Marvell Technology (NASDAQ: MRVL) confirmed on October 8, 2026 that it would showcase its full portfolio of AI data center infrastructure solutions at the OCP Global Summit 2026, which ran October 12-15 at the San Jose McEnery Convention Center. The announcement, first flagged by Yahoo Finance and picked up by outlets including CBS News, SFGATE and The Morning Call, put Marvell’s scale-out switching, memory expansion, and cryptographic security technology in front of the hyperscale operators and cloud builders who gather each year at the Open Compute Project’s flagship event.
The centerpiece of Marvell’s booth (#B31) was the Teralynx T100, a monolithic 102.4 Tbps data center switch built on a 3 nm design with 512-port radix for next-generation AI and cloud networks. Marvell says the T100 typically draws under 1,000 W and claims up to 25% lower power than rival switches. Alongside it, Marvell brought its Structera CXL memory expansion and acceleration line, the Ara DSP optical retimer family, and a set of connectivity and storage products aimed squarely at the bottlenecks now showing up inside AI clusters: bandwidth, memory capacity, and power.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Marvell Confirms AI Data Center Showcase at OCP 2026
Marvell’s October 8 announcement lays out a broad set of demonstrations rather than a single headline product launch. According to the announcement, the company planned to demonstrate scale-out switching, memory expansion and acceleration, cryptographic security, and related connectivity technologies across its OCP 2026 presence, which ran during the summit on October 12 to 15. That breadth matters: OCP Global Summit has become the venue where hyperscalers like Meta, Microsoft and Google compare notes on the open hardware specifications that underpin their AI data centers, and silicon vendors use the show to signal which parts of the stack they are prioritizing for the next build cycle.
Marvell Teralynx T100 is the clearest signal of where the company sees the most urgent need. At 102.4 Tbps, it is positioned by Marvell as a switch purpose-built for AI and cloud data center infrastructure, aimed at both frontend networking (the traffic that moves data in and out of a cluster) and the backend scale-up fabrics that stitch GPUs and accelerators together inside a single AI system. Marvell is also showing 400G-per-lane PAM4 signaling and 1.6T active electrical cables, two pieces of the same puzzle: getting more bits through the same physical footprint as AI clusters scale from thousands to hundreds of thousands of accelerators.
What OCP Global Summit 2026 Means for AI Infrastructure
The Open Compute Project exists to standardize the hardware that hyperscale data centers run on, and OCP Global Summit is where those standards get debated, demonstrated, and in many cases adopted into procurement roadmaps. For a company like Marvell, whose business depends on being designed into custom silicon programs at Amazon, Microsoft, Google and Meta, a strong OCP showing is less about selling a single box and more about proving its switching, memory, and optical building blocks fit the open specifications those customers are standardizing around.
Marvell’s decision to bring 14 separate presentations to OCP 2026, delivered by its executives and engineers, underscores how granular that sales motion has become. AI infrastructure is no longer sold as a single chip; it is sold as a set of interlocking subsystems; switching fabric, memory pooling, optical interconnect, storage controllers, and hardware-based security; each of which needs its own technical case made to a skeptical audience of hyperscale infrastructure architects.
Inside the Teralynx T100: A 102.4 Tbps Switch Built for AI
Marvell’s own product materials describe the Teralynx T100 as a 102.4 Tbps, UEC-compliant, full-duplex switching processor (part-number family MV-TX93xxx) built for fixed spine-and-leaf systems used in AI data center frontend and backend networks, as well as scale-up fabric applications. The reference to UEC, the Ultra Ethernet Consortium, is notable because it signals Marvell is building toward an industry effort to make Ethernet a credible alternative to proprietary interconnects for linking tens of thousands of AI accelerators, a space currently dominated by Nvidia’s InfiniBand and NVLink technologies.
