Two parts of the model checked against published 2026 benchmarks: the NeoCloud/GPU tenant
thesis at 6–12MW, and the cost assumptions (shell vs. turnkey split, timeline, recovery, staffing).
Compiled August 2026. Every figure is cited; where sources disagree, both are shown.
18 sources1 source disagreementResearch basis, not vendor quotes
1. Tenant realism: does a NeoCloud/GPU tenant fit a 6–12MW site?
The model's central assumption, checked against named deals: at this site's scale a
single specialist NeoCloud/GPU anchor tenant is more realistic than a diversified multi-tenant retail
mix.
Consistent with modelNeoclouds take sites in this
size range. CoreWeave holds a 9MW
allocation at a Flexential facility in Hillsboro, Oregon [2]. Crusoe has a
5MW deployment at Equinix's Culpeper, Virginia campus [3], and is separately leasing 12MW
(with 36MW more in development) at Polar's DRA01 site in Norway and 20MW (45MW more in development) at
Polar's DAL01 site [3]. These are single-digit to low double-digit MW deals signed by
recognisable neoclouds, not 50–300MW hyperscale campuses, and show the tenant archetype is plausible
at Lynnway Park's scale.
Tenant
Size
Site
What it shows
CoreWeave
9MW
Flexential, Hillsboro, OR
Neocloud taking a single-digit-MW colocation allocation, not a full campus
Crusoe
5MW
Equinix Culpeper, VA
Neocloud at a shared multi-tenant campus, comparable in scale to 0 Circle Avenue's 6MW default
Crusoe
12MW (+36MW planned)
Polar DRA01, Norway
Phased neocloud deal starting in Lynnway Park's size range
Crusoe
20MW (+45MW planned)
Polar DAL01, Norway
Same pattern: starts small, phases up
CoreWeave
36MW
Digital Realty, Hillsboro, OR
Reference point for CoreWeave's larger, single-building "one InfiniBand fabric" deals
AMD / unnamed neocloud
152MW
Core Scientific, Auburn AL & Dalton GA
Reference point for the large end of the market; not comparable to a 6–12MW site
The caveat. Market commentary states that 5MW and
larger blocks in primary AI markets have been largely pre-committed by hyperscalers, neoclouds and AI
labs through 2028, and the 1–5MW capacity that remains "does not sit on a public website
or a provider's marketing page". It is placed through direct relationships, not open listings
[1]. Two readings follow: (a) at 6–12MW, just above that 1–5MW niche, Lynnway Park
may have to compete for the same off-market, relationship-driven demand; (b) as a secondary market outside
the primary AI hubs, Lynn may see less pre-commitment pressure and more availability for a smaller anchor
tenant. Neither is confirmed for this site. The evidence shows the archetype exists and comparably sized
deals happen; it does not show a specific tenant is coming to Lynn, MA.
2. Shell vs. turnkey cost split: is 15% the right shell share?
The model's Shell-Only deal structure assumes the landlord pays ~15% of a full
turnkey build cost. Independent sources point to a similar range, one with a direct $/sqft
comparison.
Source
Shell cost
Full turnkey cost
Implied shell share
getcostidea.com [9]
$105–235/sqft
$625–1,135/sqft (total incl. shell)
≈17–21%
Turner & Townsend, via gigacapacity.com [6], air-cooled
14% of total cost
14%
Turner & Townsend, via gigacapacity.com [6], liquid-cooled
9% of total cost
9%
cortexconstruct.com [7]
"Building shell and site" 15–20% of total
15–20%
iesmartsystems.com [10]
"Land and shell" 15–20% of budget
15–20%
dgtlinfra.com
"10–20% of total cost"
10–20%
Model assumption
15%
Possible refinementFive independent sources
cluster in the 14–21% range for the shell/civil share of a full build. The model's
flat 15% for both archetypes sits at the low end of that range. The larger finding is that the
shell share differs between the two tenant archetypes.
Turner & Townsend's air-cooled vs. liquid-cooled split (14% vs. 9%) shows that liquid cooling shifts
budget share away from shell/civil and into mechanical (22% air-cooled → 33%
liquid-cooled) [6]. The model uses the same 15% shell share for Standard Colo
(air-cooled) and NeoCloud GPU (liquid-cooled); a closer fit would be about 16–18% for Standard and
9–12% for NeoCloud. Not applied in the current model.
3. CapEx/MW benchmarks: is the NeoCloud CapEx premium (+15%) large enough?
The model applies a flat +15% CapEx/MW premium for NeoCloud's liquid-cooling
infrastructure. Published 2026 benchmarks put the AI-optimised premium over standard colocation
higher.
