MicroLink × NVIDIA · Working Session Reference
← Back to Eight Zones
Prepared 4 May 2026 / Version 0.2 Draft / Confidential
Zone 01 · Build · Cluster A

Grid independence andheat recovery

For the energy team and Jared Carl

A diesel-free 2N power topology built on freed digester biogas, sized to AI compute envelopes, with server reject heat returning into the host's anaerobic process loop. The closed thermodynamic loop is the unit-economics differentiator and the EU EED 2027 template.

Owner
NVIDIA energy team + Jared Carl
Reference architecture lead
MicroLink lead
Shane Pather
CTO · Engineering
Site
San José Regional WWTP
11.2 MW · live deployment
Third parties
FuelCell Energy · Acculon
Named partners engaged
Working session
5 May 2026
30 min · Teams · 13 attendees
01 The Thesis
Conventional AI data centres are grid-bound. Ours is biogas-prime, hydrogen-ramp, battery-transient, diesel-never. Server heat returns to the host digester. The result is negative net new grid capacity at 11.2 MW of compute, on a site that produces its own fuel.
01
Heat recovered
8.5MW
85% of IT load returned across plate heat exchanger to digester loop.
Engineered Memo §04
02
Net grid impact
−21%
Site reduces local grid draw versus baseline cogen plus civic loads.
Modelled Site v0.4
03
MCFC nameplate
2.3MW
Biogas-fed, 47% electrical efficiency. Freed from existing 14 MW host cogen.
Quoted FCE 3000
04
Diesel hours
0
2N redundancy via fuel cell, LFP battery transient, PEM hydrogen ramp.
Locked SB 253 ready

The topologyin one diagram#

Four sources stacked by role. Server heat into the digester loop. A dry-cooler rejection path for any thermal that exceeds host demand. This is the picture the energy team will probe.

Every diagram with cooling shows a rejection path. Every power topology shows the role each source plays. There is no diesel anywhere.

The site runs 11.2 MW of IT load at PUE 1.12. Prime power is a 2.3 MW MCFC fuel cell stack on freed digester biogas, supplemented from the existing 14 MW on-site cogen during peak. Transient response sits with an LFP battery bank sized for sub-second to multi-minute coverage. Ramp and longer-duration backup are handled by PEM hydrogen generated and stored on-site. The grid is present but opportunistic, not load-bearing.

The thermal path is what makes the topology valuable. Server primary loop sits at 35–65 °C, sealed. A coolant distribution unit hands heat across to a secondary loop at 45 °C glycol-water. A plate heat exchanger transfers across to the host's tertiary loop, which feeds the digester sludge lift from 15 °C to 37 °C. LMTD sits at 12–18 °C, squarely inside standard PHE design windows. No heat pump required.

Figure 01 · Power and thermal topology
Four sources by role.Two loops by ownership.
Power flow left to right. Heat flow top to bottom. Dry-cooler rejection path always present. Host boundary marked.
Confidence · medium-high
POWER · BY ROLE COMPUTE HOST · THERMAL PRIME · BASE LOAD 2.3 MW MCFC FuelCell Energy · biogas · 47% LHV TRANSIENT · SECONDS LFP battery bank Acculon · ~140 MWh failover RAMP · MIN TO HOURS PEM hydrogen on-site electrolyser, stored OPPORTUNISTIC · ~12% PG&E grid 250 MW committed, not load-bearing DIESEL BACKUP None. By design. 2N BUS 11.2 MW COMPUTE POD · 11.2 MW IT 576 GPUs · Quantum-X800 + AC Scalable Unit · 1.2 MW canonical pod NVL72 · 142 kW / rack · liquid-cooled · PUE 1.12 SERVER HEAT · 35–65 °C · SEALED CDU + PLATE HEAT EXCHANGER SECONDARY 45 °C → TERTIARY 37 °C 8.5 MW · 85% recovery · primary value path HOST · TERTIARY Digester sludge 15 °C → 37 °C lift ~6 MW thermal demand REJECTION · DRY COOLER FALLBACK Partial-load · ~30–40% nameplate envelope FREED BIOGAS · DIGESTER → MCFC .
Source · MicroLink site engineering v0.4 · WWTP Thermal Memo §04 Method · Power topology and thermal balance schematic
§
The diesel-free topology is not a feature, it is the architectural premise
Removing diesel forces every other source into a defined role. MCFC carries continuous base load. LFP holds transient. PEM hydrogen handles ramp. Grid stays opportunistic. The result: zero on-site combustion, zero diesel logistics, and a topology that maps cleanly to SB 253, SB 261, and EU EED 2027 disclosure regimes from day one.