Why Switching Capacity Is the New Bottleneck
As AI training and inference clusters grow, the network fabric connecting GPUs often becomes the limiting factor on how fast a job can run, not the compute silicon itself. A switch that can push 102.4 Tbps gives data center architects more headroom to add accelerators to a single fabric without introducing the latency penalties that come from stacking multiple generations of lower-capacity switches. Marvell’s broader Teralynx lineup already supports 800 GbE and includes congestion-management features intended to keep AI training traffic; which tends to arrive in sudden, synchronized bursts as accelerators finish a computation step together; from overwhelming the fabric.
Structera CXL: Solving the AI Memory Wall
The second pillar of Marvell’s OCP 2026 showcase is Structera, its Compute Express Link (CXL) memory expansion and acceleration platform. The announcement describes Structera’s role at OCP as memory expansion and acceleration for AI data centers, and Marvell announced the second-generation Structera X 3406 and 3506 on October 12, 2026, CXL 3.2 memory expanders with industry-first six-channel DDR4/DDR5 support and PCIe Gen 6 connectivity, scaling AI inference with up to 9TB capacity and 256 GB/s bandwidth. It is a category that has taken on new urgency as large language model training and inference jobs increasingly run out of memory capacity before they run out of raw compute.
Marvell’s own CXL product documentation, published ahead of OCP 2026, frames Structera as addressing what the company calls the “AI memory wall”, the gap between how fast accelerator compute has scaled and how slowly memory capacity and bandwidth per accelerator have kept pace. CXL switches let a cluster pool memory across multiple CPUs or GPUs rather than stranding unused DRAM capacity on individual nodes, which matters as agentic AI inference workloads; the kind that hold long conversational or reasoning contexts; push memory requirements well past what a single accelerator’s onboard memory can hold.
DDR4, DDR5 and the Cost of Memory Pooling
Marvell’s CXL blog materials note that its Structera platform is designed to work across both DDR5 and DDR4 memory generations, which gives data center operators a path to repurpose older DIMM inventory into pooled memory capacity rather than retiring it outright. That flexibility carries real financial weight this year: memory pricing has climbed sharply through 2026 as AI-driven DRAM demand collides with constrained fab capacity, making every strategy that stretches existing memory further more valuable to hyperscale buyers watching their capital budgets.
Optical Connectivity: Ara DSP and the 1.6T Push
Marvell’s OCP 2026 demonstrations also include the Ara DSP, described in the company’s announcement as 1.6T transmit-retimed optics, alongside 400G-per-lane PAM4 signaling. Retimed optics sit between a switch’s electrical interface and the fiber that carries data across a data center, cleaning up and amplifying signals so they can travel further without errors. As AI clusters spread across more racks and, increasingly, more buildings, that optical layer becomes just as important as the switching silicon itself.
Marvell is pairing the Ara DSP with its COLORZ 800G ZR/ZR+ QSFP module, which adds MACsec encryption directly at the optical layer, and with 200G and 300G active copper connectivity for shorter links inside a rack. This is a recurring theme in Marvell’s 2026 product cadence: the company has used earlier events this year, including a 2nm optical technology demonstration at ECOC 2026, as covered by Marvell’s own investor relations team, to build toward the same goal of pushing more bandwidth through AI data center interconnects while keeping power and cost in check. On October 13, 2026, Marvell added the Alaska E 3.2T protocol-aware retimer and a second-generation 1.6T AEC DSP to that connectivity push.
Security and Storage: Bravera and the Hardware Security Module
Beyond networking and memory, Marvell’s OCP 2026 lineup includes the Bravera PCIe 6.0 SSD controller and an integrated hardware security module. The storage controller addresses a less glamorous but increasingly critical piece of AI infrastructure: moving massive training datasets and model checkpoints in and out of storage fast enough to keep expensive GPU clusters from sitting idle waiting on I/O. PCIe 6.0 roughly doubles the per-lane bandwidth of PCIe 5.0, which matters as checkpoint files for large models have grown into the hundreds of gigabytes.
