MicroLink × NVIDIA · Working Session Reference
← Back to Eight Zones
Prepared 4 May 2026 / Version 0.2 Draft / Confidential
Zone 07 · Contributions · Cluster C

Open-sourcecontributions

For Danny Zaidifard and Wendy Zhu

The contribution lane in the NCP Reference Architecture is empty. Seventeen contributions across software, hardware, datasets, and deployment guides, license- and foundation-home-mapped from day one, are the case for accelerated NCP onboarding, not the post-approval marketing.

Owner
Danny Zaidifard
BD, Strategic Partnerships · ecosystem flywheel
Adjacent owner
Wendy Zhu, PhD
Validation · partner-adjacent
MicroLink lead
Shane Pather
CTO · contribution governance
Foundation homes
LF Energy · CNCF · OCP
OpenTelemetry · AOUSD
Working session
5 May 2026
30 min · Teams · 13 attendees
07 The Thesis
The contribution corpus is the case for accelerated NCP onboarding. There is no published upstream contribution requirement in the NCP RA, which means the contribution lane is empty. Being the partner that fills it converts geographic adjacency to NVIDIA HQ, the only WWTP-coupled NCP candidate in California, and a $17-22M/MW build into long-duration influence. Standards contributions are strategic IP placement: when MicroLink's heat-coupled control loop, sovereign multi-tenant policy bundle, and EU EED template become defaults in DSX Blueprints and NCP Reference Architectures, every subsequent NCP candidate adopts MicroLink's patterns by following the reference design.
01
Software
5
HCCL, FacIT, DSX Max-P, digital-twin schema, sovereign multi-tenant policy bundle.
Apache 2.0 LF Energy + CNCF
02
Hardware specs
4
Secondary loop, heat exchanger, fuel cell + biogas + H2 + battery, edge-to-core.
OCP CLA ASHRAE TC 9.9
03
Datasets
4
Thermal Lab, closed-loop thermodynamic, grid-independent, GPU-hour carbon attribution.
CC-BY 4.0 NGC + HF
04
Deployment guides
4
Municipal sovereign AI, waste-heat commissioning, grid-independent siting, EU EED 2027 template.
CC-BY 4.0 17 total

Backblaze, Equinix, Cloudflarethe pattern is consistent#

Three precedents covering hardware, neutrality, and software. Each one published the working internal architecture before the competitive set did, and each one parlayed that publication into structural advantage.

The pattern repeats across three different categories of company and three different decades. None of them is the largest player in their category. All three converted contribution into compounding advantage that scale alone could not buy.

Hardware lane · 2009
Backblaze
Storage Pod 1.0 published September 2009: 67.5 TB in 4U at $0.117/GB versus competitors at 10x cost. Open-sourced full BOMs and STEP CAD files for six generations. Drive Stats datasets since 2013 became the most-cited HDD reliability reference in industry. Storage Pod 1.0 entered the Computer History Museum's permanent collection in 2026.
Neutrality lane · 2020
Equinix
Built an $8.7B revenue business by being carrier-, cloud-, and partner-neutral rather than the largest hyperscaler. Internet Exchange and Equinix Fabric became de facto interconnect standards. Tinkerbell open-sourced May 2020, accepted into CNCF Sandbox November 2020.
Software lane · 2024
Cloudflare
Pingora open-sourced February 2024 under Apache 2.0 after replacing NGINX at quadrillion-request scale with 70% lower CPU and 67% lower memory. quiche and tokio-quiche followed. A Cloudflare engineer co-chairs the IETF QUIC Working Group that produced RFC 9000.

The five-element patterndistilled across all three

Every one of the three follows the same operating logic. The differences are in domain and in license, not in mechanism.

→
The five-element pattern
(1) Publish the working internal architecture before competitors do. (2) Open the non-differentiated layer while keeping operations and data proprietary. (3) Convert publication into developer mindshare that drives recruiting and sales pipeline. (4) Pursue standards-body leadership rather than membership: chairs of working groups, not just members. (5) Use open contribution to commoditise complementary inputs.

The leverage moment is realand time-bounded

NVIDIA's NCP program sits inside the NPN with a tiered structure including specialised "Reference Platform NCPs" announced October 2024. Andria Zou is responsible for cultivating roughly 33 DGX-Ready partners and 20+ NCPs. At twenty-something NCPs globally, the standards corpus is still small enough that one operator's contributed reference design can become the default template. Once the partner population crosses one hundred, contribution becomes one of many and the leverage decays.

