material pipeline rather than exploratory proof-of-concept. The ultimate end buyers of Photonic Fabric are hyperscale cloud and AI infrastructure operators: Amazon Web Services, Microsoft Azure, Google Cloud, and Meta. These companies build custom AI accelerators (XPUs) for training and inference workloads, and Photonic Fabric addresses the bandwidth bottleneck between processing units and memory pools—the AI Memory Wall—by providing optical scale-up connectivity delivering 16 Tbps per chiplet per direction. Celestial AI's customer engagement model required deep design-in collaboration with both the hyperscaler and the ASIC packaging partner, explaining why engagement periods spanned years before revenue. Post-acquisition, Marvell Technology is both the primary custodian and the principal internal customer of Photonic Fabric technology, integrating it into its Data Center Group's custom silicon programs for delivery to hyperscaler end customers. [CU009, CU018, CU017, CU021, CU001, CU014]

6.2 Strategic Investor Signals as Customer Proxies

In the semiconductor industry, strategic corporate venture investments reliably signal supply-chain alignment and potential customer or manufacturing-partner intent. Celestial AI's investor roster from the Series C (March 2024) and Series C1 (March 2025 through August 2025) rounds contains multiple strategic investors who are simultaneously likely customers, manufacturing partners, or critical ecosystem participants. AMD Ventures' participation in the Series C signals AMD's interest in optical interconnect technology for future XPU architectures. AMD is a major AI accelerator supplier and its XPU products could directly integrate Photonic Fabric for pooled-memory scale-up. This alignment makes AMD a high-probability design partner and possible indirect customer through hyperscaler-commissioned custom chip programs that use AMD silicon. TSMC VentureTech Alliance participated in both the Series C and the final Series C1 close in August 2025, confirming foundry-level manufacturing alignment. TSMC is the sole credible volume manufacturer for Celestial AI's silicon photonics chiplets at advanced nodes, and this investment ensures preferential capacity allocation, process co-development, and supply chain readiness as Photonic Fabric transitions to volume production. Samsung Catalyst's participation in both rounds signals HBM memory integration interest; Samsung's advanced packaging and HBM expertise are directly required for Photonic Fabric's memory disaggregation architecture. Amazon's role is the most direct and most commercially significant. As part of the Marvell acquisition announcement, Marvell issued Amazon a warrant to purchase approximately 1,045,171 MRVL shares at an exercise price of $87.0029 per share, with vesting tied entirely to Amazon's cumulative purchases of Photonic Fabric products through December 2030, exercisable until December 2031. This structure legally and financially commits Amazon as a forward customer with equity-aligned incentives to drive Photonic Fabric adoption, providing Marvell with multi-year demand visibility that analysts have described as effectively locking in Amazon as a foundational customer. The ecosystem also includes Broadcom (custom silicon and ASIC design services) and Samsung (HBM and advanced packaging), forming the supply chain required to serve hyperscaler customers at production scale. [CU002, CU003, CU004, CU005, CU006, CU007]

6.3 Marvell Customer Channel and Named End Customers

Marvell's established custom AI silicon relationships provide the primary commercial channel through which Celestial AI's Photonic Fabric will reach hyperscaler end customers post- acquisition. Marvell currently serves three hyperscaler custom silicon programs with confirmed or near-confirmed commercial relationships. Microsoft is Marvell's largest and most strategically significant custom AI silicon customer. Microsoft's Project Maia300—the successor to the Maia200—is being developed with Marvell as the primary ASIC design partner, targeting a 2nm process node with HBM4 memory integration. Estimated shipments of 300,000 to 400,000 units are expected by end of calendar 2026, ramping to 1.2 to 1.5 million units in 2027. The Maia300 program requires exactly the kind of co- packaged optical scale-up connectivity that Photonic Fabric provides, making it the clearest near-term integration pathway for initial Photonic Fabric revenue contribution. Amazon is Marvell's second major custom silicon customer via the Trainium AI training and Inferentia AI inference programs. Amazon's special warrant for MRVL shares—granted as part of the Marvell-Celestial AI acquisition agreement—confirms Amazon as the most explicitly committed future Photonic Fabric customer among all publicly identifiable parties. While competitive pressure from Alchip on next-generation Trainium variants introduces some uncertainty about Marvell's long-term role in the core XPU, the warrant structure specifically ties Amazon equity to Photonic Fabric purchases, ensuring incentive alignment. Google was reported in early 2026 to be in late-stage discussions with Marvell to become its third major custom AI silicon partner, supplementing its Broadcom and TPU infrastructure with Marvell-designed inference-focused chips. This would create a third distribution channel for Photonic Fabric. NVIDIA's NVLink Fusion partnership announced March 31, 2026, creates a fourth pathway: enabling third-party custom silicon integrating with NVIDIA's NVLink interconnect fabric, with NVIDIA investing $2 billion in Marvell to co-develop optical networking and silicon photonics capabilities at scale. [CU010, CU011, CU012, CU004, CU013, CU014]

