Clouded Judgement 5.22.26 - The Neocloud Boom
Clouded Judgement · Jamin Ball
The AI data center buildout points to roughly 150GW of new capacity and $7.5T of spend over 4.5 years, making a Neocloud boom — and trillions in enterprise value for operators beyond the hyperscalers — close to inevitable.
The scale is the story: $1.7T a year, about 5% of US GDP, flowing into physical infrastructure for the next half-decade. At $90B of enterprise value per gigawatt, 150GW implies $13.5T of value creation, and even a 20% slice for Neoclouds is $2.5T — far too much for a handful of hyperscalers to absorb alone. That gap is what makes Neoclouds structurally interesting rather than a niche bet, though the whole thesis rides on OpenAI and Anthropic actually representing the 30-40% share of demand it assumes.
The data center infrastructure buildout is shaping up to be one of the largest physical-infrastructure wealth creation events ever, making a boom in Neoclouds the natural consequence.
150GW at $50B/GW equals $7.5 trillion across 4.5 years, or roughly $1.7 trillion annually — about 5% of US GDP.
If Neoclouds capture only ~20% of the pie, that is still over $2.5 trillion of enterprise value created — too big for a small number of players to serve.
If OpenAI and Anthropic represent 30-40% of new builds, the implied total is more than 150GW of new data center capacity coming online by 2030.
Extending the $90B-per-GW benchmark to 150GW of new capacity implies $13.5 trillion of enterprise value created over the next 4.5 years.
Open
- · What happens to the math if OpenAI and Anthropic end up well below 30-40% of new builds?
- · Which players actually capture the ~20% Neocloud slice, and on what terms relative to hyperscalers?
- · Is $50B/GW of capex and $90B/GW of enterprise value sustainable as supply scales?
Pipeline
- source kind
- url
- generated by
- anthropic+voyage
- candidates
- 22 (selected 5)
- embeddings
- voyage-3.5
Coverage
100% covered
Each block is one paragraph of the source. Darker means the decomposition captures it well; lighter means it was left out — the part of the document the summary doesn’t cover.
Considered candidates (17)
Below top-k · 12
- contextAI buildout compared to historical megaprojectsc 0.70
At ~5% of GDP, the AI buildout is smaller than 1880s US railroads (~6%) but dwarfs the early-2000s telecom buildout (~1.2%) and the Manhattan Project (~0.4%).
- evidenceCurrent Neoclouds trade at ~$90B EV per live GWc 0.70
CoreWeave and IREN both sit around $90B of enterprise value per live gigawatt, with Nebius higher — a rough yardstick for value creation per GW.
- exampleAnthropic is paying SpaceX $15B/year for ~500MWc 0.70
Per the S-1, Anthropic is paying SpaceX roughly $15B annually for an estimated 500MW of capacity across Colossus 1 and part of Colossus 2.
- exampleSpaceX just spun up a business as big as itselfc 0.70
SpaceX took 23 years to reach ~$19B in annual revenue, and effectively created a new ~$15B revenue line via this neocloud-style deal overnight.
- implicationNeoclouds as the natural hedge in a venture portfolioc 0.60
If AGI ends up capturing all the value at the model layer, somebody still has to build and operate the compute — making Neoclouds a structural hedge against AI investment risk.
- caveatThe whole thesis rests on 150GW actually getting builtc 0.60
If AI demand falls short or the cost share of GDP proves unsustainable, the 150GW assumption — and everything downstream of it — collapses.
- implicationFinancing $7.5T of buildout will mint fortunesc 0.50
Someone is going to make enormous amounts of money simply financing the capital required to deliver this much capacity.
- caveat$30M/MW pricing looks unusually highc 0.50
The Anthropic-SpaceX deal implies ~$30M per MW versus a ~$10-12M industry comp, suggesting either a premium one-time arrangement or different accounting.
- mechanismA 90-day out justifies the premiumc 0.50
Because the contract has a 90-day cancellation clause and Anthropic is severely compute constrained, paying a premium for flexible short-term access can be rational.
- implicationChip residual value is the next big swing factorc 0.50
The upside case for Neoclouds gets dramatically larger if GPUs retain meaningful value beyond years 4-5 of their useful life.
- contextAuthor is openly AGI-pilledc 0.40
The piece frames its forecasts with the disclosure that the author takes the over on essentially every estimate of tokens, model progress, and data center demand.
- evidencePublic Neocloud valuation multiples todayc 0.40
CoreWeave trades at ~6x NTM revenue and ~10x NTM EBITDA, while Nebius and IREN are at ~10x revenue and ~20x EBITDA.
Redundant with selected · 5
- evidenceOpenAI and Anthropic alone target ~55GW by 2030c 0.70 · sim 0.93
Rumors put each at ~3-3.5GW by end of 2025, with OpenAI targeting 30GW by 2030 and Anthropic assumed similar — together roughly 55GW of new capacity in 4.5 years.
overlapped with: Over 150GW of new capacity in the next 4.5 years
- implicationMarket structure will mirror independent power providersc 0.70 · sim 0.84
Expect a handful of very large Neoclouds plus a long tail of single-site operators, much like the IPP market, with private equity playing a major role.
overlapped with: A Neocloud boom feels inevitable
- mechanism1GW of capacity costs roughly $50B to buildc 0.60 · sim 0.83
A Blackwell-dominated 1GW data center breaks down to ~70% chips and compute, ~20% shell and power infrastructure, ~5% cooling, and ~5% land and permitting.
overlapped with: Over 150GW of new capacity in the next 4.5 years
- caveatThe business model still has to be provenc 0.60 · sim 0.83
The revenue opportunity is clearly massive, but Neoclouds still need to demonstrate that they can turn it into a durable, profitable business model.
overlapped with: Even 20% capture leaves $2.5T for Neoclouds
- caveatMuch of the buildout is captive to hyperscalersc 0.50 · sim 0.87
A large share of new capacity will be built by hyperscalers for their own internal use rather than sold to others, so not all of the implied enterprise value accrues to Neoclouds.
overlapped with: Even 20% capture leaves $2.5T for Neoclouds
Janitor
Non-content spans (acknowledgements, references, footnotes, headers, boilerplate) are dropped before the decomposition runs.
- total spans
- 74
- kept
- 64
- dropped
- 10
- outliers
- 2
- content · 64
- toc · 4
- boilerplate · 2
- noise · 2
- metadata · 1
- references · 1