Why MCFCnot SOFC, not PEM, as prime#

Engineering rationale, vendor lineage, and the November 2026 council vote constraint that makes commissioning timelines load-bearing.

Molten Carbonate Fuel Cell is the right technology for this site for three reasons: it tolerates the biogas composition without external gas upgrading, its high-grade exhaust heat is directly usable at the digester, and FuelCell Energy has a deployable product line with the certification stack already in place.

SOFC would offer higher electrical efficiency at peak, but lower current TRL and longer commissioning windows make it incompatible with a November 2026 council vote tied to the Mayor's California Governor campaign. PEM fuel cells require pure hydrogen, which means treating biogas as a feedstock for upstream reforming. That doubles the system count and removes the closed-loop simplicity that makes this site exportable as a template.

The 2.3 MW configuration sits inside FuelCell Energy's FCE 3000 module envelope. Above that, the 12.5 MW Block System is the productisation that matters for the global rollout. Datasheet figures: 50% electrical / 85% overall LHV, 7,580 BTU/kWh, NOx and CO at 0.01 lb/MWh, fuel blending up to 50% hydrogen. Independently certified to ANSI/CSA FC-1, NFPA 70, UL1741, IEEE1547, and CA Rule 21.

Table 01 · FuelCell Energy product envelope · 12.5 MW Block System Source · FCE March 2026 datasheet
Parameter Value San José config Note
Net power output12.5 MW2.3 MWSingle FCE 3000 stack
Electrical efficiency LHV50%47%Module-level vs Block
Overall efficiency LHV85%~85%With heat recovery
Heat output max31.0 MMBTU/h~5.7 MMBTU/hCooling exhaust to 120 °F
H₂ blending range0% to 50%0%Path to ramp role expansion
NOx and CO emissions0.01 lb/MWh0.01 lb/MWhCombustion-free
CO₂ electric only886 lb/MWh886 lb/MWhPre heat recovery
CO₂ with heat recovery554 lb/MWh~330 lb/MWhBiogas adjustment, project-dependent
Noise at 30 ft62 dBA62 dBACivic site compatible
Standards stackANSI/CSA FC-1 · NFPA 70 · UL1741 · IEEE1547 · CA Rule 21 · CARB 2013 Biogas

The closed thermodynamic loopis the value engine#

A WWTP does not heat its incoming water. It heats sludge to 37 °C, year-round, continuously. That is the only thing about this site that matters to the unit economics.

Make-up water is required only at sealed maintenance events. The host's process water stays inside the host's mass balance. We deliver thermal across a plate heat exchanger. We do not consume host water.

The digester sludge lift from 15 °C to 37 °C is a continuous, year-round, ~407 kW thermal demand at 100 MGD plant scale, which scales linearly with site size. Our secondary loop sits at 45 °C, eight degrees above the target. The plate heat exchanger LMTD lands at 12–18 °C, squarely inside standard product windows. No heat pump required, which keeps round-trip losses out of the value chain.

At an 11.2 MW IT load with 85% IT-side recovery and a 60% delivery factor through the loop, the site delivers approximately 179,000 MMBtu/year of thermal into the digester. At avoided gas displacement values between $15 and $25 per MMBtu (host-dependent), that converts into $2.69M to $4.48M of annual thermal value. The asymmetry against the cost of a dry-cooler-only configuration is roughly 3×.

