Celestial AI

Phase A — Understand the business

Company Overview

Celestial AI (Santa Clara, CA; founded April 2020 by Dave Lazovsky, CEO, and Preet Virk, COO) builds the Photonic Fabric™ — a silicon-photonics optical-interconnect platform whose job is to break the AI "memory wall" and the copper "scale-up wall." Plain-terms business model: it sells optical I/O as IP + chiplets + interposers + a switch ASIC, co-packaged into a customer's XPU/accelerator and into the scale-up switch, so that thousands of accelerators behave like one "virtual Super XPU" with all-to-all optical connectivity, and so HBM/DDR5 memory can be disaggregated off-package and addressed optically.

Two product lines:

• Photonic Fabric interconnect — optical link for XPU↔XPU, XPU↔switch, XPU↔memory. The August 2025 Photonic Fabric Module (PFM) is billed as the world's first SoC with optical I/O integrated in the middle of the die (not just at the beachfront edge), on a TSMC 5nm ASIC + photonic interposer (PIC) + fiber array unit in a 2.5D package, ~7.2 Tbps per module.
• Optical memory expansion — a module fusing 2× HBMHigh-bandwidth memoryMemory stacked in layers beside a processor so data reaches it faster. The usual bottleneck in AI chips. stacks (~72 GB) with 4× DDR5 DIMMs (up to ~2 TB), bridged by a 5nm switch ASIC that turns HBM into a direct-write cache for DDR5 — "HBM bandwidth, DDR capacity/cost".

Customers: undisclosed by name, but the company claims a design-win with "one of the world's largest hyperscalers" who will co-package PF chiplets into its next-gen custom XPUs and scale-up switches; AMD is both a strategic (AMD Ventures backer) and an architecture-level interlocutor. Post-acquisition, Marvell named Amazon as the initial customer for the Celestial assets. Suppliers/partners: TSMC (5nm + advanced packaging), imec (venture support / process), custom ESD IP partners on TSMC 5nm. Contract structure: semiconductor design-win economics — long design-in cycle, revenue lagging tape-out by ~2 years; no recurring/take-or-pay. Concentration is extreme (single anchor hyperscaler) — see Lens 13.

Supply Chain

Upstream → Celestial → end customer, with named stakeholders:

• Process / wafers: TSMC — 5nm logic for the switch ASIC; advanced 2.5D packaging (PIC + ASIC + fiber array). The whole roadmap is TSMC-gated. imec provided early process/venture support.
• IP inputs: Rockley Photonics patent estate (optical switch, electro-absorption modulators/EAMs, optoelectronic system-in-package) — acquired outright by Celestial for $20M, Oct 2024 — plus internal patents (>200 globally). EAM modulation is the technical choice that gives thermal stability (see Lens 3).
• Optical assembly: fiber-array-unit coupling, photonic interposer assembly — historically the silicon-photonics chokepoint (fiber-to-chip coupling is low-throughput, high-cost).
• The company itself: designs the ASIC + PIC, sells as chiplet/IP/module.
• Downstream buyers: Tier-1 hyperscaler(s) (now routed through Marvell's custom-silicon engine post-close); the chiplets go into the customer's XPU package and the scale-up switch.

Chokepoints / single-source dependencies: (1) TSMC for both 5nm and the advanced packaging — a true single point; (2) the anchor hyperscaler — one customer is most of the early revenue; (3) fiber-coupling/assembly yield — the historically hard, non-commoditized step in silicon photonics, and the place a "works in the lab" demo dies at volume. Post-acquisition, Marvell's own packaging/supply relationships and balance sheet de-risk (1) and (3) materially — that is much of the deal's industrial logic.

Competitive Advantages (moats)

The differentiator bulls pay for: Celestial's EAM-based Photonic Fabric is thermally stable enough to sit in the middle of a high-TDP (multi-kilowatt) ASIC die, not just at the cooler edge "beachfront." Most co-packaged-optics (CPO) approaches — including ring-resonator designs — are temperature-sensitive and have to live at the package edge; Celestial claims it can embed optical I/O anywhere on the die, freeing beachfront and raising achievable off-package bandwidth. It claims a roadmap to 390 Tbps off-package, >20× current CPO state of the art, and >2× the power efficiency of copper with far longer reach.

