Transition to Optical-Based Scale-Up

Copper-based interconnects used in today’s scale-up systems are nearing their limit in terms of reach and bandwidth and are thus creating a compelling need for optical solutions. In contrast to copper, optical interconnects have much higher bandwidth and enable physically longer cables, which enables AI clusters to be distributed across larger areas.

At its Q3 FY2026 earnings call, Marvell announced that it is acquiring Celestial AI, one of a host of silicon photonics companies developing solutions for AI data centers. Celestial AI’s Photonic Fabric (PF) technology platform is purpose-built for the transition to optical based scale-up. It enables large AI clusters to scale both within and across racks using a cost-effective optical fabric. Celestial AI claims that its solution provides more than twice the power efficiency of copper interconnects, but with significantly higher bandwidth and far longer reach.

Celestial AI Products and Solutions

Celestial AI’s first-generation product is a PF chiplet, which integrates the required electrical and optical components - drivers, TIAs, equalizers, SerDes, microcontrollers, modulators, photodiodes and waveguides - into a compact form factor. According to the company, this is the industry’s first scale-up optical solution providing an unprecedented 16Tbps of bandwidth in a single chiplet. As a result of the compact form factor, multiple PF chiplets can be co-packaged with XPUs on one side of the link and scale-up switches on the other.

Key technical innovations include:

• 3D/Vertical Integration – in addition to high bandwidth, Celestial AI’s solution has exceptionally low power consumption, nanosecond-class latency and excellent thermal stability. This latter point is a particularly important feature and a significant competitive differentiator as it enables reliable operation in extreme thermal environments inside large, multi-kilowatt XPUs.

Most co-packaged optics (CPO) designs attach the optical light engines around the edges of a XPU on a substrate. Celestial AI claims that its technology is sufficiently stable thermally to be co-packaged vertically in a 3D package, with the photonic connection made directly onto the XPU rather than at the edge of the die. This avoids the “beachfront” problem, i.e. the lack of space at the edge of a chip to attach wires, resulting in ten times the I/O density compared to other CPO designs, and overall, a more compact and integrated solution. The freed die edge beachfront can thus be repurposed to increase the amount of HBM in the XPU package and also increase the amount of package bandwidth possible for XPUs and switch systems.

• Memory disaggregation – Celestial AI’s PF technology supports optical CXL memory pooling, which means that HBM and DRAM memory no longer need to be located close to the XPU (Table 1). Using Celestial AI’s Optical Multi-Chip Interconnect Bridge (OMIB), the PF chiplet enables a XPU to connect to a CXL memory pool located in a separate memory appliance located several metres away - thus overcoming the problem of memory stranding. As the connection is optical, the latency is low enough (<150ns) for the XPU to assume that the remote memory is local. Another major benefit is that less expensive DDR5 memory can be used instead of HBM4 memory.

Table 1:  Comparison of HBM4 versus Optical CXL Memory Pools

Source: Counterpoint Research

Market Opportunities

Celestial AI is engaged with several hyperscalers and ecosystem partners. In particular, the company has secured a major design win with “one of the world’s largest hyperscalers” that plans to use Celestial AI’s PF chiplets in its next-generation scale-up architecture. The PF chiplets will be co-packaged into both the hyperscaler’s custom XPUs and its scale-up switches. This is expected to be the industry’s first large-scale commercial deployment of optical interconnects for scale-up connectivity. However, using optical interconnects for scale up is just one of a broad range of new applications that could be enabled by Celestial AI’s silicon photonics technology. Other notable applications include:

• Die-to-Die Optical Attach – electrical die-to-die interconnects in multi-die packages limit speeds and increase power. With Celestial AI’s PF, compute and memory chiplets can be connected optically within the same package. This means bandwidth is no longer dictated by package pin limits and power is reduced, thus lessening the burden on the cooling system.

• CXL 4.0-based Memory Pooling – combining Marvell’s Structera memory controllers and accelerators with Celestial’s optics provides Marvell with the technology base to build 100TB+ memory pools. These memory pools could be shared across an entire rack of XPUs and CPUs with near-zero latency. Released in November 2025, CXL 4.0 aligns with PCIe 7.0, doubling the performance of previous CXL generations whilst introducing architectural smart features for larger scale deployments.

Key benefits of a Marvell-developed optical memory appliance include: a 4-5x reduction in power compared to traditional electrical interconnects and the elimination of memory stranding by allowing XPUs to “borrow” RAM dynamically from a memory pool. According to Marvell, this could mean that hyperscalers would need up to 30% less HBM4 memory as they could use less expensive DDR5/CXL memory with almost no latency penalty. There are potentially many applications for these optical memory appliances - used with or without UALink. For example, they could be used in an Nvidia cluster to provide terabytes of shared CXL memory that Nvidia’s copper-based NVLink compute fabric cannot reach.

Figure 1 shows details and architectural layout of a photonic scale-up network using Celestial AI’s PF technology. 