The window closes when NVIDIA publishes its own complete reference for the relevant pattern. NVIDIA already released the Vera Rubin DSX AI Factory Reference Design and Omniverse DSX Blueprint at GTC March 2026 with co-design partners Cadence, Schneider, Vertiv, Eaton, Trane, Siemens. But no published heat-export-to-WWTP, biogas-MCFC-microgrid, or municipal sovereign multi-tenant reference exists yet. MicroLink's pre-commissioning window is approximately twelve months.

Seventeen contributionsacross four lanes#

The corpus is inventoriable. Each contribution carries its own license, its own foundation home, and its own NVIDIA strategic interest. Read it as a taxonomy, not a wishlist.

SW
Software · Apache 2.0 default · LF Energy and CNCF as foundation homes
5 of 17
SW·01 Apache 2.0
Heat-coupled cooling control loop (HCCL)
Closes the loop between server thermal output and host process thermal demand at sub-second resolution. BACnet/IP, Modbus TCP, OPC-UA on the facility side; DCGM Exporter and NVML on the IT side. No direct equivalent exists.
Foundation home · LF Energy
SW·02 Apache 2.0
FacIT monitoring schema
OpenTelemetry-aligned semantic convention extension producing unified telemetry across GPU, rack, pod, facility, host process, grid interface. Designed to plug directly into NVIDIA Mission Control's OTel pipeline.
Foundation home · OpenTelemetry SIG
SW·03 Dual-license
DSX Max-P module (the inverse of Max-Q)
Where Max-Q optimises performance per watt at maximum efficiency, Max-P maximises sustained thermal output when the host process demands heat. Captures energy-reuse credit under EU EED's ERF metric.
Co-development · NVIDIA + MicroLink
SW·04 Apache 2.0 + CC-BY 4.0
Digital-twin schema for thermally coupled industrial computing
Omniverse Blueprint extension layered on the existing AI Factory Digital Twin and DSX Blueprint. USD asset libraries for digesters, biogas cleanup, MCFC and SOFC stacks, plate-and-frame exchangers, dry-cooler fallback paths.
Foundation home · AOUSD + Omniverse
SW·05 Apache 2.0
Sovereign multi-tenant policy bundle on Run:ai + BlueField-3
Reference Helm chart and OPA policy bundle for public-sector multi-tenant deployment with hardware-attested isolation. Layered on KAI Scheduler, BlueField-3 SuperNIC, DOCA Platform Framework. NVIDIA's NCP RA does not publish a sovereign-municipal multi-tenant Helm chart today.
Foundation home · CNCF
HW
Hardware specifications · OCP as the primary home · filling the gaps NVIDIA's HGX/GB200 OCP work does not cover
4 of 17
HW·06 OCP Spec
Liquid cooling secondary loop reference design (CDU to host process)
Castrol ON DLC PG 25 chemistry, 316L stainless tubing, EPDM gaskets, HNBR/FKM elastomers, ASHRAE TC 9.9 W2 Liquid Water Quality Class. Defines flow rates, pressure differential, quick-disconnect redundancy.
Foundation home · OCP Cooling Environments
HW·07 OCP Spec
Heat exchanger architecture (data centre secondary loop to WWTP digester process loop)
Two-stage train: shell-and-tube first stage, plate-and-frame second stage. Failure-mode coverage for siloxane SiO2 deposition, particulate fouling, dry-cooler fallback. No public reference for WWTP-coupled data centres at this temperature pairing.
Foundation home · OCP + ASHRAE TC 9.9
HW·08 CC-BY 4.0 + OCP
Fuel cell + biogas + H2 + battery configuration spec for 1 MW grid-independent
Single-line and BOM for FuelCell Energy SureSource 1500, 4 MWh LFP, on-site H2 storage per CFC and NFPA 2 setbacks, microgrid controller per IEEE 2030.7-2017, BESS-led black-start sequence.
Foundation home · LF Energy + DOE
HW·09 OCP + Apache 2.0
Edge-to-core deployment hardware spec for sub-MW host sites
Jetson AGX Orin Industrial reference design with industrial temperature range, fanless thermal interface, BACnet/IP gateway for facility integration, OPC-UA northbound, FacIT-aligned telemetry.
Co-author · NVIDIA-led
DATA