The hardware security module, meanwhile, reflects growing hyperscaler demand for cryptographic protection built directly into data center silicon rather than bolted on at the software layer, covering everything from securing model weights in transit to authenticating the hardware itself inside a shared cloud environment.
Marvell’s 14 OCP Sessions: What’s on the Agenda
Marvell confirmed that its executives and engineers were scheduled to deliver a full slate of sessions at OCP Global Summit 2026, which ran October 12 to 15. While the company’s announcement did not break down each session’s exact title or specific technical performance claims beyond what it has already published, the breadth of the lineup signals how far Marvell’s AI infrastructure story has expanded beyond a single chip pitch. A company showing up with that many distinct technical sessions is effectively arguing that it belongs in nearly every layer of the AI data center stack: switching, memory, optics, storage, and security.
That volume of technical content is also a competitive signal aimed at the hyperscale engineering teams in the audience, who increasingly design their own custom AI silicon and shop among merchant-silicon suppliers like Marvell, Broadcom, and Astera Labs for the pieces they don’t build in-house.
Marvell’s OCP 2026 Showcase at a Glance
| Product | Category | Confirmed Detail |
|---|---|---|
| Teralynx T100 | Data center switch | 102.4 Tbps switching capacity |
| Structera | CXL memory expansion | Memory expansion and acceleration for AI clusters |
| Ara DSP | Optical connectivity | 1.6T transmit-retimed optics |
| PAM4 signaling | Electrical interconnect | 400G per lane |
| COLORZ 800G ZR/ZR+ | Optical module | QSFP form factor with MACsec encryption |
| Active electrical cables | Rack-level connectivity | 1.6T throughput |
| Active copper connectivity | Short-reach connectivity | 200G and 300G variants |
| Bravera | Storage controller | PCIe 6.0 SSD controller |
| Hardware security module | Cryptographic security | Integrated data center security |
Marvell OCP 2026 quick spec sheet
----------------------------------
Teralynx T100........102.4 Tbps switch, AI/cloud data center fabric
Structera.............CXL memory expansion and acceleration
Ara DSP................1.6T transmit-retimed optics
PAM4 lane rate.........400G/lane
COLORZ 800G ZR/ZR+.....QSFP module, MACsec-enabled
Active electrical cable 1.6T
Active copper..........200G / 300G
Bravera.................PCIe 6.0 SSD controller
Security................Integrated hardware security module
Booth...................#B31, OCP Global Summit 2026
Sessions................14 presentationsWhy AI Data Centers Need Scale-Out Switching Now
The timing of Marvell’s OCP push lines up with a broader industry conversation about whether AI data center buildouts are running into physical limits faster than expected. TechRadar’s reporting on the grid, thermal, and interconnect challenges facing AI data centers has documented how power delivery and cooling have become as central to data center design as the chips themselves, and networking sits right alongside those constraints: every watt spent moving data between accelerators is a watt not spent on computation, and every nanosecond of added latency in the fabric is time an expensive GPU cluster spends waiting instead of working.
That context helps explain why Marvell is pitching the Teralynx T100 and its surrounding connectivity portfolio as efficiency plays as much as capacity plays. A single higher-capacity switch that can replace several lower-capacity units reduces the number of active components drawing power and generating heat inside a rack, a meaningful consideration as data center operators increasingly describe power, not capital, as their binding constraint on AI growth.
Market Impact: MRVL Stock and the AI Infrastructure Trade
Marvell Technology trades on the Nasdaq under the ticker MRVL, and the company has positioned itself for much of 2026 as a pure-play bet on the custom silicon and networking side of the AI buildout, distinct from GPU makers like Nvidia, with its Investor Day 2026 materials targeting $80B in FY 2031 revenue from custom silicon and optics. Investors have treated Marvell’s quarterly updates and marquee event appearances, including its OCP 2026 showcase, as data points on how much of the hundreds of billions of dollars hyperscalers are spending on AI infrastructure this year will flow toward switching, memory, and optical suppliers rather than concentrating entirely in GPU silicon.
