jain.com — data centers, connectivity and security, analyzed daily

Tag: Marvell

  • Marvell’s 102.4Tbps 2nm Optics Make AI Scale-Out a Watts-Per-Bit Problem

    Marvell’s 102.4Tbps 2nm Optics Make AI Scale-Out a Watts-Per-Bit Problem

    TL;DR · 30-second read

    The Short Version

    Computers that train artificial intelligence have to talk to each other constantly, and today most of that traffic still travels over copper wiring that burns electricity and fades over distance.

    Marvell has shown chips that move the traffic as light instead, built on the most advanced chipmaking process now available. One demonstration moves data fast enough to copy well over a thousand high-definition movies every second.

    The point is not speed. Light uses less power per unit of data, and electricity is the hard limit on how big these computing sites can get.

    Marvell Technology has demonstrated optical interconnect silicon built on a 2nm process for AI data centers, including a co-packaged optics engine running at 102.4 terabits per second and supporting 200G and 400G per electrical lane. 24/7 Wall St. reported on September 21 that the company described the work as an industry first and said it is “pushing the envelope of what is possible in optical tech,” with shares trading around $248 in Monday’s first hour, up 1.5% on the session.

    The demonstration sits on top of a quarter Marvell has already reported: record fiscal second-quarter revenue of $2.739 billion, up 37% year over year, with data center revenue of $2.17 billion, up 46%, now 79% of the total. Guidance filed August 27 put the current quarter at $3.150 billion plus or minus 5%, and raised the fiscal 2027 and fiscal 2028 outlooks. Management has scheduled an Investor Day for October 6 in New York.

    Executive Summary

    The headline number is 102.4 Tbps of co-packaged optics, but the operational story is the denominator. Co-packaged optics — CPO — moves the optical engine off the front panel of a switch and onto the same substrate as the switch or accelerator silicon. That shortens the electrical path a signal has to travel before it becomes light, and shortening that path is primarily a way to spend fewer watts per bit, not a way to go faster.

    That distinction matters because a rack has a fixed power envelope set by the building, not by the silicon roadmap. Every watt the network consumes inside that envelope is a watt unavailable to the accelerators the network exists to connect. As Marvell steps its lane rates from the 800G generation deployed today, past the 1.6T generation now ramping, toward 3.2T, copper’s usable reach shrinks and more of the interconnect budget migrates to optics — which is exactly where the energy cost concentrates.

    The demo also gives silicon substance to forecasts management has already made. On the earnings call, CEO Matt Murphy said “the magnitude of our scale-up optics opportunity next year is much larger than we thought just a quarter ago,” and that Marvell expects scale-across bandwidth requirements to exceed 10 times current front-end data center interconnect networks. What a demonstration does not establish is yield, production timing, field serviceability or a named customer.

    Optics Stopped Being a Bandwidth Problem and Became a Power Budget

    Here is the mechanism the 102.4 Tbps figure is really about. In a conventional switch, optical modules plug into the faceplate, and the switch chip’s electrical lanes have to drive signals across the package, the circuit board and a connector before they reach the laser. Each of those hops costs energy and frequently requires retiming circuitry to clean the signal up. Co-packaged optics collapses that journey by placing the optical engine on the same substrate as the switch or accelerator die — in Marvell’s framing, shortening data pathways and cutting energy per bit.

    The lane rates are what force the issue. By demonstrating 200G and 400G per lane, Marvell is aiming past the 800G optics deployed across hyperscale fleets today and the 1.6T generation now ramping, at the 3.2T generation beyond it. As per-lane signalling rates climb, the distance a passive copper cable can carry a clean signal falls, which is why the release notes copper runs out of reach at the speeds these clusters require. Once copper is out, the interconnect is optical by default — and its power draw becomes a line item in the rack budget rather than a rounding error.

    Who this lands on is specific. A facility commits a fixed number of megawatts to a hall; a rack design commits a fixed number of kilowatts to a cabinet. If the network layer connecting tens of thousands of accelerators across multiple racks consumes a smaller share of that envelope, the operator either fits more compute per rack or fewer racks per hall. That arithmetic is the reason a demo with no customer attached still moves the roadmap conversation for switch vendors, optical module suppliers, rack integrators and the people underwriting the power contracts underneath them.

