CScale and the Five Conditions

On September 30, CScale exited stealth with $145 million in Series C funding, bringing its total funding to $188 million. NVIDIA and Intel Capital joined as the company’s first strategic investors. The company is building an optical interconnect for AI scale-up networks — the high-bandwidth, low-latency connections that let thousands of accelerators work together as one computer — aimed at gigawatt-class AI data centers. 1 2

What the announcement actually says

Strip away the funding and the announcement makes one technical claim: at the scale these data centers are heading for, optical links will fail often enough that failure handling becomes the design problem. CEO Martin Lund: “Lasers will fail. Compute shouldn’t.” The company says its interconnect is architected to contain optical failures without interrupting compute — an integrated light engine that keeps the compute running when a laser dies.1The market context is straightforward. Copper is fast and reliable but short-reach, which is why chip designers pack processors tightly and spend on liquid cooling. Replacing those copper links with light is a crowded field: Lightmatter, Ayar Labs, Celestial AI (now part of Marvell), Lightelligence, and others are already shipping or close to it. CScale is the new entrant, and its differentiator is reliability at fleet scale — the argument that a link that looks reliable in isolation becomes an operational problem when an AI factory contains enormous numbers of links. 2 3Three facts about the company, all from the announcement and coverage: founded 2023 (formerly named CSpeed); CEO Martin Lund, who led Cisco’s Common Hardware Group after building Broadcom’s switching business; founder and CTO Sanjai Kohli, who co-founded GPS-chip maker SiRF. About 85 employees in Palo Alto and four other sites. The round was led by Atreides Management, with Valor Equity Partners and Premji Invest as co-leads; Sutter Hill Ventures has backed the company since its inception, Maverick Silicon also participated, and no valuation was disclosed.1 3 5## What the announcement does not say

The company has disclosed no product, no specifications, and no price. Reuters reported that Lund declined to give specifics of the technology, and that CScale plans to ship its chips by 2028. 2SiliconANGLE found a job posting describing a system-on-chip with an onboard CPU, and reported that the company explained nothing about how the failure containment actually works. 4Fierce Network reported that proof-of-concept systems exist with customers under NDA. 3None of that is unusual for a stealth exit, and it is not criticism. It is the fact that determines how a capacity planner should read this news: there is nothing to measure yet. Funding, leadership pedigree, and strategic investors are signals of seriousness. They are not evidence of performance.

How the five conditions apply

The book’s evidence standard for emerging architectures asks for five kinds of evidence before an architecture enters a capacity plan: quality, boundary, comparison, supply, and obligations.(chapter 11) Applied to this announcement, every condition comes back empty:

Quality — no measured results on any workload. Proof-of-concept systems exist under NDA; nothing is published.

Boundary — the announcement describes the ambition (thousands of accelerators across dozens of racks) but no figure for what the interconnect covers, at what bandwidth, latency, or energy per bit.

Comparison — no measurement against the copper link, or against any competing optical link, that a buyer could check.

Supply — no price, no delivery date beyond a press-reported “ship by 2028,” no supported software, no repair process. On laser failure handling, the company’s whole pitch, the how is undisclosed.

Obligations — unknowable while the company is private and pre-product; supplier concentration risk is visible but not yet contractual.

So CScale goes on the watch list, with every condition named as missing. That is the honest reading, not a rejection. The five-condition standard exists to prevent two opposite mistakes: writing a check before anything is measurable, and dismissing a serious, well-funded effort because it hasn’t shipped. Funding and talent buy the right to keep working; they don’t buy evidence.

Two things would change the entry. A published specification — bandwidth, latency, energy per bit at the link boundary — would fill Boundary and start Comparison. Shipping product with a price and delivery date would fill Supply. The watch list should say what would change a decision and who will recognize the change: when a spec or a shipping product appears, review the entry; when Reuters’ 2028 date arrives, check whether the chips did. Until then, the funding tells you who believes in the problem; only measurements will tell you whether the answer works. When something ships and can be measured against the copper and optical alternatives, CScale becomes a candidate for the photonic section on the same terms as everyone else.

Sources

1. CScale press release, September 30, 2026: CScale Exits Stealth With $145 Million to Build the Interconnect for Gigawatt-Scale AI — funding, investors, quotes, company background.

2. Stephen Nellis, Reuters, September 30, 2026: CScale raises $145 million for AI networking, with backing from Nvidia and Intel — formerly named CSpeed; Lund declined technical specifics; ship-by-2028; copper-reach context.

3. Dan Jones, Fierce Network, October 5, 2026: CScale bets AI factories can ride out laser failures at gigawatt scale — competitive field (Ayar Labs, Lightmatter most comparable); proof-of-concept systems with customers under NDA; Lund’s background and the five-site footprint.

4. SiliconANGLE, September 30, 2026: Scale-up interconnect startup CScale launches with $188M in funding — round co-led by Atreides, Valor Equity Partners, Premji Invest; job posting describing a system-on-chip; no disclosed reliability mechanism.

5. CScale press release and Reuters coverage both confirm the full investor list: Atreides Management, Valor Equity Partners, Premji Invest, Sutter Hill Ventures (backing since inception), Maverick Silicon, and new strategic investors NVIDIA and Intel Capital. No valuation was disclosed.