A small office on Zanker Road. A team of Marvell veterans. An open chiplet platform that wants to be the unglamorous plumbing under the world's biggest GPU clusters.
Walk into the bottom of any modern AI training cluster and you find roughly the same fight: too many GPUs, not enough usable bandwidth between them. The model wants to talk; the network wants a nap. DreamBig Semiconductor's bet is that this is the real bottleneck of the decade, and that the fix is not bigger boxes, it is better, smaller, hot-pluggable pieces of silicon.
The company occupies the kind of building in San Jose where the parking lot tells you nothing. Inside, about 220 people - many of them refugees from the Marvell era - are building open chiplets, an 800 Gbps SuperNIC called Mercury, and a chiplet hub called MARS that stacks 3D HBM with about 20 nanoseconds of latency. Sober names. Loud numbers.
Every credible AI roadmap on Earth assumes more silicon in the same box and more boxes in the same rack. Both assumptions strain at the same seam: how do you get data from chip A to chip Z, in nanoseconds, without melting the floor? The industry's answer for thirty years has been "build one giant monolithic die." It worked. Until reticle limits, yields, and power budgets all started laughing.
So the industry turned to chiplets - small, specialized dies wired together inside a single package - and to UCIe, an open standard for talking between them. Lovely in slides. Brutally hard in practice. Most companies will not build their own packaging stack, their own die-to-die PHY, their own RDMA engine, and their own SuperNIC. That is, conveniently, what DreamBig has spent six years doing.
It is tempting to call this unsexy work. It is. That is also the point. Foundational infrastructure is rarely glamorous and almost always lucrative when it lands.
DreamBig was founded in 2019 by Sohail Syed - now CEO, previously founder of Questarium and FIRQuest, both acquired - alongside Sehat Sutardja and Weili Dai. The latter two carry the kind of resume that opens doors quietly: they co-founded Marvell Technology in 1995 and ran it for two decades. Sutardja, an IEEE Fellow, holds more than 440 issued patents. Dai was named on Forbes' "World's 100 Most Powerful Women" list in 2012. Sutardja is Executive Chairman. Dai is Executive Vice Chairwoman.
The supporting cast is unflashy in the way you want it to be. SVP Engineering Steve Majors previously ran NetEffect, the RDMA startup Intel bought. VP Hardware Dave Maguire has, by company account, a perfect tapeout record - the kind of detail engineers list and marketers skip. Chief Product Architect Brian Hausauer holds twenty-plus patents in RDMA. There is no celebrity-CEO story arc here, which is, on balance, reassuring.
DreamBig is not pitching a single chip. It is pitching the connective tissue around AI accelerators - a SuperNIC, a chiplet hub, and the services to glue them into customer silicon. Each piece can be sold on its own. The interesting orders are the ones that take all three.
800 Gbps discrete SuperNIC with fully HW-offloaded RoCE v2 and a UEC RDMA engine. As a chiplet, it scales to 12.8 Tbps inside a single AI SuperChip.
An open chiplet platform with 3D HBM stacking (~20 ns latency) and UCIe / BoW die-to-die interfaces. Designed to be the spine other people's accelerators clip onto.
Tailored chiplet-based system design, RDMA networking integration, and packaging work for customers building DPUs, accelerators, and automotive SoCs.
Skepticism is the right default with any pre-revenue silicon company. DreamBig has enough public receipts to earn a second look. The Series B - $75M, July 2024 - was co-led by Samsung's Catalyst Fund and the Sutardja Family, with Hanwha's Next Generation Opportunity Fund, Event Horizon, Raptor, UMC Capital, and BRV alongside. Total raised to date sits around $92.5M.
The named partners are the other half of the receipts. Synopsys collaborated on the 3D HBM integration. Athos Silicon selected DreamBig's Chiplet Hub as the basis for an autonomous-vehicle platform. None of this is the same as a Fortune 50 production order. All of it is the kind of move that comes a year or two before one.
"Open chiplets, fully offloaded RDMA, 3D HBM. Sober shopping list. Loud implications."
DreamBig's stated mission is "comprehensive low-cost, low-latency, high-throughput chiplet and networking solutions for AI, data centers, edge compute, and automotive markets." Underneath the corporate language is a more interesting idea: that custom silicon should not be a privilege of the three hyperscalers with their own ASIC teams. Pick chiplets off a shelf, plug them into a hub, ship a product. UCIe and BoW are how the chiplets talk. MARS is the shelf.
It is, in other words, a marketplace argument dressed as a chip company. If DreamBig is right, the next decade of AI hardware looks less like one company designing the entire stack and more like a Lego bin in which a Tier-1 automotive vendor, a robotics startup, and an enterprise inference cloud all reach for parts of the same chiplet ecosystem.
The casual observer of AI in 2026 thinks about chatbots, image generators, the occasional autonomous vehicle. The serious observer thinks about cabling - literal copper, optical interconnects, the slow tyranny of latency between racks. Every leap forward in model capability is being paid for in network engineering nobody wants to read about. Whoever lowers the cost of that fabric, by an order of magnitude, owns a real piece of what comes next.
DreamBig might not be that company. The list of plausible competitors is long. But of the small group of independent startups attempting to ship a full open-chiplet networking stack, with a SuperNIC and 3D HBM integration and tapeouts to point at, DreamBig is one of the few you can name from public information. That alone is unusual.
Back to the parking lot on Zanker Road. From the outside, nothing has changed - same low building, same San Jose afternoon, same anonymous campus among many. Inside, the chiplets keep going through tapeout. The Mercury samples ship to a small list of customers. The MARS hub starts to look less like a slide and more like something with screws in it. Whether DreamBig ends as a category-defining company, an acquisition target, or a footnote in the chiplet era depends on the next two product cycles. None of those outcomes are boring. All of them start with the same thing: getting the network right.