Copper Kings Buy the Fiber Layer: Credo and Molex Lock Down Silicon Photonics in 48 Hours

Two of the largest electrical-interconnect incumbents just moved on the optical layer in 48 hours. On April 13, Credo Technology agreed to acquire DustPhotonics for $750M cash plus roughly $123M in stock, with performance earnouts pushing the headline up to $1.3B. Two days later, on April 15, Molex announced a definitive agreement to acquire Teramount, the Jerusalem-based developer of the TeraVERSE fiber-to-chip interface. Molex did not publish a price. The figure has been reported at roughly $430M, but it is not in the filing.

Read separately, each deal is a quarterly strategy note. Read together, in the same week, at the same layer of the stack, they extend a consolidation wave that Marvell began with Inphi in 2021 and Cisco continued with Acacia - but narrow it. This round is not about absorbing pluggable-transceiver vendors. It is about owning the in-package fiber interface, the single hardest step in making co-packaged optics economical, and doing it before the 1.6T standards cycle closes. The copper-era incumbents have decided they would rather buy that IP than build it.

The acquirers didn't have a PIC. Now they do.

Credo's pre-acquisition stack is a clinic in copper-era AI networking: ZeroFlap Active Electrical Cables, OmniConnect 112G VSR SerDes, 28G–224G SerDes IP and chiplets, PCIe/CXL retimers, and optical DSPs up to 224 Gbps/lane PAM4 driving ZeroFlap optical transceivers on the PILOT platform. What Credo did not own was the silicon itself, the photonic integrated circuit at the end of the lane. That gap is exactly where Broadcom leans on TSMC's Compact Universal Photonic Engine and where Ayar Labs raised $500M. Buying DustPhotonics closes it in one step.

DustPhotonics brings four product families spanning 400 Gb/s DR4 to 1.6 Tb/s with integrated lasers (Oz, Tamar, Carmel, Kfir), plus a 3.2 Tb/s roadmap and a patented technique called L3C (Low-Loss Laser Coupling). L3C relaxes the optical budget enough that the PIC can be fed by an external laser source over fiber rather than depending on hybrid-bonded laser die inside the package, the same architectural direction as the OIF ELSFP standard. That matters because hybrid laser integration is where CPO yield dies. Moving the laser off the photonic engine, and keeping enough coupling efficiency to make it work, is worth real money.

Credo CEO Bill Brennan framed the transaction as moving the company to "a vertically integrated connectivity platform" spanning SerDes, DSP, silicon photonics, and system integration for both scale-out and scale-up AI fabrics. Credo also put a number behind it: combined optical revenue above $500M in fiscal 2027. Against a current optical run-rate in the low nine figures and DustPhotonics' last disclosed raise, a $34M Series B in 2023 implying revenue likely in the $20 to $30M range, the target assumes roughly a doubling of the existing DSP-and-transceiver book plus design-win ramp on 1.6T PICs that haven't been publicly disclosed. It also assumes IEEE 802.3dj lands on its July 2026 target. Any slip on the standard pushes a meaningful fraction of that number into FY28. It is the first hard optical-revenue target any merchant CPO-adjacent supplier has given the market, and it is a forward bet, not a run-rate.

Molex's buy is a narrower bet in a narrower layer, and arguably higher leverage for it. Teramount builds the fiber-to-chip interface. Not the PIC, not the laser, the coupler. Its TeraVERSE product is a detachable, field-serviceable fiber interface built on a universal photonic coupler and a wafer-processable Photonic-Bump self-aligning optical element. The company's own spec claims 100× better tolerance than incumbent active-alignment methods, a passive scheme designed to survive first-pass assembly without someone tuning each module under a microscope. Molex's Aldo Lopez called TeraVERSE "a crucial gap in the CPO stack." He is being precise.

The mechanism matters. Teramount's Photonic-Bump expands the optical mode at the chip edge so alignment tolerances relax from sub-micron, where every fiber attach requires active alignment under feedback, to roughly ten microns, where passive pick-and-place succeeds on the first try. At a 102.4T switch with hundreds of fiber terminations per package, Broadcom's Tomahawk 6 scale, active alignment is not a cost problem, it is an existence problem. You cannot build the assembly line. Molex's Aldo Lopez called TeraVERSE "a crucial gap in the CPO stack." He is being precise. Fiber attach is the single step where CPO yield most often breaks. Teramount's Jerusalem site becomes Molex's CPO engineering center post-close.

Why the fiber interface is the pinch point

The short version: SerDes reach is collapsing and lasers don't like being close to compute silicon.

At 224G-PAM4, the copper link budget inside a rack is already thin enough that pluggable optics sit at the faceplate and electrical lanes run only inches of substrate. Pushing past 1.6 Tb/s per port, which is the target of the IEEE P802.3dj Ethernet task force (draft at D2.0 in public comment, final ratification targeted for July 2026), means either pushing 448G-PAM on copper or moving the optics inside the package. The industry chose the second option years ago. The component suppliers are now catching up.

