Outside the carrier hotel. GPU research racks run in the 50–70 kW band (StackedAI analysis, Tier-2 conversion density roadmap citing Goldman Sachs and Dell’Oro, Aug 2026), above the 30–50 kW air-cooling ceiling of legacy halls (StackedAI analysis, Cooling Solutions engineering note, Aug 2026). The cluster belongs in a Tier-2 site priced 15–30% below primary markets (StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026), linked to the trading cabinet by fiber.
Why does a GPU cluster not belong in the carrier hotel?
Density. Carrier-hotel halls are interconnection buildings, not AI halls. Legacy halls were engineered for 5–15 kW per rack, and air cooling reaches its practical ceiling at roughly 30–50 kW (StackedAI analysis, Cooling Solutions engineering note, Aug 2026). AI training racks sit in the 50–70 kW band and rack-scale systems such as GB200 NVL72 draw 120–130 kW (StackedAI analysis, Tier-2 conversion density roadmap citing Goldman Sachs and Dell’Oro, Aug 2026). A research cluster in the hub will be refused the density, charged a premium for reserved cooling and circuit capacity, or spread thinly across many cabinets.
Cooling technology. The liquid-cooled share of AI servers rose from 15% in 2024 to 54% in 2025 and is expected to reach about 76% in 2026 (StackedAI analysis, Tier-2 conversion density roadmap citing Goldman Sachs and Dell’Oro, Aug 2026). Direct-to-chip cooling supports 60–130 kW or more per rack but needs coolant distribution units, facility piping and a chilled-water plant, at a capex premium of about $2,500–4,500 per kW over air (StackedAI analysis, Cooling Solutions engineering note, Aug 2026), and a liquid retrofit runs $1.5–1.6 million per MW (Data Center Knowledge, Neocloud Storm Gathers as Data Center Deals Stall, Apr 2026). No carrier-hotel operator will install that for one tenant’s research cluster.
Price. Carrier-hotel power is the most expensive the firm buys: $200–240 per kW per month in the NYC Tri-State market with under 1% vacancy (StackedAI analysis, carrier-hotel market analysis, Apr 2026). Paying that for a workload that need not be near a matching engine is a pure loss, and it consumes cabinet capacity the firm may later need for execution; see the low-latency colocation cost page.
Where should the research cluster go instead?
A Tier-2 market site that already supports liquid cooling, or has the electrical and chilled-water headroom to add it, in a low-cost utility territory. Tier-2 colocation prices 15–30% below primary averages (StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026), and utility cost is the largest operating line for a cluster running near full load.
| Placement option | Rack density supported | Power cost benchmark | Fit for research clusters |
|---|---|---|---|
| Carrier hotel (Tri-State, Chicago) | Legacy halls 5–15 kW/rack; air ceiling ~30–50 kW (StackedAI analysis, Cooling Solutions engineering note, Aug 2026) | $200–240/kW/mo, Tri-State (StackedAI analysis, carrier-hotel market analysis, Apr 2026) | Execution and market data capture only |
| Primary-market AI colocation (e.g. Northern Virginia) | Liquid-cooled halls at the AI band | Utility ~9.25 cents/kWh (StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026); vacancy 0.3% (CBRE, Global Data Center Trends 2026, Jun 2026) | Workable but no price advantage and little availability |
| Tier-2 liquid-ready site (e.g. Dallas-Fort Worth, Austin, Salt Lake City, Columbus) | Liquid-cooled halls at the AI band and above | Colocation 15–30% below primary; utility DFW ~7.15, Austin ~7–8, SLC ~7.77, Columbus ~8.5–9.5 cents/kWh (StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026) | Preferred: density, price and availability align |
| Public cloud GPU rental | Not applicable | H100 on-demand median $4.19/GPU-hour on dedicated clouds (GetDeploying, GPU Price Index, Aug 2026); one-year contracts $2.35 (Akash Network, H100 Rental Price 2026, Aug 2026) | Suits bursty research; sets the ceiling on what an owned cluster is worth per hour |
Two qualifications. The utility figures are territory benchmarks, not a site’s tariff; operators pass power through at their own rate plus a margin, and StackedAI benchmarks the effective rate per engagement. And “liquid-ready” means the coolant distribution, piping and chilled-water headroom exist today; a site that must be retrofitted adds 12–18 months of conversion time (StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026). The liquid cooling retrofit page and the Tier-2 brownfield conversion pillar cover what makes a site convertible.
How is the research site linked to the trading cabinet?
By dedicated fiber or wavelength circuits from the Tier-2 site into the carrier hotel, terminating on a cross-connect in the execution cabinet. The traffic is batch-oriented: raw market data is captured at the hub and shipped out in bulk; models are trained and backtested at the cluster, and the resulting parameters are shipped back on a schedule.
Latency across that link is not a constraint. Training and backtesting are batch jobs whose runtime is measured in hours; a wide-area circuit adding milliseconds of round-trip time is invisible to them, whereas the execution path is engineered in microseconds. The design rule: nothing on the execution path crosses the inter-site link, and nothing that tolerates milliseconds sits in the hub.
The exception is live inference that feeds execution: a model scoring market data in real time, with its output going into an order, belongs in the hub, sized to a standard cabinet allocation. That is a server or two, not a cluster.
