트렌드
구축 현황의 이면에 있는 수치: 자본적 지출, 전력 수요, 랙 밀도, 캠퍼스 규모, 학습 컴퓨트, 원자력 전환.
Hyperscaler capital expenditure
USD billions
The four largest cloud and AI platform companies have roughly tripled their combined capital spending in three years, with the overwhelming share going to AI data centers, accelerators and power. 2025 figures are company guidance, not final actuals.
| 2023 | 2024 | 2025 (g) | |
|---|---|---|---|
| Microsoft (FY) | 32.2 | 55.7 | 88.2 |
| Alphabet | 32.3 | 52.5 | 92.0 |
| Amazon | 52.7 | 83.0 | 125.0 |
| Meta | 28.1 | 39.2 | 71.0 |
출처 및 근거: Company 10-K/10-Q filings and stated guidance, as of late 2025. Microsoft figures are fiscal years ending June and include finance leases; Amazon figures are purchases of property and equipment; 2025 values marked (g) are company guidance.
Data-centre electricity demand
TWh per year, global
The IEA estimates global data centres consumed about 415 TWh of electricity in 2024 — around 1.5% of world demand — and projects roughly 945 TWh by 2030 in its base case, with AI the dominant driver of the growth. Projections carry wide uncertainty bands; the 2030 figure is a scenario, not a fact.
| TWh per year, global | |
|---|---|
| 2024 (IEA estimate) | 415 |
| 2030 (IEA Base Case projection) | 945 |
출처 및 근거: IEA, "Energy and AI" (April 2025): 2024 estimate and Base Case 2030 projection. Scenario ranges in the same report span materially higher and lower outcomes.
Rack power density by generation
kW per rack (typical / vendor spec)
A conventional enterprise rack draws under 10 kW. An air-cooled H100 rack runs around 40 kW; a GB200 NVL72 rack is specified around 120 kW and is liquid-cooled by requirement; NVIDIA's announced Rubin Ultra NVL576 targets roughly 600 kW per rack for 2027. Cooling and power distribution, not floor space, now define data-center design.
| kW per rack (typical / vendor spec) | |
|---|---|
| Enterprise rack (typical) | 8 |
| HGX H100 rack, air-cooled (typical) | 40 |
| GB200 NVL72 (vendor spec) | 120 |
| Rubin Ultra NVL576 (announced target, 2027) | 600 |
출처 및 근거: Typical enterprise figure: industry rule of thumb. H100/GB200 figures: NVIDIA system specifications and partner deployment documentation. Rubin Ultra: NVIDIA GTC 2025 announcement — a stated target, not a shipping product.
Largest announced AI campuses
GW, announced ultimate capacity (reported)
Announced ultimate capacities of the flagship AI campuses. These are targets reported by the companies or credible press — build-outs are phased over years and plans change; first-phase capacities are far smaller.
| GW, announced ultimate capacity (reported) | |
|---|---|
| Meta Hyperion (Louisiana) | 5.0 |
| Amazon New Carlisle / Rainier (Indiana) | 2.2 |
| Stargate Abilene (Texas) | 1.2 |
| xAI Colossus 2 (Memphis) | 1.0 |
출처 및 근거: Company announcements and reporting by mainstream outlets, as of the site's snapshot date. Each campus's record page lists its specific sources.
Training compute of notable models
training FLOP (floating-point operations)
Frontier training compute has grown roughly 4–5× per year for over a decade (Epoch AI estimate). Because the range spans nine orders of magnitude, it is shown as a table rather than a chart.
| Model | Year | Training compute | Basis |
|---|---|---|---|
| AlexNet | 2012 | ~4.7×10¹⁷ FLOP | Epoch AI estimate |
| GPT-2 | 2019 | ~1.5×10²¹ FLOP | Epoch AI estimate |
| GPT-3 | 2020 | 3.1×10²³ FLOP | disclosed in paper |
| Llama 3.1 405B | 2024 | ~3.8×10²⁵ FLOP | computed from disclosed GPU-hours |
| GPT-4 | 2023 | ~2×10²⁵ FLOP | Epoch AI estimate |
| Grok 3 | 2025 | ~4–5×10²⁶ FLOP | Epoch AI estimate |
출처 및 근거: Published papers where disclosed (GPT-3, Llama 3.1); Epoch AI estimates otherwise. Estimates carry significant uncertainty.
Nuclear power deals signed for AI demand
Hyperscalers have signed the first wave of nuclear offtakes in a generation — restarting shut reactors, buying existing output and backing SMR fleets. Restart and deployment dates are company targets and can slip.
| Deal | Announced | Scale | Status / target |
|---|---|---|---|
| Microsoft – Constellation, Crane Clean Energy Center (Three Mile Island Unit 1 restart) | Sep 2024 | 835 MW, 20-year PPA | Restart targeted 2027; work under way |
| Amazon – Talen Energy, Susquehanna | Mar 2024 (expanded Jun 2025) | Up to 1,920 MW carbon-free supply arrangement | Operating plant; supply ramps to 2032 |
| Meta – Constellation, Clinton Clean Energy Center | Jun 2025 | 1,121 MW, 20-year agreement | Begins 2027 when state support expires |
| Google – Kairos Power SMR fleet | Oct 2024 | Up to 500 MW across multiple reactors | First deployment targeted ~2030 |
| Microsoft — Holtec Palisades context: first US restart | — | 800 MW (Holtec restart, DOE loan; not a Microsoft deal) | Returned to service Oct 2025 — first restarted US reactor |
출처 및 근거: Company announcements (Constellation, Talen, Holtec, Kairos Power, Amazon, Microsoft, Meta, Google), as reported 2024–2025.
High-bandwidth memory
HBM is the scarcest input in AI hardware: every leading accelerator depends on it, three companies make it, and supply has been sold out generations in advance.
출처 및 근거: Vendor statements (SK hynix, Samsung, Micron) and TrendForce market estimates, as of the site's snapshot date.