Lattice Semiconductor's FPGA Business Set to Double This Year as AI and General-Purpose Server Demand Both Surge
KeyBanc Technology Leadership Forum, August 11, 2026
Lattice Semiconductor's CFO Lorenzo Flores delivered a striking data point at KeyBanc's Technology Leadership Forum on Tuesday: the company's FPGA business, which generated roughly $500 million last year, is on track to roughly double in 2026 and continue growing into 2027. The disclosure came alongside a broader narrative about surging server demand, expanding content per rack, and early traction for the AMI Group acquisition that closed earlier this year.
Server Unit Estimates Keep Getting Revised Higher
Flores offered a candid look at how quickly the total addressable market has moved in 2026. At the start of the year, industry expectations called for roughly 15 million total server units with an AI mix near 20%, a figure that crept up to 16 million units as the year progressed. "Now, like the latest past 3 weeks, we've seen the estimate for overall servers about 20 million units and with a more -- higher AI mix," Flores said. That is a meaningful upward revision late in the year, and Flores noted Lattice was initially surprised by how strong its own demand signals were running relative to those industry numbers, prompting the company to dig into what was actually driving the growth.
The answer, according to Flores, is that it's not just AI servers. General-purpose, CPU-based server infrastructure is also being built out aggressively to support agentic AI workflows, and Lattice benefits from both trends. The company has historically had a strong, pervasive footprint in standard servers, while AI server racks offer substantially more attach points — Flores reiterated the figure of up to 100 or more attach points per AI rack given the added complexity.
Attach Rates and ASPs Both Moving Up Together
Perhaps the most important structural detail from the session was Flores's breakdown of the attach-rate trajectory. Four years ago, Lattice's attach rate per server was roughly 1x; it moved to 2x, then 3x, and is "now even higher." At the same time, average selling prices have been climbing as customers move up the product stack — from the MachXO2 and XO3 families to XO3D, XO5, and Nexus parts — pushing blended ASPs from roughly $3 to $4 and above. Flores described this as a "multiplier effect": unit volume growth from higher attach rates, layered on top of rising ASPs, all against a backdrop of industry unit growth. He said unit volume will remain the primary growth driver for the standalone FPGA business over a three-to-five-year horizon, with ASP expansion as a secondary tailwind.
On the current mix, Flores said AI-related revenue has grown into the "high 20%" of Lattice's overall business, though he was careful to note the company doesn't have full visibility into how every device is ultimately deployed by customers. "We know where our design wins are. We know where customers are shipping to... but it's not yet fully transparent to us. More importantly, we want to ship all the units to them and drive the growth," he said.
Attach Drivers Are Shifting From I/O to Management and Cooling
Flores traced the evolution of what functions are actually driving attach rate growth. Early attach was primarily bridging and I/O expansion. That was followed by security functions, including Root of Trust. More recently, Lattice has seen expansion into manageability applications such as power and cooling — a category that dovetails directly with the AMI acquisition's firmware capabilities. Flores specifically called out leak detection as an example of an application category that "18 months ago nobody thought about," and is now considered a critical rack-level function. Lattice believes it is well positioned to capture that opportunity given its existing footprint.
The AMI Deal: "1 Plus 1 Equals 3"
Flores gave the clearest articulation yet of the strategic logic behind the AMI Group acquisition. AMI's boot and management firmware, layered on top of Lattice's "first-on/last-off" position in server hardware, creates what he called a "1 plus 1 equals 3 phenomenon." He said Lattice has already engaged with 100 different customers — hyperscalers, OEMs, ODMs and ecosystem partners — since closing the deal, with feedback centered on uptime, security, power and cooling optimization, and the ability to manage heterogeneous supply chains that mix old and new infrastructure from different ODMs. Flores was explicit that the combined solution captures more value than the sum of its parts: "The solution may look like the same, but I think the overall value will be much greater than the individual components." He characterized the opportunity as partly greenfield — the underlying problems hyperscalers face are the same ones FPGAs and firmware have long addressed, but integrating the two into a unified offering unlocks new functionality and pricing power that neither business could capture alone.
Why Lattice Believes It's Insulated From ASIC Displacement
Asked directly about the classic FPGA-to-ASIC displacement risk — the pattern seen in networking base stations, where Xilinx-style FPGA content eventually got designed out — Flores argued Lattice's position in AI racks is structurally different. The company has deliberately avoided competing for core processing capability, instead positioning its FPGAs around I/O expansion, sensor aggregation, and management functions where dedicating advanced-node silicon doesn't make economic sense. "You don't want to spend a 2-nanometer silicon, your silicon die budget, on I/O expansion. It makes way more sense to do it in a 65-nanometer FPGA," he said. He also pointed to compressed rack design cycles — down from two years to 18 months to 12 months — as an inherent advantage for programmable flexibility, along with the power efficiency of Lattice's parts relative to the aggregate power budget of hundreds of devices per rack.
Physical AI Is a Real but Not-Yet-Material Opportunity
Flores described physical AI and robotics as an extension of the same "companionship" strategy Lattice has deployed in data centers — aggregating sensor data, preprocessing it, and driving physical actuation. The company's sensor fusion and vision capabilities, including its role in Nvidia's Holoscan platform, give it a foothold across "almost all" robotics companies, according to Flores, though he cautioned that most robotics customers are less mature in FPGA design than traditional industrial automation players, requiring Lattice to build more software-like programming environments. On timing, Flores was measured: "We're not sure. We'll probably start seeing it at the end of next year." Underlying industrial demand, meanwhile, is already recovering, with Lattice — now largely de-exposed to auto — posting roughly 20% year-on-year industrial growth and expecting similar growth again next year.
Lead Times Normalizing After Spiking Above a Year
On the supply side, Flores acknowledged that lead times ballooned to over a year earlier in 2026 as demand accelerated, but have since been brought back down into the 40-week range. The current bottleneck sits in assembly, test and packaging rather than wafer fabrication, since Lattice uses older process nodes with more comfortable fab capacity. The company has been investing in tester capacity, reflected in its capital spending, and is actively expanding its supply chain to support a business that is roughly doubling in size this year. Flores emphasized that Lattice's order book has grown both in dollar value and in duration, with bookings now extending through the fourth quarter of 2027 — a level of forward visibility he attributed to tighter channel inventory management rather than double-booking risk, which the company continues to monitor closely. Importantly, he clarified that being booked through late 2027 does not mean capacity is capped: "We are not booked out in the sense that we can't support more demand. We can support more demand."