Source: Celestial AI

Figure 1:  Photonic Fabric Scale-Up Network Using Celestial AI

Strategic Rationale behind Acquisition

Marvell will pay $3.25 billion once the transaction has been approved by regulators. However, the total amount paid could extend to $5.5 billion if Celestial AI reaches specified revenue targets by the end of Marvell’s fiscal year 2029. The acquisition should close in the first quarter of 2026, subject to the usual regulatory reviews. Key strategic reasons behind the acquisition include:

• Full-stack connectivity solution – acquiring Celestial AI will provide Marvell with a “full-house” of interconnect solutions. Celestial AI will enable Marvell to add optical scale-up connectivity in addition to its copper-based interconnect products. Combining Marvell’s existing strengths (custom silicon, switches, packaging, networking, etc.) with Celestial AI’s photonics gives Marvell a more comprehensive, vertically integrated platform for custom AI/cloud infrastructure. When combined with its scale-out and scale-across solutions, this strengthens Marvell’s position as a one-stop-shop for data center connectivity – an attractive proposition for large cloud providers and hyperscalers.

• Technology readiness and differentiation – Celestial AI’s photonic bridge is reportedly viable for practical packaging now and is thermally stable with today’s high-power AI chips, an important factor potentially enabling Marvell to lead the transition from copper to optical scale-up connectivity and enable it to differentiate itself from competitors lacking their own photonic IP.

• Open alternative to Nvidia’s NVLink – with the acquisitions of Celestial AI, Marvell is clearly positioning itself as an open alternative to Nvidia and a leader of two emerging open standards: UALink and CXL-based memory pooling. Celestial AI is thus Marvell’s bet that optical interconnect and photonic scale-up fabrics represent the next inflexion point in AI/cloud infrastructure.

• New Market Opportunities – acquiring Celestial AI will enable Marvell to target the emerging multi-billion dollar optical scale-up market, a market considerably more valuable than its existing copper interconnects business. However, this is just one new opportunity. Silicon photonics will provide Marvell with key technologies that will enable new applications - such as D2D optical attach, disaggregated compute and memory, etc. - that can be leveraged to strengthen its custom silicon business.

• Relationships with hyperscalers - Celestial AI already has engagements with multiple hyperscalers and ecosystem partners that plan to use its PF for next-generation multi-rack scale-up architectures. This is likely to bring in design wins resulting in high-volume production. Marvell also has its own relationships with cloud providers and a footprint in the AI hardware market giving a pathway to scale Celestial AI’s photonics commercially.

• Revenue enhancing - Marvell forecasts Celestial AI will reach a $500 million annualized revenue run rate by the end of fiscal year 2028, rising to $1 billion by the end of fiscal year 2029. In time, Celestial AI could add multi-billion dollar increments to Marvell’s revenues every year.

Analyst Viewpoint  

Celestial AI brings a new disruptive technology, a photonic fabric platform purpose-built for next-generation scale-up interconnect, which will enable multi-rack scale-up networks with hundreds and thousands of XPUs to be tightly coupled together:

• Celestial acquisition - broadens Marvell’s addressable market in scale-up connectivity, potentially providing it with a leadership position in optical scale-up connectivity. In addition, the Celestial AI acquisition provides deeper integration with cloud customers while providing access to new multi-billion dollar market opportunities and accelerating its roadmap to deliver a complete one-stop-shop connectivity platform for AI and cloud customers.

• Risks and Challenges - as with any acquisition, there are risks and challenges to overcome. These include the successful manufacturing, integration and customer adoption of optical scale-up networks technologies. Transitioning from prototype/early production to high-volume manufacturing will not be easy as embedding optical I/O alongside high-power XPUs and memory and aligning with switching, system-level design, software orchestration for memory/compute pools will be challenging. In addition, photonic packaging, yield, thermal management and supply-chain issues may pose significant execution risk and customers will need to commit to new architectures - optical scale-up fabrics - a shift that takes time, validation and ecosystem support.

• UALink traction - Marvell’s future success to a large extent depends on UALink becoming the Android equivalent in AI/cloud infrastructure. This relies on continued support from major players such as AMD, Intel and the hyperscalers, particularly AWS and Microsoft, seemingly the most committed to UALink.

• Marvell’s Differentiation - Marvell is cleverly and uniquely hedging its bets by participating in three ecosystem plays: the open-source UALink and Ethernet-based E-SUN/UEC ecosystems as well as Nvidia’s proprietary NVLink ecosystem. Whilst Nvidia also supports E-SUN/UEC, Marvell’s big rival Broadcom is firmly in the Ethernet camp, claiming that its E-SUN/SUE technologies can be used for both scale up and scale out. Another key differentiator is Marvell’s CXL-based memory pooling technology. If it can successfully leverage its leadership position in CXL and associated technologies, memory pooling could well become the icing on top of Marvell’s AI infrastructure cake.

Note:

[1] Ultra Accelerator Link

[2] Ethernet for Scale Up/Ultra Ethernet Consortium

[3] Compute Express Link