    A 79% Data Center Mix Makes the Optical Roadmap the Whole Company

    Marvell’s reported quarter explains why an engineering milestone carries this much weight. Revenue of $2.739 billion, up 37% year over year, came with data center revenue of $2.17 billion, up 46% and equal to 79% of the total. Non-GAAP earnings were $0.94 per share, and the company guided the following quarter to $3.150 billion plus or minus 5% with earnings of $1.10 plus or minus $0.05, while raising its fiscal 2027 and fiscal 2028 outlooks in the August 27 filing.

    A business that derives roughly four dollars in five from data centers has effectively converted itself into a single-cycle instrument. That is an advantage while hyperscale capital expenditure expands and a concentrated exposure when it does not. It also means the optical and custom silicon roadmaps are not adjacent product lines — they are the earnings model, which is why a 2nm demonstration is treated as corroborating evidence for guidance rather than as a lab result.

    The market has priced that trajectory rather than waiting on it. The stock has advanced 192% year to date, carries a market capitalisation near $217.4 billion and a forward price-to-earnings multiple around 58, against an average analyst target of $289 with 39 buy or strong-buy ratings and five holds. A multiple set that far ahead of current earnings is a statement about execution risk: the roadmap has to arrive roughly on the schedule management has sketched, and the October 6 Investor Day is where that schedule gets specified.

    What Being First Actually Buys — and What It Costs

    The optical push runs alongside a custom silicon business that has been given an unusual structure. Management disclosed an expanded partnership with Google that includes a warrant allowing Google to acquire up to 7% of Marvell’s shares, tied to revenue milestones, and Murphy said custom revenue should more than double in fiscal 2028. Linking a customer’s equity upside to the volume it buys aligns incentives efficiently, but it also converts a commercial relationship into potential dilution for existing holders, on terms whose triggers Marvell has not published. A separate expanded NVLink Fusion arrangement with NVIDIA positions Marvell in scale-up switching, and the earlier Celestial AI photonic fabric acquisition, closed February 2, 2026, and the XConn Technologies chiplet connectivity deal, closed February 10, 2026, supply the building blocks.

    Against Broadcom, the incumbent in high-speed networking silicon, first-to-2nm is a claim on the next design cycle rather than on current sockets. Hyperscalers qualify interconnect years ahead of deployment, so the practical value of a demonstration is optionality in roadmap reviews happening now for clusters that power up later. It converts into revenue only if the process yields at volume, the thermal and reliability behaviour survives qualification, and a named platform adopts it.

    The demand side of that equation is funding itself in size, if not always for named projects: Amazon’s Form 8-K filed September 14 records the close of £4.242 billion of notes maturing between 2029 and 2045, with net proceeds of roughly £4.235 billion and no specific use of proceeds identified in the filing. Capital of that scale is available to the companies buying interconnect; whether it reaches co-packaged optics specifically is the question Marvell has not yet answered with a customer name.

    Background

    Marvell Technology designs custom silicon and networking chips for hyperscale data centers. Its portfolio spans custom AI accelerators built to a single customer’s specification, electro-optics, the digital signal processors that drive high-speed optical links, and data center interconnect products that carry traffic between facilities. Over the past two years the company has reweighted itself decisively toward that market, with the data center segment now accounting for 79% of revenue.

    Interconnect has become one of the more consequential bottlenecks in AI infrastructure. Training and inference clusters link tens of thousands of accelerators, and the network that connects them has to scale in step with the compute — first inside a rack, then across racks, then between buildings. Marvell has assembled capability for that through acquisition as well as internal development, closing the Celestial AI photonic fabric deal on February 2, 2026 and the XConn Technologies chiplet connectivity deal on February 10, 2026, while expanding partnerships with NVIDIA on scale-up switching and with Google on custom silicon. Its principal competitor in high-speed networking silicon is Broadcom.

    Sources

    Source: Marvell Smashes the AI ‘Power Wall’ with Industry-First 2nm Optical Chips — 24/7 Wall St. report on Marvell’s 2nm optical interconnect demonstrations, its fiscal second-quarter results and its custom silicon partnerships.