Figure 02 · Annual thermal value
Three sinks ranked.One sink dominates.
Same 8.5 MW recovered heat, three host-side use cases. Digester sludge lift is continuous and year-round; the others are seasonal or partial.
Confidence · medium · placeholder recovery factor
$0 $1M $2M $3M $4M $0 Dry cooler only No host sink $1.1M Building heat Seasonal · 200 kW–1 MW $2.69M Digester @ $15 Continuous · 22 °C ΔT $4.48M Digester @ $25 Highest tier ~3× asymmetry ANNUAL THERMAL DELIVERED 179,000 MMBtu / year 8.5 MW recovered 8,760 hours 60% recovery factor PLACEHOLDER · LOCK AT ENGINEERING REVIEW
Source · WWTP Thermal Memo §07 · Internal model v0.3 Method · 8.5 MW recovered · 60% delivery · gas displacement at $15–$25 / MMBtu
WUE · this site
0.10–0.30L/kWh
Sealed liquid loops · make-up at maintenance only.
WUE · US average
1.80L/kWh
LBNL 2024 benchmark · evaporative cooling pattern.
Reduction
10–15×
Versus US average. 50–60× versus hyperscale evaporative.

The diesel-free decisionand what it costs#

A capex premium of roughly 60% buys a deployment archetype that the conventional configuration cannot reach. For a public-sector reference design, this is the architectural choice.

A diesel-backed AI data centre on a wastewater site is a non-starter under California's combustion air permitting and the planning environment around environmental services land. A diesel-free configuration removes the constraint, but the topology has to do real work to be 2N without diesel: the MCFC carries continuous base, the LFP bank covers transient response, the PEM hydrogen system holds ramp duty, and the grid stays opportunistic.

The capex premium is real and we state it openly. $17–22M per MW for the diesel-free configuration against a roughly $12M per MW baseline for a conventional diesel-backed AI data centre. The premium is concentrated in the fuel cell stack, the hydrogen subsystem, and the larger battery bank. The closed thermodynamic loop offsets a portion through annual thermal value (~$3.6M / year at the highest tier on this site), but the strategic case is not payback-driven. It is the unlock case.

Figure 03 · Capex stack · per MW · diesel-free vs conventional
The premium is real.The unlock is bigger.
Indicative breakdown for a 11.2 MW deployment. Project-dependent at engineering and procurement lock.
Confidence · medium · indicative
CONVENTIONAL · $12M / MW Shell + IT systems $5.0M Cooling + MEP $2.9M Power + UPS $2.1M Diesel gensets $2.0M CANNOT DEPLOY ON THIS HOST SITE Combustion permitting precludes deployment on civic land. DIESEL-FREE · $17–22M / MW Shell + IT systems $5.5M Liquid cooling + PHE $3.4M MCFC stack · 2.3 MW $4.0M LFP battery bank $2.8M PEM H₂ subsystem $2.5M Power + DC $1.8M Premium offset annually by ~$3.6M thermal value
Source · MicroLink internal model v0.3 · Vendor RFQ envelope Method · Per-MW illustrative breakdown · 11.2 MW reference site
§
The unlock case, not the payback case
A 60% capex premium pays back over time through thermal value at the highest tier, but that is not the argument. The argument is that the diesel-free configuration is the only way this kind of site reaches deployment. Conventional cannot get on the property. That is what makes the topology exportable as a public-sector pattern.

The askand what we bring#

A working-session asks for a defined set of things. This is what we are asking the energy team for, and what we contribute in return.

Tier · Reference architecture co-development
A canonical patternfor grid-constrained public-sector sites

San José is the live deployment. The pattern is the export. Both depend on a tight engineering loop with the NVIDIA energy team through Q3 2026.

From the energy team
  • Topology review against NVIDIA reference patterns
  • Secure design portal access for the 1.2 MW pod
  • 576 GPU Quantum-X800 + AC Scalable Unit sized into envelope
  • Biweekly cadence through Q3 2026
  • Spec sheet shared, locked items reviewed
From Jared Carl
  • Internal advocacy for the public-sector archetype
  • Co-authored NVIDIA-recognised reference design
  • San José as the worked example
  • Path to DGX-Ready Colocation entry
  • Bridge to Andria Zou for sustainability layer
What MicroLink contributes
  • Live deployment data from Q1 2027 commissioning
  • Closed-loop thermodynamic dataset for white paper
  • Civic procurement pattern, deployable template
  • FuelCell Energy + Acculon engaged
  • Council vote secured November 2026