Everyone Knows Nvidia. Nobody's Talking About the Company Wiring It All Together.
Marvell has quietly become the connective tissue of the AI data center: custom silicon for every hyperscaler, optical interconnects between every GPU cluster, and a $75 billion lifetime revenue pipeline.
Marvell Technology has quietly transformed from a storage chip company into the connective tissue of the entire AI data center. Custom silicon for every major hyperscaler. Optical interconnects linking every GPU cluster. And a $75 billion lifetime revenue pipeline that most investors haven't discovered yet. When people think about the AI infrastructure buildout, they think about Nvidia. GPUs, data centers, Jensen Huang on stage in a leather jacket. Nvidia is the face of the trade and deserves to be — $96.2 billion in quarterly revenue and $89 billion from data centers alone speaks for itself. But GPUs don't work in isolation. They need to talk to each other. They need custom silicon tailored to each hyperscaler's specific architecture. They need optical interconnects that move data at the speed of light between racks, between buildings, between continents. They need memory controllers, PCIe switches, CXL expansion, and networking infrastructure that makes the GPU cluster function as a single coherent system. That's Marvell. And the market is only beginning to understand what that position is worth. The Transformation Nobody Noticed A decade ago, Marvell was a storage and networking chip company. Data center revenue was less than 10% of the business. The company made hard drive controllers, enterprise switches, and broadband processors. Solid business. Boring stock. Today, data center revenue is on track to exceed 80% of total company revenue. AI data center revenue alone now tops $2 billion per quarter. The company has design wins with every major U.S. hyperscaler — Alphabet, Amazon, Microsoft — and a $75 billion lifetime revenue pipeline from custom silicon programs that are either in production or ramping. CEO Matt Murphy described the current moment at this week's Six Five Summit as an "era of abundance" in AI infrastructure — one where the market is not zero-sum and multiple suppliers can grow simultaneously. The data supports him. Marvell's stock has delivered a 224% return over the past year and 203% in just six months, yet the company is still guided to grow data center revenue above 25% sequentially through fiscal year 2027, with a path toward $10 billion in organic revenue and management guiding for $15 billion by FY2028. This isn't a company riding a one-quarter tailwind. It's a business that has been methodically repositioned over five years to sit at the center of the AI infrastructure stack — and is now harvesting the results. Three Businesses Inside One Company Marvell's AI infrastructure story rests on three product pillars that are each independently significant. Custom Silicon (XPU/ASIC) Every hyperscaler wants custom chips optimized for its specific AI workloads. Google has TPUs. Amazon has Trainium and Inferentia. Microsoft and Meta have their own custom accelerators. These chips are not off-the-shelf GPUs — they're purpose-built processors designed in collaboration with silicon partners. Marvell is one of only two companies (alongside Broadcom) with the capability to design and deliver these custom ASICs at scale. The company has announced 18 XPU and XPU-attach socket design wins, many already in volume production. These programs span full custom XPU designs, XPU-attached silicon, electrical I/O chiplets integrated inside multi-die packages, and related products. The custom silicon business currently contributes roughly one-quarter of data center revenue and is expected to double in the coming years. The $75 billion lifetime revenue pipeline from these design wins represents contracted or near-contracted revenue that will flow over multiple years as programs ramp from sampling to volume production. Morgan Stanley recently flagged Google's Frozen v2 custom AI chip as a potential Marvell opportunity — a signal that design win momentum is continuing beyond existing programs. Optical Interconnects As AI clusters scale from hundreds to tens of thousands of GPUs, the bottleneck shifts from compute to connectivity. Moving data between GPUs, between racks, and between data center buildings requires optical interconnects running at 800G and emerging 1.6T speeds. Electrical copper connections can't carry data fast enough or far enough at the bandwidth AI workloads demand. Marvell is a leading supplier of optical DSPs (digital signal processors) and PAM4 SerDes technology that powers these connections. Its 800G solutions are in volume production, and 1.6T products are ramping. When Jensen Huang talks about NVLink rack-scale computing, the optical interconnects linking those racks together are overwhelmingly supplied by companies like Marvell. The recent acquisition of Celestial AI adds photonic fabric technology — the ability to route data optically within the rack, not just between racks. This extends Marvell's addressable market from inter-rack connectivity into intra-system connectivity, where the bandwidth requirements are even higher and the growth trajectory is steeper. The acquisition of XCON Technologies further strengthened the photonics portfolio. At FMS 2026 this month, Marvell unveiled its Photonic Fabric shared-memory architecture — a platform that connects memory pools to compute clusters using photonic interconnects rather than electrical ones. If adopted at scale, this technology could fundamentally change how AI data centers are designed. Storage and Memory Infrastructure AI data centers don't just need compute and connectivity. They