Eliyan: The Wall Isn't the Chip. It's the Wire.
A quiet Santa Clara startup just hit a $1 billion valuation solving the bottleneck strangling every AI data center. Read closely and it's also a power story.
Last week I promised a deep-dive on why silicon and electricity have become the same story. Most people expected me to write about a reactor. Instead I want to write about a wire, because the wire is where the two halves actually meet.
Company Overview
In plain English, Eliyan builds the connective tissue between chips. Modern AI accelerators stopped being single slabs of silicon years ago. They’re now clusters of smaller chips, called chiplets, stitched together inside a package, then packages stitched into racks, then racks stitched into clusters. Every one of those stitches is a place where data can get stuck. Eliyan makes the interconnect, the physical and electrical link, that moves data between those pieces faster and, critically, with less power per bit.
On July 29, Eliyan announced it completed a Series C totaling $145 million at a $1 billion valuation, crossing into unicorn territory. The oversubscribed round was led by Seligman Ventures, whose managing partner Umesh Padval joins the board, with two notable new strategic backers: Cisco Investments and Lumentum. Worth flagging up front: strategic investment is capital, not a purchase order. No joint products or supply agreements with either company have been disclosed, so read their participation as a signal of technical fit, not confirmed adoption.
The team is the tell. CEO and co-founder Ramin Farjadrad was CTO and VP of networking at Marvell, one of the deepest benches in high-speed connectivity anywhere. He co-founded Eliyan in 2021 with Syrus Ziai and Patrick Soheili. This isn’t a first rodeo. Eliyan raised a $40 million Series A in 2022 and a $60 million Series B in 2024, and separately took $50 million in strategic investment in January 2026 from a roster that reportedly included AMD, Arm, Coherent, and Meta. A quick compliance note for anyone doing the math: the company describes the Series C as “completed at a total of $145 million” and hasn’t said whether the January strategic money sits inside that figure, so don’t simply stack the numbers into a single total.
Its products carry the house branding: NuLink PHYs and NuGear chiplet families, the silicon-proven building blocks. Days before the raise, Eliyan unveiled NuLink-XD, a 224G interface built on TSMC’s 3nm process. The company claims its interconnect delivers up to four times the performance at half the power of conventional approaches. That’s a company claim, not an independent benchmark, so hold it accordingly.
The Market
Here’s the problem almost nobody outside the industry talks about. The world’s most expensive AI chips spend a large share of their lives waiting. Industry estimates put GPU utilization in large clusters at roughly 30% to 40%, and it’s often not a software problem. The chips are idle because they’re waiting for data to arrive. When you can’t feed silicon fast enough, you’ve bought a Ferrari to sit in traffic.
That problem is about to get worse in a very specific way. Inference, the work of running trained models in production, is projected to grow from roughly one-third of AI compute in 2023 to about two-thirds by the end of 2026. Inference is bandwidth-hungry and latency-sensitive. As the industry shifts from training a handful of giant models to serving billions of queries, the value migrates from raw compute toward moving data between chips without stalling. The bottleneck stopped being how fast a chip thinks and became how fast you can talk to it.
Now connect that to the thing I wrote about last week. Interconnect efficiency isn’t only a speed story, it’s a power story. Every wasted bit of data movement is wasted electricity, and electricity is the input the entire AI buildout is now short on. S&P Global estimates only about half of the roughly 50 gigawatts of US data centers the industry wants over the next three years can actually plug into the grid. When power is the binding constraint, the company that moves more data per watt isn’t selling a nice-to-have. It’s selling headroom inside a fixed power budget. That’s where silicon and electricity become one problem, and it’s why an interconnect startup belongs in the same conversation as a reactor company.
Timing is the last piece. The industry is mid-transition from monolithic dies to chiplet architectures, and from purely electrical links toward co-packaged optics for the longer runs. Transitions are when incumbents’ advantages reset and new standards get set. Eliyan is positioning to span that whole continuum, from electrical die-to-die inside a package out to optical rack-to-rack fabrics.
Business Model and Moat
Eliyan makes money two ways: licensing its interconnect IP, and selling chiplet products, to hyperscalers, chip developers, memory suppliers, and optical module makers. That model has a specific shape. Design cycles are long and painful to win, but once your IP is designed into a customer’s chip, you’re locked in for years. The company’s Chief Strategy and Business Officer, Patrick Soheili, has said Eliyan expects revenue to grow from low single-digit millions in 2025 to hundreds of millions by the end of 2027, with initial chiplet shipments beginning in 2026. Treat that as a company projection, not a fact on the ground. It’s the number the whole thesis rides on, and it hasn’t happened yet.
The competitive field is crowded and includes the most powerful company in the industry. Nvidia has NVLink for chip-to-chip and, through its Mellanox acquisition, InfiniBand for cluster networking. Both are excellent and both are proprietary, which is precisely the opening. There’s also UCIe, the Universal Chiplet Interconnect Express, an open industry standard backed by a consortium that wants die-to-die connection to become a commodity. And there are the SerDes incumbents like Broadcom and Marvell, plus a wave of optical interconnect startups such as Ayar Labs, Celestial AI, and Lightmatter chasing the same co-packaged-optics future.