Durable moats:

• IP / process know-how — >200 patents + the Rockley estate; EAM-in-silicon manufacturability is genuinely hard and is the stated barrier ("make optics manufacturable, testable, at volume in hyperscaler environments").
• Design-in switching cost — once a hyperscaler co-designs PF chiplets into its XPU and switch silicon, ripping it out is a multi-year re-spin. This is the strongest moat: it converts a single design-win into multi-generation lock-in.
• Time / "10 years ahead" claim — marketing, discount heavily; the field is moving fast and the claim is unverifiable.

Bargaining power: weak as a standalone private (one hyperscaler held most of the leverage; that asymmetry is exactly why selling to Marvell made sense). Stronger inside Marvell — bundled with Marvell's custom-ASIC franchise (Trainium, Maia IP, Axion), Celestial's optics become a feature of a sticky platform rather than a point product begging for a socket.

Segments

n/a — private, not disclosed. No segment revenue/EBITDA exists (our figures is an empty stub; company never reported). Qualitatively the business is pre-revenue / nominal revenue with two product vectors (interconnect; optical memory expansion) and effectively one geography of demand (US hyperscalers). Marvell's modeled ramp is the only forward "segment" signal: $0 meaningful revenue until 2H FY28 → ~$500M annualized run-rate exiting FY28 → ~$1B annualized run-rate exiting FY29. Treat those as acquirer projections, not actuals.

Phase B — Measure performance

+private overlay applied: Lens 5 → Funding & valuation trajectory; Lens 7 → Cap table & secondary marks; plus a Traction & unit-economics note. Lens 6 → founder interviews; Lens 8 → funding/product/deal events.

Funding & Valuation Trajectory

Round history, seed → exit (all, unaudited):

Total private capital raised since founding: >$515M. The valuation step from a $2.5B private round (Mar 2025) to a $3.25B upfront cash+stock acquisition (Dec 2025) is a ~30% markup in ~9 months on the upfront alone, before the earnout — a clean, fast result for late-stage backers.

Deal mechanics (the numbers that matter): upfront ≈ $1.0B cash + 27.2M MRVL shares (~$2.25B stock); contingent up to 27.2M more MRVL shares ($2.25B) on revenue milestones — ⅓ of the earnout vests at ≥$500M cumulative Celestial revenue by end of Marvell FY29, the full earnout at >$2.0B cumulative by end of FY29. At close Marvell issued 24,601,976 shares and assumed ~3M Celestial options (≈+27M diluted shares).

Founder/Management Commentary

No earnings calls (private). Signal comes from founder interviews and conference talks. Recurring, consistent messaging across 2024–2025 (Laser Focus World, eeNews Europe, ITF World, Hot Chips): the thesis is "the memory wall and the scale-up wall are the binding constraints on AI, and only optics fixes them"; the repeated proof-point is manufacturability ("not just making optics work — making them testable and at volume in hyperscaler environments"). Tone shifted over 2024→2025 from capability demo ("we proved the link in summer 2023") to commercial traction ("tape-out of production-grade chiplets; a Tier-1 hyperscaler design-win"). The most informative "sentiment" event is the founders/board agreeing to sell rather than push for an IPO in Dec 2025 — a tell that the team judged the standalone path (one customer, huge capex-to-volume, optical yield risk) riskier than a Marvell exit. Read that as honest risk-pricing, not weakness.

Cap Table & Secondary Marks / Mechanism Comps

Syndicate quality (an IPO-proximity / validation tell): Celestial pulled multiple crossover funds — Fidelity (lead C1), BlackRock, T. Rowe-adjacent late money via the category, Tiger Global, plus strategics AMD Ventures, Koch, Temasek/Xora, Porsche SE and Lip-Bu Tan personally. A Fidelity-led round at $2.5B is normally a "12–24 months from a tradeable event" signal — here the event turned out to be M&A, not S-1.

Mechanism / peer comps (optical interconnect, by valuation — ):

Read: Celestial's upfront ($3.25B) sits below Lightmatter's last private mark ($4.4B) and near Ayar's ($3.75B) — but Celestial holders also get a $2.25B earnout and liquidity now, while Lightmatter/Ayar holders still carry IPO/execution risk on paper. The optical-interconnect cohort is clearly being marked $3–5B each; Celestial is the first to convert paper into a strategic's cash+stock.

