A Palo Alto startup most people had never heard of until this week just picked up a $145 million check with Nvidia and Intel attached to it. CScale, which exited stealth mode on September 30, 2026, says it has built a way to connect AI chips with fiber-optic cables instead of the copper wiring that has run inside server racks for decades. The Series C round, reported first by Reuters and The Economic Times, pushes the company’s total funding to $188 million since it was founded in 2023.

The timing matters. AI data centers are running into a wall that has nothing to do with GPU speed: the copper links that move data between accelerators inside a server rack can’t keep up with how fast those accelerators now compute. CScale’s bet is that optical interconnect, not another generation of faster copper, is what breaks the bottleneck. Nvidia and Intel Capital, two companies that normally compete rather than co-invest, both put money behind that bet.

Add Tech Insider once in the Google app and our stories appear in your news suggestions.

What CScale Actually Raised and Who’s Behind It

According to CScale’s own announcement and corroborating coverage from Unite.AI and The Next Web, the $145 million Series C was co-led by Atreides Management, Valor Equity Partners, and Premji Invest. Existing backers Sutter Hill Ventures and Maverick Silicon returned for the round, and two new strategic names joined as first-time investors: Nvidia and Intel Capital.

That combination is the headline here. Nvidia and Intel rarely appear on the same cap table, especially not for a networking startup that could eventually compete with pieces of both companies’ own interconnect roadmaps. Reports describe the new capital as earmarked for accelerating development and commercialization of CScale’s optical interconnect technology, with commercial shipments targeted for 2028 according to Reuters’ reporting.

CScale was founded in 2023 by Sanjai Kohli, who serves as chief technology officer. Kohli previously co-founded SiRF, a GPS chip company, and later founded Inovi, which Facebook acquired in 2014. The company’s chief executive, Martin Lund, joined after senior roles at Broadcom, Microsoft, and Cadence, where he was associated with building out Broadcom’s switching franchise into a major business line. That pedigree is part of why the round drew attention: this isn’t a pair of first-time founders pitching a slide deck, it’s a team that has already shipped networking silicon at scale once before.

The Copper Problem CScale Says It’s Solving

Every large AI training cluster is built from racks of GPUs or accelerators that need to talk to each other constantly, trading gradients and activations as they train a model together. For years, that intra-rack and intra-system communication has run over copper traces and cables because copper is cheap, reliable, and good enough at short range. The problem, as CScale’s own materials and multiple outlets covering the announcement describe it, is that accelerator clusters have grown so large and so power-hungry that copper is running out of runway. Signal loss over copper gets worse as data rates climb, which forces engineers to either shorten the distance between chips or spend more power pushing signal through the wire.

CScale’s pitch is to replace those copper links with fiber-optic connections inside the server itself, not just between buildings or between racks across a data center floor. That distinction, often called scale-up networking as opposed to scale-out networking, is why the company frames its target market narrowly: it isn’t trying to replace the Ethernet or InfiniBand fabric that connects whole clusters together. It’s trying to replace the wiring that lets a few thousand accelerators inside one AI supercomputer behave like a single, much larger machine, a description CScale uses directly in its own announcement.

Optical links, in theory, carry more bandwidth per connection over longer distances with less power loss than copper. That’s a known physics advantage, which is why photonics has been creeping toward the processor for over a decade. What has kept the technology out of mainstream AI servers so far is cost, manufacturing complexity, and reliability at scale. CScale’s investors are betting the company has cracked enough of that equation to matter before 2028.

Why Nvidia and Intel Both Wrote Checks

It’s worth sitting with how unusual it is for Nvidia and Intel to back the same networking startup. Nvidia already sells its own interconnect stack, including InfiniBand and the Spectrum-X Ethernet platform, both covered in tech-insider.org’s prior reporting on Nvidia’s hardware roadmap. Intel has its own networking silicon ambitions too. Strategic investment rounds like this one are rarely about near-term competitive threat; they’re about optionality. If optical interconnect becomes the standard way accelerators talk to each other inside a rack, both companies would rather have a seat at CScale’s table than get locked out of the technology entirely.

There’s also a simpler reading. Nvidia’s GPU shipments have outpaced the networking ecosystem’s ability to keep those GPUs fed with data at full speed, a dynamic that shows up across the AI chip supply chain. Reports on Broadcom’s AI chip revenue climbing to $16.7 billion and AMD’s data center revenue jumping to $6.7 billion both point to the same underlying trend: every major chip vendor’s AI business is now gated as much by system-level bottlenecks, including interconnect, as by raw compute output. Backing a company trying to solve the wiring problem is a hedge against that bottleneck becoming Nvidia’s own growth ceiling.

