Building a Datacenter (for Dummies) Part I

Below top-k · 33

• claimBandwidth, not compute, is the next decade's binding constraintc 0.75

  Austin Federa argues that AI, blockchain, and other high-performance systems will be gated by network capacity rather than server count — making bandwidth the new foundation layer for infrastructure scale.

• contextAI infrastructure spend is reaching scales that dwarf historical capexc 0.70

  Jensen Huang projects $3-4 trillion in AI infrastructure spend by decade's end, and Morgan Stanley forecasts $2.9 trillion in domestic datacenter spend through 2028 — annualizing to $900bn per year, roughly the entire S&P 500's 2024 capex.

• contextRack densities of 600 kW make reliable baseload power non-negotiablec 0.70

  Modern rack-scale systems pull up to 600 kW per rack, so any serious datacenter needs redundant baseload — yet Morgan Stanley projects a 60 GW domestic grid deficit by 2030.

• mechanismOn-site natural gas is the fastest BTM option but supply-chain constrainedc 0.70

  Gas-fired BTM plants cost $722-$1,677/kW and deploy faster than grid upgrades, but turbine lead times of 2-5 years and community opposition over emissions create real bottlenecks.

• mechanismSMRs offer carbon-free baseload but won't be commercially ready until post-2027c 0.70

  Small modular reactors cost $4,000-$6,000/kW upfront with $33-$45/MWh operating costs and offer scalable carbon-free baseload, but NRC approvals and safety testing push commercial deployment to the late 2020s.

• claimData center financing has shifted from hyperscaler self-funding to a deep capital stackc 0.70

  Hyperscalers historically self-funded capex from free cash flow, but PE deals in the sector nearly doubled to $107.7B over four years, and private credit now offers tailored debt for high-capex projects across the full structure.

• evidenceEquity returns hit 12-19%+ IRR while debt prices at 9-14%c 0.65

  Levered equity in datacenter development clears 12-19%+ IRR, and CoreWeave's public filings show financing rates between 9.25% and 14.12% depending on terms, with H100s sometimes pledged as collateral.

• evidenceNIMBY opposition has already blocked $64bn of datacenter projectsc 0.60

  Data Center Watch documented $64 billion of projects blocked across 28 states between May 2024 and March 2025, making local opposition a material risk in site selection.

• implicationBringing your own substation can leapfrog stressed utility infrastructurec 0.60

  Developers can pre-empt grid congestion by offering to build their own substation in the load request, though this means overcoming the transformer supply bottleneck — a gap startups like Giga are trying to fill.

• exampleTexas is emerging as the epicenter of gas-powered BTM datacentersc 0.60

  Stargate in Abilene permitted 360 MW with plans for 4,500 MW more, and Sailfish in Tolar proposed a 5,000 MW cluster with on-site gas specifically to bypass ERCOT interconnection queues.

• exampleHyperscaler SMR deals are pre-booking nuclear capacity years outc 0.60

  Equinix prepaid Oklo $25M for 500 MW of SMR power by 2030, while AWS committed to 5 GW of X-energy SMR projects by 2039 — long-dated capacity reservations placed years ahead of commercial availability.

• mechanismRenewables plus storage are cheap but intermittentc 0.60

  Solar PV runs $24-$39/MWh and onshore wind $33-$46/MWh with no fuel costs, but intermittency forces oversizing of generation plus expensive long-duration batteries that push 24/7 solar to roughly $104/MWh.

• implicationBTM economics are state-specific because ISO pricing variesc 0.60

  There's no blanket answer on whether BTM beats FOM — every Independent System Operator prices power differently, so the calculus depends on local rates and regulation, with Ohio, Texas, and Virginia emerging as BTM-friendly.

• claimThe Uptime Institute Tier system is the industry's redundancy languagec 0.60

  Tiers I through IV map cleanly to uptime guarantees — from 99.671% (Tier I, ~29 hours of downtime per year) up to 99.995% (Tier IV, under 25 minutes) — and these ratings drive how operators talk about and certify resilience.

• caveatBuilding fiber to a greenfield site is slow and expensivec 0.60

  Extending fiber from the nearest route costs roughly $100k per mile and takes 6 to 24 months, with a one-mile extension running up to $2M in capex once labor and permits are included.

• mechanismTier ratings escalate by adding independent power paths and fault tolerancec 0.55

  Tier II adds N+1 components, Tier III adds concurrently maintainable independent paths, and Tier IV requires 2N fully redundant systems with automatic failover — each step adds capex but eliminates a class of failure mode.

• implicationCarrier-neutral sites near IXPs and subsea landings win on connectivityc 0.55

  Beyond raw availability, latency to internet exchange points, proximity to subsea cable landings, and the presence of multiple carriers are what separate a viable site from a single-point-of-failure liability.