Traction & unit economics (Phase-B add): revenue is n/a — private, not disclosed and almost certainly de minimis today (Marvell models first meaningful revenue 2H FY28). The "unit economics" that matter are the acquirer's: ~$50M/yr added non-GAAP opex now for ~$0 revenue until FY28, i.e. Marvell is funding 2+ years of burn for an option on $1B run-rate.

Catalysts / Events (funding & product, replaces Stock-Price Catalysts)

No public stock to chart, so the catalysts are funding/product/deal events (each materially re-rated the company's worth or de-risked the tech):

• Summer 2023 — Photonic Fabric silicon link validated → moved company from "concept" to "designed-in candidate".
• 2024-04 — $175M Series C (USIT) → scale to manufacturing.
• 2024-10 — Rockley patent buy ($20M) → IP moat hardened.
• 2025-02/03 — $250M Series C1 at $2.5B (Fidelity) → crossover validation.
• 2025-08 — Photonic Fabric Module at Hot Chips (in-die optical I/O) → flagship product proof.
• 2025-12-02 — Marvell acquisition announced (up to $5.5B) → the terminal catalyst.
• 2026-02-02 — Deal closed → Celestial ceases to be an independent entity; thesis migrates to MRVL.
• Forward (now MRVL catalysts): first silicon tape-outs with Celestial IP by end-2026; early-access samples to clouds early 2027; revenue 2H FY28; earnout test end FY29.

Phase C — Judge people & books

Management

• Dave Lazovsky (Founder/CEO). 30 yrs semis; ran >$1B of equipment business at Applied Materials; founded Intermolecular (2004), took it public on NASDAQ (2011). Track-record caveat — material: Intermolecular IPO'd at $10/share (Nov 2011, below the $12–14 range) and was ultimately sold to Merck KGaA for $1.20/share in 2019 (~$62M equity value) — an ~88% destruction of IPO-price value over its public life. He left Intermolecular in 2014, did a stint as Khosla Ventures Venture Partner, then founded Celestial. So the pattern is: technically credible operator, strong fundraiser, but his one prior public company was a poor equity outcome. Celestial is the redemption data-point — a clean ~$3.25B+ strategic exit — but it is an acquisition, where the moat/lock-in did the work, not years of public-market compounding.
• Preet Virk (COO). Ex-MACOM, ex-Mindspeed; drove early 100G electro-optical transceiver adoption — directly relevant operating pedigree for optical I/O at scale.
• Phil Winterbottom (CTO). 10+ yrs Bell Labs; photonics + AI hardware + ASIC architecture.
• Capital allocation: as a private, the notable decision was the $20M Rockley IP buy (cheap, accretive to the moat) and the decision to sell rather than IPO — defensible given the risk profile. No buybacks/dividends frame applies.
• Skin in the game / founder archetype: founder-led, deep-tech archetype; founders + early VCs converted to MRVL cash+stock + an earnout that keeps them aligned to the revenue ramp (the ⅓/full milestones explicitly tie payout to delivering $500M–$2B cumulative revenue) — a well-structured retention/alignment mechanism.
• Red flags: the Intermolecular equity outcome (above); standard private-company opacity (no audited financials sourced); "10 years ahead" promotional framing typical of the category. Nothing fraudulent surfaced.

Forensic Red Flags

As a forensic analyst, with the caveat that no audited financials exist (our figures empty; no 10-KForm 10-KA company’s audited annual report to the US regulator. The most complete thing it publishes.; private). There is therefore no income statement, balance sheet, or cash-flow statement to forensically test — that absence is itself the headline limitation. What can be said:

• Revenue recognition: not testable; revenue is ~nil pre-FY28. The risk migrates to Marvell, where Celestial revenue will be recognized inside MRVL's data-center segment and the earnout milestones create an incentive to pull-in/recognize "cumulative revenue" aggressively to hit the $500M/$2B gates by FY29 — a thing to watch on MRVL's filings, not Celestial's.
• Capitalized R&D / intangibles: post-close, Marvell will book substantial goodwill + intangibles on a pre-revenue asset (~$3.25B+ for ~$0 current revenue) → future impairment risk if the ramp slips. This is the single biggest accounting red flag, and it lands on MRVL's balance sheet.
• SBC / dilution: ~3M Celestial options assumed + 24.6M shares issued + up to 27.2M earnout shares → real MRVL dilutionDilutionIssuing new shares, so each existing share owns a smaller slice of the same company.; UBS modeled ~$0.20 FY27 / ~$0.16 FY28 EPS drag.

