Wiwynn unveils CPO rack with Ayar Labs at Computex 2026

Wiwynn, the Wistron subsidiary that has become a bellwether for hyperscale data center hardware, unveiled a rack-scale AI server at Computex 2026 that marries co-packaged optics with liquid cooling in a single integrated design. The system, on display at booth J0106 in TaiNEX1, incorporates Ayar Labs' TeraPHY optical engines and SuperNona remote light sources alongside a custom AI ASIC packaged using GUC's 2.5D advanced packaging technology. This is the first time Wiwynn has combined optical scale-up interconnects with high-power liquid cooling in a production-ready rack form factor, signaling that CPO has moved from lab curiosity to deployable infrastructure. The announcement comes as hyperscalers face a fundamental bottleneck: electrical interconnects cannot keep pace with the bandwidth demands of trillion-parameter models without consuming prohibitive power. Wiwynn's solution directly addresses that constraint by replacing electrical SerDes with optical links that reduce per-bit energy by an order of magnitude while enabling rack-scale memory pooling. Why this matters now: CPO adoption is the single most consequential hardware shift in AI infrastructure since the GPU, and Wiwynn's Computex debut provides the first concrete proof that the technology can be integrated at rack scale without compromising thermal or power delivery.

Optical interconnects break the electrical bottleneck

The core innovation in Wiwynn's rack design is the substitution of traditional electrical interconnects with Ayar Labs' TeraPHY optical engines, which convert electrical signals directly into light at the chip edge. These engines sit on the same 2.5D advanced package as the AI ASIC, eliminating the need for pluggable optical modules that consume rack faceplate space and add latency. The SuperNona remote light source, an external laser array, provides the continuous-wave light that the TeraPHY modulators encode with data, separating the laser's thermal load from the compute die. This architecture allows each optical link to carry 512 Gbps per lane while consuming less than 5 picojoules per bit, compared to 10??5 pJ/bit for electrical SerDes at equivalent reach. The rack design also incorporates high-voltage direct-current power delivery, which reduces conversion losses by 3??% compared to traditional AC distribution. Wiwynn's liquid cooling system handles the thermal load of both the ASIC and the optical components, a non-trivial engineering challenge given that the SuperNona laser source operates best at a stable 25?C while the compute die can spike to 85?C. The company has not disclosed the exact thermal design power of the system, but the inclusion of direct-to-chip liquid cooling suggests per-rack power densities exceeding 100 kW, a threshold that forces data center operators to rethink facility-level cooling and power distribution. The optical engines eliminate the need for electrical retimers and signal conditioning circuits, further reducing component count and board complexity. Wiwynn's engineers validated the thermal interface between the laser source and the compute die using a shared cold plate design that maintains both components within their operating ranges.

How the P&L shifts for hyperscale operators

The financial case for CPO rests on three line items: capital expenditure on optics, operating expenditure on power, and the opportunity cost of rack space. Wiwynn's rack eliminates pluggable optical transceivers, which currently account for 15??0% of the bill of materials in a typical AI server rack. Ayar Labs' TeraPHY engines, while more expensive per unit than traditional transceivers at scale, reduce the total number of optical components because each engine handles multiple lanes. The power savings are more dramatic: at 5 pJ/bit versus 12 pJ/bit for electrical interconnects, a rack operating at 100 kW and 50% utilization saves approximately 438 MWh per year, translating to roughly $44,000 at the US average industrial electricity rate of $0.10/kWh. For a hyperscaler deploying 10,000 racks, that annual power savings alone reaches $440 million. The rack-space benefit is harder to quantify but equally significant: eliminating pluggable modules frees up the front panel for additional compute or storage, increasing per-rack compute density by an estimated 10??5%. Wiwynn's integration of HVDC power delivery further improves efficiency by reducing the number of AC-to-DC conversion stages from three to two, cutting conversion losses by roughly 1.5 percentage points. These savings compound across the data center fleet, making CPO not just a technical improvement but a direct lever for improving data center unit economics at a time when hyperscalers are under pressure to show returns on their AI capital expenditure. The total cost of ownership reduction for a single rack over a five-year lifespan exceeds $200,000 when factoring in power, space, and maintenance savings.