Datasets · the highest-leverage category · non-replicable, only an operating site produces them
4 of 17
DATA·10 CC-BY 4.0
Thermal Lab dataset
51 instrumentation variables x 6+ months of GPU thermal field data per site, anonymised by tenant. Per-GPU junction temp, memory temp, hot-spot temp; SM and memory clock under throttle; coolant inlet/outlet, flow, pressure differential, secondary-loop chemistry pH and conductivity. Backblaze Drive Stats analogue.
Foundation home · NGC + Hugging Face
DATA·11 CC-BY 4.0
Closed-loop thermodynamic operational dataset
Hourly resolution x 1 year. Data centre IT load, thermal export with delivery-temperature curves, digester process temperature, biogas yield, methane-to-MCFC fuel rate, MCFC electrical output, dry-cooler fallback engagement frequency. Target ERE under 0.7 paired with PUE under 1.2.
Foundation home · NGC + LF Energy + SJSU
DATA·12 CC-BY 4.0
Grid-independent operational dataset (MCFC ramp + LFP cycle)
MCFC ramp behaviour (NREL's LADWP Scattergood study explicitly excluded SOFC for backup because of extended start-up time; this dataset quantifies how the LFP battery bridges that gap). Biogas variability, hydrogen storage cycles, LFP cycle life under data centre load profiles.
Foundation home · LF Energy + DOE NREL
DATA·13 CC-BY 4.0
Carbon attribution at GPU-hour level
Per-workload Scope 1, 2, 3 dataset, anonymised by tenant, applying SCI Specification (ISO/IEC 21031:2024) at GPU-hour granularity using marginal grid carbon intensity from CAISO. Directly addresses EU AI Act Article 53(1)(d) energy-consumption disclosure.
Foundation home · Green Software Foundation
DOC
Deployment guides · lowest IP risk, highest pipeline value · what unblocks the next 30 partners
4 of 17
DOC·14 CC-BY 4.0
Municipal sovereign AI deployment guide
Co-authored with NVIDIA Public Sector. Legal (Bayh-Dole, sponsored research IP), financial ($17-22M/MW capex, value-share), technical (NCP RA compliance, Run:ai + BlueField-3), political (community engagement, environmental justice review, GovAI Coalition templates). No US city has built dedicated municipal GPU compute for residents and startups.
Foundation home · NVIDIA Public Sector + GovAI
DOC·15 CC-BY 4.0
Waste-heat-recovery commissioning playbook
Stockholm Data Parks, Meta Odense, Yandex/Nebius Mäntsälä, Equinix Helsinki HE5, IBM Aquasar, all generalised. Failure modes (heat-pump COP collapse under fouling, gasket failure, dry-cooler fallback timing). Cutover sequence: pre-commissioning calibration, soft-start, ramp, steady-state tuning across two seasonal cycles.
Foundation home · LF Energy
DOC·16 CC-BY 4.0
Grid-independent AI data centre site selection playbook
Biogas availability (US WWTPs over 50 MGD typically produce sufficient biogas for over 1 MW MCFC, with FuelCell Energy's Tulare CA WWTP at 2.8 MW SureSource 3000 as the operational precedent). Electrical interconnection, water access, community acceptance, regulatory environment.
Foundation home · LF Energy
DOC·17 CC-BY 4.0
EU EED 2027 compliance template
Article 12 reporting, Article 26 waste-heat utilisation. Climate Neutral Data Centre Pact targets, WUE_max with K1, K2, K3 multipliers. The template lets every NVIDIA EMEA NCP partner produce a compliant report from Day 1 without bespoke engineering.
Foundation home · EUDCA + LF Energy
§
The corpus is the case
No two of these seventeen items are duplicative. Each closes a specific gap NVIDIA has not closed (DSX Blueprint silences, OCP Cooling Environments gaps, Mission Control schema extensions). The argument is not "let MicroLink in", it is "the contribution lane is empty and MicroLink is the partner ready to fill it."

Apache 2.0 is the right defaultand AGPL would defeat the purpose#

Section 3 patent grant auto-terminates against litigants. Distro-uniformity is preserved. The trade-secret floor is the operational tuning, the per-tenant data, and the host commercial economics.

