Hayagreeva Inc builds all-copper, direct-to-chip cold plates for AI compute. A validated 5,000W+ thermal envelope in a single 80 by 80 mm plate, engineered as a drop-in for the single-phase liquid loops data centers already run, at a materially lower cost than incumbent options.
The accelerators powering modern AI dissipate more than a kilowatt each, and rack densities are climbing past 100 kW. Fans and finned heat sinks, the architecture that served computing for forty years, cannot move that much energy out of a few square centimeters of die.
The industry settled on single-phase liquid cooling as the bridge: water-based coolant, familiar facility loops, no exotic fluids or phase-change complexity. But inside that loop, one component decides how cool the silicon actually runs. The cold plate sitting on the die owns the largest remaining thermal resistance in the chain, and every degree it gives back is a degree of headroom for clocks, for coolant temperature, and for rack density.
Heat flux per unit die area rises with each silicon generation while the footprint available for a cold plate stays fixed. Performance has to come from within the same volume.
A single AI site can need tens of thousands of plates. A component priced for prototypes becomes a line item that shapes the entire build budget.
Cooling capacity now gates deployment schedules. Processes that only a handful of vendors can run worldwide are a structural risk to any build-out plan.
A fully brazed, all-copper assembly built around our engineered Copper Core. It mounts where a conventional cold plate mounts, plumbs into the same single-phase loop, and returns a materially lower thermal resistance for the same footprint.
Performance figures are from our validation program. Detailed test data, pressure-flow curves, and the evaluation protocol are available to qualified partners under NDA.
Total junction-to-coolant resistance is a sum, and most cold plates optimize one term at a time. Our architecture is designed against the whole budget: conduction into the plate, convection at the wetted interface, and the caloric rise as coolant carries the load away.
A continuous copper path from the mounting face into the transfer structure. No dissimilar interfaces, no interstitial layers to add resistance where the flux density is highest.
The Copper Core presents the equivalent of 110 channels per inch in traditional microchannel geometry, arrayed in all three dimensions rather than as parallel grooves on one plane. Area density does the work that pumping power would otherwise have to.
Flow distribution across the full active area keeps the coolant temperature rise uniform, which protects the die from the hot-corner gradients that flat channel arrays produce.
| Specification | Value |
|---|---|
| Validated thermal envelope | 5,000W+ on a single plate |
| Measured thermal resistance | ~0.009 to 0.012 °C/W |
| Active plate footprint | 80 by 80 mm |
| Transfer area density | ~6,600 m²/m³, close to 21x footprint |
| Construction | All-copper, fully brazed, no gaskets in the flow path |
| Loop compatibility | Standard single-phase, water-based facility loops |
| Intellectual property | 14 patents filed |
High-performance cold plates are typically precision-machined devices, and their prices reflect it. That economics works for prototypes and strains badly at AI scale, where cooling is bought by the rack and the megawatt.
We took the other route. Our plates are built from copper and joined by brazing: two of the most mature and widely available processes in manufacturing. No exotic materials, no single-source tooling, no process that only a handful of vendors on earth can run. That is what lets the same plate deliver top-of-class thermals at a materially lower cost, and it is why capacity can follow demand instead of gating it.
Each generation tightened the same thesis: raise transfer area density inside a fixed footprint, keep the heat path in a single material, and remove every part that could leak, creep, or add resistance. The validated 5,000W+ plate is the commercial baseline; the program continues toward the mounting standards and lower stack heights our OEM conversations call for.
Flow-visualization and welded all-copper builds established that the Copper Core approach held its thermal advantage in a manufacturable assembly.
An 80 by 80 mm all-copper plate carrying the validated 5,000W+ envelope at ~0.009 to 0.012 °C/W measured, with no gaskets in the flow path.
Samples to partner mounting and manifold specifications, lower stack heights, and joint validation programs with OEM and laboratory partners.
We are engaging server OEMs, cooling-system makers, and data-center operators. Evaluation units are built to your specification, and our test protocol is designed to sit alongside your incumbent on the same bench.
Samples to your mounting, keep-out, and manifold specifications, with pressure-flow and resistance-versus-flow data at stated heater sizes for side-by-side benching.
Plate-and-CDU pairing evaluations on single-phase loops, including joint validation programs and co-engineering on manifold integration.
Evaluation units and deployment planning for high-density AI halls, including retrofit paths onto existing facility water.
No. The plate is designed for the single-phase, water-based loops already deployed in liquid-cooled data centers, and integrates with standard CDUs and manifolds. The thermal headroom it returns can be spent on higher coolant temperatures if you prefer that to lower junction temperatures.
On instrumented thermal test vehicles at stated heater sizes, reporting resistance against flow rate rather than a single headline number. We share the full protocol and the underlying data with evaluation partners under NDA, and we distinguish clearly between measured results and simulation-predicted figures in every document we issue.
You send your mounting specification and target thermal load. We build samples to that specification, ship with a documented test protocol, and support a side-by-side comparison against your current plate on your own bench. Joint validation programs with shared reporting are available where the volumes justify them.
The architecture was chosen for that reason. Because the plates use commodity materials and mature joining processes rather than bespoke machining, capacity expansion is a matter of adding conventional lines. We scope volume commitments during evaluation.
14 patents have been filed covering the architecture and its manufacture. Hayagreeva Inc holds an independent, unconsolidated intellectual property position in this area.
Tell us the accelerator, the loop, and the constraint you are designing against. We will come back with what an evaluation would look like.