← All docs

The bigger picture — Project BOSC and the data-center boom

Project BOSC is one campus in a national build-out. This page places it against the broader pattern: what its disclosed plant implies about compute capacity and workloads, whether its water problem generalizes across the Maumee basin, and where Lima is typical of the boom versus an outlier.

Evidence discipline as elsewhere: [verified] (read from a cited record or live connector), [inference] (a labelled reading), [open] (a question, not a finding).

1. Compute capacity & workloads

The disclosed plant. The Ohio EPA air permit (P0138965) lists 114 backup generators at ~2,750 ekW → ~313 MW of backup, and the design carries 36 cooling towers. [verified: data/extracted/permits/3987141.epa.yaml] Backup is sized N+1 to the critical IT load, so the campus IT load is on the order of ~275 MW. [inference]

What ~275 MW buys. Rack power density is the swing factor. At the conventional 5–15 kW/rack the data appendix cites, 275 MW of IT is on the order of ~18,000– 55,000 racks; at the 40–140 kW/rack of dense AI/accelerator halls, it is ~2,000–7,000 racks. [inference: data appendix rack-density bands] The 36 cooling towers and the 1.8–1.9 L/kWh WUE put the consumptive draw at 3.1–10 MGD — the hydrology headline. [verified via HYDROLOGY]

Which workloads — the GovCloud question, substantially answered. Asked directly about the campus, Liz Schwab answered in indirect language — “classification levels” — which is the tell: this is regulated / classified-data compute (GovCloud-style), not generic commercial hosting. GovCloud capacity carries a ~20– 30% premium over commercial, and the dedicated, hardened, anti-ram / security-fenced site plan is consistent with that tier rather than a cost-optimized commercial build. [verified: data appendix; site plan] / [inference: workload class] The end user is Google (Dossier §1); the workload class is the regulated-compute tier its government cloud serves. [inference]

2. The Maumee watershed comparison — does the dilution thesis generalize?

Lima’s finding. At design low flow, the county plants discharge more than the entire 7Q10 of their receiving streams — effluent effectively undiluted (Hydrology; water balance). The distinctive cause is what they discharge into: tiny tributaries — Dug Run, Pike Run — not a mainstem.

Basin context (EPA ECHO). The committed Maumee NPDES inventory holds 129 POTWs. Ranked by design flow: [verified: data/reference/echo/maumee-wwtp.potw.yaml]

Does Fort Wayne mirror Lima? Partly, and instructively. Fort Wayne is the basin’s largest discharger (74 MGD, ~4× Lima) and ECHO flags it on an effluent monthly-average limit — so even the big mainstem plant runs against its permit. [verified: ECHO compliance_status] But Fort Wayne discharges to Baldwin Ditch → the Maumee mainstem, which carries far more dilution than Lima’s under-flow tributaries. So Lima is distinctive not in size but in receiving-water choice: small county plants on intermittent tributaries, where 7Q10 ≈ the discharge itself.

The shared constraint. All 129 sit under the 2023 Maumee Watershed Nutrient TMDL’s binding phosphorus cap (Lake Erie’s largest tributary; the basin future-growth reserve is only ~1.4–1.5 mt P/spring). [verified: data/reference/hydrology/maumee-tmdl-budget.yaml] So a new data-center sanitary load enters a fully-allocated basin regardless of which plant takes it.

Extended basin-wide [v]. The per-plant 7Q10 dilution screen now runs over the full ECHO POTW inventory (watermark basin-screen), not just Lima’s plants. The denominators are the cited fact-sheet 7Q10s plus a derived LP3 7Q10 for the four major USGS-gaged mainstems (Maumee, Auglaize, St. Marys, St. Joseph; data/reference/hydrology/low-flow-7q10.derived.yaml). The honest result: only 8 of 129 POTWs are screenable — 76 have no receiving water in ECHO, 44 discharge to ungaged tributaries/ditches (reported “no 7Q10” rather than screened against a downstream river’s larger flow), and 1 has no design flow on record. Of the screenable, two are violations — Lima’s American Bath → Pike Run (0.01:1) and Van Wert → Town Creek (0.03:1) — while Decatur → St. Marys (3.1:1) and Defiance → Maumee (6.2:1) screen “tight.” (The Defiance 6.2:1 uses the basin-wide Waterville proxy; the reach-specific denominator — USGS 04192500, Maumee near Defiance, LP3 7Q10 139.24 cfs — gives 7.5:1, still “tight” and slightly looser than the proxy, since the near-Defiance mainstem gage already includes the Auglaize and reads higher than Waterville ~50 mi downstream; #391.) The wide data gap is itself the finding — Lima’s plants are unusually well-documented, and a basin-wide answer needs each tributary’s own cited/gaged 7Q10.

The network view [v]. The screen is one dimension of a wider cross-site comparison: see The BOSC network for the full scorecard — the eight onboarded watershed points as nested nodes on one connected basin (all draining through Defiance → Toledo to Lake Erie under one TMDL cap), compared across receiving-water regime, grid, economy, and toxics. The headline holds across every dimension: Lima is distinctive only in its receiving-water choice — its economic shape (manufacturing-heavy, information-sector-absent) and its place in the basin are typical of the network.

3. Boom patterns — where Lima is typical, where it is the outlier

Typical of the boom [inference]:

• Secrecy architecture — a Delaware shell counterparty, NDA-by-default, and an economic-development disclosure exemption (ORC §9.66(D)). The thin record is a pattern, not a Lima quirk (Dossier §1).
• Incentive shape — a 15-yr/75% property-tax abatement and publicly-financed roadwork for a multi-hundred-MW load and ~50 jobs: the load-not-jobs subsidy the Economics page anatomizes is the boom’s standard trade.
• Demand drivers — GovCloud premium, the AI rack-density jump, and a sales-tax exemption on fast-refreshed hardware are national, not local.

Where Lima is the outlier [inference]:

• Water — siting hundreds of MW of evaporative cooling on a basin whose design low flow is zero (Ottawa 1Q10 = 0 cfs) and whose tributaries are already undiluted is a sharper water constraint than most boom sites face.
• The local economy — the localized baseline shows a county whose employment shrank 2.5% (2018→2023), concentrated in manufacturing (location quotient 2.08) with a near-absent information sector (LQ 0.37). The campus lands a regulated-compute use onto a shrinking industrial base — not onto an existing tech cluster. [verified: BLS QCEW]

These tracks deepen as the basin 7Q10 ingest and the capacity model mature; the Research course §1.5 carries them forward.