Why Investors Watch Event Cadence Closely
Marvell has used a steady drumbeat of events through 2026, including AI Infra Summit in September and ECOC in the optical space, to keep its AI infrastructure narrative in front of both customers and investors between quarterly earnings calls. OCP Global Summit 2026 is the highest-profile of these so far this quarter because of its direct audience of hyperscale infrastructure buyers, making it a closely watched proof point for whether Marvell’s design-win pipeline is translating into the kind of broad-based demand that would show up across switching, CXL, and optical revenue lines simultaneously rather than in just one category.
Competitive Landscape: Marvell vs Broadcom vs Nvidia vs Astera Labs
Marvell’s OCP 2026 pitch lands in a crowded field. Broadcom remains the dominant force in merchant Ethernet switch silicon for data centers and has its own aggressive AI networking roadmap, while Nvidia continues to push its InfiniBand and NVLink ecosystems as the preferred fabric for its own GPU clusters, giving it a built-in advantage when customers buy Nvidia accelerators and networking as a bundle. Astera Labs has carved out a focused position in PCIe and CXL connectivity retimers and controllers, competing most directly with the memory-expansion side of Marvell’s Structera line rather than its switching business.
| Company | Core AI Data Center Focus | Primary Competitive Position vs Marvell |
|---|---|---|
| Marvell Technology | Switching, CXL memory, optical DSPs, storage controllers, security | Broadest horizontal portfolio across networking, memory, and storage |
| Broadcom | Ethernet switch silicon, custom AI accelerator ASICs | Larger incumbent in merchant switching, direct T100 rival |
| Nvidia | GPUs, InfiniBand/NVLink interconnect, networking | Owns the accelerator plus proprietary fabric bundle |
| Astera Labs | PCIe/CXL retimers and smart connectivity controllers | Closest rival to Structera in CXL memory expansion |
Historical Context: From Merchant Silicon to AI-Era Custom Chips
Marvell’s path to an OCP 2026 keynote-level presence did not happen overnight. The company built its data center silicon business over more than a decade of supplying merchant Ethernet switch chips and storage controllers to networking equipment makers, long before “AI infrastructure” became a standalone category that investors tracked on earnings calls. The shift toward custom silicon programs for hyperscale customers, where Marvell designs application-specific chips tailored to a single cloud provider’s AI workloads rather than selling off-the-shelf parts, has reshaped how the company is valued and how closely its product roadmap tracks the AI buildout cycle.
Compute Express Link itself followed a similar arc. CXL launched in 2019 as a cache-coherent interconnect standard aimed mostly at connecting accelerators to CPU memory inside a single server. It has since evolved, through CXL 2.0 and 3.x, into a technology capable of pooling memory across many servers in a rack, which is exactly the memory-expansion use case Marvell’s Structera platform targets. That evolution from a single-server interconnect to a rack-scale memory fabric mirrors the industry’s broader pivot from traditional cloud computing architecture toward AI-specific infrastructure designed around accelerator clusters rather than general-purpose servers.
Analysis: What This Means for Hyperscalers and Cloud Buyers
For the hyperscale operators in OCP’s audience, Marvell’s showcase functions as a menu of building blocks rather than a single purchase decision. A cloud provider designing its next-generation AI cluster can, in principle, mix Marvell’s switching silicon with memory pooling from Structera, optical links built around the Ara DSP, and storage controllers from the Bravera line, all while sourcing accelerators from Nvidia, AMD, or an in-house custom chip program. That modularity is precisely what OCP’s open hardware specifications are designed to enable, and it is why a horizontal supplier like Marvell has an incentive to show up with 14 sessions covering nearly every layer of the stack rather than one flagship announcement.
