OCP Global Summit 2026: From Cold Plate Selection to Co-Design of AI Server Cooling Systems

Systemic Approach Replaces Individual Component Selection

October 2026, San Jose, California. The theme of OCP Global Summit 2026 is "Scaling Innovation for the AI Era." The exhibition and technical sessions featured Microsoft, Google, Meta, Rittal, Corintis, TE Connectivity, MiTAC Computing, ZutaCore, Wistron, Wiwynn, Indium Corporation, CoolIT Systems, Boston Materials, actnano, and other companies.

The key shift visible on the floor is the move from the question "which cold plate is better" to "does a system of hundreds to thousands of chips work at the specified heat flux." Google Cloud, at its booth with Mt Diablo and Ironwood racks, emphasizes the "Co-designed infrastructure" principle: chip, server, rack, power supply, and thermal management are designed jointly, not separately and then integrated.

Rittal demonstrates how a rack for high-density AI servers combines a rear-door heat exchanger (RDHx) and an in-rack CDU. The exhibition lists the In-Rack 250 model (rated cooling capacity of 250 kW at ΔT 4 K) and In-Rack 150 (150 kW at ΔT 6 K), as well as DC power, N+1 redundancy, filters, and replaceable sensors/control modules. The author emphasizes: the stated values are rated specifications, not evidence of actual heat rejection from a specific rack.

Single-Phase Liquid Cooling: Microchannel and Cellular Architectures

Microsoft presented the "Unitary Cell" approach: instead of one long microchannel section, the cold plate is divided into small cells where high heat transfer is utilized at the inlet zone. In a prototype adapted to the chip's thermal map, the transition temperature was reduced by up to ~10 °C. The author notes that channel subdivision increases contact area but simultaneously increases hydraulic resistance and clogging risk.

Corintis presented a comparison of a microfluidic cold plate with a standard Skived Cold Plate on an NVIDIA H200 GPU at the same coolant and a flow rate of 1 L/min: the transition temperature on the GPU decreased from 29.5 to 23.9 °C (~19 %), on HBM — from 42.0 to 32.4 °C (~23 %). The company also presented a silicon test stand with heat flux control over 2×2 mm elements and 2 kW power, which allows reproducing the non-uniform heating of a real chip.

TE Connectivity reported on a 64-channel cold plate for a 448G optical module at ~500 W heat dissipation: maximum temperature 74.9 °C. The hydraulic resistance of the plate itself is <2 psi, while the small quick-disconnect connector adds 6–10 psi, which, according to the presenter's assessment, can significantly affect overall system power consumption.

Two-Phase Cooling: Potential and Systemic Limitations

MiTAC Computing displayed a rack based on AMD Instinct MI355X with single-phase liquid cooling (CDU Nidec Max 200 kW In-rack) and in parallel — a two-phase solution in collaboration with ZutaCore. Externally the racks are similar, but the controlled state of the coolant is fundamentally different.

Wistron described the "Negative-pressure Water-boiling" technology: at an absolute pressure of ~1.64 psia, the boiling point of water drops to ~48.3 °C. At a heat flux of 2000 W, the heat transfer resistance over a 40 cm² area was ~7 % lower, and over 6.5 cm² — ~15 % lower compared to single-phase mode. The author points out: low pressure reduces the risk of coolant ejection during a leak, but does not eliminate the accumulation of non-condensable gases and performance degradation during prolonged operation.

Wiwynn detailed the systemic problems of a two-phase loop: at boiling, density and volumetric flow rate change sharply, causing flow redistribution; in parallel channels, a hotter chip generates more vapor, its hydraulic resistance increases, and flow in that channel decreases (an effect related to Ledinegg instability). At the outlet, vapor requires expansion of collectors, separators, buffer tanks, condensers, and low-NPSH pumps. The author's conclusion: reducing the cold plate does not automatically simplify the system.

Thermal Interface Materials and Contact Losses

Indium Corporation and CoolIT Systems emphasized that even with ideal heat transfer on the liquid side, a large temperature drop at the chip–plate contact limits the overall effect.

Boston Materials proposed the Fit-Freeze-Measure (FFM) method: a liquid-metal TIM preform under load fills the deformed chip package shape, then is fixed and measured. In the free state, package deformation is ~180 µm; under 29.4 psi load — ~99 µm. Taking into account the actual load and TIM/seal selection at 2600 W, 6 L/min, and 40 °C inlet temperature, the "hot spot" temperature decreased by ~6.1 °C (~8 %).

actnano proposed a gel protective coating for liquid-metal TIM that suppresses metal migration and retains the protective layer after cleaning. In some samples after 14 days at 100 °C, insulation resistance was >100 MΩ. The author emphasizes: confirming thermophysical properties and suitability for long-term operation in a server are different tasks.

Refrigerants and Coolants

The presentations discussed R134a, R1234yf, R515B, R1233zd(E) — with various saturation pressures, safety, and environmental characteristics. An example was given: at a hydraulic resistance of 100 kPa, the saturation temperature of a low refrigerant changes from ~50 to ~37 °C, of a medium one — from ~50 to ~46 °C. The author reminds: comparison of refrigerants is only valid under equivalent boundary conditions and accounting for system power.

For single-phase and two-phase immersion cooling, the importance of not only thermophysics but also electrical insulation, material compatibility, viscosity, flash point, oxidative stability, and property changes during prolonged operation was emphasized. In water-based loops with 25 % propylene glycol (PG25), the adsorption of corrosion inhibitors, formation of microparticles, and clogging of ultra-fine channels were discussed; topics included flushing, charging, condition monitoring, sampling, filtration, and nitrogen purging.

Manufacturing and Standardization

3D metal additive manufacturing opens the possibility of creating curved channels and three-dimensional distribution structures; however, the author points out: complex geometry does not guarantee high heat transfer and reproducibility. Surface roughness, residual powder, channel blockage, hermeticity, process repeatability, and cost must be considered.

OCP working groups are advancing the standardization of test conditions for cold plates, quick-disconnect connectors, CDU, and coolants: heat transfer resistance, hydraulic resistance, leaks, corrosion, erosion, durability. According to the author's assessment, without agreed-upon protocols, interchangeability and reliability of multi-component systems remain limited.

Author's Conclusion

Based on the visit, the author formulates the key question: not "how well is the chip cooled" but "how long, stably, and with minimal energy consumption does the system operate, and can the rejected heat be usefully utilized." Single-phase cooling continues to develop through precise flow distribution and co-design of microchannels; two-phase cooling promises operation at very high heat fluxes but requires solving vapor management, pressure, and flow rate issues. Competitiveness is determined not by the principle but by the quality of integration into a real server, rack, and data center.

For domestic developers, the author recommends moving from testing single samples to system test stands with simulation of local chip heat flux, flow distribution across branches, pump power consumption, corrosion, and material compatibility, as well as linking rejected heat with industrial consumers (heat pumps, low-temperature heat networks).

The material is based on presentations and observations at OCP Global Summit 2026. All numerical values refer to specific test or calculation conditions specified by the presenters; rated specifications of exhibits are not equated with actual operational results. Some interpretations and assessments belong to the author.