ai/ml data center projects analysis

overview

ai/ml workloads are driving the most significant transformation in data center infrastructure history. 140 ai/ml projects represent 23.2% of the us data center database but account for disproportionate investment and power demand due to fundamentally different infrastructure requirements.

key statistics

why ai/ml is different

traditional data centers operate at 15-20 kw per rack. ai/ml infrastructure requires:

• power density: 100-140 kw per rack (5-10x increase)
• cooling: liquid cooling mandatory (direct-to-chip or immersion)
• networking: 400-800 gb/s gpu interconnects (nvidia quantum-2 infiniband, spectrum-x ethernet)
• scale: gigawatt-scale projects increasingly common
• speed: compressed construction timelines (18-24 months typical)

leading ai deployments

these projects represent the cutting edge of ai infrastructure, demonstrating the scale and technical requirements driving the industry:

technology requirements

gpu specifications

power density evolution

cooling technology

direct-to-chip liquid cooling (dominant):

• cold plates attached directly to gpu/cpu
• 70-98% heat capture at source
• vendors: vertiv (cdu 70-2300kw), supermicro (dlc-2), lenovo (neptune), dell, hpe cray
• coreweave: all new facilities liquid cooling foundation (130kw+ racks)
• lambda labs: dallas-fort worth facility liquid-cooled for highest-density gpus

immersion cooling (emerging):

• servers submerged in dielectric fluid
• single-phase or two-phase systems
• vendors: grc, liquidstack, asperitas
• best for 80-100kw+ rack densities
• market: $4.87b (2025) →$11.10b (2030), 17.91% cagr

networking requirements:

• nvidia quantum-2 infiniband: 400gb/s (ndr), up to 51.2tb/s aggregate
  • deployments: coreweave (3.2tbps per vm), lambda labs, meta 24k clusters, azure nd h100 v5
• nvidia spectrum-x ethernet: 800gb/s with rdma
  • deployment: xai colossus (100,000 h100 on single rdma fabric)
• nvlink 4.0: 900gb/s per gpu (3.6tb/s bisectional in 8-gpu systems)

geographic distribution

top 10 states for ai/ml projects

clustering patterns

texas dominance (10 projects, 11.0 gw total state capacity):

• deregulated power market enables flexibility
• oncor interconnection queue: 186 gw pending
• major projects: stargate abilene ($40b, 1.2 gw), crusoe abilene ($15b, 1.2 gw), lambda plano, coreweave austin/plano

pennsylvania emergence (8 projects, 16.9 gw total state capacity):

• marcellus shale natural gas advantage
• major project: coreweave lancaster ($6b, 100 mw scalable to 300 mw)
• homer city energy campus: 4.5 gw planned

emerging ai hubs:

• north dakota: applied digital ellendale (180 mw, 50k gpu capacity), coreweave lease ($7b)
• tennessee: xai colossus memphis (300 mw, 230k gpus operational)
• ohio: meta prometheus new albany (1,020 mw, 500k+ gpus, 2026)
• arizona: vermaland la osa (3 gw, $33b), tract buckeye (1.8 gw,$20b)

project status distribution

key insight: 67.9% of ai/ml projects are under construction or planned, indicating massive near-term capacity additions (2025-2027).

top sponsors

analysis: why ai drives different infrastructure

training vs inference workloads

training (high compute intensity):

• massive parallel gpu clusters (10k-500k+ gpus)
• high-bandwidth networking critical (400-800gb/s gpu interconnects)
• power density: 100-140kw/rack
• examples: meta prometheus (500k gpus), xai colossus (230k gpus)
• cooling: liquid cooling mandatory
• location: near power sources, less latency-sensitive

inference (lower per-request compute, high throughput):

• distributed across multiple sites
• latency-sensitive (user-facing)
• lower power density possible but still 3-5x traditional
• examples: openai api infrastructure, google tpu v7 ironwood (inference-optimized)
• cooling: direct liquid or advanced air
• location: near users/edge locations

the gpu density challenge

rack-scale systems (nvidia nvl72 example):