    Primary sources: Amazon.com, Inc., Form 8-K filed September 14, 2026 (close of £4.242 billion of notes due 2029 through 2045; net proceeds of approximately £4.235 billion).

  • Marvell’s $5.5B AI Optics Deal and the Interconnect Bottleneck

    Marvell’s $5.5B AI Optics Deal and the Interconnect Bottleneck

    A widely syndicated item from retail-investor research site simplywall.st, circulating through Google News, asks what Marvell Technology (Nasdaq: MRVL) gains from a $5.5 billion AI optics deal. Marvell is a US-based fabless chip designer whose largest end market is data center silicon, including the optical components that move data between AI servers.

    The syndicated text available to us consists of the headline and link only. It does not name a counterparty, state whether Marvell is the buyer or the seller, describe the consideration mix, or give a closing date. The $5.5 billion figure and the “AI optics” framing are the only substantive details carried in the source, and neither is accompanied in that material by a quote or a primary company disclosure.

    Executive Summary

    The headline points at a genuinely important shift, even though the source itself is thin. For most of the current AI build cycle, the constraint operators talked about was compute: how many accelerators could be bought, powered and cooled. Increasingly the binding constraint is the fabric between those accelerators. A training or inference cluster is only as fast as its slowest link, and the links are now measured in hundreds of thousands of optical connections per site.

    That is why a $5.5 billion transaction attached to “AI optics” is worth attention regardless of its direction. Optical interconnect sits at the intersection of two things that are hard to replicate: high-speed mixed-signal silicon, where Marvell has a strong franchise inherited from its Inphi acquisition, and photonics manufacturing, where supply has been tight through the AI cycle. A deal of this size in that space either consolidates a defensible position or monetises one.

    The honest caveat is that the material in front of us does not establish which. Readers evaluating the transaction should treat the $5.5 billion number as reported by a third-party analysis site and verify structure, counterparty and timing against Marvell’s own filings before drawing conclusions about accretion, market share or roadmap.

    Why the Wires Became the Bottleneck

    Modern AI clusters are not single computers. They are thousands of accelerators stitched together so tightly that software treats them as one machine. Two networks do that stitching. Scale-up connects a handful to a few dozen chips inside a rack at extremely high bandwidth and very low latency. Scale-out connects racks to each other across the hall. Both have had to grow roughly in step with accelerator performance, and accelerator performance has been growing faster than copper cabling can comfortably follow.

    Beyond a metre or two at current data rates, copper runs out of headroom and the signal degrades. That pushes traffic onto optics: lasers, fibre and the transceiver modules that convert electrical signals to light and back. Inside those modules sit digital signal processors, or DSPs, which clean up a distorted waveform so the receiving end can read it. Each generational jump, 400G to 800G to 1.6T per port, roughly doubles the data a single link carries and forces a redesign of that signal chain. Marvell’s electro-optics business, built largely on its 2021 Inphi acquisition, is one of the small number of places that silicon comes from.

    The economic consequence is that optics have moved from a rounding error to a meaningful share of cluster capital cost, and from a background concern to a live operational one. Optical modules consume power and they fail; at hundreds of thousands of links per site, even a low failure rate becomes a staffing and spares problem. Any vendor that can cut watts per bit or improve link reliability is selling something operators will pay for.

    What a $5.5 Billion Number Implies, in Either Direction

    Read as an acquisition, $5.5 billion is large but not transformative for a company of Marvell’s scale. It would signal that management sees interconnect as the durable part of the AI stack, and the questions that follow are conventional: what revenue and gross margin come with the assets, whether the consideration is cash, stock or both, how it affects the balance sheet, and how long integration takes relative to the eighteen-to-twenty-four-month cadence at which optical generations turn over. In fast-moving silicon markets, an acquired roadmap can age before it closes.

    Read as a divestiture, the same number tells a different story: capital recycled out of a components business and toward custom accelerator silicon, where Marvell designs bespoke chips for individual hyperscale customers. That path trades a broad merchant franchise for deeper exposure to a small number of very large buyers. Neither reading is inherently better. They imply different risk profiles, and the source material does not let us choose between them.

