Ayar Labs adds $150M: copper is the bottleneck

The cheap tokens you are building on are not a software achievement. They are a hardware bet, and this week it got another $150 million. Ayar Labs raised that much on September 10 to push co-packaged optics into volume manufacturing, taking its 2026 haul to roughly $650 million. The company's pitch is one sentence long, and it explains more about your 2027 AI bill than any model release this month: copper has run out of room.

What actually happened

Per Unite.AI's writeup, the $150 million is an extension of the $500 million Series E that closed in March at a $3.75 billion valuation, led by Neuberger Berman. The new round brings in Wiwynn as a strategic investor, alongside existing strategic backers that read like a who's-who of everyone selling AI silicon: Alchip, AMD, Intel, MediaTek and Nvidia.

The money goes to high-volume manufacturing, product validation, scaling the supplier ecosystem, and a new design center in Bengaluru. CEO Mark Wade framed the problem directly: copper interconnect is becoming the limiting factor for AI scale-up.

That is the whole story. As models get served across multiple racks, the cost of moving a token between chips starts to dominate the cost of computing it. Optical interconnect replaces the electrical wiring between packages with light, which is why chipmakers that compete with each other on everything else are all funding the same supplier.

Why your inference bill has a physics problem

Every AI pricing decision you make assumes a cost curve that keeps bending down. For the last two years it did, and most operators assumed the reason was model efficiency. Increasingly it is not — it is memory bandwidth and interconnect. DeepSeek's newest model cut its price largely by cutting KV cache footprint. Ayar Labs is attacking the same constraint one layer lower, in the wires.

Price your product on today's rates, not next year's. If your margin only works at a token price that requires hardware still in "transition to volume manufacturing," you have taken a supply-chain bet without writing it down. Hardware schedules slip. Price the offer at rates you can buy this quarter.

Keep AI contracts short. The vendors who win the interconnect race will have real cost advantages, and they may not be the vendors you signed with. A twelve-month commit on inference capacity in a market repricing this fast is a bet against your own flexibility. We have said this about long-dated compute backstops and it applies at your scale too.

Make the model layer swappable, because the cost leader will change. The only way to benefit from someone else's hardware win is to be able to route to them. That is a config file and an eval set, not a rewrite — and it is the cheapest hedge available against a market where the cost floor is set by fabs.

Sources: Unite.AI, Ayar Labs Series E announcement.