Cooling Portfolio · AI Accelerator Thermal Management · Pieta IP Engine

Adaptive Nunome
one part. zero compromises.

One monolithic drop-in part that replaces the cold plate, TIM layer, and lid — passive transient buffering to 500 W/cm², matched to silicon, electronics-free flow telemetry on the OCP interface. Flagship of the Pietarien cooling portfolio. The materials and geometry that make it work are disclosed under NDA only.

Domain AI Accelerator Liquid Cooling (>10 kW/unit)
Maturity Prototype-ready
Patent status 25 packages drafted, August 2026 priority date
Next step Hyperscaler pilot qualification
82
◆ Diamond
IPS Gate
500
W/cm² transient flux
Buffered passively — how is under NDA
3
Assembly steps eliminated
Cold plate + TIM + lid → one part
~7
ppm/K body CTE
Silicon-matched — composition withheld
5 yr
Anti-fouling lifetime
Coating system disclosed under NDA
0
Embedded electronics
Sensing is passive and integral
01

The AI Cooling Bottleneck

Next-generation accelerators exceed 1,000 W per package with local heat fluxes of 500 W/cm² at tensor-core and HBM hotspots. Conventional cold plates fail in three compounding ways.

Thermal throttling

Steady-state cold plates have no buffer for transient load steps. Every workload spike hits the thermal ceiling and downclocks the chip — lost inference throughput billed at cloud rates.

CTE fatigue failures

Copper cold plates run ~17 ppm/K against silicon's ~3 ppm/K. Cyclic stress delaminates joints, pumps out TIM, and generates RMA events at fleet scale.

Assembly complexity

Separate cold plate, TIM application, and lid bonding are three precision steps — three yield-loss points. Lifecycle replacement cost exceeds hardware cost over a 5-year loop.

02

The Architecture

One monolithic part integrates four functional layers that previously required separate components, qualification cycles, and supply chains.

Adaptive channel architecture

A channel topology that re-routes coolant to wherever the heat map demands it — not uniformly across the die. Geometry disclosed under NDA.

Passive transient buffering

Workload spikes are absorbed as latent heat before junction temperature moves — no pumps, no valves, no active control. Buffer design disclosed under NDA.

Electronics-free flow telemetry

Flow telemetry encoded in the part itself — no sensors, no firmware, no calibration expiry — readable on the OCP interface. Method disclosed under NDA.

Fatigue-proof body + anti-fouling surface

A silicon-matched composite body that kills thermomechanical fatigue, plus a surface system that prevents scaling and biofouling across a 5-year loop life. Materials disclosed under NDA.

03

The Run

2026-08-21 · Full autonomous pipeline — 2 × 8-hour runs over 2 days. No manual intervention between input and 25 attorney-ready patent packages.

72
Forge ideas in
A–K
11 anchor sections
66
Survived E1.5 trim
6
Killed by trim
3
SYNTH batches
52
Hybrid concepts
25
Patent packages
8.8
Top SYNTH score

Concept withheld — converged independently with an external AI reviewer (see below)

82
Top IPS gate

Unified Monolithic — the only DIAMOND of the run. Strongest claim fence.

41
Survivors in pool

All hybrids ≥7.0 auto-tagged — the next run starts from the best.

11
Anchor sections

A–K, including four new sections (H–K) for hydraulics, materials, coatings, method.

Independent Convergence — 8.8 / 8.8

Pieta SYNTH and a DeepSeek external review were run independently on the same domain, without communication. Both surfaced the same priority concept and scored it identically — 8.8. The concept itself is disclosed under NDA; the convergence is the point.

Pieta SYNTH — autonomous
A tapered adaptive channel topology with passive damping buffers at each flow deceleration node — no moving parts. Full construction under NDA.
8.8 / 10 SYNTH
=
DeepSeek — external review
Flagged a non-monotonic channel geometry with a passive damping buffer at the flow deceleration node as the priority concept. Full review text under NDA.
8.8 / 10 external
04

Patent Portfolio — 25 Packages

Each package: attorney-ready application draft, structured claims, patent drawings, commercial kit, prior-art search. Scores are SYNTH commercial viability (0–10, five pillars: MFG / C-B / Rel / TTM / CPP). Package mechanisms and claims are withheld — this page shows the shape of the portfolio, not the claims.

◆ Diamond IPS 82 · Strongest fence

Unified Monolithic Cold Plate

82
IPS Gate
8.2
SYNTH

A single-piece substrate with a silicon-matched CTE, additively manufactured. One integrated part combining four functional layers that today are separate components — channel network, transient buffering, integral sensing, anti-fouling surface system. Ships as one drop-in unit replacing cold plate, TIM, and lid. Full construction disclosed under NDA.

Why Diamond: five distinct patentable elements fence the claim — competitors must circumvent all five simultaneously. No prior art anticipates the full combination in a single device. The five elements are mapped in the NDA disclosure package.

4 prior-art refs · mapped · under NDA
Score 8.8 1 package

Passive Damped Adaptive Plate

8.8
SYNTH

The concept both reviewers converged on independently at identical 8.8 — the run's strongest external validation. Mechanism disclosed under NDA.