6.4 Customer Qualification and Procurement Process

Hardware qualification for new interconnect chiplets in hyperscaler AI systems follows a multi-stage, multi-year process. Celestial AI navigated several of these stages prior to acquisition, placing it in the design-in and early qualification phases with at least one confirmed hyperscaler partner. The process begins with an NDA Technical Evaluation stage in which hyperscalers review technical specifications and receive first silicon samples for internal lab evaluation. CEO Dave Lazovsky confirmed that multiple hyperscalers were engaged in NDA-based technical evaluations prior to the Series C1 close in March 2025. The March 2024 announcement confirmed that at least one hyperscaler had formally passed the Design-In Commitment stage, incorporating Photonic Fabric into a specific next-generation AI processor architecture. Silicon Sampling and Characterization follows design-in, during which the hyperscaler receives pre-production silicon for electrical, optical, and thermal characterization—typically six to twelve months. This is followed by Qualification Testing and System Integration, the longest stage at twelve to twenty-four months, involving full integration testing in multi-rack AI cluster environments with co-packaging yield validation. Revenue recognition begins only at Production Ramp, when volume chiplet manufacturing commences and purchase orders translate into recognized revenue. The cumulative timeline from design-in to production is typically 24 to 36 months for hyperscaler hardware, explaining why Marvell projects Celestial AI revenue beginning only in H2 FY2028 despite the March 2024 design-win announcement. The qualification process is gatekept by hyperscaler internal programs that operate under mutual NDAs, meaning no publicly confirmable production commitment existed at acquisition time. The earn-out structure in the Marvell acquisition—contingent on cumulative revenue milestones by FY2029—implies Marvell's due diligence confirmed a credible customer pipeline sufficient to support $500M ARR by late FY2028. [CU021, CU001, CU015, CU009, CU009, CU014]

6.5 Revenue Ramp and Customer Concentration Risk

Marvell explicitly guided to de minimis revenue contributions from Celestial AI for fiscal years 2026 and 2027, consistent with the company's pre-production status at acquisition and the multi-year hyperscaler qualification timeline. The first material revenues are expected in H2 FY2028, with a projected $500M annualized run rate by Q4 FY2028 and $1B ARR by Q4 FY2029. These projections represent Marvell management's forward guidance and carry standard forward-looking statement caveats. Analysts expect revenue to begin in H2 calendar 2027 (aligning to Marvell's H2 FY2028) when the first production deployments commence. The earn-out structure embedded in the acquisition agreement provides an important signal about the implied customer pipeline: one-third of the up-to-$2.25B earn-out vests if cumulative Celestial AI revenue reaches $500M by FY2029 end; full earn-out requires cumulative revenue exceeding $2B. These thresholds imply Marvell reviewed specific contracted or highly probable customer commitments during due diligence that supported its confidence in the ramp trajectory. Customer concentration risk is a key structural vulnerability. Early-stage semiconductor programs with hyperscaler customers typically generate highly concentrated initial revenues— the first production program likely represents near-total revenue for FY2028. A delay, capex cycle deceleration, program cancellation, or competitive substitution at the lead hyperscaler could materially defer the $500M ARR milestone and threaten earn-out vesting. Analysts specifically identified this single-hyperscaler concentration as the primary near-term financial risk from the acquisition. The Amazon warrant partially mitigates concentration risk by ensuring a second committed demand signal, but does not constitute a binding purchase commitment with a fixed quantity or timeline. Diversification across Microsoft, Amazon, Google, and NVIDIA ecosystem channels is expected to reduce concentration risk in FY2029 and beyond as additional programs reach production. [CU018, CU017, CU019, CU020, CU015, CU009]

