The hardest problem in an AI data center right now isn't compute. It's cabling.

Here's the deal: AI infrastructure news is usually about chips. How many GPUs, how much compute, what HBM bandwidth. But among the people actually building these facilities, a different topic keeps surfacing — what do you connect all those chips with?

Point2 Technology co-founder and CEO Sean Park put it plainly in the company's announcement: "As AI systems scale and bandwidth demands reach terabit-per-second speeds, interconnect has become the defining bottleneck."

On August 10, the company said it had completed its Series B at a total of $136 million. The latest extension was led by LB Investment, with Arm joining as a new strategic investor and existing backer Maverick Silicon continuing to participate.

The cap table is the interesting part. Nvidia's venture arm NVentures, UMC Capital, Molex, and Bosch Ventures were already in; Arm now joins them. A GPU company, a foundry, a connector manufacturer, an automotive supplier, and a CPU architecture licensor all sitting on the same shareholder list tells you roughly where this technology is expected to land.

Point2: from KAIST to San Jose

Point2 Technology is a deep-tech startup spun out of KAIST, headquartered in San Jose, California. Technology from Korea, business from the US — a common structure in semiconductors.

Its first major moment of Korean visibility came in April 2026, when Nvidia's NVentures participated in a funding round, reported as the first time a KAIST deep-tech spinout received Nvidia investment. That round was led by Maverick Silicon with UMC also participating.

The Series B itself started much earlier. Its first public tranche was $23 million from Bosch Ventures and Molex, pitched around "multi-terabit interconnect for AI and automotive." Compared with today's AI-datacenter-first framing, automotive weighed more heavily at the outset.

The structure is worth noting: this round grew in stages. $23 million to start, an extension in April, another in August, arriving at $136 million total. Serial extensions are a familiar pattern in deep-tech hardware, where each technical milestone attracts a different strategic investor.

What e-Tube actually is — a third option that's neither copper nor fiber

To understand the technology, start with how cables work inside a data center.

Connecting chip to chip within a rack currently comes down to two choices. Copper is cheap and power-efficient but short-reach — signal attenuates fast as speeds climb, and at the highest rates you can't go more than a few meters. Optical goes much farther but requires converting electrical signal to light and back. That conversion costs power, costs money in components, and adds latency.

Point2's e-Tube is a third approach. It never converts to light. Instead it turns the electrical signal into RF — radio frequency — and sends it down a plastic waveguide. Radio inside a cable, effectively. Products built this way are called Active RF Cables, or ARCs.

The company's stated comparisons:

Versus e-Tube's claim
Copper 10x reach, 5x lower weight, 2x lower cable volume, comparable cost
Optics 3x lower power, 3x lower cost, 1000x lower latency

Read as a spec sheet, it appears to take the best of both. These are company-provided figures and independent verification is a separate matter — but the physics story is coherent. Removing the optical-electrical conversion removes its power draw and latency, and a plastic waveguide is simpler to manufacture than optical fiber.

The weight and volume claims aren't cosmetic either. A single AI rack carries hundreds of cables, and that bundle causes real physical problems. Weight constrains rack structural design; volume obstructs airflow and degrades cooling. Halving cable volume is a change that propagates through the whole data center design.

Two applications are named: accelerator-to-accelerator compute fabric, and rack-scale interconnect. Both sit exactly where AI infrastructure is most constrained today.

Who gets what

Point2 gets money and credibility together. Hardware startups face a long, expensive road to volume production, and $136 million buys distance on that road. Arguably more valuable is that the investor list functions as sales material. Being the interconnect company with Nvidia and Arm on the cap table meaningfully lowers the barrier to a first conversation with a data center customer.

Arm gets system-level visibility. It licenses CPU architecture, but it has spent recent years extending into full data center systems. Holding a stake in chip-to-chip connectivity is one way to earn a voice in rack-scale design.

Nvidia's calculus is more direct. It owns NVLink and designs rack-scale systems itself. More options at the cable physical layer means more design freedom. Taking small equity positions to hold technology options open is a familiar Nvidia move.

Molex and Bosch are the manufacturing and distribution angle. Molex is a major connector supplier; Bosch is automotive. High-speed in-vehicle data transport is also an e-Tube application, so those two investors are looking at a market that isn't the data center at all.