Martin Lund, CScale’s CEO, put the company’s design philosophy this way, as quoted in The Next Web’s coverage of the funding round: “We’re designing the interconnect for continuity. Lasers will fail. Compute shouldn’t.” That line gets at the core engineering challenge CScale is tackling. Optical components, including the lasers that convert electrical signals into light, are historically less reliable at data-center scale than copper connections. A single failed laser in a cluster with thousands of optical links could, in theory, take down an entire training run. Lund’s comment signals that CScale’s architecture is built around tolerating individual component failures without losing the whole system, which is a different engineering problem than simply making optical links faster.

Scale-Up Networking vs. Scale-Out Networking

The distinction between scale-up and scale-out networking is the single most important technical detail in CScale’s pitch, and it’s worth spelling out clearly because the two markets have very different competitive dynamics. Scale-out networking connects separate servers and racks to each other across a data center floor, or across buildings. That’s the realm of Ethernet switches, InfiniBand fabrics, and the kind of gear sold by Broadcom, Arista, and Cisco. Scale-up networking, by contrast, connects accelerators to each other within a single tightly coupled computing domain, often inside one server chassis or one rack, where latency has to be as close to zero as physically possible.

CScale is explicitly targeting the scale-up layer. That’s a narrower market than general data-center networking, but it’s also one where the physics problem is most acute, because the distances are shorter and the bandwidth demands per link are higher than almost anywhere else in a data center. It’s also a market segment where Nvidia’s own NVLink and similar proprietary interconnects currently dominate, which is part of why Nvidia’s investment reads as strategic hedging rather than pure financial upside-seeking.

The Competitive Field CScale Is Stepping Into

CScale is not the only company chasing optical interconnect for AI systems, and it’s worth being honest about how crowded this corner of the chip industry has become. Ayar Labs has spent years developing co-packaged optical I/O that puts light-based connections directly next to the compute die. Celestial AI is pursuing an optical fabric aimed specifically at linking memory and compute, a related but distinct problem from CScale’s scale-up networking focus. Lightmatter has pushed photonic interconnect technology from a different angle, with an emphasis on photonic computing more broadly.

Then there’s the incumbent layer: Broadcom, which has built a dominant position in Ethernet switch silicon and optical transceivers for data centers, and which reported AI-driven revenue growth that outpaces almost every other semiconductor company’s AI segment. Arista Networks and Cisco round out the traditional networking vendors with exposure to AI cluster buildouts, though neither is pursuing the specific scale-up optical replacement that CScale is targeting. Marvell and Astera Labs also sit adjacent to this market, building connectivity silicon such as PCIe and CXL infrastructure that AI accelerators depend on.

What separates CScale from most of this field, according to the reporting around its stealth exit, is the narrowness and specificity of its target: replacing copper inside the server, not selling switches or transceivers for the broader data-center fabric. That narrow focus is either a strength or a weakness depending on how the market evolves. If scale-up optical interconnect becomes a standard component that every AI server needs, CScale could become a default supplier the way merchant switch-silicon vendors became default suppliers for Ethernet networking. If hyperscalers instead build the capability in-house or fold it into GPU vendors’ own packaging, CScale’s addressable market shrinks fast.

A Brief History of Optical Interconnect’s Long Road to the Data Center

Optical interconnect isn’t a new idea. Fiber optics have carried long-haul telecom traffic since the 1980s, and optical links have connected data centers to each other, and often connected racks within the same data center, for well over a decade. What’s new is pushing optical connections further down the stack, closer to the chip itself, replacing the copper that has traditionally handled the “last few meters” or even the “last few centimeters” of a signal’s journey.

That push has accelerated specifically because of AI training workloads. A decade ago, the compute demands of a typical data-center application didn’t come close to saturating copper’s bandwidth ceiling at the distances involved inside a server. Large language model training changed that math within just a few years, as accelerator clusters scaled from hundreds of chips to tens of thousands, all needing to exchange enormous volumes of data continuously during training runs. The pressure that created is visible across the industry: Micron’s leadership has described memory supply running roughly 75% sold out heading into 2027 in reporting on the company’s outlook, a sign of how aggressively the entire AI hardware stack, not just networking, is straining against physical limits.