Competitive reshuffle: who gains and who loses

The CPO rack positions Wiwynn as a first-mover in optical interconnect integration, a category that threatens established players in both the server and networking markets. Cisco Systems, which dominates the pluggable optical module market through its Acacia subsidiary, faces the most direct disruption: if CPO adoption accelerates, the $4 billion pluggable transceiver market will shrink by 30??0% within three years. Ayar Labs, the startup supplying the TeraPHY engines, gains a marquee customer and a reference design that other server OEMs can replicate. GUC, the advanced packaging house, benefits from the increased complexity of 2.5D integration that CPO requires, a trend that raises barriers to entry for smaller packaging competitors. For Wiwynn's parent Wistron, the CPO rack represents a strategic pivot from commodity server assembly to value-added integration, potentially improving gross margins from the low teens to the mid-20s. The losers extend beyond optics: traditional interconnect vendors like Broadcom and Marvell, which derive significant revenue from electrical retimers and SerDes IP, face a structural demand decline if CPO becomes the dominant interconnect topology. Memory vendors Samsung and SK Hynix also face pressure, as CPO-enabled memory pooling reduces the need for high-bandwidth memory modules by allowing direct optical access to remote DRAM pools. The cyclical nature of memory demand, which investors have flagged as a risk despite AI enthusiasm, compounds this threat. Optical module manufacturers such as Finisar and Lumentum will need to pivot their product lines toward CPO-compatible components or risk losing their position in the data center supply chain.

Downstream effects on hyperscalers, fabs, and enterprise buyers

The downstream implications of Wiwynn's CPO rack ripple through the entire AI infrastructure stack. Hyperscalers like Amazon Web Services, Microsoft Azure, and Google Cloud, which collectively account for 70% of AI server procurement, will evaluate the design against their internal CPO roadmaps. Microsoft has invested in its own optical interconnect research through the Microsoft Research Silicon Valley lab, while Google has deployed optical circuit switches in its data centers for years. Wiwynn's rack gives these operators a turnkey alternative to in-house development, potentially accelerating CPO adoption by 12??8 months. For foundries, the CPO rack increases demand for advanced packaging capacity, particularly 2.5D and 3D integration that TSMC offers through its CoWoS and InFO platforms. The SuperNona remote light source, which requires III-V semiconductor manufacturing, creates new demand for compound semiconductor fabs at a time when capacity is already constrained by photonics and RF applications. Enterprise buyers, which have lagged hyperscalers in AI infrastructure deployment, benefit from the simplified cabling and reduced power consumption that CPO enables, lowering the barrier to deploying rack-scale AI in colocation facilities. However, the liquid cooling requirement remains a hurdle: most enterprise data centers lack the chilled-water loops needed for direct-to-chip cooling, limiting initial adoption to hyperscalers and large colocation providers. Industry analysts estimate fewer than 20% of North American enterprise data centers currently have infrastructure to support liquid-cooled AI racks, constraining near-term CPO adoption to cloud-native environments. Colocation providers such as Equinix and Digital Realty are already retrofitting their facilities to support liquid cooling, which will expand the addressable market for CPO racks over the next two years.

Policy and strategy signal from the Computex debut

Wiwynn's Computex 2026 announcement reads as a strategic signal that the Taiwanese server ecosystem is betting on CPO as the next dominant interconnect standard, a move with policy implications for both trade and technology standards. Taiwan's server OEMs, which manufacture roughly 85% of the world's servers, have historically been fast followers rather than technology pioneers. By integrating Ayar Labs' cutting-edge optical engines at rack scale, Wiwynn is asserting technological leadership that shifts the center of gravity for AI hardware innovation from Silicon Valley to Taipei. This has implications for US export controls: if CPO-enabled servers become the standard for AI training clusters, the US will face pressure to restrict access to the underlying photonics technology, particularly the III-V semiconductor lasers that SuperNona uses. The announcement also signals to the market that the CPO ecosystem has reached a level of maturity that justifies production-scale investment. Ayar Labs, which has raised over $300 million from investors including Intel Capital and Applied Ventures, now has a clear path to revenue through Wiwynn's supply chain. For regulators, the CPO rack raises questions about data center energy efficiency standards: if CPO reduces interconnect power by 60%, regulators will mandate its adoption in new data center builds, accelerating the technology's deployment curve.

The CPO rack will face its first production-scale test in the second half of 2026, when Wiwynn begins shipping evaluation units to hyperscale customers. If those trials validate the power and density claims, the technology will trigger a multi-year replacement cycle in which every new AI server rack incorporates optical interconnects. The winners will be the companies that own the photonics IP and the advanced packaging capacity; the losers will be those that bet on incremental improvements to electrical interconnects. For hyperscalers, the calculus is straightforward: CPO reduces both capital expenditure and operating expenditure while increasing compute density, a trifecta that makes adoption inevitable. The open question is whether the supply chain can scale fast enough to meet demand, particularly for the III-V semiconductor lasers and the 2.5D packaging capacity that CPO requires. Wiwynn's Computex debut provides the first concrete evidence that the answer is yes, and that the era of electrical interconnects in AI data centers is drawing to a close.