Facility type
$/MW
Source
Enterprise / Tier II
$8–10M
irecruit.co [8]
Colocation / Tier III
$10–12M
irecruit.co [8]
Enterprise / Small Colo
$7–10M
cortexconstruct.com [7]
Large Colocation
$8–12M
cortexconstruct.com [7]
Model: Standard Colo (turnkey)
$12.9M
Model assumption
Hyperscale (AI-optimised)
$15–20M+
irecruit.co [8]
Hyperscale (AI-optimised)
$12–18M
cortexconstruct.com [7]
Model: NeoCloud GPU (turnkey, +15%)
$14.84M
Model assumption
Possible refinementThe model's Standard Colo
figure ($12.9M/MW) sits slightly above the top of the sourced $7–12M range for enterprise and colocation
facilities [7][8]. The NeoCloud figure ($14.84M/MW, a +15% premium) sits below
the sourced AI-optimised range of $15–20M+ and in the middle of the $12–18M range
[7][8]. AI-optimised builds appear to run +16% to +55% above standard rather than
a flat +15%, so the NeoCloud CapEx premium may be understated; correcting it would weaken the NeoCloud
archetype's economics (higher CapEx, same rent). Not applied in the current model.
4. Construction timeline by site size: sources disagree
The model uses a 14-month construction period for Standard Colo and 12 months for
NeoCloud, based on a facility-size timeline table. A second, separately sourced benchmark
differs materially for small sites; both are shown.
Source
Size bucket
Construction duration
cortexconstruct.com [11]
Edge/Small, 1–5MW
5–8 months
cortexconstruct.com [11]
Mid-Size, 5–15MW
8–12 months
cortexconstruct.com [11]
Large Enterprise, 15–30MW
12–18 months
Model: Standard Colo / NeoCloud
6–12MW
14 / 12 months
hamminstitute.org [12]
Enterprise/Labor-Intensive, 5–20MW
30–36 months
hamminstitute.org [12]
Regional/Standard, 20–50MW
24–30 months
hamminstitute.org [12]
Hyperscale/Efficient, 100+MW
18–24 months
Sources disagreeTwo sourced
benchmarks disagree sharply on how construction time scales with site size. A construction-industry
cost and timeline source (Cortex Construct) says smaller sites build faster: 8–12
months for 5–15MW, which is the basis for the model's 12–14 month assumption
[11]. A labour and workforce economics source (a data center employment forecasting model)
says the opposite: smaller "labor-intensive" sites take longer (30–36 months for
5–20MW). They gain less from the standardisation, prefabrication and repeatable design of large
hyperscale campuses, and more often involve custom permitting or brownfield/conversion complexity
[12]. Both mechanisms are plausible; the sources measure different things (a generic
construction-cost timeline vs. a labour-intensity-driven schedule). This is an open uncertainty in
the model. The 12–14 month assumption is at the optimistic end of the range found; a 30+
month build would weaken both archetypes' returns by delaying revenue and extending the loss-making early
years. A contractor quote for this site is needed before relying on either figure.
5. Lease structure & cost recovery: is 98% recovery realistic for a landlord deal?
The model assumes near-100% cost recovery (the operator reimburses the
landlord for almost all operating costs) at both 0 Circle Avenue Shell-Only and the 626 Lynnway MEDC
conversion, since Lynnway Park is landlord in both. This matches a well-documented commercial real
estate lease type.
Consistent with modelNear-100% recovery matches a
named data center lease structure, the Triple Net (NNN) lease, where the tenant pays a base price (quoted in $/kW) plus its
full share of operating expenses (property tax, maintenance, staffing), with recoverable costs
"often around $35/kW" [13]. NNN and the related Modified Gross lease are the leases wholesale
data center developers typically use [13], which fits the powered-shell landlord role the model
assumes for Lynnway Park. 626 Lynnway conversion cost recovery. Benchmark: ~85%; model: 98%
NNN. The same source describes a Gross/Full-Service/All-In
lease with lower recovery as "best suited for smaller, multi-tenant deals like retail colo"
[13]. The model uses 98% NNN for the conversion because Lynnway Park acts as landlord, not
operator. The ~85% (modified gross) recovery applies only to the 0 Circle Avenue Full-Turnkey case
(see Assumptions §10).
6. Shell-Only landlord rent: is the model's yield on cost in line with powered-shell deals?
The model's Shell-Only landlord rent ($28.37/kW/mo by default, ~12.0% of the
operator blend) is an input and yield on cost is the result. This section checks that yield
against published powered-shell development benchmarks.
