What we work on
The four layers, and the seams between them
Where the practice operates, and why the gaps between layers are where the commercial problems collect.
Every layer of the AI infrastructure stack has competent people selling within it. Almost nobody owns the seams, which need synchronization and agreement between the parties. Each of those parties can describe their own layer precisely, and with little accountability for the handoff. That is where programs stall, and it is the real work.
Thermal
Direct-to-chip cold plates, fluid-to-die and in-package cooling, coolant distribution units, rack integration and thermal interface materials. Single-phase and two-phase, and the honest trade between them. The key question is when the thermal design must be finalized before the chip design is complete.
Silicon
Merchant accelerators and their rack-scale reference architectures, custom ASIC programs, chiplet and multi-die packages, high-bandwidth memory and co-packaged optics. Access at this layer is designed into each program, not purchased through a traditional sales process.
Platforms
Tier-one leading server manufacturers, specialized platform builders, rack integrators, systems integrators and distributors. Knowing where each decision is made, and who makes it, can save a year.
Power and site
Generation and interconnect, behind-the-meter power, electrical distribution, controls and automation, and the developers, colocation operators and the data centers who have to buy all of it. The engineering firms often design these campuses before the operator frames its requirements.