• 72 gpus + 36 cpus per rack
• 120-140kw power per rack
• 13.5-21tb gpu memory per rack
• must deploy in multiples of 18 nodes
• requires liquid cooling infrastructure

cluster-scale challenges:

• 100,000 gpu cluster (xai colossus):
  • 1,389 racks of 72 gpus each
  • ~173 mw for gpu compute alone
  • single rdma fabric connectivity
  • 122 days construction time
• power delivery at this scale requires utility partnerships or on-site generation

geographic clustering drivers

power availability (primary constraint):

• texas: deregulated market, 186 gw queue
• pennsylvania: marcellus shale gas, 16.9 gw capacity
• ohio: competitive market, ample generation
• north dakota: low-cost power, minimal constraints

alternative power sources:

• on-site natural gas: 9 of 11 gigawatt projects
• nuclear partnerships: 24+ gw smr commitments (microsoft-constellation, amazon-x-energy, google-kairos, switch-oklo)
• renewable + battery: xai colossus (150 mw megapack battery backup)
• hydrogen: lambda terrasite-tx1 (1 gw campus, hydrogen-powered)

construction speed:

• xai colossus: 100,000 gpus in 122 days
• standard timeline: 18-24 months for ai facilities
• traditional hyperscale: 24-36 months

investment patterns

hyperscaler build-out (vertical integration):

• microsoft: 10 ai/ml projects
• google: 13 ai/ml projects
• meta: 8 ai/ml projects (including 1 gw prometheus)
• amazon: 4 ai/ml projects

specialized ai infrastructure operators:

• coreweave: 250k gpu fleet, 2.2 gw contracted capacity
• crusoe: 1.6 gw operational/construction, 10+ gw pipeline
• applied digital: ellendale hpc campus
• lambda labs: gigawatt-scale gpu cloud

financial sponsors entering ai:

• nvidia: $8.13b portfolio (applied digital, coreweave investments)
• blackstone: airtrunk $16b acquisition
• digitalbridge: 5.4 gw portfolio
• blue owl capital: crusoe $15b joint venture

project economics:

• average ai/ml project: $6.9b investment
• average power: 717 mw per project
• contrast traditional: $2-3b average, 100-200 mw typical
• premium pricing: ai colocation commands 2-3x traditional rates

complete ai/ml projects table

mega-projects (≥$10b or ≥1gw)

data access

full project list: all 140 ai/ml projects available in database

• location: support/datacenters/analysis/ai_ml_projects_simplified.json
• includes: project name, state, city, sponsor, power capacity, investment, status
• sortable by investment, power, sponsor, status, geography

dimensional research:

• gpu inventory: 1m+ deployed gpus tracked
• technology specs: h100/h200/b200/b300/gb200/gb300
• cooling vendors: vertiv, supermicro, grc, liquidstack, lenovo, dell, hpe
• networking: quantum-2 infiniband, spectrum-x ethernet specifications
• operators: coreweave, lambda, crusoe, applied digital, hyperscalers

future outlook (2025-2030)

projected growth

• ai/ml share: 23.2% (2025) → 40-50% (2030)
• average project size: $6.9b (2025) →$10-15b (2030)
• typical power: 717 mw average → 1-2 gw standard
• gpu clusters: 100k-500k gpus → 500k-2m gpus
• cooling: liquid cooling universal → immersion increasing for 300+ kw racks

technology evolution

gpu generations:

• current: h100/h200 (80-141gb, 350-700w)
• 2025-2026: b200/b300 (180-288gb, 600-1,400w)
• 2027+: next-gen blackwell, 2x performance/watt improvements
• rack systems: gb200/gb300 nvl72 (120-140kw) becoming standard

cooling infrastructure:

• direct-to-chip liquid cooling: dominant for 100-140kw racks
• immersion cooling: growing for 300+ kw racks
• hybrid systems: liquid for gpus/cpus, air for ancillary
• vendors scaling: vertiv cdu 2300 (2.3mw capacity), supermicro in-row cdu (1.8mw)

networking bandwidth:

• 2025: 400gb/s quantum-2, 800gb/s spectrum-x
• 2026-2027: 800gb/s, 1.6tb/s standards
• 2028+: 3.2tb/s and beyond

power solutions

grid constraints driving alternatives:

• on-site generation: natural gas increasingly common (9 of 11 gw+ projects)
• nuclear smrs: 24+ gw committed (2027-2035 deployment)
• battery integration: tesla megapack (xai 150mw backup)
• hydrogen: lambda terrasite-tx1 (1 gw hydrogen-powered)

interconnection queue crisis:

• oncor (texas): 186 gw queue
• pjm (mid-atlantic): 270+ gw queue
• multi-year delays standard
• forcing geographic diversification to less constrained markets

market dynamics

consolidation likely:

• specialized ai infrastructure operators (coreweave, crusoe, lambda) gaining share
• hyperscalers building captive capacity (meta prometheus, google project mica)
• traditional colocation adapting (equinix 100 liquid-cooled dcs, cyrusone intelliscale)
• financial sponsors backing buildout (nvidia, blackstone, digitalbridge, kkr)

pricing premium:

• ai colocation: $200-300/kw/month (vs$100-150 traditional)
• gpu cloud: premium for h100/h200 vs commodity compute
• long-term contracts: 3-7 year commitments typical
• pre-leasing common: projects leased before construction

regulatory evolution:

• permitting: ai-specific review processes emerging
• power allocation: utility commissions prioritizing economic development
• environmental: renewable energy requirements increasing
• community impact: noise, water use, thermal pollution concerns

using this analysis

for infrastructure operators

• capacity planning: 717 mw average project size, 100-140kw rack density
• technology roadmap: liquid cooling mandatory, gb200/gb300 deployment timeline
• geographic strategy: cluster analysis shows power availability primary driver
• competitive intelligence: 140 projects tracked with investment, sponsor, status

for investors

• market size: $538b disclosed investment across 140 projects
• growth trajectory: 67.9% projects under construction/planned (near-term capacity)
• key players: hyperscalers (microsoft, google, meta, amazon) + specialists (coreweave, crusoe, lambda)
• deal flow: mega-projects ($10b+) increasingly common

for technology vendors

• cooling demand: liquid cooling mandatory for 140 projects (vertiv, supermicro, grc, liquidstack)
• gpu sales: 1m+ deployed, 500k-2m+ needed by 2030
• networking: 400-800gb/s interconnects (nvidia quantum-2, spectrum-x)
• power infrastructure: smr partnerships, battery storage, on-site generation

for policymakers

• economic development: average $6.9b investment, 717 mw power per project
• infrastructure planning: 65.2 gw disclosed capacity from ai/ml alone
• grid impact: concentrated demand requiring utility coordination
• incentive design: competition for mega-projects intense

related resources

• gpu infrastructure - detailed gpu specifications and deployment tracking
• liquid cooling - cooling technology vendors and requirements
• power infrastructure - nuclear partnerships, grid constraints, alternative solutions
• project jupiter/stargate - $165b new mexico mega-project analysis
• state rankings - geographic distribution and competitive analysis
• mega-projects - projects >$10b or >1gw

---

ai/ml infrastructure represents the most significant data center buildout in history. the 140 projects tracked here demonstrate fundamentally different requirements: 5-10x power density, mandatory liquid cooling, gigawatt-scale projects, and compressed timelines. success requires solving the power delivery challenge through grid upgrades, on-site generation, or nuclear partnerships. the next 5 years (2025-2030) will determine which regions and operators can deliver the infrastructure needed for ai leadership.