Apache 2.0 is the right default for software because of the explicit Section 3 patent grant that auto-terminates rights against litigants, because it is what NVIDIA itself uses across NeMo, Megatron-LM, TensorRT-LLM, KAI Scheduler, OSMO, and infra-controller-core, and because the alternative copyleft path (AGPL or SSPL) creates the same distro-rejection problem MongoDB experienced when Debian, Red Hat, and Fedora dropped MongoDB in 2018-2019.

Source-available licenses (BSL, Elastic License v2, SSPL) would also fail the Apache 2.0-uniform expectation of NVIDIA's ecosystem and would invite forks the moment MicroLink's market position changed. The Terraform to OpenTofu fork in August 2023 (joining Linux Foundation September 2023, 9.8M downloads, 300% YoY growth within months) is the most recent precedent for that failure mode.

Hardware and patent strategyOCP CLA with grant, OIN for defence

OCP CLA with patent grant for hardware specs. The OCP project gates contributions through a CLA with explicit patent grants and trademark license. Open Invention Network membership (4,100+ members, 2.7M+ patents, January 2026 OIN 2.0 with paid enterprise tier and free small/startup tier) is the correct defensive patent hub for MicroLink to join, providing cross-licensing protection against patent assertion in the open-source ecosystem.

Trademark strategyfollows Databricks-Spark and Confluent-Kafka

Keep MicroLink-named research outputs (Thermal Lab, Carbon Ledger) on data products where the brand vector is the dataset itself, but adopt community-neutral names on software contributions targeting foundation donation. A schema named "MicroLink Telemetry Schema" will not become an OpenTelemetry semantic convention. A schema named "Industrial Computing Telemetry Conventions" with MicroLink as originating contributor will.

The trade-secret residualwhat is held back

The non-negotiable retained moat is the three-loop thermal architecture, the Attestation-to-Heat patent, the operational tuning parameters and control loop gains, the host-process bridging implementation specifics for active archetypes, and the host commercial economics. Schemas, interface specs, control loop architecture (not tuning), curriculum, and anonymised aggregate datasets are contributable. Everything below that line is retained.

Trade-secret floor by element Status as of v0.2 draft
Element Status
Three-loop thermal architectureTrade secret
Attestation-to-Heat IPPatent (defensive)
Heat-coupled control loop architectureContributed (Apache 2.0)
Heat-coupled control loop tuning gainsTrade secret
Host-process bridging architecture (general)Contributed (OCP)
Host-process bridging implementation (active archetypes)Trade secret
Telemetry schemaContributed (Apache 2.0)
Analytics layer running against the schemaTrade secret
Thermal Lab dataset (12-month lagged, anonymised, aggregated)Contributed (CC-BY 4.0)
Thermal Lab dataset (real-time, per-tenant)Trade secret
Operator training curriculumDual-license (CC-BY playbook + paid certification)
Host commercial economicsTrade secret
!
The 12-month-lag-plus-anonymisation pattern
The publicly released subset is anonymised, aggregated to a useful-but-not-operationally-identifying granularity, and lagged 12 months from real-time. Full real-time stream stays inside MicroLink and is shared with NVIDIA under bilateral NDA. The dataset is published; the model trained on it, and the operational decisions taken from it, are retained.
"
The contribution lane is empty. Being the partner that fills it is the cheapest way to convert geographic adjacency, the only WWTP-coupled NCP candidate in California, and a 576-GPU build into long-duration influence.
Cluster C synthesis
v0.1 working draft

Each contribution maps toa named NVIDIA pipeline interest#

The flywheel is mutual. NVIDIA gets a pipeline of qualified operators, contributable patterns that compound across the next thirty NCP candidates, and chip-thermal-envelope data the chip team cannot otherwise generate. MicroLink gets early NCP approval, influence over the standards layer, and a defensible non-scale moat.