Metric
Benchmark
Source
Powered shell: development yield on cost
7.00–8.50%
MaxLife Development [16]
Powered shell: stabilised cap rate
5.25–6.50%
MaxLife Development [16]
Powered shell scope (per source)
Building envelope, raw utility power, fibre access; tenant fits out cooling/generators/IT
LandGate [18]
Powered shell CapEx share of full turnkey
10–20%
LandGate [18]
Model: Shell-Only landlord rent
$28.37/kW/mo (default)
Model input; the default is what a 7.75% yield produced before the rent became an input
Model: Shell-Only yield on cost (first stabilised year, default rent)
8.2% (Standard) / 7.7% (NeoCloud)
Model output
Model: Shell-Only CapEx share (Standard, at 6MW)
~16.7% of full turnkey
Model output (build share + grid/flood/fibre ÷ turnkey CapEx)
Consistent with modelThe model's realised yield on cost
(8.2% Standard, 7.7% NeoCloud) falls inside the sourced 7.00–8.50% powered-shell development band.
That is expected at the default rent, which was set from a 7.75% yield; the check becomes independent
once a quoted rent replaces it. The independent check today is the CapEx share: Shell-Only CapEx (~16.7% of
full turnkey, with grid, flood and fibre included alongside the shell build line) falls within LandGate's
sourced 10–20% range for powered-shell development cost. It also covers the scope (site power,
fibre, building envelope) the yield benchmark is quoted against, so the CapEx share and yield band
describe the same scope (see
Assumptions §3).
7. Maintenance & staffing: how do the model's figures compare with size-scaled benchmarks?
The landlord model carries no facility staff in any case; the operator or tenant
employs them. 626 Lynnway's landlord carries only a 24/7 guard post (about $0.18M a
year). For reference, the team needed if Lynnway Park ran the facility itself is a 17-FTE Tier III
crew costing about $1.49M a year (see
Assumptions §9). Facility maintenance ($0.30M/MW/yr) applies to the landlord only in the
Full-Turnkey case. Both are checked against published benchmarks.
Tier III crew build-up, Assumptions §9 (not in the model)
Gap to reviewMaintenance is close:
the model's $300K/MW/yr sits just above the sourced $100–250K/MW/yr range [14], a
slightly conservative fit. Staffing is not close. The size-scaled industry
benchmark for a facility this small implies only 1.8–2.1 FTE[12],
against the 17-FTE reference crew for running the facility. Part of the gap is facility type: a multi-tenant colocation facility needs more
security, compliance and access-control staff than the single-tenant "enterprise" sites the benchmark is
drawn from. The multiple is still large enough to check against an actual staffing plan, though it matters
only if Lynnway Park chose to run the facility; the landlord model carries no operating staff.
Sources
Sources accessed August 2026. Dates below are publication dates
where available on the source page.
Inflect, Wholesale Colocation for AI Workloads: Power Density, Cooling, and Deployment Considerations. inflect.com, 2026-05-21.
Dgtl Infra, CoreWeave: Data Center Regions, Locations, and GPU Cloud. dgtlinfra.com, 2026-04-27.
Data Center Dynamics, Crusoe to lease Polar data center in Norway for AI cloud. datacenterdynamics.com, 2026-08-05.
BuilderMuse, Data Center Costs 2026: $1000-$2000/sqft Breakdown. buildermuse.com, 2026-04-11.
gigacapacity.com, Data Center Construction Cost per MW in 2026 (cites Turner & Townsend 2025-2026 index). gigacapacity.com, 2026-08-14.
Cortex Construct, Data Center Construction Cost | 2025 Cost Breakdown. cortexconstruct.com, 2026-04-09.
irecruit.co, Data Center Cost Per MW: 2026 Benchmarks for Owners. irecruit.co, 2026-08-16.
getcostidea.com, Data Center Construction Cost Guide [2026]. getcostidea.com, 2026-04-14.
iesmartsystems.com, Data Center Costs: From Building to Maintaining. iesmartsystems.com, 2022-12-22.
Cortex Construct, How Long Does It Take to Build a Data Center? Timeline Breakdown. cortexconstruct.com, 2026-04-05.
Hamm Institute, Data Center Employment Forecast Analysis (PDF). hamminstitute.org, 2025-12-09.
Data Center POST, Triple Net vs. Modified Gross vs. All-In Data Center Leases. datacenterpost.com, 2020-06-11.
CAE LED, Data Center Maintenance Budgeting: Cost Benchmarks, Predictive Strategies, and Modeling Tools. caeled.com, 2025-06-11.
Thunder Said Energy, Economic costs of data-centers? (secondary source, less consistent figures; not cited above). thundersaidenergy.com, 2026-08-18.
MaxLife Development, Data Center Cap Rates 2026: What Investors Pay (powered shell development Yield-on-Cost, 7.00–8.50%). maxlifedevelopment.com, 2026-06.
MaxLife Academy, Cap Rates, Lease Structures, and Pricing Data Centers (cross-references the same powered-shell Yield-on-Cost figure). maxlifedevelopment.com.
LandGate, Powered Shell Data Centers: Everything You Need to Know (scope definition; 10–20% of full turnkey cost). landgate.com, 2026-01-10.