6.6 Exhibits

7.1 Technology Execution Risks

Celestial AI's Photonic Fabric is a silicon photonics optical chiplet using electro-absorption modulators (EAMs) manufactured on TSMC's advanced process nodes and co-packaged using CoWoS advanced packaging. Four interlocking technology risks could impair the revenue ramp. Manufacturing yield and capacity. Silicon photonics yield for EAM-based chiplets at high volumes has not been publicly disclosed. Celestial AI has reported at least four successful tapeouts and a pre-FEC bit error rate below 1e-8 at -10 dBm OMA input, which is consistent with manufacturing-ready photonic yields, but production-volume yield curves remain private. The more acute near-term constraint is TSMC CoWoS advanced packaging capacity: TSMC's CoWoS lines are sold out through 2026, with NVIDIA reserving an estimated 60–70% of CoWoS-L capacity for its Blackwell and Rubin GPU programs. Even as TSMC targets expanding from approximately 70,000 to 130,000 CoWoS wafers per month by late 2026, independent analysts project demand will continue to outpace supply. This creates reservation-timeline risk for Marvell: any delay in securing a CoWoS slot will directly delay the Photonic Fabric volume ramp and the $500M ARR target. TrendForce reported in April 2026 that CoWoS advanced packaging wafer ASP is reportedly approaching 7nm wafer levels, significantly adding to unit cost pressure. Thermal management. Co-packaged optics place the optical engine in close physical proximity to hot XPU die, creating a thermal co-management challenge that is non-trivial at scale. EAMs offer a notable advantage here: Celestial AI's CTO Phil Winterbottom demonstrated at Hot Chips 2025 that EAMs tolerate temperature excursions of up to 200°C per second and operate over swings exceeding 85°C without degrading signal integrity — a direct advantage versus ring modulators, which require fine thermal control. However, managing the combined thermal envelope of XPU die and the laser pump source in the same advanced package remains an active engineering problem. Integration software and technology obsolescence. The firmware and software stack for managing optical I/O across multi-rack AI clusters is immature industry-wide. Configuration management, fault isolation, and dynamic reconfiguration of optical links add complexity compared to copper. In addition, alternative photonics platforms — lithium niobate-on-insulator (LNOI), indium phosphide (InP), and III-V compound semiconductors — continue to advance, with LNOI particularly attracting research interest for its ultra-low insertion loss and high bandwidth density. Should any alternative platform achieve CMOS-compatible volume manufacturing at competitive cost, EAM silicon photonics could face technology obsolescence risk within a 5–7 year horizon. [CR021, CR022, CR023, CR023, CR011, CR019]

7.2 Market Timing and Adoption Risks

Marvell's earn-out structure — with $500M ARR targeted by fiscal year 2028 and $2B ARR for the maximum payout — is tightly coupled to the pace of hyperscaler CPO adoption. Three factors introduce meaningful timing risk. CPO adoption timeline. LightCounting's July 2025 analysis of the first virtual CPO conference characterized initial large-scale CPO deployments as beginning in 2026, but high-volume rollout as more likely in 2027. This is consistent with the typical 24–36 month design-in to production cycle for custom silicon at hyperscale, and with Marvell's own guidance for de minimis Celestial AI revenue through fiscal year 2027. Any further slippage — hyperscaler program delays, integration challenges, or procurement re-prioritization — would push first material revenue toward FY2028, compressing the earn-out window. Copper infrastructure longevity. Copper-based NVLink, InfiniBand, and Ethernet interconnect systems have demonstrated reliability at 400G and are being upgraded to 800G. If hyperscalers conclude that copper solutions are adequate for AI systems through the late 2020s — or if AI cluster scaling plateaus below the level that requires optical-only interconnects — the total addressable opportunity for Photonic Fabric could be materially smaller or arrive later than Marvell's internal models assume. AI capex concentration risk. Marvell's $500M ARR target implies that two to three hyperscaler customers will account for the vast majority of early revenue. Any slowdown in AI infrastructure capex by the leading hyperscalers — driven by macroeconomic pressures, capex discipline cycles, or shift in AI workload characteristics — would directly reduce Photonic Fabric demand. The concentration means that a single hyperscaler's production schedule delay is effectively the company's revenue delay. [CR005, CR020, CR021, CR021, CR006, CR019]

7.3 Competitive and Strategic Risks

The optical chiplet market is nascent but rapidly contested. Marvell/Celestial AI faces three credible competitive threats and one structural strategic risk from the NVIDIA relationship. Ayar Labs. Ayar Labs in March 2025 announced the TeraPHY, described as the industry's first UCIe optical interconnect chiplet, achieving 8 Tbps bandwidth using a 16-wavelength SuperNova light source. NVIDIA, AMD, and Intel have all invested in Ayar Labs. Critically, NVIDIA's investment alignment means Ayar Labs has the potential to become the preferred optical chiplet supplier for NVIDIA-ecosystem deployments. Ayar's collab with Global Unichip Corp (GUC) — a key TSMC-ecosystem ASIC design services firm — targets hyperscale customers and is designed to embed TeraPHY into reference AI designs. Ayar Labs targeted mass production by mid-2026 with volumes scaling toward hundreds of millions of chiplets by 2028. If Ayar Labs achieves this ramp, Marvell/Celestial AI could face competition at scale earlier than the earn-out milestones require. Intel OCI and Broadcom CPO. Intel's own optical connectivity (OCI) silicon photonics platform serves Intel Gaudi customers, creating a captive competitive alternative. Broadcom's Tomahawk 6 switch with integrated CPO represents a competitive product at the AI switching layer. Neither Intel OCI nor Broadcom CPO currently targets the same scale-up interconnect position as Photonic Fabric, but the competitive pressure at adjacent layers constrains Marvell's pricing power and ecosystem footprint. NVLink Fusion strategic dependency. NVIDIA's $2B investment in Marvell (announced April 2026) and the NVLink Fusion partnership simultaneously validates the optical interconnect thesis and creates strategic dependency. If NVIDIA shifts its partnership terms, acquires an alternative optical chiplet supplier, or advances its in-house photonics program, Marvell's NVLink Fusion channel would be at risk. The partnership's commercial terms are not publicly disclosed, creating opacity around exclusivity, pricing, and termination provisions. [CR011, CR007, CR007, CR002, CR008, CR019]

