Everybody is trying to build an AI portfolio these days. Vendors are trying to vertically integrate product lines, or at least lock down multiple sources of components with long-term supply commitments, and private equity companies are trying to build out a portfolio of companies that they are invested in that supply these companies or sell assembled AI gear to OEMs or ODMs for final manufacturing.

All of these component suppliers need money – lots of money – if they are to ramp up production. This is as true of established parts like storage controllers, network interface controllers, or PCI-Express or CXL switches. But the funding for these can come from established product lines. For upstarts – say, those in the finally-emerging silicon photonics fields – they have a more pressing need for money because they have to demonstrate the high volume manufacturing of their products as well as the technical merits of whatever they have invented. If a part can’t be made consistently and as cheaply as possible, it hardly matters how good it is.

This is why funding rounds are so important to startups, and particularly the later funding rounds that come as an idea is taking off and the task of making the product and ramping it fast, as the GenAI boom demands, is upon them. I talked about this at length in a webinar with Ayar Labs, Alchip, and Astera Labs last fall.

This week, Ayar Labs has hauled in $500 million in Series E funding, bringing its total haul to $870 million and nearly quadrupling its valuation to $3.75 billion.

The Series E round was led by Neuberger Berman, which is looking to build an AI portfolio and which has invested in Groq’s $640 million Series D round in August 2024. (That was some serious paydirt, given Nvidia’s $20 billion acquisition of Groq as 2025 was coming to a close.) Neuberger Berman also took part in a $650 million minority investment in neocloud CoreWeave back in November 2024, which had a Who’s Who list of private equity firms lining up with bags of money ahead of the neocloud going public.

Alchip Technologies, which is a complex chip shepherd like Broadcom and Marvell and a partner of Ayar Labs, also kicked in money for the Series E round, as did ARK Invest, Insight Partners, MediaTek, Qatar Investment Authority, Sequoia Global Equities, and 1789 Capital. The latter is a private investment firm that has Donald Trump Jr as a partner. MediaTek, which is a chip maker based in Taiwan, had to disclose to the Taiwanese regulators that it has invested $90 million in Ayar Labs this time around. A bunch of prior investors, including AMD and Nvidia, also got out the checkbooks for the Series E.

With the money, Ayar Labs has opened an office in Hsinchu, which is the Silicon Valley of Taiwan and where foundries Taiwan Semiconductor Manufacturing Co and United Microelectronics Corp have massive facilities and hundreds of chip assembly, packaging, and testing companies are located. This is, in fact, where advanced technologies ramp on Earth these days.

The Alchip partnership announced last fall is interesting, and it resulted in a test XPU chip being created with eight of Ayar Lab’s TeraPHY optical engines on it, showing how one might use silicon photonics to provide the I/O into and out of a complex compute engine and ditch the copper. Take a look:

The chip above has more than 8 Tb/sec of bandwidth coming off of each of those TeraPHY optical engines, with less than 25 nanoseconds of latency across the optical I/O and greater than 5 decibels of optical margin across the links.

The TeraPHY optical engines are etched using the COUPE technique from TSMC, and in a sense, the partnership between Ayar Labs and Alchip allows compute engine and switch chip vendors to get silicon photonics links from the TSMC-Ayar-Alchip collective as a completed and tested unit. TSMC uses its SoIC-X process to etch the optical waveguides on a wafer to create the photonic integrated circuit, or PIC, designed by Ayar Labs that modulates optical signals. Alchip is providing a UCI-Express converter chip that then talks to a separate electronic integrated circuit, or EIC, both based on normal CMOS transistors. The EIC has a SerDes that takes the parallel data coming off the UCI-Express port of the compute engine and then boosts the electric signals and serializes them so they can modulate the light signals coming off the external laser source. These chips are bonded together using the SoIC-X process using copper-to-copper bonds rather than trying to make a single chip that can do both jobs well.

TeraPHY represents the big revenue opportunity for Ayar Labs, Mark Wade, co-founder and chief executive officer at the company, tells The Next Platform. But the company is also buying semiconductor lasers and building its SuperNova remote light sources to drive the TeraPHY links so it has end-to-end expertise to control the optical links. Wade says that it may manufacture 10,000s to 100,000s of the ELSFP form factor lasers a year within the next few years, but the opportunity for TeraPHY optical engines is many orders of magnitude larger.

“This funding round came together very, very quickly,” says Wade. “The message I drive inside of our company is that the funding rounds get a lot of attention, and they are exciting moments. But ultimately, we are using this capital to drive execution, and the hard thing about the kind of company we are building is we are going to live or die based on whether we truly deliver on the volume ramp. And certainly the hard thing about high volume production is that every single step along the production process flow, and with every single component in that flow, you are only as good as your weakest link along that whole path.”

The idea is to mitigate all of those risks, and to do so in a relatively short period of time for chip design. A year ago, Wade says, he was worried about getting enough materials, such as indium phosphide, to manufacture SiPho laser light sources at the high volumes that are necessary for the GenAI buildout. The news from Nvidia this week that it was investing a combined $4 billion in laser makers Lumentum and Coherent (who along with Sumitomo are suppliers of laser dies to Ayar Labs for its SuperNova external light sources) helps alleviate some concerns here, and signals to the market that Nvidia is going to be adding more silicon photonics products.

Nvidia had better do that, and for its own good. Let’s say that each GPU or XPU has ten optical engines. If AI clusters have hundreds of thousands to millions of compute engines, and there are dozens of companies building such behemoths, that is tens of millions to hundreds of millions of optical engines and laser light sources. Each laser source has eight laser dies on it, so it is hundreds of millions to billions of laser dies. This is a scale that laser light source makers have never tackled before.

If you haven’t got the message, silicon photonics is going mainstream in 2028, which is what we all talked about in that webinar back in October 2024.

“In the last year or so, we kind of graduated from low volume technology showcases to actually getting designed into volume intercepts, says Wade. “And volume intercepts start to change the kind of optimization that you have to aim at. It's not just about having the highest performing, hot off the press research result that that gets us to a paper at the OFC conference or something. It's about these technologies are getting selected in to going into the AI accelerators, and the huge challenge is the ramps that we are looking at with our customers are faster and steeper than anything the optics world has ever seen.”

In addition to staring down that challenge with the help of Alchip and TSMC, it looks like this is also the last funding round before Ayar Labs goes public.

“You could think of this as our final round before we would be trying to position ourselves for a potential IPO,” Wade confirms when asked about that. “The right way to think about it is this: Our customers are aiming at volume ramps that are that are happening in the 2028 timeframe. We have to get our end to end products selected, validated, and qualified by the second half of 2027, and in the semiconductor world, that's not that far away. And with this funding, we are really gearing up and driving towards having everything ready to support those ramps in in 2028.”