Regulatory findings (required sub-section).

• SEC (EDGAR LR + AAER): total_sec_findings: 0 — Celestial has no CIK, is private, and is not required to file; no EDGAR enforcement search is possible.
• Non-SEC (web search): ran "Celestial AI" (FTC OR DOJ OR lawsuit OR litigation OR settlement OR fine OR penalty OR "consent decree") enforcement — no enforcement action, litigation, or settlement naming Celestial AI surfaced. Hits were unrelated FTC "Operation AI Comply" cases against other companies (Cleo AI, DoNotPay, etc.).
• Item 3 (Legal Proceedings): n/a — no 10-K exists (private; never filed).
• Conclusion: No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR, AAER, 0 hits), web search (no hits), and the absence of any 10-K Item 3, as of 2026-06-18. Unaudited per public sources.

Phase D — Project & stress-test

IPO-Readiness & Path-to-Tradeable

This lens resolves to a closed event, which is the whole point. Celestial will not IPO — the path-to-tradeable already executed as M&A: Marvell acquired it, closed 2026-02-02. There is no S-1 to wait for and no private-watch.json entry to update (the name is no longer a private-frontier watch — it is inside MRVL). The relevant "tradeable" is MRVL common stock, and the question flips to: was this a good acquisition, and does the embedded option pay off?

The milestone ladder is the forward model (all, treat as acquirer guidance):

• Base path — Celestial hits ~$500M annualized run-rate exiting FY28 and ~$1B exiting FY29; clears the ⅓ earnout ($500M cumulative by FY29) but likely not the full $2.0B cumulative gate → Marvell pays ~$1B cash + $2.25B stock upfront + ⅓ of the $2.25B earnout ($0.75B) ≈ ~$4B all-in.
• Bull path — design-win proliferates beyond Amazon to a 2nd/3rd hyperscaler, optical scale-up ramps with Rubin-class systems late-decade → full $5.5B earned, and the optics become a durable Marvell franchise feeding a $39B-by-2030 CPO TAM.
• Bear path — optical yield/assembly stalls volume, or the anchor hyperscaler designs it out (cf. Marvell's Trainium3 chiplet loss to a monolithic die / Alchip) → revenue misses FY28, no earnout, and Marvell impairs a chunk of the ~$3.25B goodwill.

No our model create in --watchlist mode (per SKILL). If logged later, the scoreable binary would be: "Celestial AI cumulative revenue ≥ $500M by end of Marvell FY29 (Jan 2029)" — the ⅓-earnout gate — rather than an EPS line.

Bull vs Bear

Bull case. AI has blown past the single rack; multi-rack scale-up is the next binding bottleneck and copper is out of headroom on reach/power. Optics is the only physics that scales, and Celestial owns the hardest version of it — in-die, thermally-stable EAM optical I/O with a credible 390 Tbps roadmap and a >200-patent + Rockley moat. A Tier-1 hyperscaler already designed it in; switching cost locks multi-generation revenue. Inside Marvell — which already co-designs the industry's custom XPUs (Trainium, Maia IP, Axion) and shares ~95% of the custom-ASIC market with Broadcom — Celestial's optics stop being a point product and become a feature of a sticky platform. The CPO/NPO TAM goes ~$100M (2025) → $39B+ (2030) with scale-up dominating; a single-digit share of that funds the earnout several times over.

