CUDO Observatory connects the critical factors that determine where and how AI infrastructure can be deployed, giving decision-makers a clearer view of the trade-offs, dependencies and constraints behind their infrastructure strategy.
Numbers that define the AI infrastructure landscape entering the 2028–2029 supply crunch. Each resolves to one canonical claim — open its evidence receipt for source, period and rights.
~$1.15tn
Hyperscaler capex, 2025–27
Goldman Sachs estimate · more than double the prior three years
2,060+
US interconnection queue (GW)
Proposed generation & storage seeking connection — not delivered power
~945
Global data-centre demand by 2030 (TWh)
IEA projection · 415 TWh in 2024
19 + 13
EU AI Factories + Antennas
Operational under the EuroHPC programme
~$90bn (2026)
AI infrastructure market, 2026
Analyst forecast · toward ~$465bn by 2033
<5%
Europe's share of AI compute
US hyperscalers hold ~70% of the European cloud market
Every blank is a stated reason, not an omission. The striped cells are where a source exists but we cannot reach it — or where the thing itself does not exist.
Manually reviewed developments across power, policy, capacity, supply and pricing since v0.7. Each card cites its source; only reviewer-approved entries appear.
Regulation (EU) 2026/1744 preserves the AI Act general application date of 2 August 2026, with Annex III high-risk obligations from 2 December 2027 and Annex I high-risk obligations from 2 August 2028.
Regulated workloads must plan against staged obligation dates, not a single 2026 high-risk enforcement date.
NVIDIA states that full-scale production of 800 VDC data centres will coincide with the Kyber rack generation in 2027, not the 2026 Vera Rubin platform.
Facility electrical design timelines shift: 800 VDC is a 2027 Kyber-era assumption, correcting the earlier 2026 expectation.
NVIDIA's GB300 raises HBM3e to 288 GB per GPU with the NVL72 rack cited at around 1.5x GB200, and the next rack-scale platform is confirmed as the Vera Rubin NVL72.
Cluster planners should design against Vera Rubin NVL72, replacing stale NVL144 and exaFLOPS figures.
The IEA projects global data-centre electricity consumption rising from about 415 TWh in 2024 toward roughly 945 TWh by 2030, replacing the earlier '1,000 TWh by 2026' framing.
The corrected trajectory sets a realistic horizon for AI-driven electricity demand and grid planning.
Unit token prices have fallen roughly 99.7% since the GPT-3 era, yet total enterprise AI bills have risen as agentic workloads multiply token consumption per task.
Falling unit prices do not reduce total inference spend; volume growth dominates the economics.
Specialised neocloud providers rent comparable GPUs materially below hyperscaler list pricing, with reported buyer savings in the 40–70% range.
Placement and provider choice, not just GPU model, drive a large share of cluster cost.
“The window to lock favorable, power-anchored AI infrastructure economics is closing. Whoever secures firm power, sovereign-compliant capacity, and reserved compute at 2026 pricing — while modeling depreciation and goodput honestly — will hold a durable cost and capability advantage through the 2028–2029 supply crunch.”
Executive summary · Bottom line
Each watch tracks one structural force shaping the AI factory. Dive into interactive data, charts and analysis.
The decisive shifts across power, build, silicon, sovereignty and economics.
Survey coverage, interview depth and collection period behind this release.
n=701
Survey respondents
15
In-depth interviews
3
Regions · UK, US, Europe
February 12–25, 2026
Data collection · Censuswide
Discuss power, location, cluster architecture and TCO with a CUDO AI infrastructure specialist.
