CScale
Scale-up optical interconnect — a Palo Alto systems-silicon start-up exiting stealth with a thesis that the binding constraint on gigawatt-class AI factories is link reliability, not raw bandwidth
Thesis in one line: CScale is a bet that in gigawatt-class AI clusters, where a single scale-up domain spans thousands of accelerators across dozens of racks, optical failure must be treated as a continuity problem for the fabric rather than a field-replacement problem for the operator. There is no observable financial or product evidence yet; the underwriting rationale rests almost entirely on team pedigree, syndicate quality, and the identity of the strategic investors.
- Valuation, revenue, customers and ship dates are all undisclosed. The primary source (company press release, Sep 30, 2026) contains no valuation, revenue, named customer, launch timing, or performance specification (bandwidth, latency, power, port count). Every qualitative judgement below is made against that blank.
- Prior financing history is opaque. The round is labelled Series C, yet the timing, size and participants of earlier rounds have not been disclosed. The “~$43M raised pre-round” figure is derived arithmetic (company-stated $188M total less the $145M round), not a company disclosure.
- Lead attribution varies by outlet. The press release describes a three-way co-lead (Atreides, Valor, Premji Invest); some trade coverage describes Atreides as sole lead. This report follows the primary source.
- Single-source product detail. The suggestion that CScale will ship a system-on-chip with an on-board CPU rests on a job posting cited by SiliconANGLE; it is not confirmed by the company.
- Peer data are third-party. Marvell/Celestial AI figures come from SEC filings; Ayar Labs, Lightmatter and NVIDIA figures come from trade-press reporting. None are independently audited.
CScale was founded in 2023 in Palo Alto, California by Sanjai Kohli, who serves as founder and CTO. Martin Lund, a three-decade veteran of silicon and networking, is CEO. The company operated in stealth for roughly three years before formally launching on September 30, 2026. The pairing of a founder-CTO who owns architectural direction with a CEO drawn from large-cap networking silicon is a familiar governance template in systems semiconductors: technical vision and commercial execution are split across two principals.
Serial communications-semiconductor founder. Co-founded SiRF, which brought GPS to mass-market consumer devices, and shared the 2010 European Inventor Award from the European Patent Office for the underlying GPS receiver invention. He subsequently founded Inovi, acquired by Facebook in 2014. The company frames him as a multi-exit operator in communications. Public reference sources (Wikipedia) also list a tenure as CEO of Wave Computing (now MIPS Technologies) from September 2019; that item does not appear in the press release and should be treated as unverified. Our read: Kohli is the architectural author of CScale’s reliability-first interconnect philosophy.
Thirty years across Cisco, Microsoft, Cadence, Broadcom and Intel. Credited with building Broadcom’s switching business into a billion-dollar franchise; most recently led Cisco’s Common Hardware Group, which covers silicon, hardware systems and optics, including Silicon One. That résumé maps directly onto the commercialization gating items for CScale: foundry and packaging relationships, switch-ASIC and XPU ecosystem access, and design-in negotiations with hyperscalers.
Investor commentary stresses depth across photonics, electronics and software. Headcount is low relative to a $145M round, which we read as capital earmarked for hiring and for pre-production and validation spend (analyst inference).
Dependence on two principals is high, and no additional C-suite (CFO, manufacturing or operations lead) has been disclosed. Whether operations and supply-chain leadership is added as the company approaches volume is a watch item.
CScale is developing an integrated light engine for AI scale-up connectivity, the high-bandwidth, ultra-low-latency fabric that lets thousands of accelerators behave as a single, much larger computer. We classify the company as pre-commercial. Management says proceeds will fund technology development and commercialization; no product specification, customer, or shipment timeline has been published.
(2023→2026)
(NVIDIA, Intel Capital)
Market context: The company expects gigawatt-class AI data centers by the end of the decade, in which a single scale-up domain couples thousands of accelerators across dozens of racks. Copper faces physical limits on bandwidth, reach and power in that regime, and investor commentary consistently treats optics as a necessity rather than an option. This is also a segment where strategic value is being repriced in real time: Marvell closed its acquisition of Celestial AI in February 2026.
An integrated optical engine that moves data between accelerators over light instead of copper, targeting reliable connectivity at gigawatt scale. A job posting cited by SiliconANGLE hints at an SoC form factor with an on-board CPU; unconfirmed.
An interconnect architected to contain optical failures without interrupting compute, optimizing for continuity rather than ease of part replacement. The mechanism (link redundancy, laser sparing, etc.) has not been disclosed.
Premji Invest highlighted the team’s combined photonics, electronics and software expertise and its deployment experience as the basis for turning optical capability into dependable infrastructure; system-level architectural judgement is the stated differentiator.
Note: the three cards above are the analyst’s structuring of company-described attributes; CScale has not announced a formal product line-up.
CScale has raised $188M in total, of which $145M (roughly 77%) arrived in this Series C. The shape is a capital-efficient formation phase funded by a small group during stealth, followed by a large growth-capital injection at public launch. Two features stand out: this is the company’s first public financing event, and NVIDIA and Intel Capital joined as its first strategic investors.
Founded by Sanjai Kohli in 2023. Sutter Hill Ventures has backed the company since inception and continued to support it through the Series C. Size, valuation and syndicate composition of the initial financing are not public.
Subtracting the $145M Series C from the stated $188M total implies roughly $43M raised before this round. The company has not disclosed round names, timing or participants for this period; Maverick Silicon is the one confirmed existing investor. This is the largest information gap in the capital stack, and the pricing of these rounds is the key variable for sizing the step-up embedded in the Series C.