Korea's deep-tech ecosystem gains a reference path: technology out of KAIST, US headquarters, staged strategic investors. Whether that counts as a win is contested, since a US-domiciled parent means much of the eventual value accrues there.

Interconnect startups usually end in acquisition

Anyone who has watched this space knows the pattern: interconnect startups get acquired far more often than they grow into independent companies.

The largest case is Mellanox. The Israeli InfiniBand networking company was acquired by Nvidia for $6.9 billion in 2019, and that deal became a core pillar of Nvidia's data center business. It was decisive in Nvidia's transition from selling GPUs to selling entire racks.

Optical interconnect repeated the pattern. Silicon photonics startups were absorbed by Cisco, Marvell, and Broadcom in succession, and the reason was consistent: good technology, no viable independent route to market. Large incumbents already own the customer relationships in data center equipment.

There are failures too — companies that brought a new physical layer and vanished after losing a standards fight. Standards dominate this market. However good the technology, customers have no reason to adopt something that isn't in the spec, and cables in particular require equipment at both ends to support them, which makes ecosystem formation mandatory.

Two implications for Point2. Having large strategic investors on board is genuinely useful for standards and ecosystem positioning. Those same investors are also the most likely acquirers.

The competitive picture

Copper isn't standing still. Active Electrical Cables (AECs) — which embed signal-conditioning silicon in the cable to extend reach — have spread quickly. Companies like Credo grew on this, and AECs already ship in volume to hyperscale facilities. The realistic benchmark for e-Tube may be AECs rather than fiber.

Optics is answering with co-packaged optics (CPO), putting optical engines in the same package as the switch ASIC to cut power and latency. Broadcom and Nvidia have both shipped product and hyperscale deployment has begun. At longer reach, optics still wins.

Nvidia is simultaneously investor and competitor. NVLink and its extensions are becoming the de facto standard for in-rack connectivity. If e-Tube gets adopted as a physical layer within that, the opportunity is large; if not, market access itself becomes difficult.

Incumbent cable and connector vendors are ambivalent. Molex came in as an investor, but most have incentive to defend existing product lines, since a new physical layer displaces them.

Chinese suppliers are investing here as well. Data center interconnect overlaps with US export-control scope, which creates strong incentive for domestic alternatives — and raises the possibility of the market fragmenting regionally.

So what actually changes

If you design data centers, you have one more option on the horizon — but not today. The company hasn't disclosed volume production wins, and the sensible sequence is to wait for deployment evidence. Two things to watch: adoption by a major operator, and where this lands in industry standards work.

If you follow Korea's startup ecosystem, this is a template worth studying. Deep tech rarely reaches volume production on domestic capital alone, and Point2 chose to layer in strategic investors stage by stage. The shifting character of those investors — automotive, connectors, foundry, GPU, CPU architecture — is itself informative.

If you invest in AI infrastructure, this illustrates the bottleneck migrating. Capital moved from GPUs to memory and is now spreading into interconnect and power. Standards risk makes picking winners at this layer genuinely hard.

If you're an engineer, the approach is technically interesting: skipping optical-electrical conversion rearranges a familiar set of tradeoffs. RF transmission brings its own challenges around interference and signal integrity, though, and judging it requires field data rather than datasheets.

If you're a general reader, this is a window into the physical constraints under AI. Hundreds of cables per rack, their weight and bulk choking airflow, and consequently hundreds of millions of dollars flowing into a new kind of cable. The layer beneath an industry that looks like software is extremely physical.

🥄 Three Things You're Probably Wondering

— Is $136 million a lot? For deep-tech hardware, it's a meaningful figure. Worth noting that it's a cumulative total built through successive extensions rather than a single close. Getting a semiconductor product to volume production frequently requires more.

— Should I trust those performance numbers? They're company-provided, with no published independent verification yet. The physics explanation holds up, but how the technology behaves in a live data center only becomes clear with deployments. Lab figures and field figures diverge often in hardware.

— Does this replace fiber? Not really. The targets are relatively short runs — inside racks and between accelerators. Optics still wins at distance, and at short reach there's already an incumbent alternative in active electrical cables. Three technologies dividing the problem by segment is the more realistic picture.

Sources

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!