So what’s defensible? Three things, in order of durability. First, silicon-proven PHY IP and the patents around it. Farjadrad’s team has been taping out real interconnect on leading-edge nodes since 2022, and interconnect physics is genuinely hard to copy. Second, neutrality. Eliyan is foundry-agnostic, memory-agnostic, and increasingly optics-agnostic, which makes it the natural choice for any hyperscaler or chip developer who wants to mix best-of-breed parts without being taxed by a single vendor’s ecosystem. Third, design-win stickiness. Once you’re in the chip, you’re in for the product’s life.
The moat, honestly stated, is “be the neutral standard in a market defined by one dominant proprietary player.” That’s a real position. It’s also a position that lives or dies on whether an open standard commoditizes it first.
Spence’s Take
Two mental models here, one for the bull case and one for the bear.
First Principles for the bull case. The lazy way to scale an AI chip is to make the die bigger. Physics says no: you hit the reticle limit, yields collapse, costs explode. Eliyan didn’t accept the inherited assumption that a processor should be one big monolithic thing. They went back to the physics and asked a cleaner question: if the future is many small chiplets, what’s the best possible way to connect them? Rebuild from that question and interconnect stops being an afterthought bolted onto a chip and becomes the architecture. That’s why the “half the power” claim matters more than the “four times faster” claim. Faster wins benchmarks. Lower power per bit wins in a world where every data center is running into a power wall. A silicon company whose core value proposition is fewer watts is, quietly, an energy company. That’s the whole reason it belongs in this week’s deep-dive.
Inversion for the bear case. Don’t ask what makes Eliyan win. Ask what has to be true for it to fail, then check how likely each of those is. It fails if UCIe matures fast enough that die-to-die connection becomes a free commodity and the IP premium evaporates. It fails if Nvidia’s proprietary NVLink plus its captive ecosystem simply keeps the volume, because customers would rather buy the whole stack from one vendor than assemble best-of-breed. It fails if the design wins slip and that low-millions-to-hundreds-of-millions revenue ramp stalls out in the gap between “designed in” and “shipping.” And it fails if the optical transition rewards a fundamentally different architecture than the one Eliyan is building toward. None of those is far-fetched. At least two of them are actively being pursued by companies with more money than Eliyan will ever raise. The bull case is real. So is the list of ways it ends.
Why It Matters
For investors and VCs. This is a deal worth watching, not for the valuation, which is a company-reported, point-in-time private figure, but for what it signals about where AI capital is rotating: away from models and toward the plumbing. If you’re tracking the story rather than the stock, the signals that actually matter are concrete. Watch for design-win announcements with named customers. Watch whether initial chiplet shipments actually land in 2026 as promised. Watch whether Cisco and Lumentum convert from investors into customers, because that’s the difference between a strategic bet and a strategic relationship. And watch the UCIe roadmap, because the pace of that standard is the single biggest external variable in the thesis. None of this is a recommendation to do anything. It’s a list of what would confirm or break the story.
For potential customers. If you’re a hyperscaler, a chip developer, or a system builder, what’s coming is an alternative to buying your entire scale-up architecture from one vendor. Eliyan is building a foundry-neutral, memory-agnostic interconnect that lets you combine the best chiplets from different sources and squeeze more usable work out of the same power envelope. If GPU utilization in your clusters is stuck in the 30s and 40s, the problem you’re feeling is very likely the wire, not the chip, and the fix is better interconnect, not more accelerators you can’t fully feed.
For competitors and builders. The lesson here is portable and it’s the same lesson as last week’s reactor. Find the constraint everyone is ignoring and own it. Compute got all the attention while interconnect and power quietly became the real limits. Eliyan’s second lesson is about positioning: in a market with one dominant proprietary player, the durable move is often to be the neutral layer everyone else can build on, not to fight the giant head-on. And the third lesson is temperamental. Eliyan sells the boring connective tissue, the wires and the PHYs, in an industry obsessed with the glamorous part. Boring, load-bearing, and hard to replace is a very good place to build a business.
The Bottom Line
Eliyan is a credible, well-led bet on the least glamorous and most load-bearing problem in AI infrastructure, and the fact that its core advantage is measured in watts is what makes it matter beyond semiconductors. The bull case is that it becomes the neutral standard for connecting a chiplet world that’s also power-constrained. The bear case is that an open standard or a proprietary giant gets there first, and the revenue ramp everyone is underwriting never fully arrives.
This is Startup Spotlight. Every week I break down a company worth watching. Subscribe for $7.99/mo to get the full analysis.
Spence
This post is for informational and educational purposes only and is not investment advice, a recommendation, or a solicitation to buy or sell any security or asset. Funding figures, valuations, revenue projections, and company metrics are as reported by the company, its investors, or the cited outlets; private-company valuations are point-in-time snapshots that do not reflect public-market prices or predict future performance. Company performance claims and forward-looking projections are attributed to the company and are not independently verified. Strategic-investor participation does not imply any product, adoption, or supply agreement. Always do your own research and consult a licensed professional before making financial decisions.