    What holds in both cases is that the buyers are concentrated. A handful of hyperscalers and large AI labs account for the bulk of demand for high-speed optics. Concentration is pleasant on the way up, because a single design win can move a quarter, and unpleasant on the way down, because a single deferred build can do the same. Any assessment of this transaction that ignores customer concentration is incomplete.

    Custom Silicon Plus Photonics: A Real Moat With Real Erosion Risk

    The strategic case for combining custom accelerator design with optical interconnect is coherent. A vendor that designs a customer’s chip and also supplies the links between those chips can co-optimise the two, and it becomes harder to displace because switching costs compound across the design cycle. That is a genuine moat, not a slogan.

    It is also under pressure from several directions at once, and an even-handed analysis has to say so. Broadcom competes across switching silicon, optical DSPs and custom accelerators simultaneously. Nvidia has strong incentives to keep its scale-up fabric proprietary and in-house. Specialists such as Credo and Astera Labs attack adjacent slices of the connectivity problem, and module manufacturers in the United States and Asia compete hard on cost. Meanwhile hyperscalers keep expanding their own silicon teams, which makes today’s supplier a candidate for tomorrow’s insourcing.

    The most interesting technical risk is co-packaged optics, or CPO, which moves the optical engine onto the same package as the switch or accelerator instead of into a pluggable module at the faceplate. Done well, CPO saves power and board area. It also changes which components carry value and could reduce the role of the standalone DSP that anchors part of Marvell’s franchise. CPO has been arriving more slowly than its advocates predicted, partly because pluggable modules are serviceable and CPO largely is not, but the direction of travel is worth watching. A $5.5 billion commitment in optics is a bet on how that transition resolves.

    Reading a Headline-Only Story Responsibly

    This is a case where the analysis is more substantiated than the news. The industry context is well established: interconnect is a real bottleneck, optics is a real chokepoint, and consolidation there is a rational strategy. The specific transaction, as carried in this source, is a dollar figure in a headline from a third-party research site.

    That is not a criticism of the publisher, whose format is short-form investor commentary rather than primary reporting. It is a caution about how such items propagate. A number repeated across aggregators acquires an authority its original sourcing may not support, and AI summarisation tends to accelerate that effect. The appropriate response is to anchor on primary documents: a company press release, an SEC filing, or a counterparty confirmation.

    For practitioners, the practical takeaway is independent of the deal’s details. If you are procuring capacity or designing clusters, interconnect supply, roadmap alignment and vendor concentration deserve the same diligence you already apply to accelerators and power. Consolidation among optics suppliers, whichever way this transaction runs, narrows the field you are negotiating with.

    Background

    Marvell Technology is a fabless semiconductor company, meaning it designs chips and outsources their manufacture to foundries. Founded in 1995 and headquartered in Santa Clara, California, it spent its early years in storage controllers and consumer connectivity before reorienting around infrastructure silicon under chief executive Matt Murphy. A sequence of acquisitions built that position: Cavium in networking processors, Aquantia in Ethernet, Innovium in switching, and Inphi in high-speed electro-optics, its largest deal to date.

    The data center is now Marvell’s principal end market, spanning custom accelerator silicon for hyperscale customers, Ethernet switching, storage controllers and the optical components that connect servers. The company has also been pruning: in 2025 it agreed to sell its automotive Ethernet business to Infineon, a move consistent with concentrating capital on AI infrastructure. That context is why a multibillion-dollar transaction in AI optics reads as strategy rather than opportunism, whichever side of it Marvell turns out to be on.

    Source: What Does Marvell Technology (MRVL) Gain From Its $5.5 Billion AI Optics Deal? — a short-form investor analysis item from simplywall.st, distributed via Google News, whose syndicated text carries the $5.5 billion figure without accompanying transaction details.

  • Google’s $12.2B Marvell Deal Reshapes the Custom AI Chip Race

    Google’s $12.2B Marvell Deal Reshapes the Custom AI Chip Race

    Google has expanded its custom AI chip partnership with Marvell Technology in a deal reported at $12.2 billion, according to multiple Yahoo Finance reports published this week. Broadcom — long regarded as Google’s incumbent partner for custom AI accelerators — saw its shares fall 6.2% on the news, while analyst fair-value estimates for Marvell edged higher.