What is the placement checklist for a research cluster?
- Separate the workloads. Classify every server as on the execution path, feeding it in real time, or batch; only the first two go in the hub.
- Size the density for the target hardware generation and the next: AI band 50–70 kW today, rack-scale systems 120–130 kW (StackedAI analysis, Tier-2 conversion density roadmap citing Goldman Sachs and Dell’Oro, Aug 2026).
- Confirm liquid cooling is installed, not planned: coolant distribution capacity, facility loop and chilled-water headroom.
- Benchmark the effective power rate against the territory figure (StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026), and compare the owned cluster’s cost per GPU-hour against rental ($4.19 median on demand, GetDeploying, GPU Price Index, Aug 2026), with depreciation covered on the GPU depreciation page.
- Secure the link. Contract fiber or wavelength capacity to the hub with a cross-connect landing in the execution cabinet, on diverse paths, and negotiate density step-ups and adjacent rack rights at the Tier-2 site.
- Keep the hub lean. Audit the carrier-hotel cabinet after the move so nothing batch-oriented is still drawing hub power at $200–240 per kW per month (StackedAI analysis, carrier-hotel market analysis, Apr 2026).
Key terms
- Execution path: the servers and network segments between market data arrival and order transmission, engineered in microseconds.
- Rack density: the electrical load per rack in kW, which determines the cooling technology a hall needs.
- Direct-to-chip liquid cooling: coolant circulated through cold plates on processors, supporting 60–130 kW or more per rack (StackedAI analysis, Cooling Solutions engineering note, Aug 2026).
- Tier-2 market: a secondary data center market priced 15–30% below primary averages (StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026).
How StackedAI applies this
StackedAI designs the two-site split for trading firms: a lean execution footprint in the carrier hotel and a liquid-cooled research site in a Tier-2 market, joined by contracted fiber. The firm screens sites for installed liquid-cooling capacity and electrical headroom, benchmarks effective power rates, and models the own-versus-rent decision for the GPU fleet. Placement engagements are described on the advisory services page; the hub map is in the pillar on colocation for trading firms in carrier hotels.
Frequently asked questions
Can a quant fund run GPU inference in a carrier hotel?
A small inference footprint that feeds live signals can sit in the hub if it fits a standard cabinet’s power allocation. A training or backtesting cluster cannot: AI racks run at 50–70 kW while legacy halls were built for 5–15 kW per rack (StackedAI analysis, Aug 2026), and carrier-hotel power is the most expensive the firm buys.
What rack density does a GPU research cluster need?
The AI band is 50–70 kW per rack, with rack-scale systems such as GB200 NVL72 at 120–130 kW (StackedAI analysis citing Goldman Sachs and Dell’Oro, Aug 2026). Air cooling tops out around 30–50 kW per rack, so the AI band needs direct-to-chip liquid cooling, a facility decision rather than a cabinet decision.
Where is the cheapest place to put research compute?
A Tier-2 market site priced 15–30% below primary averages, in a low-cost utility territory: about 7.15 cents per kWh in Dallas-Fort Worth, 7 to 8 in Austin and 7.77 in Salt Lake City, against about 9.25 in Northern Virginia (StackedAI analysis, Aug 2026). The site must already support liquid cooling or have the headroom to add it.
How is the research site linked to the trading cabinet?
By dedicated fiber or wavelength circuits from the Tier-2 site into the carrier hotel, terminating on a cross-connect in the execution cabinet. Market data captured at the hub flows out; trained model parameters flow back. Training and backtesting are batch jobs that tolerate wide-area latency measured in milliseconds, so distance is not a constraint.
What is the biggest mistake firms make with research clusters?
Buying hardware before securing a site that can cool it. The liquid-cooled share of AI servers is expected to reach about 76% in 2026 (StackedAI analysis citing Goldman Sachs, Aug 2026), and a liquid retrofit runs $1.5–1.6 million per MW, 20–25% above air-cooled build (Data Center Knowledge, Apr 2026). Site selection must precede the hardware order.
Sources
- StackedAI analysis, Tier-2 conversion density roadmap citing Goldman Sachs and Dell’Oro, Aug 2026 (internal; rack density bands, average rack density, liquid-cooled server share)
- StackedAI analysis, Cooling Solutions engineering note, Aug 2026 (internal; legacy hall density, air-cooling ceiling, direct-to-chip range and capex premium)
- StackedAI analysis, Tier-2 brownfield conversion thesis, Aug 2026 (internal; Tier-2 pricing discount, utility rates, conversion timeline)
- StackedAI analysis, carrier-hotel market analysis, Apr 2026 (internal; Tri-State carrier-hotel pricing and vacancy)
- Data Center Knowledge, Neocloud Storm Gathers as Data Center Deals Stall Over Credit Risk, Apr 2026, https://www.datacenterknowledge.com/cloud/neocloud-storm-gathers-as-data-center-deals-stall-over-credit-risk
- CBRE, Global Data Center Trends 2026, Jun 2026, https://www.cbre.com/insights/reports/global-data-center-trends-2026
- GetDeploying, GPU Price Index, Aug 2026, https://getdeploying.com/gpu-price-index
- Akash Network, H100 Rental Price 2026, Aug 2026, https://akash.network/the-bid/h100-rental-price-2026-cost-per-hour/