KI-Rechenzentren

AI has split the data center industry into two products: legacy space built for 5-15 kW racks, and AI-grade capacity delivering 60-130+ kW per rack with liquid cooling — and the second category is being absorbed as fast as it is built. This hub covers the density and cooling economics, the rent-vs-colocate decision, and how to find GPU-ready capacity, backed by our live GPU price tracker and aiReady-tagged facility catalog.

Market context: AI is eating the capacity pipeline

The scale numbers are unambiguous. McKinsey projects global data center capacity demand of roughly 219 GW by 2030, with AI-ready capacity required to grow about 33% annually and AI workloads representing from half to roughly 70% of total demand by decade’s end. That demand is landing on a market already at record tightness: Northern Virginia vacancy near 0.5%, roughly three-quarters of US under-construction capacity preleased, and North American wholesale rates at a record $196.25/kW/month per CBRE — with operators of AI-optimized, liquid-cooled facilities capturing explicit rent premiums over conventional space.

On the hardware side, the market has matured fast. H100 rental pricing collapsed from $8+/hour at the 2023 peak to roughly $2-3/hour spot and ~$2.35/hour on 1-year committed contracts as supply normalized and B200/GB200 generations shipped — a repricing that transformed the build-vs-rent math for everyone below hyperscale size. Tracking that curve daily is precisely what our GPU tracker does; our H100 rental price guide explains the contract structures behind the headline rates.

The consequence for buyers: AI infrastructure decisions are now two-market decisions. You are simultaneously pricing physical capacity (colocation, $/kW) and compute (GPU rental, $/GPU-hour), and the optimal answer shifts with utilization, duration and hardware generation. Most advisory sources cover one market. We track both.

Density and cooling: why 100 kW racks changed everything

A modern AI training rack is a different physical object from what data centers were built for. An NVIDIA GB200 NVL72 system draws roughly 120-132 kW in a single rack — versus the 5-15 kW enterprise standard — with cluster architectures pushing toward 250 kW+ per rack on upcoming generations. Three engineering consequences cascade from that number:

Cooling becomes liquid. Above ~50 kW/rack, moving enough air is physically impractical. The industry has standardized on two approaches:

	Direct-to-chip (DTC)	Immersion
How it works	Cold plates on GPUs/CPUs; coolant loop via CDU	Servers submerged in dielectric fluid
Rack density supported	80-132+ kW (current gen)	100-250+ kW
Heat captured in liquid	~70-80% (air handles remainder)	~95-100%
Facility retrofit burden	Moderate — CDUs, piping, water loop	High — tanks, fluid handling, new ops model
Hardware compatibility	Standard OEM racks (NVL72 ships DTC-ready)	Requires immersion-adapted hardware
2026 market position	Dominant choice for new AI builds	Niche: HPC, crypto, extreme density

DTC wins most deployments because it preserves the rack-and-row operating model while handling current GPU thermals; it typically adds 7-10% to mechanical/electrical capex but cuts cooling energy enough to bring PUE from a legacy ~1.5 to 1.1-1.2 — a permanent 20-25% saving on total power cost.

Power distribution scales up. 60-130 kW racks need overhead busway at 415V+, high-amperage feeds, and floor loading rated for 1,500-2,000 kg per rack. Utility headroom becomes strategic: a 1,024-GPU H100 cluster draws roughly 1.5-2 MW all-in; a 16,000-GPU training build approaches 25-30 MW — sizes that instantly make you a wholesale tenant negotiating power, not a retail buyer picking racks. For requirements at that scale, see our hyperscale hub and power hub.

The screening implication: most of the world’s installed colocation base cannot host modern AI clusters without retrofit. Retrofitting legacy air-cooled space for liquid-cooled density costs $2-4M per MW. This is why “GPU-ready” is now the single most important filter in facility selection — and why our catalog tags facilities with an aiReady flag covering density, liquid cooling support and power headroom.

Colocate or rent GPUs: the decision framework

The most consequential AI infrastructure decision under ~$50M scale is whether to own hardware in colocation or rent GPU capacity. The honest answer is arithmetic, not ideology:

Rent when: utilization is uncertain or spiky; the project horizon is under ~12-18 months; you need frontier hardware immediately (B200/GB200 rental beats a 6-month procurement cycle); or team bandwidth for infrastructure operations is zero. At ~$2.35/hour committed, an H100 costs about $20,600/GPU-year with power, cooling, networking, and hardware risk entirely on the provider.

Colocate when: you will sustain high utilization (60%+) for 2+ years; workload data has residency or security constraints; or you are large enough that rental margins exceed your operating costs. Owning an H100 node runs roughly $25-30k/GPU in capex plus $3-5k/GPU-year in colocation power and space (at ~$0.08-0.12/kWh markets — one reason Southeast Asian colocation is attractive for inference fleets). Over three years of hard utilization, ownership typically lands 30-50% below rental — if utilization holds.

The breakeven sits around 12-18 months of sustained use. Two factors push it later: GPU generational turnover (an owned H100 fleet competes with next-gen rentals in year two) and falling rental prices (the market repriced ~60-70% downward in two years). One factor pulls it earlier: rental capacity for the newest generation is scarce and premium-priced at launch, exactly when training demand peaks.