3 prior-art refs · under NDA
Score 8.6 5 packages

Adaptive Hexagonal Plate

8.6
SYNTH

Hexagonal adaptive network with passive buffering at each node. Mechanism under NDA.

3 prior-art refs · under NDA

Hex Shear Plate

8.6
SYNTH

Anti-fouling by geometry rather than chemistry — no coatings, no additives. Mechanism under NDA.

3 prior-art refs · under NDA

Convergent Damper Plate

8.6
SYNTH

Channels converge toward the hotspot zone with passive damping at each deceleration point — profile matched to the chip's heat map. Mechanism under NDA.

3 prior-art refs · under NDA

Self-Cleaning Branch Plate

8.6
SYNTH

A self-cleaning branching geometry that tracks the thermal load. Mechanism under NDA.

3 prior-art refs · under NDA

Graded Nozzle Plate

8.6
SYNTH

Pressure-drop budget distributed by thermal load, not uniformly. Mechanism under NDA.

3 prior-art refs · under NDA
Score 8.4 · 8.2 · 8.0 14 packages

Skeleton Channel

8.4
SYNTH

A single-pass channel whose internal treatment escalates with local heat flux. Mechanism under NDA.

2 prior-art refs · under NDA

Graduated Hex Plate

8.4
SYNTH

Surface treatment graduated along the flow path by branch generation. Mechanism under NDA.

2 prior-art refs · under NDA

Porous Insert Plate

8.4
SYNTH

A graded porous insert at the highest-flux zone. Mechanism under NDA.

2 prior-art refs · under NDA

Junction-Damped Plate

8.4
SYNTH

Pressure disturbances absorbed at each junction — never reaching the hotspot zone. Mechanism under NDA.

2 prior-art refs · under NDA

Pin-Fin Graded Plate

8.2
SYNTH

Pin-fin array graded outward from the thermal centroid. Mechanism under NDA.

2 prior-art refs · under NDA

Self-Balancing Nozzle Plate

8.2
SYNTH

A self-balancing manifold that equalises flow with no active control. Mechanism under NDA.

2 prior-art refs · under NDA

Transient Buffer Node

8.2
SYNTH

Thermal spikes absorbed as latent heat where the coolant's capacity is first exhausted — while the liquid loop responds. Mechanism under NDA.

2 prior-art refs · under NDA

Junction Buffer Array

8.2
SYNTH

Buffers sized to one workload spike duration, placed where flow naturally decelerates. Mechanism under NDA.

2 prior-art refs · under NDA

Undulating Channel Plate

8.2
SYNTH

Vortex intensity tracks local heat flux with no moving parts. Mechanism under NDA.

2 prior-art refs · under NDA

Self-Regulating Shear Plate

8.0
SYNTH

A geometry that holds target shear across the full GPU power envelope — no sensors, no actuators. Mechanism under NDA.

2 prior-art refs · under NDA

Tri-Zone Graded Plate

8.0
SYNTH

Three thermal zones, three surface treatments — boundaries set by the actual chip thermal map. Mechanism under NDA.

2 prior-art refs · under NDA

Uniformity-Optimised Splitting

8.0
SYNTH

Method claim. An optimisation algorithm that drives terminal-channel heat flux to maximal uniformity, scored by a proprietary uniformity metric — the algorithm is the patentable contribution. Method under NDA.

2 prior-art refs · under NDA

Balanced-Thermal-Load Plate

8.0
SYNTH

Holds wall shear and coolant heat pickup within bounds simultaneously across the die. Mechanism under NDA.

2 prior-art refs · under NDA

+ First-Run Packages

6.6–7.6
SYNTH

Four first-run concepts re-packaged and re-scored. Titles anonymised.

2 prior-art refs · under NDA
05

Market

TAM: global spend on direct-to-chip liquid-cooling thermal stacks for AI accelerators, growing 22.4% CAGR to 2030 (Grand View Research, MarketsandMarkets). SAM narrows to >10 kW hyperscale with a standardised OCP interface.

TAM
22.4% CAGR
$7.8B
2030 projection
~90% liquid-cooling attach × ~7M AI accelerator shipments × $1,000 avg thermal-stack price. Cross-checked against $17B DC liquid-cooling market (Dell'Oro, IDC).
SAM
~60% of TAM
$4.7B
Hyperscale + HPC, OCP interface
>10 kW/unit accelerators in hyperscale and HPC with a standardised OCP-compatible interface. Excludes edge, retrofit, immersion, non-standard.
SOM
Year 3
$127M
~4% share 2028 SAM
~127k units at $1,000 midpoint ASP. Assumes 2–3 hyperscaler/OEM design wins, 10k+ annual production ramps.
54%
Gross margin at pilot volume

ASP $780 blended · COGS 46% · validated at pilot

5.0×
LTV / CAC

CAC $850K · LTV $4.25M · 14-month payback

3
Buyer personas mapped

Hyperscaler thermal architect · accelerator OEM packaging director · Tier-1 supplier VP

06

Status & Access

Full technical whitepapers and the 25-document strategy package (GTM playbook, pitch deck, financial model, battlecards, PRD, data-room index) are available under NDA.

Talk to us

Adaptive Nunome is prototype-ready and seeking hyperscaler pilot qualification. Patent packages are drafted and ready for attorney review — August 2026 priority date.

david@pietarien.com