CScale’s emergence from stealth fits a broader pattern that has played out repeatedly in chip history: a bottleneck appears, a wave of well-funded startups chase it, and a handful survive to become standard infrastructure while the rest get acquired or fade. Photonics has been “five years away” from mainstream data-center adoption for most of the last fifteen years. What’s different now is the sheer dollar scale of AI infrastructure spending creating enough urgency, and enough capital, for optical-electrical convergence to actually happen rather than remain a research curiosity.

What the Funding Round Signals About the Broader AI Chip Market

CScale’s raise lands in the middle of a broader wave of AI infrastructure investment that has little precedent in the chip industry’s history. Synopsys recently signed a silicon IP deal with Amazon worth more than $1 billion, covered in tech-insider.org’s report on that agreement, underscoring how much money is flowing into the tooling and IP layer that sits underneath custom AI chip design. CoreWeave’s move to offer Nvidia’s Vera CPU, detailed in prior coverage of that rollout, shows cloud providers racing to lock in next-generation silicon ahead of competitors.

What makes CScale’s round distinct from most of that activity is the strategic-investor pairing. Most AI infrastructure deals this year have involved one dominant chip vendor backing a supplier or customer. Nvidia and Intel jointly backing a single startup is a far less common pattern, and it suggests both companies see enough uncertainty in how scale-up networking will shake out that neither wants to bet against the other’s preferred outcome. It also reflects how even companies that compete fiercely at the GPU and CPU layer, a dynamic visible in coverage of who actually makes GPUs beyond Nvidia and AMD, can still find common cause one layer down the stack, where a shared bottleneck threatens everyone’s roadmap equally.

The Risks CScale Still Has to Clear

A $145 million round and a roster of credible investors doesn’t guarantee CScale ships anything. The company’s own target of commercial chips by 2028, as reported by Reuters, leaves roughly two years for CScale to move from a funded concept to a manufacturable, qualified product that hyperscalers are willing to deploy inside production AI clusters. That’s an aggressive timeline for any hardware company, and it’s especially aggressive for optical interconnect, a category where reliability at scale has tripped up well-funded startups before.

No valuation for the round has been disclosed in any of the public reporting reviewed, which makes it hard to gauge how much pressure CScale is under to deliver versus how much runway its investors expect the $188 million total to buy. The absence of a disclosed total addressable market figure in CScale’s own materials or in coverage of the deal is also notable; the company and its investors appear to be betting on the category’s growth rather than pointing to an existing, sized market opportunity.

There’s also execution risk tied to the specific failure mode CScale’s design is meant to solve. Lund’s comment about lasers failing while compute shouldn’t implies the company has built redundancy into its architecture, but redundancy at the scale of a gigawatt-class AI cluster, the scale CScale references in its own stealth-exit announcement, is a genuinely hard distributed-systems problem on top of being a hard photonics problem. Both have to work simultaneously for the product to matter.

Five Predictions for CScale and the Optical Interconnect Race

First, expect at least one more strategic chip vendor, likely AMD or a major cloud provider, to make a similar investment in an optical interconnect startup within the next 12 months, following the pattern Nvidia and Intel just set. Second, CScale’s 2028 shipment target will likely slip by at least a few quarters, consistent with how most novel interconnect hardware timelines have played out historically. Third, expect consolidation among the smaller optical-interconnect players, Ayar Labs, Celestial AI, and Lightmatter among them, as capital concentrates around the two or three companies that demonstrate working silicon first rather than spreading evenly across the field. Fourth, Broadcom and the established Ethernet-switch vendors will likely respond by accelerating their own optical-module roadmaps rather than ceding the scale-up layer entirely, since the line between scale-up and scale-out is likely to blur as cluster architectures evolve. Fifth, expect CScale or a close competitor to announce a design partnership with a named hyperscaler within the next year, since that kind of validation is typically the next milestone venture-backed interconnect startups pursue after a funding round of this size.

How the Optical Switch Would Actually Work Inside a Server

Stripped of marketing language, the engineering problem CScale is tackling comes down to converting electrical signals into light at one end of a connection and back into electrical signals at the other, fast enough and cheaply enough that doing so makes more sense than just running a copper trace. That conversion happens through components called transceivers, which sit at the edge of a chip package or on a small daughter card near the accelerator. The tighter CScale can integrate that conversion step with the accelerator package itself, the less latency and power the system loses in the process, which is why companies in this space talk about “co-packaged optics” as the end goal rather than optical modules that simply plug into a server’s front panel the way a conventional network interface card does.