Sophisticated teams increasingly split the difference: owned, colocated capacity sized to baseline inference load; rented capacity for training bursts and generation experiments. Pricing both sides of that portfolio requires live data on both markets — which is the core of what this site publishes at /index/ and /gpu/.

The GPU market has stratified — price the generation, not “a GPU”

“GPU pricing” is now four distinct markets moving on different curves, and conflating them is the most common — and most expensive — analytical error we see:

Tier	Hardware	Market behavior (2026)
Frontier	GB200/B200 class	Scarce, premium-priced, allocated more than sold; multiples of H100 rates
Workhorse	H100/H200	Liquid, competitive; ~$2-3/hr spot, ~$2.35/hr on 1-year commitments
Value	A100 and prior	Deeply discounted; viable for fine-tuning, smaller models, batch inference
Inference-optimized	L40S and specialized parts	Priced per served token economics, not training FLOPs

Three dynamics govern the curves. Generational cascade: each new flagship launch pushes the prior generation down 30-50% within roughly a year — the pattern that took H100s from $8+/hour to $2.35 committed. Contract-term spread: spot, monthly, and 1-3 year committed rates for identical hardware can differ 40%+, so matching contract length to workload certainty is worth as much as provider selection. Provider dispersion: hyperscaler list prices for the same GPU run 2-4x above specialized GPU cloud rates, with the gap widest at the workhorse tier. Our GPU tracker maintains the current picture across all tiers; treat any single quoted number — including ours from last quarter — as a snapshot of a fast-moving curve.

Networking: the third pillar buyers discover too late

AI clusters fail on networking more often than on power or cooling, because training is a synchronized workload: thousands of GPUs exchange gradients every step, and the slowest link paces the entire cluster. Practical implications for capacity buyers:

• East-west dominates. A 1,024-GPU cluster needs a non-blocking 400/800G fabric (InfiniBand or RoCE Ethernet) whose cost — often $2,000-4,000 per GPU in switching and optics — must sit in your capex model alongside the accelerators themselves.
• Physical locality is non-negotiable for training. Fabric latency budgets keep training clusters within a hall or adjacent halls; you cannot stitch a training cluster across two buildings, let alone two facilities. This constrains facility choice to sites with large contiguous GPU-ready space — another dimension our catalog’s aiReady screening addresses.
• Inference distributes; training concentrates. Serving traffic tolerates ordinary metro latency, which is what makes hub-and-spoke architectures work — heavyweight clusters in wholesale space, inference nodes closer to users (see our edge hub).

When comparing colocation offers for GPU deployments, ask specifically about structured cabling support for high-fiber-count fabrics, cross-connect pricing at volume (dozens of 400G links change the math — see the hidden-cost section of our colocation hub), and cloud on-ramps if your architecture is hybrid.

Where to put AI capacity: the geography is shifting

AI workloads are more latency-tolerant than user-facing applications — a training job doesn’t care if it runs in Ashburn or Johor. That tolerance, against sub-1% vacancy and $215+/kW pricing in core US markets, is redirecting AI capacity toward power-rich, lower-cost markets: US secondary metros, the Nordics, the Middle East, and above all Southeast Asia, where Thailand’s BOI approved ~THB 746 billion of data center investment in 2025, Johor’s pipeline reached roughly 4-5 GW, and Malaysia now explicitly prioritizes AI-grade projects. Industrial power at $0.08-0.12/kWh and build costs 20-40% below US levels translate directly into lower $/GPU-hour for anyone running fleets at scale — the full regional picture is in our development hub and Thailand analysis.

How Coloprice helps: the colo-to-GPU bridge

Every other resource covers half of this market. Colocation directories don’t track GPU pricing; GPU marketplaces don’t cover physical capacity. Since the actual decision — rent compute versus deploy hardware versus something in between — spans both, we built the bridge:

• Live GPU pricing, free. The GPU tracker monitors rental rates across providers and generations — H100 from ~$2.35/hour committed through current B200/GB200 premiums — so you can benchmark any provider quote in seconds. The H100 pricing guide decodes contract structures.
• GPU-ready facilities, filterable. Our 202-facility catalog carries an aiReady filter covering high-density power and liquid cooling support — turning “which of the world’s facilities can actually host 100 kW racks?” from weeks of RFI emails into a two-minute query.
• Both markets on one benchmark. Colocation rates and GPU rates side by side, plus market supply data on stats — the complete input set for the rent-vs-colocate model above.
• Matched quotes in one business day, free. Describe your requirement at /quote/ — a 20-rack liquid-cooled cluster, 2 MW of GPU-ready space in Southeast Asia, or a GPU rental commitment you want benchmarked — and we return matched options from facilities and providers we’ve vetted within one business day. No listing fees shaping the shortlist, no “request a quote” black box.
• Deep-dives. Start with the colocation pricing guide for the $/kW side and the investment guide if you’re funding AI capacity rather than buying it.

AI infrastructure is the fastest-repricing market in technology — GPU rates moved 60%+ in two years while colocation moved 6-12% the other direction. Decisions made on stale data are expensive in both directions: overpaying for compute on one side, stranding capital in the wrong facility on the other. Check /gpu/, screen /data-centers/, benchmark against /index/, and let /quote/ put real, comparable offers on your desk within one business day.