    Executive Summary

    The reported agreement deepens Google’s relationship with Marvell for custom silicon — chips designed to a single customer’s specification rather than sold off the shelf. In AI infrastructure, these custom accelerators (often called XPUs or ASICs) are the hyperscalers’ primary lever for reducing dependence on Nvidia’s general-purpose GPUs, and the design partner that wins the engagement captures years of high-visibility revenue.

    The market reaction tells the story in one frame: Broadcom, which has been widely credited as the co-design partner behind Google’s Tensor Processing Units (TPUs), dropped 6.2%, while Marvell’s bull case strengthened. A $12.2 billion figure, if it represents committed or expected purchases, would be one of the larger custom-silicon engagements publicly reported — though the source articles leave the deal’s structure, duration, and scope largely undefined.

    For the broader AI infrastructure market, the significance is less about one stock move and more about confirmation of a trend: hyperscalers are dual-sourcing their chip design partners the same way they dual-source power, fiber, and data center capacity — to control cost, schedule risk, and negotiating leverage.

    Why Hyperscalers Refuse to Depend on One Chip Partner

    Custom AI accelerators are multi-year commitments. A hyperscaler like Google picks a design partner, co-develops a chip over 18–36 months, then ramps production across successive generations. That timeline creates lock-in — and lock-in creates pricing power for the partner. Broadcom’s custom-silicon business has been a major beneficiary of exactly that dynamic. By expanding work with Marvell, Google gains a credible second source, which pressures pricing on every future generation and insulates its TPU roadmap from any single vendor’s execution stumbles.

    This mirrors how large infrastructure buyers behave everywhere in the stack. No serious operator single-sources grid power, network transit, or construction contractors for a multi-gigawatt buildout. As custom silicon becomes as strategically important as the data centers that house it, the same procurement discipline is arriving in chip design.

    Broadcom’s 6.2% Drop: Signal Versus Substance

    A one-day 6.2% decline reflects what investors fear, not necessarily what Google has decided. The reports do not state that Google is reducing its Broadcom engagement — only that it is expanding Marvell’s. Those are different things: Google’s total accelerator demand is growing fast enough that two partners could both see rising volumes. The bearish reading is about share and leverage, not necessarily absolute revenue.

    That said, the concern is not irrational. In custom silicon, the design win for generation N strongly influences who builds generation N+1. If Marvell’s expanded role includes compute (the accelerator itself) rather than adjacent components such as networking or interconnect silicon, the competitive implications for the incumbent are materially larger. The source reporting does not settle that question — and it is the single most important unknown in this story.

    What $12.2 Billion Does — and Doesn’t — Tell Us

    Headline deal values in semiconductors deserve careful reading. A $12.2 billion figure could represent firm purchase commitments, a cumulative multi-year revenue expectation, or an analyst’s sizing of the opportunity — each with very different levels of certainty. The reports cited here frame it as changing Marvell’s bull case, which suggests investors are treating it as durable pipeline, but the articles do not disclose contract structure, timeline, or margin profile.

    Custom silicon also carries structurally lower gross margins than merchant chips, because the customer funds the design and captures much of the value. Marvell’s win is real in revenue-visibility terms; whether it is equally attractive in profitability terms depends on details not yet public.

    Downstream Effects on AI Infrastructure Buyers

    For enterprises and operators who buy cloud AI capacity rather than chips, this competition is quietly good news. Every credible alternative to Nvidia GPUs — and every second source within the custom-silicon supply chain — adds capacity to a market that has been supply-constrained for years. More TPU supply at better economics ultimately shows up as more available accelerated compute, and potentially better pricing, for Google Cloud customers. It also intensifies demand on the physical layer: more accelerator volume means more high-density data center space, more power procurement, and more advanced cooling — the parts of the stack where constraints now bind hardest.

