By Sanjoy Maity, SVP of AMI Business Unit, Lattice Semiconductor
Roughly 1,200 sensors measure fan speed, coolant flow and optical link health inside a single AI server. A software console collects their readings and displays them to the operator, but it cannot send a response back at the speed the hardware requires. That response must come from the hardware itself, next to the sensor.¹ ²
Someone has to boot, secure and manage that hardware, and everything above it. That job now takes two layers working as one: deterministic logic at the sensor, and firmware that boots, secures and manages the platform. Lattice builds the first. AMI builds the second. Together, they form the foundation AI runs on, without closing it off. That is why the joining of these two companies adds up to more than an acquisition.
Why hardware and firmware stopped being separate purchases
That arrangement, with sensors reporting up to a console and commands coming back down, served operators for most of the past two decades. They bought hardware and firmware from different vendors. They updated each on its own schedule. They secured each with its own tools. The console tied the pieces together, and for the workloads of the time, the round trip was fast enough.
AI workloads run on a different clock. Power and thermal conditions inside an AI rack change faster than a software dashboard can respond. A voltage droop (VDROOP) under a training burst comes and goes before software checking on a schedule can read it. The response that comes back through the console now arrives too late. The management plane has to close the loop at the source, with integrated control that lives where power is delivered, converted and consumed.
What the handshake does
The two layers do different jobs. Lattice FPGAs work at the hardware layer, next to the sensors. An FPGA is a small programmable chip that provides real-time control at that spot. Lattice’s FPGAs sequence power, control boot, manage security functions and fuse sensor data. Two years ago, a rack held tens of them. Today, it holds hundreds.¹
Ford Tamer, president and CEO of Lattice, described the division of labor at the Six Five Summit in August.
“We provide these deterministic, real-time control, low latency, parallel processing capabilities near these sensors… AMI is also very low-level firmware, booting, managing, securing all of this hardware,” he said.¹
AMI firmware works at the layer above. It boots the platform. It verifies the code the platform loads. It manages the platform through the baseboard management controller, or BMC, and up to the fleet. AMI Data Center Manager 6.2, announced in August, extends that management to BIOS and firmware provisioning across hundreds or thousands of servers at a time.
The handshake is the point where the two layers meet. The FPGA reads a signal at the sensor and acts on it locally. The firmware turns that action into a system-level decision the operator can trust. Neither layer delivers the full result alone.
A complete stack is not a closed stack. The value is pre-validation. When the silicon and the firmware arrive already tested together, customers spend fewer engineering cycles on integration and more on their own IP and their own business.¹
Pre-validation does not come with a bundle requirement. AMI remains silicon-neutral. Customers can pair AMI firmware with the processors, silicon and platforms of their choice, exactly as before.³
Why this matters now
The cost of a failure has changed. When a $10,000 server went dark, the operator lost one server’s worth of work until it came back. When a $2.5 million NVIDIA system goes dark, the operator pays for idle capacity and loses revenue every minute it stays down.⁴
The number of systems that can fail has grown as well. Single clusters now pass 100,000 GPUs. Training racks that once drew a handful of kilowatts now pull 100 kW to 200 kW and more.⁴ Across a fleet that size, firmware that is inconsistent or unverified creates risk that is large and mostly invisible.²
Designing the layers together also simplifies updates. When the FPGA logic, the firmware and the management software come from separate vendors, each update needs its own validation and its own maintenance window. As the three layers are designed together, the goal is a single validated sequence that carries an update through all of them.
Why 1+1 = 3
A complete stack has to do three things: boot the platform, secure it and manage it. Lattice and AMI now do all three under one roof, with the hardware layer and the firmware layer built to work as one, and open to the silicon and platform customers already use.
We’ll continue this conversation at the Open Control Plane Summit we’re hosting on Monday, Oct. 12 in San Jose, “Day Zero” of the 2026 , where operators, OCP contributors and silicon partners will work through how the hardware, firmware and management layers fit together. AMI will bring its firmware as one reference point, not a required path.
To learn about the full program that both AMI and Lattice Semiconductor have prepared for OCP Global Summit, including a keynote address, technical presentations, live and virtual product demos showcasing our latest innovations, and much more, please visit our event page: https://go.ami.com/visit-lattice-and-ami-at-2026-ocp-global-summit.
Endnotes
- Ford Tamer and Sanjoy Maity, Lattice Semiconductor at The Six Five Summit: AI Unleashed 2026, Aug. 27, 2026 (transcript via Investing.com).
- Sanjoy Maity, “The 5-Headed Hydra of AI Data Center Control,” AMI, April 3, 2026.
- Sanjoy Maity, “AMI Joins Lattice Semiconductor: An Open Letter to Customers,” AMI, July 27, 2026.
- Earle Philhower, “The $7 Trillion Control Problem: Why AI Infrastructure Can’t Scale Without a Unified BMC,” AMI, Aug. 20, 2026.