Once the optics are inside the package, the laser becomes the dominant reliability liability. Lasers hate heat, and the inside of a switch ASIC is very hot. NVIDIA's headline metric for Quantum-X and Spectrum-X Photonics, announced at GTC 2025 (Quantum-X at 144× 800 Gb/s InfiniBand, Spectrum-X up to 512× 800 Gb/s), is not the line rate. It is "4× fewer lasers." That phrase is the whole thesis of CPO economics. Every laser you eliminate is a failure mode you don't have to design around.

The industry's answer is to socket the laser. OIF's External Laser Small Form-factor Pluggable IA defines a detachable laser module that pairs with the 3.2T Co-Packaged Module IA. The thermal physics is also under active academic attack. A 2025 Nature Photonics paper on frequency-stable nanophotonic microcavities demonstrates an integrated thermometer stabilizing a locked DFB laser to within 0.5 pm over 50 hours, a direct answer to the wavelength drift that plagues co-packaged lasers.

Teramount's coupler and DustPhotonics' L3C are engineering answers to the same question: how do you keep the laser-yield problem from killing the economics of a 100+ Tbps switch? Credo bought the PIC side of the answer. Molex bought the fiber side. That is the market naming its pinch points out loud.

Koch is consolidating its own portfolio

One wrinkle worth flagging. Koch Industries has owned Molex outright since 2013, and Koch Disruptive Technologies was among the lead investors in Teramount's Series A. The Molex–Teramount deal is, in substance, Koch moving a portfolio company inside an operating company it already owns. TeraVERSE had already appeared in Molex's "one-stop CPO solution" demo at OFC 2026; the commercial collaboration predated the paperwork by months. The same pattern (financial-arm investment, then operating-company acquisition once the tech de-risks) has played out across Samsung, Hitachi, and SoftBank portfolios for a decade. Expect more of it.

The bear case these deals ignore

CPO is not the only way to get to 1.6T, and it may not be the winning way outside the highest-end scale-up fabrics. Linear-drive pluggable optics have already eaten a meaningful chunk of the power argument, with LPO modules running the DSP-less path at roughly a third of the power of a traditional pluggable. Hyperscaler operations teams prefer field-replaceable modules for reasons CPO advocates consistently underestimate. A bad laser inside a switch ASIC is a $30,000 RMA. A bad pluggable is a ten-second swap. NVIDIA's own roadmap still uses pluggable optics alongside Quantum-X and Spectrum-X Photonics for the near-term generations.

If that's how the market splits, CPO for NVL-class scale-up and pluggable LPO/LRO for most scale-out, Credo's DustPhotonics buy still works because PIC IP feeds both form factors. Molex's Teramount buy is a narrower bet. Fiber-attach economics at that scale only matter if CPO volumes materialize. Watch whether the next wave of hyperscaler RFPs asks for CPO explicitly or accepts linear pluggable alternatives. That answers the question these two deals are silently asking.

The wider consolidation wave

Credo and Molex aren't moving alone. Broadcom's Tomahawk 6 "Davisson" at 102.4 Tbps, the first CPO switch at that scale to actually ship, went to customers in October 2025 on TSMC's COUPE platform. Ayar Labs closed a $500M Series E in March 2026 at a $3.75B post-money, with strategic participation from AMD, NVIDIA, MediaTek, and Alchip. The CHIPS NAPMP program names photonic interconnect as an explicit R&D target, and $1.4B in final awards went out in January 2025. DARPA's PIPES program is chasing 100 Tbps per package under 1 pJ/bit.

Capital, federal funding, foundry capacity, and the standards bodies are all assuming the same thing: the AI factory's interconnect layer is photonic from the package outward. When the two biggest electrical-interconnect vendors in the world buy silicon-photonics pure-plays inside 48 hours, they are telling the market they already finished that calculation.

What to watch next

A few things will tell you whether this wave is pricing in real revenue or pricing in expectations.

Start with Credo's fiscal 2027 optical-revenue number. Above $500M combined is the guidance, the first concrete figure from any merchant supplier, and the external KPI the Street will use to grade whether DustPhotonics integration actually produces the "vertically integrated connectivity platform" Brennan promised.

Then IEEE 802.3dj ratification. The standard is not ratified yet; July 2026 is a target. If the working group hits it, 1.6 Tb/s Ethernet becomes the volume line rate for the next hyperscale refresh, and Credo's consolidated 224G-lane DSPs plus 1.6T PICs become the reference design. A slip pushes revenue recognition to the right for everyone downstream.

And the laser-yield question. External-laser sockets per the OIF ELSFP IA solve one version of it. On-chip wavelength generation, the target of DARPA's DWDM SBIR, solves another. Whoever first ships a co-packaged engine that doesn't need a per-wavelength external laser resets the bill of materials.

Two acquisitions in 48 hours, roughly $1.3B in combined headline value, targeted not at the transceiver layer the last consolidation wave absorbed but at the two steps inside the package, the PIC and the fiber coupler, where CPO economics actually break. The copper incumbents have decided the fiber layer is not something they are willing to source. Whether the AI factory rewards that decision depends on a standards ratification in July and a form-factor fight that has not finished.