7.4 Post-Acquisition Integration Risks

Marvell completed the Celestial AI acquisition on February 2, 2026 after receiving FTC early termination. The integration introduces four distinct risk vectors. Earn-out milestone ambiguity. The acquisition included up to $2.25B in earn-out payments, but the specific milestone structure — whether $500M ARR alone triggers full payout, whether partial ARR triggers proportional payment, and what constitutes qualifying revenue — has not been publicly disclosed. This ambiguity creates an information asymmetry between Marvell and former Celestial AI shareholders and employees who hold earn-out interests. Under California employment law, earn-out and vesting terms must be clearly defined, and amendments require employee consent. Disputes over milestone measurement methodology or revenue classification could generate litigation, increasing integration costs and management distraction. Talent retention. The Celestial AI team of 141–152 silicon photonics specialists was assembled over several years in a talent market where fewer than 3,000 researchers globally have deep EAM-based silicon photonics expertise. Post-acquisition equity conversions typically accelerate vesting, reducing the forward retention incentive for key engineers. Marvell's compensation structure, engineering culture, and product prioritization may differ materially from the startup environment that attracted Celestial AI's founders and senior technical staff. R&D pace and organizational complexity. Integrating a 141-person team into Marvell's Data Center Group means Celestial AI's roadmap must compete for engineering resources, budget cycles, and executive attention within a larger organization. Marvell's prior acquisitions — Inphi (2021), Innovium (2021), Avera (2019) — each absorbed significant integration effort. The Inphi integration is widely cited as successful, but the pace of innovation in optical chiplets is faster than in the Inphi-era DSP business, and any slowdown in the Photonic Fabric roadmap could cede ground to Ayar Labs. Competing priorities. Marvell's portfolio spans custom XPU programs (for Amazon, Microsoft, Google), AsteraLabs-style interconnect products, storage controllers, carrier Ethernet, and now Photonic Fabric. Resource allocation decisions could deprioritize Celestial AI programs if near-term revenue opportunities elsewhere appear more certain. [CR009, CR010, CR021, CR004, CR011, CR012]

7.5 Financial and Investment Risks

Marvell paid $3.25 billion in upfront consideration — $1B cash plus 27.2 million shares at $2.25B value — for a pre-revenue company with no production customers at the time of close. The acquisition was priced at an estimated 6.5x–10x Celestial AI's last post-money valuation, reflecting a substantial control premium. Financial risks cluster around three themes. Return threshold. Breaking even on the $3.25B upfront investment likely requires Photonic Fabric to contribute sustained ARR well above $500M with strong gross margins over a 3–5 year horizon. No pre-revenue optical chiplet company has previously generated this level of ARR within the implied timeframe, making the investment dependent on scenarios with no direct comparable precedent. If the technology adoption timeline slips by even 12–18 months, Marvell would need to carry the goodwill on its balance sheet through an extended period of de minimis revenue. Earn-out opacity. The earn-out milestone terms are not publicly disclosed. It is unclear whether the $500M cumulative ARR threshold triggers the full $2.25B earn-out, whether it triggers a proportional tranche, or what the measurement period is. This opacity makes it difficult for Marvell's shareholders — or former Celestial AI stakeholders — to model expected total acquisition cost. The AWS warrant (1,045,171 MRVL shares vesting on Photonic Fabric cumulative purchases through December 2030) adds a further variable cost if Amazon becomes a major customer. Goodwill impairment risk. If AI capex contracts broadly — due to macroeconomic pressures, hyperscaler capex discipline cycles, or a slowdown in frontier model training demand — Marvell could be required to recognize a material goodwill impairment on the $3.25B purchase price. Given Celestial AI was pre-revenue at acquisition, even modest adverse changes in the CPO market outlook could trigger an impairment review under ASC 350. Marvell's own 10-K risk factors explicitly cite integration, achieve-expected-benefits, and market-timing risks for acquisitions. [CR013, CR013, CR014, CR005, CR019, CR015]

