TL;DR
Following frenzied endorsements by Huang Renxun, Marvell once again confirmed exceptionally strong AI-related orders at the Evercore TMT Summit and raised its 2027 revenue guidance to around $11.5 billion, with the interconnect business growth outlook increasing from around 50% to over 70%. Custom ASIC is expected to reach about $2 billion this year and double next year. The company's stock price has surged by nearly 140%-158% since 2026, far outperforming peers and the market average. There have been multiple single-day spikes of over 30% during this period, accompanied by significantly increased trading volume.
Many reports interpret this series of signals as Marvell firmly holding the leading position in AI interconnects, with the data center business becoming an absolute core. However, the current growth driver is still the market-proven scale-out business, relying on the mass production advantages of PAM DSP, TIA drivers, and pluggable optical modules. Meanwhile, scale-up optics and CPO (Co-Packaged Optics), representing higher unit cluster value and greater strategic significance, are still preparing for volume production in 2027. From early deployment to forming a several hundred million dollars' annual revenue by 2028, this execution certainty is key to assessing whether the growth momentum can continue until 2028, whether the company can transition from a component supplier to a strategic cloud partner, and whether the current high valuation already fully reflects optimistic expectations.
Summit Update Shows Strong AI Orders
Marvell's guidance increase at the summit is based on existing AI orders and locked-in supply chains, rather than long-term vision. The 2027 revenue guidance has been raised to around $11.5 billion, with the upper limit for 2028 further increased to around $16.5 billion. The interconnect business's year-on-year growth rate in 2027 has been raised to over 70%, with custom ASIC contributing around $2 billion this year and targeting over $4 billion next year. The external CXL plus custom NIC target for 2028 is over $3 billion.
These figures are substantiated. Management has repeatedly mentioned that capex-driven demand is very strong, orders have translated into a visible growth trajectory, key components like lasers, although facing tight lead times, have secured capacity, and the PAM DSP has maintained its pace of being the first to mass-produce in every generation (1.6T deployed at scale this year, 3.2T sampling next year). This is consistent with the paths indicated in previous earnings conference calls, pointing towards real demand rather than merely optimistic forecasts.
However, the stock price has already priced in this story well in advance. The returns this year have far exceeded those of peers and the index. Following the surge, both volatility and option-implied volatility have increased in tandem. Market discussions have shifted from identifying beneficiaries to questions of "how much higher can it go" and the risk of a pullback. This has created a clear contrast: the momentum supported by short-term orders is clearly visible, but the market pricing has already incorporated many optimistic medium-to-long-term assumptions.
After the strong guidance, investors naturally wonder which specific businesses are driving this growth and how much each contributes to the overall scale and out-of-scale growth.
The Current Growth Driver is Mature Out-of-Scale Business
The current main driver of interconnected businesses is still the out-of-scale portion. PAM DSP, TIA drivers, and 400G, 800G, 1.6T pluggable optical modules have already formed mature product lines. Being the first to mass-produce each generation has helped the company maintain its market share. The annual revenue from the TIA plus driver business has reached approximately $1 billion. From long-reach coherent to more lightweight high-speed PAM, the complete coverage enables the company to have a favorable position in the market for interconnecting data centers.
The in-scale business is now transitioning from its initial stage to an accelerated phase. As AI clusters scale from thousands of GPUs or XPUs to millions, the front-end computing power undergoes upgrades akin to engines, while the back-end interconnect is like a high-speed highway needing to expand from single to multiple lanes. The cluster's internal bandwidth demand is approximately 10 times that of the front end. Power consumption and density bottlenecks are driving the industry to shift from traditional pluggable optical modules to near-packaged optics and CPO. This transition can significantly increase the chip loading capacity per rack and expand the overall market space.
CPO's core lies in integrating the optical engine with XPU or a switch to reduce electrical-to-optical conversion losses, thereby enhancing bandwidth density and reducing overall power consumption. This is not merely a component swap but a systemic solution to the new bottleneck in large-scale clusters. Marvell is aiming to establish an integrated advantage in this direction through the SerDes IP and DSP capabilities accumulated from previous acquisitions and the recently integrated Celestial AI technology. However, the contribution from the in-scale business is still limited at present, and real volume production is expected to wait until after 2027.
Representing higher per-machine content growth and market expansion opportunities, how is Marvell's technological roadmap, manufacturing readiness, and execution capabilities shaping up in this emerging field?
Post Celestial Acquisition, What Is the Mass Production Path for CPO?