    Background

    Google has designed its own AI accelerators — the TPU line — for roughly a decade, working with external semiconductor partners on design and production. Broadcom has long been identified in industry reporting as the principal partner behind that program, and custom accelerators for hyperscalers have become one of the fastest-growing segments in semiconductors as cloud providers seek alternatives to merchant GPUs. Marvell, meanwhile, has built its own custom-compute franchise serving hyperscale customers, making it the most frequently cited challenger to Broadcom in this market.

    The reported $12.2 billion expansion lands in that context: a two-horse race for hyperscaler design partnerships, where each win shapes multiple future chip generations and, downstream, the data center, power, and cooling infrastructure required to deploy them.

    Source: Broadcom (AVGO) Is Down 6.2% After Google Expands AI Chip Ties With Marvell — Yahoo Finance, with related Yahoo Finance coverage of Marvell’s reported $12.2 billion Google partnership expansion and its impact on analyst fair-value estimates.

  • Marvell’s 32TB Memory Tier Bets AI Inference Gains Hinge on Where KV Cache Lives

    Marvell’s 32TB Memory Tier Bets AI Inference Gains Hinge on Where KV Cache Lives

    TL;DR · 30-second read

    The Short Version

    Chipmaker Marvell unveiled new parts aimed at a quiet problem in artificial intelligence: the expensive chips that run chatbots often sit idle, waiting for data.

    When an assistant works through a long conversation or a multi-step task, it keeps a running memory of everything it has already read. That memory outgrows the chip. Marvell’s answer is a shared memory pool, reached by light over cables up to 50 meters long, that holds about as much as 32 laptops.

    Marvell says this lets the same building produce two to three times more answers on the same electricity. That figure has not yet been tested by anyone else.

    Marvell Technology announced on August 4, 2026, at the FMS 2026 memory and storage conference in Santa Clara, California, a set of products aimed at AI inference, the stage where trained models answer requests. The lineup spans three levels of the data center: the Bravera SC6 PCIe 6.0 SSD controller for individual servers, Structera X CXL memory expansion for racks, and Photonic Fabric memory modules, network cards and chiplets that form an optical shared-memory tier across multiple racks.

    Marvell says Photonic Fabric reaches up to 50 meters, supports up to 32TB of warm KV cache offload and delivers up to 2-3x higher token throughput within existing footprints and power budgets. It says Bravera SC6 doubles the performance of its Bravera SC5 predecessor and is expected to begin sampling in Q4 2026. Marvell gave no availability dates for the Structera X or Photonic Fabric products.

    Executive Summary

    Marvell is presenting memory, not processors, as the next place where AI inference performance is won or lost. All three products in the announcement target the same object: the KV cache. This is the working memory a language model builds as it reads a prompt, and it grows with longer conversations and with more simultaneous users. The company argues that expanding, pooling and sharing memory separately from GPUs reduces the time those GPUs spend idle, which it calls stalls, and raises output per watt.

    This matters to anyone who pays for AI infrastructure by the megawatt. If the claims hold, operators could get more usable work from GPU fleets they have already bought and powered, without adding floor space or grid capacity. The strongest number, the 2-3x throughput gain, is framed as an “up to” figure with no stated baseline or workload. Only one of the products, the SC6 controller, has a date, and that date is for sampling rather than volume shipment.

    Three Tiers Aimed at One Problem

    The announcement is best read as a memory hierarchy rather than three separate products. At the top sits high-bandwidth memory (HBM), the fast but scarce memory stacked beside each GPU. Below it, Structera X uses CXL (Compute Express Link, an open standard for attaching extra memory to processors over the same physical lanes as PCIe) to add and pool DRAM at the rack level. Photonic Fabric then extends a shared memory tier across racks over optical links of up to 50 meters. At the bottom, the Bravera SC6 controller makes SSDs fast enough that, in Marvell’s words, more KV cache can “move from high-bandwidth memory to SSD.”

    Every tier has the same stated purpose: holding the KV cache somewhere cheaper and larger than HBM, while keeping it close enough that the GPU does not wait. Marvell’s framing, that the portfolio helps hyperscalers “scale memory more independently from compute,” is a statement about data center design. It describes memory as a resource provisioned at the rack and pod level, not a fixed amount soldered into each server.

