Product1 publisher3 min readPublished
Celero raises $275m at a $3bn-plus valuation with its coherent DSP still short of production
The chip that carries the valuation has been validated on TSMC's 2nm process but not finalised for production, no customer has been named, and the three funds that set the new price were already shareholders.
The Product Desk · Product desk
What happened
- Celero Communications, two years old and building the chips that push data down fibre between AI chips and racks, raised $275m in a Series C that values it above $3bn, co-led by Atreides, Valor and CapitalG.
- CapitalG also led the $100m Series B last November, which took Celero's funding to $140m at the time, and that announcement disclosed no valuation, so the step up cannot be measured from outside.
- Celero says it has validated what it calls the industry's first coherent digital signal processor built on TSMC's 2nm process, with a roadmap running from 1.6 terabit connections to 3.2 terabit ones.
- Bloomberg reported that Celero will compete with Marvell, Broadcom and Cisco; its two founders are former Marvell executives, and chief executive Nariman Yousefi spent 17 years at Broadcom running its networking unit.
- Marvell last month gave Google a warrant to buy up to $12.2bn of its shares as part of a custom chip deal covering the controllers that sit around Google's TPUs.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint Because the chip has to fit inside equipment other firms build, a 2nm lead converts into revenue only at the pace of somebody else's module and switch refresh cycle.
- decision Buyers scoping optics for clusters two years out are weighing a part they can order against a promised date, which puts the weight on roadmap credibility rather than measured performance.
- exposure Alphabet capital now sits across the optical networking fight rather than behind one supplier, with CapitalG funding a challenger to a company its parent holds warrants in.
The person specifying optics for a cluster that lands in 2027 can put PAM4 modules on a purchase order this week [11]. Celero's part cannot go on that order: the design has still to be finalised, production is aimed at next year, and no buyer has been named [15][16].
The arithmetic reconciles cleanly. Total funding of $415m less this round's $275m leaves $140m, exactly the figure Celero reported after its Series B, so nothing was raised in between [1]. This one round is close to twice everything the company had taken before it, and 2.75 times the Series B that closed roughly ten months earlier [1][2]. The people who set the new price are the people who set the old one, since all three co-leads were already on the cap table [2]. The Series B carried no disclosed valuation, so how far the price actually moved is not something an outsider can check [7].
The thing being pitched is not quite the thing being bought. The pitch is that coherent optics, long reserved for links between cities, should now run at every scale down to racks inside one building, because PAM4 is running out of room as clusters grow [11][12]. What $275m buys is narrower: roadmap work, R&D and production readiness [15]. The strongest statement that networking is now the limit comes from chief executive Nariman Yousefi himself, in Bloomberg's interview and on LinkedIn afterwards [13][21]. The McKinsey figure Celero's own announcement cites, nearly $7tn of data centre investment by 2030 with about $5.2tn of it behind AI workloads, is a denominator rather than demand for coherent DSPs; AI accounts for about 74% of that projected total [22][3].
Its backers are not all-in on light, either. Atreides co-led a $700m round for Lumilens, which also builds optical interconnect for hops inside a rack and runs between racks [23], while Maverick Silicon, in Celero since the Series B, also backs Point2, which pushes radio signals over plastic and claims lower power and latency than optics [24][4]. Read together, that is conviction that the wiring between chips is where value moves, with no settled answer on which medium wins the short hops.
For any component vendor pitching into a build you own, two axes matter. One is whether the part drops into a socket you already buy or needs somebody else's module and switch redesign. The other is whether you can order it this quarter or are being sold a date. Celero sits in the corner where both answers are the harder one, which is an unremarkable place for a two-year-old chip company to be [1][15][17]. Until a part number exists, the observed behaviour holds: PAM4 keeps doing most of the work inside the building [11], and the $3bn prices a thesis; the bottleneck has not yet moved.
What to watch
- A named module or switch vendor designing in a Celero part, or a hyperscaler qualifying one, which would turn this from a funding story into a supply story.
- Whether Celero finalises the design and starts production inside the window Yousefi gave Bloomberg, namely next year.
- Any incumbent DSP supplier announcing coherent parts for links inside the building, which would test whether PAM4's limit is a real ceiling or a selling wedge.