• evidenceRoughly two-thirds of installed US fiber is darkc 0.55

  A 2007 FCC report estimated about 48 million of 73.4 million kilometers of installed US fiber was dark, suggesting a massive unused inventory waiting to be aggregated and lit.

• exampleDoubleZero aggregates dark fiber into a global low-latency backbonec 0.55

  Crucible-backed DoubleZero pools dark and private fiber from contributors and uses hardware-accelerated routing to deliver 30-70% lower latency than the public internet, scaling toward 70+ links across 26 cities at mainnet.

• mechanismLand selection turns on stability, climate, and proximity to existing infrastructurec 0.50

  Suitable sites need ground stability, favorable wind and heat conditions, and proximity to fiber, with greenfield parcels avoiding demolition but still requiring perimeter security, road access, and municipal zoning approval.

• evidenceWater demand at hyperscale is staggeringc 0.50

  A 100 MW datacenter consumes 2-4 million liters per day for direct cooling, plus another 10-50 million liters per day indirectly depending on the energy mix.

• mechanismLoad studies are the gating analytical step for grid interactionc 0.50

  Before utilities will provision power, the regional ISO or RTO runs a load study to predict demand and confirm the grid can handle it reliably — these take 1-3 months and shape every downstream procurement decision.

• caveatAlways file the load request even if you go BTMc 0.50

  Even when pursuing behind-the-meter power, developers should still submit a load request to preserve future grid optionality — especially for hyperscale tenants who want both hedged inputs and the ability to sell power back.

• exampleX Supercluster blends FOM nuclear/hydro with on-site gasc 0.50

  xAI's Memphis supercluster combines 1.5 GW from TVA's nuclear and hydro assets under a multi-year PPA with on-site natural gas turbines via VoltaGrid, illustrating a hybrid FOM/BTM strategy.

• evidenceA powered shell runs $8-15M depending on tierc 0.50

  From a large developer, the bones of a Tier 3 data center cost roughly $8-12M and a Tier 4 site $11-15M — a concrete anchor for capex planning before GPUs are even ordered.

• exampleHydra Host provides turnkey GPU monetization to de-risk financingc 0.50

  Crucible partners with Hydra Host, a top-10 NVIDIA Cloud Partner, which bundles GPU management software, a sales team monetizing equipment across 50+ data centers, and cluster design — letting financing partners underwrite on real usage data rather than generic asset models.

• contextPower redundancy is achieved through layered backup systemsc 0.45

  Data centers ensure uptime through UPS units, generators, and dual utility feeds, creating multiple power paths so that utility failures or maintenance don't take critical IT offline.

• mechanismDevelopers are contracted via EPC contracts after RFPc 0.45

  Once pre-development is done, operators issue an RFP and select a developer under an engineering-procurement-construction contract covering design, permitting, supply chain, and commissioning — taking 12-24 months traditionally or 6-12 months for modular builds.

• contextCluster scale has jumped from megawatts to gigawatts within a decadec 0.40

  A large datacenter build a decade ago was 5-10 MW; today gigawatt-plus clusters are routinely announced, reflecting an order-of-magnitude shift in what counts as standard infrastructure.

• exampleHyperscalers are co-locating data centers with on-site renewables and storagec 0.40

  Google's $20B Intersect Power deal and the Stargate project both pair data centers directly with solar, wind, and battery storage for 24/7 power, signaling that co-location is becoming the template for AI-era sites.

• mechanismFiber screening combines public maps, commercial tools, and direct provider inquiriesc 0.40

  Operators triangulate connectivity using the FCC National Broadband Map, paid aggregators like FiberLocator and Rextag, and carrier-specific availability tools from Lumen, AT&T, Verizon, Zayo, and others to confirm proximity and capacity.

• mechanismFiber builds require permits, rights-of-way, and choices between aerial and buried installationc 0.40

  New fiber needs local approvals, NEPA compliance, and utility locates, after which operators choose faster/cheaper aerial poles or more reliable but disruptive buried trenching — each with distinct timeline and cost trade-offs.

• contextOBBBA bonus depreciation makes GPU ownership newly attractivec 0.35

  Post-OBBBA Act, bonus depreciation benefits make harvesting unicorn-GPU-cloud economics through ownership particularly compelling, sharpening the case for the Hydra-style turnkey model.

Redundant with selected · 2

• claimFiber is the export channel that turns flops into productc 0.75 · sim 0.83

  After power, the next gating question for any site is fiber: high-bandwidth, low-latency, multi-carrier connectivity is what lets GPU output actually reach customers.

  overlapped with: Dark fiber is where the real performance edge lives

• claimDark fiber is becoming the preferred backbone over lit fiberc 0.40 · sim 0.92

  Lit fiber is optimized for cost rather than performance, so datacenter operators increasingly prefer dark fiber for the low-latency connectivity that compute workloads demand.

  overlapped with: Dark fiber is where the real performance edge lives