The practical effect for cloud buyers is more supplier choice at each layer of the AI data center, which should, over time, put pressure on pricing and push performance gains faster than a world where a single vendor controlled the entire stack from accelerator to network to memory. Whether that plays out depends heavily on how quickly Ultra Ethernet Consortium-compliant switching, like the Teralynx T100, can close the performance gap with Nvidia’s proprietary interconnects in real-world large-scale training clusters, a question that will likely only be settled by independent benchmarking once the hardware ships broadly, which Marvell’s available reporting targets for early 2027 while the T100 remains in sampling.
Predictions: Where Marvell’s AI Infrastructure Bet Goes Next
- Expect Marvell to lean harder into Ultra Ethernet Consortium messaging through the rest of 2026 as it tries to position the Teralynx T100 as a credible alternative to Nvidia’s proprietary scale-up fabrics for hyperscalers that want multi-vendor flexibility.
- Structera and other CXL memory-pooling products are likely to see faster adoption than switching in the near term, since memory costs have risen sharply in 2026 and give operators an immediate financial incentive to pool existing DRAM rather than buy more.
- Optical connectivity, led by the Ara DSP and COLORZ modules, will keep gaining relative importance as AI clusters spread across more racks and buildings, pushing data center operators to prioritize optical bandwidth per watt alongside raw switch capacity.
- Competitive pressure from Broadcom’s own AI networking roadmap and Astera Labs’ CXL connectivity push will likely keep margins on data center switching and memory-expansion silicon under scrutiny, even as unit volumes grow.
- Watch for Marvell to use its next earnings call to quantify how much design-win momentum from events like OCP 2026 is translating into booked revenue, since investors have shown a pattern of treating event appearances as leading indicators rather than confirmed financial outcomes.
Frequently Asked Questions
What is Marvell showcasing at OCP Global Summit 2026?
Marvell showcased its AI data center infrastructure portfolio at booth #B31 during the summit, which ran October 12 to 15. The lineup included the Teralynx T100 switch, Structera CXL memory expansion and acceleration technology, Ara DSP optical connectivity, COLORZ optical modules, active electrical and copper cabling, the Bravera PCIe 6.0 SSD controller, and an integrated hardware security module.
What is the Marvell Teralynx T100?
The Teralynx T100 is a 102.4 Tbps data center switch that Marvell describes as purpose-built for next-generation AI and cloud data center infrastructure, supporting both frontend networking and backend scale-up fabric applications.
What does Marvell Structera do?
Structera is Marvell’s CXL (Compute Express Link) memory expansion and acceleration platform, designed to help AI data centers pool and extend memory capacity across CPUs and accelerators rather than leaving unused memory stranded on individual servers.
When and where is OCP Global Summit 2026?
OCP Global Summit 2026 ran October 12-15, 2026, at the San Jose McEnery Convention Center in San Jose, California.
How many sessions is Marvell presenting at OCP 2026?
Marvell confirmed that its executives and engineers were scheduled to deliver sessions covering its AI data center infrastructure portfolio during OCP Global Summit 2026, which took place October 12 to 15.
Who are Marvell’s main competitors in AI data center silicon?
Marvell’s closest competitors in this space include Broadcom in merchant Ethernet switching, Nvidia through its InfiniBand and NVLink interconnect ecosystem, and Astera Labs in PCIe and CXL connectivity, particularly around memory expansion technology that competes with Structera.
Has Marvell announced pricing for the Teralynx T100 or Structera?
No. Marvell’s October 8, 2026 announcement did not include pricing for the Teralynx T100, Structera, or any other product in its OCP 2026 showcase.
What does this announcement mean for MRVL stock?
The OCP 2026 showcase itself is a marketing and design-win event rather than a financial disclosure, so it does not directly move reported revenue. Investors typically treat appearances like this as signals of Marvell’s positioning in the broader AI infrastructure spending cycle, with the real financial impact depending on whether design wins from events like OCP eventually convert into booked orders disclosed in future earnings reports.