Structure: Co-led by Atreides Management, Valor Equity Partners and Premji Invest, with Sutter Hill Ventures and existing investor Maverick Silicon participating. NVIDIA and Intel Capital came in as the company’s first strategic investors. Valuation was not disclosed.
Use of proceeds: Acceleration of development and commercialization of the optical interconnect. Allocation across R&D, foundry and packaging pre-commitments, and hiring is not disclosed.
Investor signaling: Atreides’ Gavin Baker framed the thesis around system reliability, arguing that at gigawatt scale optical failures are inevitable and that interconnect bandwidth is worth little without it. Premji Invest’s Sandesh Patnam pointed to the team’s architectural judgement across photonics, electronics and software, and to the discipline to build an enduring company.
Analyst view: A syndicate pairing a crossover/hedge-fund-style investor (Atreides), a growth equity house (Valor) and long-duration capital (Premji Invest) signals capital-markets access at a later pre-IPO or strategic-exit stage. That said, NVIDIA’s dual role as strategic investor and potential customer, with its own platform roadmap, introduces customer-investor circularity that warrants monitoring.
At this stage CScale’s edge is not a demonstrated product moat but a set of leading indicators across three axes: (1) a differentiated design philosophy, (2) a management team with commercialization credentials, and (3) ecosystem-aligned capital. The assessments below are analyst judgements built on public information and should be revisited once product data emerge.
Most optical interconnect competitors lead with bandwidth and power efficiency. CScale instead defines optical failure as a compute-continuity problem and absorbs it at the architecture level. Across fleets of hundreds of thousands of accelerators, even low per-link failure rates become a persistent source of degradation, so if the premise holds it translates directly into TCO and GPU-utilization arithmetic for customers. No empirical evidence has yet been published to support the claim.
Lund’s record building Broadcom’s switching franchise and leading Cisco’s hardware group, including Silicon One and optics, offers both credibility and access in design-in discussions with switch-ASIC vendors, XPU suppliers and hyperscalers. Kohli’s SiRF and Inovi history supplies an architecture-leadership and exit track record that underpins investor confidence.
The simultaneous entry of the dominant AI accelerator platform and a major semiconductor group’s venture arm is a partial external validation of technical direction and improves the odds of ecosystem access (reference-platform integration, design collaboration). The scope of any commercial or technical cooperation has not been disclosed.
Building the core team and technical base on roughly $43M (derived) before securing $145M suggests technical milestones were reached ahead of heavy cash burn. An organization of about 85 people now holds $188M in cumulative capital, giving it runway for foundry, packaging and validation investments ahead of volume production.
Celestial AI’s absorption into Marvell narrowed the field of independent scale-up optics vendors, raising the scarcity value of those that remain. With large semiconductor firms and hyperscalers seeking to internalize optical interconnect IP, CScale holds optionality as both a standalone supplier and a potential acquisition target (analyst interpretation, not a statement of company intent).
Integrating photonics, electronics and software supports delivery as an integrated light engine rather than as discrete optical components. System-level accountability is structurally consistent with a reliability-centered value proposition and can raise switching costs after a design win. As with the other claims, this can only be validated after product shipment.
Candidate re-rating triggers: (1) publication of product specifications (bandwidth, latency, power) and reliability data; (2) a first hyperscaler or XPU-vendor design win; (3) confirmation of foundry and packaging partners; (4) concrete technical cooperation with NVIDIA or Intel; (5) valuation disclosure through a follow-on round or strategic transaction. None of these is observable today, and we classify the name as pre-evidence.
Technology and market timing favor CScale, but a wide capitalization-execution gap separates it from better-evidenced rivals: peers have either raised larger sums, reached production-oriented programs, or been folded into well-capitalized parents.
Raised $500M in March 2026 to scale co-packaged optics manufacturing and subsequently added $150M, bringing 2026 primary capital to $650M (trade press). Reported as NVIDIA-backed; pursuing TeraPHY chiplet reference designs.
$400M Series D in 2024 at a reported $4.4B valuation (led by T. Rowe Price). Its optical I/O chiplet and photonic interposer compete head-to-head with Ayar Labs in the CPO segment.
Acquired by Marvell at a $3.25B headline value, closed February 2, 2026 (SEC filings), with contingent milestone consideration. Illustrates both competitive intensity and the M&A value benchmark: an independent start-up absorbed into a large-cap roadmap.
Reported in March 2026 to be investing $2B each in Coherent and Lumentum ($4B total) to secure optical manufacturing capacity (The Register). A strategic investor in CScale, but how CScale fits its own platform roadmap is undisclosed.
Ayar Labs and Lightmatter are moving toward production, and Celestial AI had guided to a first hyperscaler product with volumes in early 2027 before its acquisition (trade-press reporting). CScale has disclosed no specifications, customers or shipment timeline, so it risks missing the design-in window as a later entrant. If hyperscaler supplier selection hardens early, barriers to entry rise sharply.
NVIDIA is simultaneously an investor, a prospective customer, and a platform owner building its own optical supply chain. Aligned incentives help; a strategy shift would introduce structural uncertainty into design-in decisions. The direction of scale-up standards (UALink and others) versus proprietary platform roadmaps also bears on the commercialization path.
The fault-containment value proposition is public only at the level of design philosophy. There is no fleet-scale data on MTBF, recovery latency, or overhead in power, area and cost. If the redundancy needed for reliability erodes power or cost competitiveness, the core value proposition could be offset.
Optical engines depend on photonics foundries, laser sourcing and advanced packaging capacity, areas where rivals and platform owners are racing to lock in supply. Valuation is likely to follow a binary structure keyed to design-win outcomes, and with the current mark undisclosed, pricing of any follow-on round is highly uncertain.