Reliability is the other half of the equation, and it’s the part Lund’s comment about lasers failing speaks to directly. A copper trace either works or it’s physically broken; there isn’t much in between. A laser degrades gradually, can drift with temperature, and can fail outright with far more varied failure modes than a copper wire. Building a scale-up fabric that can route around a failed optical link without stalling a training run that might involve thousands of accelerators working in lockstep is a software and systems problem layered on top of the photonics itself, and it’s one of the main reasons optical interconnect has taken this long to reach production AI servers despite the physics advantage being understood for years.

What Cloud Providers and Enterprise Buyers Should Watch Next

For hyperscalers and large enterprises planning multi-year AI infrastructure budgets, CScale’s raise is less an immediate buying decision than an early signal worth tracking. None of CScale’s technology is shipping yet, and the 2028 commercial target reported by Reuters means any production deployment is still years away. The more immediate signal for buyers is what Nvidia and Intel’s involvement implies about their own roadmaps: if both chip vendors are hedging on optical scale-up interconnect, buyers evaluating multi-year GPU cluster contracts should expect interconnect architecture, not just chip generation, to become a bigger part of vendor comparisons over the next few procurement cycles.

The more practical near-term watch point is design partnerships. Venture-backed interconnect startups at this funding stage typically need at least one named hyperscaler or large AI lab willing to test silicon in a real cluster before the technology can be considered de-risked. CScale has not announced such a partnership as of this writing. Enterprises and infrastructure teams tracking this space should treat the current round as a funding and talent signal rather than a procurement option, and revisit the category once CScale or a competitor announces actual silicon in a customer’s hands.

Why This Matters Beyond One Startup

CScale’s $145 million round is a small number by the standards of 2026’s AI infrastructure spending, where single data-center buildouts routinely run into the billions. But the round is a useful signal of where the next bottleneck in AI hardware sits. For the last several years, the public conversation about AI infrastructure constraints has centered on GPU supply, memory supply, and power availability. Reports on memory supply running 75% sold out for 2027 and ongoing memory crunch coverage have made that part of the story familiar. Networking inside the server, the literal wiring connecting one chip to the next, has gotten far less attention, even though it’s increasingly the thing standing between a cluster of expensive chips and those chips actually working together efficiently.

Nvidia and Intel’s joint bet on CScale is a signal that both companies think this bottleneck is real enough to hedge against, even while they each maintain their own competing interconnect strategies. Whether CScale becomes the company that solves it, or simply validates the category long enough for a better-funded competitor to win the market, the fact that this round happened at all tells you AI infrastructure’s constraints are shifting, chip by chip, deeper into the plumbing that nobody outside the industry usually thinks about.

Frequently Asked Questions

What is CScale and what does it make?

CScale is a Palo Alto startup founded in 2023 that is developing optical interconnect technology to replace the copper links currently used to connect AI accelerators inside data-center servers, according to the company’s own stealth-exit announcement and reporting from Reuters and The Economic Times.

How much funding did CScale raise, and when?

CScale announced a $145 million Series C round on September 30, 2026, bringing its total funding since founding to $188 million, per the company’s press materials.

Why did Nvidia and Intel both invest in CScale?

Both companies joined as new strategic investors in the round, according to reporting on the deal. Neither Nvidia nor Intel has publicly detailed its specific rationale beyond the funding announcement, but the investment gives both companies visibility into, and potential future access to, optical scale-up interconnect technology that could affect their own AI system roadmaps.

Who leads CScale?

Martin Lund serves as chief executive officer, having previously held senior roles at Broadcom, Microsoft, and Cadence. Sanjai Kohli, the company’s founder and chief technology officer, previously co-founded SiRF and founded Inovi, which Facebook acquired in 2014.

When will CScale ship a commercial product?

Reuters’ reporting on the funding round indicates CScale is targeting commercial chip shipments in 2028.

What is scale-up networking, and how is it different from scale-out networking?

Scale-up networking connects accelerators within a single tightly coupled computing domain, such as inside one server or rack, where latency requirements are extremely tight. Scale-out networking connects separate servers and racks across a data center floor or between buildings. CScale is targeting the scale-up layer specifically.

Who are CScale’s main competitors?

Companies pursuing related optical interconnect or AI networking technology include Ayar Labs, Celestial AI, and Lightmatter at the scale-up layer, and Broadcom, Arista Networks, and Cisco in broader data-center networking and optical transceivers.

Has CScale disclosed a company valuation?

No valuation has been disclosed in CScale’s announcement or in the reporting on the Series C round reviewed for this article.