7.6 Regulatory and Legal Risks

Regulatory and legal risk for Marvell/Celestial AI spans four distinct areas: US export controls, foreign investment review, intellectual property title, and Taiwan geopolitical concentration. US export controls (EAR). The US Bureau of Industry and Security maintains Export Administration Regulations (EAR) covering advanced computing integrated circuits under ECCNs 3A090.a and 4A090.a. The Biden Administration's January 2025 AI Diffusion Rule — which would have imposed worldwide license requirements on exports of advanced computing ICs — was rescinded by the Trump Administration in May 2025 before taking effect, but BIS indicated it would strengthen targeted chip controls. As of May 2026, preexisting EAR controls remain in force for exports to China and Macau, and the regulatory framework continues to evolve. Marvell, as the acquirer of Celestial AI, must maintain a robust export compliance program to ensure Photonic Fabric chiplets and associated technology are not transferred to restricted end-users or nations without appropriate licenses. Legal analysis by Blank Rome LLP (January 2026) documents four key BIS updates on advanced computing and AI export controls that apply to this technology sector. Failure to comply could result in fines, loss of export privileges, or reputational damage. CFIUS. The Committee on Foreign Investment in the United States reviews transactions that could result in foreign control of or certain non-controlling investments in TID US businesses (Technology, Infrastructure, Data). Celestial AI's silicon photonics IP qualifies as a critical technology. Any future secondary transaction — such as a partial asset sale, licensing arrangement with a foreign entity, or change of control of Marvell itself — that involves the Celestial AI IP portfolio would likely require CFIUS filing or notice. The US Treasury CFIUS FAQ confirms that mandatory filings apply when a foreign government holds substantial interest in the acquiring party and the target is a critical-technology business. Outbound investment rules effective January 2, 2025, add an additional layer: US persons investing in Chinese entities working in semiconductors or AI are subject to notification or prohibition requirements. IP title risk (Rockley Photonics). In October 2024, Celestial AI acquired Rockley Photonics' silicon photonics patent portfolio — approximately 200 global patents covering EAM technology, optoelectronic systems-in-package, and optical switches — for $20 million in a distressed sale from Rockley's post-Chapter 11 restructuring. Distressed IP acquisitions from bankruptcy carry a risk that pre-existing licenses granted to third parties by the debtor may have survived the bankruptcy sale and encumber the acquired patents. Additionally, competing patent claims from other silicon photonics players (Intel, Broadcom, II-VI / Coherent, Lumentum) could generate interference or invalidity proceedings. Taiwan geopolitical concentration. TSMC manufactures both the Photonic Fabric silicon photonics die and performs CoWoS advanced packaging. Over 90% of the world's most advanced semiconductor capacity is in Taiwan, and any military, political, or natural-disaster scenario disrupting TSMC's operations would immediately halt Photonic Fabric production with no short-term alternative. TSMC's Arizona fab is not expected to have advanced packaging capacity at meaningful scale before 2028. US-China tensions have escalated, and the combination of export control retaliation (China's gallium and germanium restrictions), military rhetoric, and the absence of a viable alternative foundry for Celestial AI's EAM silicon photonics process node make this risk systemic and unmitigatable in the near term. [CR004, CR021, CR016, CR012, CR017]

7.7 Personnel and Organizational Risks

Celestial AI's competitive position rests disproportionately on a small, highly specialized team. Three personnel risk dimensions are material. Key-person concentration. Dave Lazovsky (CEO, co-founder) is the primary external face and strategic architect of the Photonic Fabric program. Phil Winterbottom (CTO, co-founder) is the principal technical architect; his Hot Chips 2025 presentation detailing in-die optical I/O design choices demonstrated technical depth that would take years to replace. The departure of either co-founder within the first 12–24 months post-acquisition would represent a severe organizational disruption: loss of customer relationships, technology vision continuity, and the institutional knowledge embedded in the EAM architecture choices. Marvell's standard retention mechanisms (RSU refreshes, performance bonuses) are typically less compelling for co-founders who have already achieved a significant liquidity event. Team depth and replicability. The Celestial AI team of 141–152 specialists spans EAM process engineering, silicon photonics design, advanced packaging co-design, optical I/O firmware, and system integration. This represents a substantial fraction of the global talent pool in this specialty; recruiting even 10–15 senior replacements in silicon photonics would take 18–24 months. If post-acquisition attrition exceeds normal levels — driven by earn-out frustration, cultural friction, or competitor recruiting (particularly by Ayar Labs, which is aggressively hiring) — the technology execution timeline would slip. Historical M&A precedent. Marvell's acquisition of Inphi (completed April 2021, $10B) is the closest comparable integration. Marvell retains Inphi's Dr. Ford Tamer on its board, and the Inphi DSP products are a significant revenue contributor. However, post-acquisition attrition rates from the Inphi integration were not publicly disclosed, leaving a key diligence gap for assessing Marvell's ability to retain a smaller, more technically concentrated team in a faster-moving technology category. [CR004, CR018, CR019]