Contribution to strategic interest 8 of 16 mappings shown
MicroLink contribution NVIDIA strategic interest
HCCL software open under LF Energy Every NCP candidate with a waste-heat play needs this; reduces NVIDIA's enablement burden and makes adjacent partners convertible without bespoke build.
FacIT monitoring schema in OpenTelemetry Mission Control already uses OpenTelemetry-native pipeline; FacIT extends the schema to facility-and-host coverage NVIDIA does not want to build itself.
DSX Max-P module Heat-export use cases were not in the DSX Max-Q optimisation envelope; Max-P fills the gap the DSX team would otherwise have to build.
Sovereign multi-tenant policy bundle NCP RA covers IaaS/CaaS/PaaS but does not publish a municipal sovereign multi-tenant Helm chart. Cuts municipal evaluation time from quarters to weeks.
Thermal Lab dataset NVIDIA chip team uses thermal field data for Vera Rubin and Feynman generation thermal envelopes. Direct chip-design input from operating data NVIDIA cannot otherwise generate at NCP scale.
Fuel cell + biogas + H2 + battery configuration Microsoft Quincy 3 MW PEM and Cat-Ballard-Microsoft Cheyenne 1.5 MW were demos; this would be the first commercial baseload reference. Unblocks DSX Flex's claim on 100 GW of stranded grid power.
Municipal sovereign AI deployment guide NVIDIA Public Sector pipeline depends on cities procuring with confidence. Published reference shortens every city's evaluation by months.
EU EED 2027 compliance template Direct unblock of EMEA NCP partner pipeline targeting EU EED Article 26 from 2027.

The flywheel reads in both directionsand that is the point

The contribution corpus is the case for accelerated NCP onboarding because it is the case for NVIDIA's own next thirty partner deployments. Every gap closed for MicroLink is a gap closed for NVIDIA's enablement team. Every dataset published is signal for the chip team. Every deployment guide co-authored is a city procurement that does not need a custom evaluation cycle.

⇄
The flywheel is mutual
NVIDIA gets a pipeline of qualified operators, contributable patterns that let every subsequent NCP candidate deploy faster, datasets that improve chip thermal envelopes and DSX Blueprint accuracy, and a municipal sovereign AI reference that converts US city procurement at scale. MicroLink gets early NCP approval (because the contribution corpus is itself the case), influence over the standards layer that shapes every subsequent NCP, a defensible non-scale moat through operating data and operator pipeline, and trademark equity in MicroLink-named research products.

The askand what we bring#

Two named owners, two distinct mandates. Danny on the ecosystem flywheel. Wendy on the validation pathway. The bilateral session for each is the lock that turns the corpus into a sequence.

Tier · NCP contribution corpus and standards co-authorship
A 17-contribution corpuslicense-mapped, foundation-mapped, ready to sequence

San José is the live deployment. The contribution corpus is the case. The bilateral sessions with Danny and Wendy are the locks that turn empty lanes into reference designs.

From Danny Zaidifard
  • Bilateral session within 30 days on the contribution sequence
  • Foundation-donation timing recommendation across HCCL to LF Energy, FacIT to OpenTelemetry, multi-tenant bundle to CNCF
  • Bridge to NVIDIA developer relations for the publication launches
  • Co-development of the Municipal Sovereign AI Deployment Guide (DOC·14)
  • Inception-startup integration for the Apache 2.0 contribution governance
From Wendy Zhu
  • Validation pathway for DSX Max-P co-development (SW·03)
  • Reference platform validation for the multi-tenant Helm chart (SW·05)
  • Thermal Lab dataset sampling-rate and instrumentation-variable specification (DATA·10)
  • Co-design review for the heat exchanger spec (HW·07)
  • Validation criteria for the EU EED 2027 compliance template (DOC·17)
What MicroLink contributes
  • Pre-commissioning window of approximately twelve months: the standards-shaping moment is now
  • 17 contributions across four lanes, license- and foundation-home-mapped from day one
  • Apache 2.0 default for software, OCP CLA for hardware, CC-BY 4.0 for datasets and docs
  • Open Invention Network membership for defensive patent coverage
  • Trade-secret floor preserved: tuning gains, real-time per-tenant data, host commercial economics held back
Bilateral working sessions requested · 30 minutes each
  • Danny Zaidifard · the contribution sequence and foundation-donation timeline · within 30 days of the LOI
  • Wendy Zhu · validation pathway for DSX Max-P and the multi-tenant Helm chart · within 60 days
  • NVIDIA Open Source Programme Office · CLA harmonisation across LF Energy, CNCF, OCP, OpenTelemetry · within 90 days