need storage systems capable of feeding data to GPU clusters at the speed those clusters can consume it. Traditional storage architectures create bottlenecks that leave expensive GPUs idle waiting for data. Marvell's Bravera SC6, unveiled at FMS 2026, is the industry's first PCIe 6.0 SSD controller — doubling the performance of the prior generation and sampling in Q4 2026. Structera X is a CXL-based memory expansion platform that allows data centers to pool and share memory resources across multiple servers, dramatically improving utilization and reducing the amount of memory each server needs to provision individually. CXL (Compute Express Link) memory expansion is a multi-billion dollar emerging market that barely existed two years ago. As AI workloads grow beyond what a single server's local memory can handle, CXL enables memory to be disaggregated and shared. Marvell is among the first to productize this technology at scale. The Nvidia Relationship Is Symbiotic, Not Competitive One of the most common misunderstandings about Marvell is that it competes with Nvidia. It doesn't. The two companies occupy complementary positions in the AI stack. Nvidia makes the GPUs — the processors that perform the AI computation. Marvell makes the silicon that connects those GPUs together, the optical modules that link GPU clusters across racks, the storage controllers that feed data into the compute pipeline, and the custom ASICs that hyperscalers use alongside Nvidia GPUs for specific workloads. When Nvidia sells a $3 million DGX server or a multi-million dollar NVLink rack, every one of those systems contains optical interconnects, PCIe switches, storage controllers, and networking silicon. Much of that silicon comes from Marvell. As Nvidia's data center revenue scales from $89 billion per quarter toward $108 billion next quarter, Marvell's addressable content per server scales with it. Jensen Huang on tonight's earnings call talked about "NVLink rack-scale computing" as the future of AI infrastructure. That vision requires exactly the optical and connectivity solutions Marvell provides. Nvidia scaling means Marvell scaling. The relationship is multiplicative, not competitive. The Numbers Marvell reports Q2 FY2027 earnings on August 27 — tomorrow. Here's what the market expects: Revenue: $2.70 billion (up 34% from $2.01 billion a year ago) EPS: $0.87 (up 30% from $0.67 a year ago) Data center revenue: expected to be roughly 76% of total, implying ~$2.05 billion For context, data center revenue has grown sequentially every quarter for over two years. Management guided FY2027 organic revenue toward $10 billion, with data center growing above 25% sequentially throughout the year. The $15 billion FY2028 target implies continued acceleration as custom silicon programs ramp to volume and new optical product cycles begin. RBC sees Marvell sustaining 40%+ revenue growth for three years. KeyBanc has a $400 price target. The stock is trading around $222 after rising 37% from its July lows — still well below the $400 target and with an earnings catalyst tomorrow. Why Marvell Matters for the AI Trade Here's the Stockgecko thesis on why Marvell deserves attention alongside the memory names we've been covering all month: The AI data center has four critical layers: compute (Nvidia), memory (SK Hynix, Samsung, Micron, SanDisk), connectivity (Marvell, Broadcom), and power (Bloom Energy, Constellation). The market has aggressively repriced compute and is in the process of repricing memory. Connectivity is next. As AI clusters scale from thousands to hundreds of thousands of GPUs, the connectivity layer becomes the binding constraint. You can add more GPUs, but they're worthless if you can't move data between them fast enough. The optical interconnect market is projected to grow from $10 billion to $40+ billion by 2028 as 800G gives way to 1.6T and photonic fabric penetrates inside the rack. Marvell sits at the intersection of three growth vectors — custom silicon, optical interconnects, and memory infrastructure — each of which is independently scaling at 25-60%+ annually. The combined effect is a company that can grow total revenue at 40%+ for multiple years while expanding margins as higher-value AI products become a larger share of the mix. The comparison to Broadcom is inevitable. Broadcom is the larger, more diversified incumbent with custom ASIC deals at Google (TPU) and Meta at tens of billions in revenue. Broadcom's margins are higher (77.5% gross vs Marvell's 59%). But Marvell is the faster-growing pure play with more percentage upside from a smaller base. On a percentage basis, Marvell tripled in 2026 while Broadcom climbed more steadily. Both are excellent AI infrastructure positions. Marvell is the higher-beta, higher-growth version of the same thesis. The On-Chain Angle Marvell trades as an equity perpetual on Hyperliquid. The stock moved from $218.72 at Friday's close to $222.60 over the weekend on the Hyperliquid perp — price discovery happening in real time while Nasdaq was dark. With earnings tomorrow (August 27), on-chain traders can position through the report and react in real time to the print regardless of what time zone they're in. Given the stock's 37% recovery from July lows and KeyBanc's $400 target, the setup into earnings is loaded with asymmetric potential in both directions. The AI data center buildout runs on four layers: compute, memory, connectivity, and power. Stockgecko tracks all of them. Marvell is the connectivity layer — the part of the stack that wires everything together and is just beginning to reprice. Track MRVL and every HIP-3 asset on Stockgecko. Prices and data referenced as of August 27, 2026. This is not financial advice.