7.8 Exhibits

8.1 Deal Overview and Implied Transaction Multiples

Marvell Technology completed its acquisition of Celestial AI on February 2, 2026, approximately two months after announcing the transaction alongside its fiscal Q3 FY2026 earnings on December 2, 2025. The deal structure comprises two distinct tranches. The upfront consideration totals approximately $3.25 billion: $1 billion in cash plus approximately 27.2 million newly issued shares of MRVL common stock valued at approximately $2.25 billion based on the 10-day volume-weighted average price preceding announcement. The contingent earn-out consideration—up to an additional 27.2 million MRVL shares (~$2.25 billion)—is structured in tranches tied to cumulative Celestial AI revenue milestones by Marvell's fiscal year 2029 year-end. On conventional trailing revenue metrics, the transaction cannot be benchmarked because Celestial AI was pre-revenue at acquisition—revenues were de minimis and no commercial deployments had been achieved. Marvell's management provided forward revenue guidance: a $500 million annualized run rate from Celestial AI by Q4 FY2028 (approximately H2 calendar 2027, roughly 18 months post-close) and a $1 billion annualized run rate by Q4 FY2029. Using these targets as the denominator, the implied EV/forward ARR multiples are: (1) $3.25 billion upfront divided by $500 million 2-year-forward ARR = 6.5× NTM ARR; (2) $5.5 billion full earn-out value divided by $500 million ARR = 11.0× at the first milestone; and (3) $5.5 billion divided by $1.0 billion FY2029 ARR = 5.5×. The last VC valuation was $2.5 billion at the March 2025 Series C1, implying a 30% premium on the upfront price and a theoretical 120% maximum premium at full earn-out. Total capital invested across all five funding rounds was approximately $515 million, placing the upfront multiple at approximately 6.3× invested capital. Financial advisors provide an arm's-length pricing signal. Celestial AI was advised by Morgan Stanley (financial) and Latham & Watkins (legal). Marvell was advised by Citigroup (financial) and Wilson Sonsini Goodrich & Rosati (legal). Both advisory pairings are consistent with top-quartile competitive M&A processes in the semiconductor sector. MRVL shares rose approximately 9–13% on the announcement day, reflecting investor approval of the strategic rationale. The market was simultaneously reacting to strong Q3 FY2026 earnings and raised guidance, so the share move cannot be fully attributed to the acquisition announcement alone; however, the absence of any negative market reaction to a $3.25 billion outlay for a pre-revenue company is a signal that buy-side consensus viewed the price as fair or strategically justified. [CV013, CV014, CV015, CV016, CV017, CV018]

8.2 Comparable M&A Transactions

Valuing a pre-revenue semiconductor photonics company requires reference to comparable M&A transactions in silicon photonics, optical interconnects, and AI-adjacent chip design. The most directly relevant precedents are acquirer-to-acquirer consistent and technology-to-technology aligned. Marvell's own acquisition of Inphi Corporation (announced October 2020, closed April 2021) for approximately $10 billion provides the most powerful within-acquirer benchmark. Inphi was a growing, revenue-generating silicon photonics and electrical DSP company with approximately $683 million in trailing twelve-month revenue at announcement, implying a roughly 14.6× TTM revenue multiple. Inphi served cloud data centers and 5G telecom, a very close adjacency to Celestial AI's Photonic Fabric optical scale-up platform. The Inphi deal established Marvell's willingness to pay premium multiples for leadership in optical connectivity—making the Celestial AI upfront price at 6.5× forward ARR appear conservative by comparison if the revenue ramp achieves plan. NVIDIA's $6.9 billion acquisition of Mellanox Technologies (closed April 2020) is a critical AI networking precedent. Mellanox was generating approximately $1.3 billion in annual revenue at the time, implying roughly a 5× TTM revenue multiple. The Mellanox deal validated the strategic proposition that connectivity silicon—in that case InfiniBand switching—is essential to AI cluster performance. Celestial AI extends this proposition to the optical layer, with a higher-conviction forward-looking case. Cisco's acquisition of Acacia Communications for $4.5 billion (closed March 2021) is directly relevant: Acacia produces high-speed optical interconnect silicon critical to data center and AI infrastructure. The deal was initially announced in 2019 at $2.6 billion but repriced to $4.5 billion after antitrust delays, reflecting appreciation in photonics asset values during the period. AMD's $1.9 billion acquisition of Pensando Systems (April 2022) provides a closer structural analog: Pensando was a data center networking startup focused on programmable DPU silicon with limited revenue at exit, and AMD paid primarily for engineering talent, IP portfolio, and future design-win pipeline. This deal most closely mirrors the Celestial AI transaction structure—a pre-revenue or early-revenue AI chip startup acquired for IP and roadmap rather than current cash flows. Across this comparable set, EV/TTM revenue multiples range from approximately 5× to 14.6× for revenue-generating targets. For pre-revenue acqui-hires or IP-premium deals, the multiple framework is not directly applicable—but the deal sizes ($1.1B to $10B) provide a credible range into which Celestial AI's $3.25 billion upfront price fits comfortably, especially given the strategic importance of Photonic Fabric to Marvell's $10 billion scale-up TAM narrative. [CV020, CV021, CV009, CV022, CV023, CV024]