The Celestial AI acquisition was fully integrated in early 2026, with its photonics fabric technology being incorporated into Marvell's in-scale roadmap, co-packaged with custom XPUs and in-scale switches. The management expects to enter the formal mass production preparation stage in 2027, targeting approximately $500 million in annualized revenue from photonics fabric by the end of 2028, with subsequent plans for doubling this figure. Over a 15-month period, it is anticipated to contribute approximately $1 billion (on a single-product basis). This path has been optimized compared to earlier outlooks but still depends on smooth customer validation and manufacturing ramp-up.
Supply chain details show improved certainty. Laser lead times are tight, but Marvell has secured relevant capacity, with the company focusing on silicon photonics chip design and integration rather than producing lasers themselves. The collaboration with NVIDIA on NVLink Fusion further covers silicon photonics, custom XPU, and NVIDIA cluster interconnectivity, as well as OCTEON baseband support for AI-RAN. This not only enhances technical compatibility but also strengthens credibility in hyperscaler bidding, making Marvell's heterogeneous computing solutions more likely to be incorporated into next-generation AI factory plans.
If widespread adoption of a million XPU-level clusters with CPO occurs, resulting in significantly lower power consumption and increased bandwidth density, it will directly increase Marvell's content value per rack and support revenue and gross margin expansion. Conversely, if the ramp to mass production in 2027 is slower than expected, customer validation cycles lengthen, or yield and cost curves fall below expectations, the 2028 contribution will be below the current market's implied level, and overall guidance realisation pressure will emerge. All current statements are forward-looking, and actual shipment and market share data will provide harder validation only in 2027.
Given Marvell's full-stack layout from SerDes, DSP to optical engines, and custom XPU, along with ecosystem partner support, can the company maintain differentiation in the competition and convert it into actual market share gains?
Can Marvell's Differentiation Hold Up Against Broadcom?
Marvell's leadership in the PAM DSP market, comprehensive coverage from long-reach coherent to 1.6T and 3.2T, and integrated advantages of custom ASIC and OCTEON are tangible achievements, with the NVIDIA partnership also providing additional options for heterogeneous computing. These factors enable the company to be competitive in some hyperscaler projects, especially in scenarios requiring rapid mass production and ecosystem compatibility.
However, Broadcom still maintains a significant lead in network ASICs, overall scale, and maturity in optics. Broadcom's custom silicon projects have greater scale and stronger customer stickiness. In areas of overlap beyond a certain scale, Marvell still needs higher integration and a specific technological path to compete for incremental market share. The positioning as an "interconnect strategic partner" is currently more of an aspiration, and its eventual realization depends more on the actual bidding results and shipment breakdown in 2027-2028, rather than the current breadth of the layout.
For investors, this implies that in the next generation of hyperscaler bidding, Marvell's content value per rack could potentially increase by 20-50% from current levels, thereby supporting revenue visibility and gross margin expansion potential. However, if CPO ramp-up is slower than guided, or if Broadcom has an advantage in more custom projects, growth in 2028 may be below the optimal scenario implied by current valuation, increasing the pressure for a stock price correction. The current high-beta nature amplifies gains under the narrative drive but also magnifies volatility when execution falls short of expectations.
Massive CPO Deployment Remains the Biggest Unvalidated Variable
Despite strong AI orders, leading PAM production, and NVIDIA compatibility providing real mid-term momentum and option value, the yield, cost curve, thermal management, supply chain stability, and hyperscaler capital expenditure required for the ramp-up and large-scale deployment of CPO as a new technology to sustain the current pace remain the core constraints on growth sustainability in 2028. Historical experience shows that new technologies usually require multiple iterations from early production to meaningful revenue generation, and currently CPO is still in this early stage.
Marvell has had a good track record of meeting past guidance, but the climb this time is relatively aggressive. Valuation has risen significantly by 2026, incorporating many best-case assumptions, and the increasing crowdedness of market discussions has pushed up volatility. Broadcom's scale advantage will not disappear in the short term, and net share gains still require ongoing evidence. The current stage is an acceleration period, while early signs of late-stage crowding have emerged—intense price fluctuations and a shift in focus to realization pressures, which are characteristic of this phase.
Investors should not just focus on the summit's vision, but should instead closely track the 2027 actual revenue breakdown, shipment milestones, specific customer validation progress of CPO, and hyperscaler capital expenditure pace. These variables are most likely to change current assessments. If the 2027 milestones are strongly achieved, the trend of structural enhancement in interconnect content will receive more solid support; if delays or lower-than-expected market share occur, high valuations will face correction pressure. The structural opportunity of AI interconnects does exist, Marvell's full-stack layout offers differentiated options, but the ultimate level of realization depends on execution details, with 2027 being a key validation window.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