Bear case (permanent-impairment risks). (1) Optical manufacturability at volume — fiber coupling and PIC assembly are the historical graveyards of silicon photonics; "works at Hot Chips" ≠ "yields at hyperscale," and the entire $500M-by-FY28 ramp assumes the hard step gets solved on schedule. (2) Single-customer / design-out risk — early revenue is ~one hyperscaler; Marvell just lost Trainium3 to a monolithic-die competitor (Alchip), proving these sockets are not safe. (3) The acquirer overpaid for an option — ~$3.25B+ for ~$0 current revenue means goodwill/intangible impairment if the ramp slips even a year, plus ~$0.20 FY27 EPS dilution for no offsetting revenue until FY28. Pre-mortem (18 months out, thesis broke): it's mid-2027, early-access samples shipped but yield on the in-die optical assembly is stuck below production economics, the anchor hyperscaler hedges to an internal/Broadcom CPO path, FY28 revenue guidance gets cut, and MRVL takes a partial Celestial impairment — the deal is reframed from "memory-wall masterstroke" to "expensive call option." Contrarian view the market is underpricing: the consensus treats the Marvell deal as validation that Celestial's tech won; it may instead be validation that standalone optical-interconnect startups can't reach volume alone — i.e. the exit is bullish for the category's necessity but bearish for the independence/economics of every other name (Lightmatter, Ayar) still trying to IPO into this.

Devil's Advocate (short-seller)

Dismantling the bull case (the bet is now MRVL): Where revenue is concentrated — essentially one hyperscaler at first, gated through Amazon; lose or delay that socket and the FY28 ramp evaporates. Why the moat may be weaker than bulls think — Nvidia (Spectrum-X/Quantum-X silicon photonics on TSMC SoIC) and Broadcom (Bailly CPO + Tomahawk Ultra/Scale-Up Ethernet) are pushing their own CPO; the anchor customer can co-design around Celestial, and ring-resonator/edge-CPO may be "good enough" for many sockets without paying for in-die EAM complexity. Most dangerous competitor bulls underestimate — not Lightmatter/Ayar, but the hyperscalers' own internal optics + Broadcom's open ecosystem, which can commoditize the interconnect layer. Worst capital-allocation read — Marvell paying $3.25B+ cash+stock for a pre-revenue asset, diluting ~27M shares, taking $0.20 EPS drag, and parking $3B+ of goodwill on a single-customer option. Assumptions that must hold for today's MRVL price — that custom-ASIC + optical scale-up both compound through Rubin-class systems and that Celestial yields on schedule. If Celestial revenue disappoints by 20–30% — the earnout doesn't pay (a good outcome for MRVL cash, but a signal the thesis is breaking), and the upfront goodwill faces impairment. Single scenario that permanently impairs — in-die optical assembly never reaches production yield/cost and the category settles on edge-CPO; Celestial becomes stranded IP. Plausibility: moderate — silicon-photonics-at-volume is genuinely unsolved, but Marvell's packaging muscle + a real design-win lowers it from "high."

Management Questions (ordered by information value)

(For Marvell management re: the Celestial assets, since Celestial is now a Marvell business.)

1. What is the current qualified yield on the in-die optical assembly (PIC + fiber array), and what yield is required for the FY28 $500M run-rate to be gross-margin-accretive?
2. Is the anchor hyperscaler design-win contractually committed to volume, or still a design-in that can be cancelled before production?
3. Beyond Amazon, how many additional hyperscaler/XPU programs have PF chiplets designed in today, and at what stage?
4. What cumulative Celestial revenue do you actually expect by end-FY29 against the $500M (⅓) and $2.0B (full) earnout gates — and what's the probability you assign each?
5. How does Photonic Fabric coexist or compete inside accounts that are simultaneously evaluating Nvidia's and Broadcom's CPO?
6. What is the capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. and packaging-capacity commitment (TSMC and otherwise) required to take this from samples to $1B run-rate, and is it secured?
7. Under what revenue/yield scenario would you impair the Celestial goodwill, and what's the trigger timeline?
8. How much of the in-die EAM advantage is durable IP vs. a process lead competitors close in 2–3 years?
9. What share of the $39B-by-2030 CPO TAM does your model assume Celestial captures, and is that scale-up, scale-out, or both?
10. How are the founders/key engineers retained and incentivized post-close beyond the earnout — what's the attrition risk on the photonics team?
11. What is the realistic first-revenue date if the current lead program slips one cycle?
12. How does the Trainium3 chiplet loss to a monolithic-die competitor change your confidence in chiplet-based optical I/O winning sockets?
13. What's the power-per-bit and reach advantage in shipping silicon today vs. the marketing roadmap (390 Tbps), and when does roadmap become product?
14. Does the memory-disaggregation product (HBM+DDR5 module) have a committed customer, or is the near-term story interconnect-only?
15. What did you learn in diligence about Celestial's manufacturability that the market doesn't yet appreciate — bull or bear?