8.3 Comparable Public Company and Private Round Benchmarks

Public AI silicon companies and recent private funding rounds provide additional valuation anchors. The most relevant public comps are companies with material AI chip design, custom silicon, and optical interconnect exposure. Among AI semiconductor startups, Cerebras Systems filed for a Nasdaq IPO in April 2026, targeting a valuation range of $23–35 billion against $510 million of reported 2025 revenue—implying roughly 45–69× 2025 revenue at the filing range midpoint. Cerebras reached this revenue through a large multi-year compute contract with OpenAI and an AWS partnership, making it a less directly comparable but instructive ceiling for AI chip valuations in the current cycle. SambaNova Systems has remained private at a reported valuation above $5 billion with no disclosed revenue. Groq, the AI inference chip company, was valued at approximately $2.8 billion in a 2023 Series D round. For silicon photonics specifically, Ayar Labs—a co-packaged optical interconnect company with technology overlapping Celestial AI's Photonic Fabric—raised approximately $870 million in total funding through a Series C as of early 2024, but has not disclosed a post-money valuation publicly. The absence of comparable listed silicon photonics companies at the commercial stage forces reliance on M&A comps and market-sizing proxies rather than public company EV/revenue benchmarks. Among public semiconductor companies, Marvell itself provides the most useful acquirer reference point. Marvell's AI-driven revenue grew 78% in FY2025 to an estimated $1.7–$2.0 billion, and the company reported total FY2026 revenue of approximately $8.2 billion (+42% year-over-year). At announcement, Marvell's own enterprise value was approximately $55–65 billion, implying a forward revenue multiple of 7–8× against FY2026 estimates. Marvell paying 6.5× two-year forward ARR for Celestial AI is thus consistent with Marvell's own market multiple—suggesting the board applied disciplined pricing logic rather than a strategic overpay. The MarketsAndMarkets silicon photonics market report (2025 edition) projects the global silicon photonics market at approximately $2.1–3.1 billion in 2025, growing to $9.6–11.8 billion by 2030 at a CAGR of approximately 27–30%. At Marvell's stated $500 million ARR target, Celestial AI would represent roughly 15–24% share of a 2028 market estimated at $4–5 billion. That is an ambitious but achievable share for a first-mover with hyperscaler design wins and no direct equivalent in co-packaged optics at 5 nm process node. The GrandView Research optical interconnect market report projects the overall optical interconnect market reaching $34.5 billion by 2030, providing a larger TAM backdrop within which the silicon photonics segment sits. [CV028, CV017, CV015, CV029, CV030, CV031]

8.4 Discounted Cash Flow and Revenue Ramp Analysis

A full discounted cash flow analysis of Celestial AI as a standalone entity is precluded by the absence of any disclosed revenue, gross margin, operating cost structure, or customer contract terms. However, a simplified forward-revenue discount approach anchored on Marvell's published milestones allows scenario-based valuation analysis. Base case: Marvell projects $500 million ARR by Q4 FY2028 (H2 calendar 2027, approximately 18 months post-acquisition close), scaling to $1 billion ARR by Q4 FY2029. Assuming a 50–60% gross margin profile consistent with custom silicon licensing and photonic module sales—a reasonable assumption for a fabless design company selling into hyperscaler supply chains—EBITDA contribution at the $1 billion ARR scale would be approximately $100–200 million per year after allocated R&D and SG&A. Discounting this stream at a 15–20% discount rate to reflect execution risk, technology adoption uncertainty, and semiconductor cycle exposure yields a present value range of approximately $2.5–4.0 billion—broadly consistent with the $3.25 billion upfront price. Bull case: Celestial AI achieves $2 billion in cumulative revenue by FY2029, unlocking the full $2.25 billion earn-out. This requires approximately $300–400 million per year in revenue across FY2028 and FY2029, implying successful parallel hyperscaler co-packaged optics deployments. Under this scenario, the effective acquisition price of $5.5 billion against a projected run-rate revenue of $1.5–2.0 billion yields an EV/forward revenue multiple of approximately 2.8–3.7×—extremely attractive by any AI silicon benchmark. Bear case: Revenue ramp delays by 12 months or more due to hyperscaler qualification holdups or competing optical solutions. Cumulative FY2029 revenue reaches only $100–200 million; no earn-out tranches vest. The transaction settles at $3.25 billion for an asset generating minimal near-term revenue. In this scenario the effective price represents a significant write-down risk relative to Marvell's book value for the acquired intangibles. The earn-out structure materially improves Marvell's risk-adjusted expected value relative to paying $5.5 billion upfront. By deferring 41% of total potential consideration as contingent performance-based equity, Marvell's board preserved capital while aligning management incentives with commercial execution. Probability-weighting these scenarios—assigning perhaps 30% probability to the bull case, 45% to base, and 25% to bear—yields an expected value of approximately $3.8–4.2 billion for the transaction, above the upfront consideration but below the full earn-out value. This probability weighting is an analyst estimate and should be treated with low confidence. [CV034, CV035, CV025, CV036, CV037, CV038]

