Lattice Semiconductor: AI Servers Are Only Half the Story as Attach Rates and AMI Integration Reshape the Growth Algorithm
KeyBanc Technology Leadership Forum, August 11, 2026
Lattice Semiconductor CFO Lorenzo Flores used his appearance at the KeyBanc Technology Leadership Forum 2026 to unpack a growth story that is running well ahead of what a simple AI server narrative would suggest, while also detailing how the AMI acquisition is expected to layer a materially higher-value software business on top of the company's FPGA silicon franchise. The comments, delivered in a fireside chat with KeyBanc analyst John Vinh, offer investors the most granular breakdown yet of what is actually driving the comms and compute segment's more than 80% growth this quarter, and why management believes the current cycle is structurally different from prior FPGA build-outs.
General Purpose Servers Are Quietly Doing as Much Work as AI
The headline surprise from the session is that Lattice's blowout data center growth is not simply an AI compute story. Flores explained that industry-wide server unit forecasts have been revised up sharply through the year, from roughly 15 million units with a 20% AI mix at the start of 2026 to around 20 million units with a higher AI mix now. But he was explicit that the acceleration in Lattice's own bookings has been driven by both AI servers and what he called "more standard foundational, CPU-based server infrastructure that is enabling this agentic workflow being broadly deployed." In other words, the buildout of general-purpose compute needed to support agentic AI workloads is generating just as much incremental FPGA demand as the AI accelerator racks themselves, a dynamic that broadens Lattice's addressable growth well beyond the narrower AI data center TAM that the sell-side has typically modeled.
Flores was careful to note that AI-related revenue has grown to "the high 20% of our overall business," but stressed that Lattice does not have full visibility into end-use because it ships components rather than finished systems. That lack of granularity, combined with the strength in the base business, is arguably an underappreciated positive: Lattice is capturing upside from two independent secular trends simultaneously rather than being purely levered to AI capex cyclicality.
Attach Rates Have More Than Tripled, and the Next Leg Is Manageability
The most quantifiable new data point from the session is the trajectory of Lattice's attach rate per server, which Flores said has moved from roughly 1x four to five years ago to "2.x, 3.x and it's now even higher." He walked through the functional progression driving this: early attach was almost entirely I/O expansion and bridging, followed by a wave of security and Root of Trust deployments, and now a newer wave of power, cooling and broader manageability applications. Layered on top of unit growth is an ASP mix shift as customers move up Lattice's product stack from MachXO2 and XO3 parts toward XO5 and Nexus-class devices, which Flores said is lifting blended ASPs from roughly $3 to $4 per unit. The combination of rising unit volumes, rising attach points per rack, and rising ASP per attach point creates what Flores termed a "multiplier effect," and he was direct that over a three-to-five-year horizon, unit volume growth from expanding attach points, not ASP expansion, will remain the primary driver of the FPGA business specifically. That is an important distinction for investors modeling the business, since it implies the growth algorithm is durable and volume-driven rather than dependent on continued mix-shift into higher-priced silicon.
AMI Deal Thesis: "1 Plus 1 Equals 3," With Solutions Revenue as the Real Prize
Flores devoted significant time to explaining why the AMI acquisition is strategically additive rather than a simple bolt-on. He described Lattice's FPGAs as historically "first-on/last-off" in a server's boot sequence, a position that dovetails directly with AMI's boot and system management firmware. The combined entity, he said, can now offer hyperscalers integrated solutions spanning uptime management, security, and power and cooling optimization, including leak detection capabilities that he said were essentially nonexistent as a customer priority 18 months ago and are now "a really critical application." Asked directly whether the combined opportunity is additive or genuinely new, Flores was unambiguous: "The problem statement for the hyperscalers is the same. How they would solve it by using, say, an FPGA and some of AMI's firmware, could be the same. But integrating it and adding more features and functionality, that's greenfield." He added that the packaged solutions business should generate "a much stronger revenue growth per unit" than either the standalone FPGA or firmware businesses, positioning the AMI integration as a margin and ASP lever for the combined company beyond what either asset could achieve independently. Lattice has already engaged with over 100 hyperscaler, OEM, ODM and ecosystem partners on the combined offering, according to Flores, with heterogeneous supply chain management, mixing old and new infrastructure across multiple ODMs, cited as a particularly strong pull for AMI's firmware layer.
Management Argues FPGA-to-ASIC Displacement Risk Is Structurally Lower Here
Vinh pressed Flores on the historical pattern in which FPGAs get designed out for ASICs once volumes reach sufficient scale, as happened in wireless base stations during the Xilinx era. Flores pushed back on the read-across, arguing that Lattice's "companionship strategy," deliberately avoiding head-to-head competition with core processing silicon, insulates it from that dynamic. His reasoning centered on three points: it makes little economic sense to burn leading-edge 2-nanometer die budget on I/O expansion functions that run fine on 65-nanometer FPGAs, rack design cycles have compressed from two years to as little as 12 months, which favors programmable flexibility over fixed-function ASICs, and Lattice's power consumption profile is measured in milliwatts rather than watts, which matters when hundreds of devices populate a single AI rack's power budget. Whether this holds up over multiple product generations remains to be tested, but it is a clear articulation of the bull case for content durability that investors should weigh against the historical precedent Vinh raised.
Physical AI Is a 2027 Story, Not a Near-Term Catalyst
On robotics and physical AI, Flores confirmed Lattice sees a natural extension of its sensor-aggregation and preprocessing strength, including its presence on Nvidia's Holoscan platform for multi-camera vision streams, but set realistic expectations on timing: "We're not sure. We'll probably start seeing it at the end of next year." That pushes any meaningful physical AI revenue contribution into late 2027, and Flores acknowledged that most robotics customers are less sophisticated FPGA users than traditional industrial automation companies, requiring Lattice to build more software-like programming tools to drive adoption. Separately, the core industrial business is recovering, with Flores citing roughly 20% year-on-year growth versus last year's comparable periods and improving global PMI data, while noting Lattice's reduced automotive exposure has been a relative advantage.
Supply Chain Is the Binding Constraint, Not Demand
On the operational side, Flores disclosed that lead times spiked to over a year earlier in 2026 as demand accelerated but have since been brought back down to the 40-week range, with assembly, test and packaging now the primary bottleneck rather than wafer fabrication, since Lattice relies on older, less capacity-constrained nodes. He confirmed the FPGA business is on track to roughly double this year from approximately $500 million in 2025, a scale-up that he said requires the supply chain to become "2x plus more robust." Perhaps most notably for investors modeling forward revenue, Flores confirmed comments made previously by another Lattice executive that the company has bookings visibility extending through the fourth quarter of 2027, while clarifying that this reflects strong order visibility rather than a capacity ceiling: "We are not booked out in the sense that we can't support more demand. We can support more demand." That distinction matters, since it suggests upside optionality remains if end demand continues to surprise positively, rather than a supply-constrained ceiling on 2027 growth.