    Why KV Cache Placement Drives Throughput

    The case for the headline rests on how inference works. When a model processes a prompt, it computes intermediate values for every token and stores them in the KV cache so it does not have to recompute them for each new word it generates. That cache grows with context length and with the number of concurrent sessions. Agentic AI, where a model runs multi-step tasks and carries long histories between steps, pushes both factors up at the same time. Once HBM fills, the operator has three options. It can evict the cache and recompute it later, which spends GPU cycles on repeated work. It can fetch the cache from storage, which is slow. Or it can fetch it from a faster intermediate tier.

    Marvell’s throughput claim sits on that third option. The company says Photonic Fabric enables KV cache “to be loaded from a shared memory tier rather than storage,” offering up to 32TB of warm cache (recently used but not currently active) across multiple XPUs (Marvell’s umbrella term for GPUs and custom accelerators). From that it claims up to 2-3x higher token throughput within the same footprint and power envelope. The GPUs are unchanged. What changes is how often they sit idle waiting for context to be rebuilt or retrieved. Operators most affected are those running long-context, many-user inference on power-capped sites, where adding GPUs is not an option but raising their utilization is.

    This supports a narrower claim than “memory has replaced GPUs as the bottleneck,” and the release itself does not go that far. It says memory is becoming “as critical as compute.” 650 Group analyst Alan Weckel, quoted in the release, calls memory capacity, bandwidth, latency and data movement “primary constraints.” The defensible reading is that for inference workloads with large KV caches, gains now come from memory placement as much as from raw compute.

    What the Numbers Do and Do Not Establish

    The concrete figures are the 32TB capacity, the 50-meter reach, the 2x performance step from SC5 to SC6, and the Q4 2026 sampling date for SC6. The 50-meter reach is significant for facility planners. It means the shared tier spans multiple racks, enough to cover a GPU pod, and it adds optical cabling inside the pod.

    The 2-3x figure is the least specified. It carries an “up to” qualifier and no stated model size, context length, batch size or comparison system. Throughput gains from cache offload depend heavily on how often the cache is reused, so the figure is best read as an upper bound for favorable workloads, not a fleet-wide expectation. Likewise, “doubling the performance” of SC5 fits the generational jump from PCIe 5.0 to 6.0 but does not specify which metric doubles. “Extremely low latency” for the optical tier is not quantified at all.

    Sourcing Flexibility and the Buyer’s Calculus

    One of the more practical details concerns procurement. Bravera SC6 is NAND-agnostic, meaning it works with flash memory from multiple suppliers. For cloud providers, that lowers dependence on a single flash vendor, a real consideration in a market prone to supply swings. Structera X is pitched in similar terms, as a way to extend “the value of existing memory investments” rather than forcing wholesale replacement.

    The trade-off is architectural complexity. Each tier adds devices, interconnects and software orchestration that must decide, in real time, where each session’s cache should live. The added power draw of CXL expanders, optical network cards and memory modules also counts against the power envelope the throughput gains are meant to protect. Buyers will need measured results on their own workloads before the net efficiency claim can be verified.

    Background

    Marvell Technology (NASDAQ: MRVL) is a Santa Clara-based semiconductor company focused on data infrastructure: chips that move, store, process and secure data for cloud, enterprise and carrier networks. It says it has served leading technology companies for over 30 years. Its Bravera line covers storage controllers, its Structera line covers CXL memory, and Photonic Fabric is its optical interconnect and memory technology.

    As AI shifts from training models to serving them at scale, the memory attached to each GPU has become a practical limit. Models that handle long documents, long conversations or multi-step agent tasks build large KV caches that can exceed on-package high-bandwidth memory. Chipmakers and cloud providers are responding with tiered designs that spill that cache into pooled DRAM, shared memory and fast flash storage. FMS, the annual memory and storage conference, is a common venue for these announcements.

    Sources

    Source: Marvell Advances AI Memory Infrastructure Portfolio to Accelerate Agentic AI Inference: Marvell Technology’s August 4, 2026 announcement of Bravera SC6, Structera X and Photonic Fabric memory products at FMS 2026.