8.5 Earn-Out Structure and Contingent Consideration Analysis

The earn-out mechanism is a materially important and analytically distinctive element of the Celestial AI acquisition that distinguishes it from straightforward cash or stock acquisitions common in semiconductor M&A. The structure is equity-settled: up to approximately 27.2 million additional MRVL shares (~$2.25 billion in value at announcement) are contingent on Celestial AI achieving two cumulative revenue milestones by the end of Marvell's fiscal year 2029 (approximately January 2029). First tranche: approximately one-third of the earn-out (~9.1 million MRVL shares, ~$750 million) vests when cumulative Celestial AI revenue from the acquisition date through FY2029 year-end reaches at least $500 million. Second tranche: the remaining two-thirds (~18.1 million shares, ~$1.5 billion) vest only if cumulative revenue exceeds $2.0 billion by the same deadline. Revenue accumulation begins at close (February 2026), and with meaningful revenue not expected until H2 FY2028, the effective window for reaching $2 billion in cumulative revenue is approximately 18 months—a highly aggressive target requiring outstanding commercial execution from multiple hyperscaler customers simultaneously. The earn-out structure creates several important dynamics. Celestial AI's founding team— CEO Dave Lazovsky, COO Preet Virk, and CTO Phil Winterbottom—joined Marvell as employees and are directly incentivized by these milestones, aligning them with Marvell's commercial execution goals. The equity settlement means no cash burden on Marvell's balance sheet but creates potential dilution risk for MRVL shareholders proportional to how much of the earn-out vests. An important strategic nuance is the Amazon warrant issued simultaneously: Marvell issued Amazon approximately 1,045,171 MRVL shares at a strike price of $87.0029 per share, exercisable based on Amazon's cumulative purchases of Photonic Fabric products through December 2030. This warrant functions as an indirect demand signal: if Amazon exercises the warrant in full, it implies Amazon has committed to Photonic Fabric product purchases large enough to warrant the vesting threshold, providing a structural corroboration for the earn-out achievability from Celestial AI's largest likely customer. The 41% contingent component (at announcement pricing) is at the high end for semiconductor acquisitions of this size in recent history, reflecting the combination of transformational upside potential and substantial pre-revenue uncertainty. This structure is consistent with best-practice M&A deal design for pre-commercial technology assets: it limits acquirer downside while providing seller upside if technology proves out. From a governance perspective, the contingent consideration also reduces the risk of reputational damage to Marvell management if the photonics ramp underperforms. [CV002, CV014, CV003, CV004, CV005, CV006]

8.6 Valuation Verdict and Investment Diligence Stance

Synthesizing deal structure, comparable transactions, market sizing, and risk factors, the valuation verdict for the Celestial AI acquisition is: the $3.25 billion upfront price is fair for Marvell's strategic purposes, and the full $5.5 billion earn-out scenario would represent an attractive outcome. From the perspective of a hypothetical external observer benchmarking entry price against risk-adjusted forward value, the stance is research-more pending first-revenue validation in Marvell's FY2028 earnings. Three pillars support the fair verdict on the upfront price. First, the 6.5× two-year forward ARR multiple is conservative relative to comparable photonics and AI silicon M&A precedents—the Inphi deal at 14.6× TTM and the Acacia deal at ~12× TTM frame the upper end, while Mellanox at 5× TTM on a revenue-generating asset frames the lower end. Celestial AI's implied forward multiple sits between these poles without revenue risk adjustment. Second, the earn-out structure demonstrably limits Marvell's downside: the bear-case terminal settlement is $3.25 billion, not $5.5 billion. Third, NVIDIA's $2 billion NVLink Fusion investment (announced March 31, 2026) specifically cited silicon photonics and Marvell's optical interconnect capabilities—providing independent third-party corroboration that the Photonic Fabric technology is commercially credible and will be integrated into the dominant AI infrastructure platform. Three factors argue for caution and support a research-more posture for outside observers. First, Celestial AI has zero confirmed revenue at acquisition, and Marvell's forward guidance carries standard forward-looking-statement caveats without contractual revenue commitments from hyperscaler customers. Second, competitive risk from Broadcom's internal photonics program, NVIDIA's own silicon photonics R&D, and Intel's optical compute interconnect program is material and could erode the addressable market faster than the 18-month ramp window allows. Third, Marvell's customer concentration— approximately 73% of revenue from data center customers with over 50% estimated from two hyperscalers—means any hyperscaler CapEx deceleration would simultaneously compress Marvell's core revenue trajectory and Celestial AI's earn-out ramp. The thesis-break trigger is identifiable: failure to show any disclosed Celestial AI revenue contribution in Marvell's FY2028 earnings (reporting in early calendar 2028) would signal a ramp delay of at least 12 months, rendering the $500 million Q4 FY2028 ARR target unachievable and raising earn-out forfeiture risk. That single data point— the first earnings call with a Celestial AI revenue line—is the most important near-term validation event for the valuation thesis. [CV007, CV008, CV022, CV009, CV010, CV011]

8.7 Exhibits