--- title: "What Would Superintelligence Need? Power, Compute, Capital" slug: what-resources-do-frontier-labs-need-for-superintelligence description: "No ASI exists and no source says what it would need. What labs use today: power, clusters, data, capital and security, with dated, sourced figures." date: 2026-11-03 last_reviewed: 2026-11-03 status: published type: knowledge-base author: [liyam-flexer] tags: [superintelligence, ASI, compute, energy, data centers, AI security] --- # What Would Superintelligence Need? Power, Compute, Capital **In short:** No artificial superintelligence (ASI) exists, and no public source gives a reliable figure for what building one would take; anyone who quotes an exact chip count, wattage or price for ASI is guessing. What can be sourced is the scale of today's frontier AI projects and published projections for the end of this decade. Epoch AI (August 2024) projected that a 2e29 FLOP training run, far above today's largest runs, could be feasible by 2030, needing roughly 6 GW of power, with power the first constraint likely to bind. Whether that is the right scale for ASI is unknown. ## What this post is, and is not ASI is hypothetical. The encyclopedia covers the concept in [Artificial Superintelligence](../articles/vol06_asi/ASI-001_artificial_superintelligence.md) and the proposed routes in [Pathways to Superintelligence](../articles/vol06_asi/ASI-PATH-001_pathways_to_superintelligence.md). Nobody knows whether ASI would come from more scale, from new methods, or from something else, so nobody knows what resources it would need. A new algorithm could change the answer by orders of magnitude in either direction. So this post does something narrower. For each resource category it gives what frontier labs use or plan today, with a date and a link, and then says what is unknown. Where a number is a projection, the source that made it is named. Treat every "by 2030" figure as a forecast, not a plan any lab has committed to. ## 1. Power and cooling Electricity is the most concrete constraint, and the best-sourced. - **Grid-level demand.** The International Energy Agency's *Energy and AI* report (2025) puts global data centre electricity consumption at about 415 TWh in 2024, roughly 1.5% of global electricity, and projects around 945 TWh in 2030 in its base case, just under 3%. It projects US data centre consumption rising by about 240 TWh (about 130%) from 2024 to 2030. These cover all data centres, not only AI training. - **United States.** Lawrence Berkeley National Laboratory's report for the US Department of Energy (December 2024) estimates data centres used about 4.4% of US electricity (176 TWh) in 2023 and projects roughly 6.7% to 12% by 2028, an increase of 325 to 580 TWh. The wide range reflects scenario uncertainty. - **Single sites.** OpenAI, Oracle and SoftBank announced on 23 September 2025 that their Stargate programme had nearly 7 GW of planned capacity and over $400 billion of planned investment over three years, against a stated $500 billion and 10 GW commitment. These are announced plans, not delivered capacity. - **Projection for one training run.** Epoch AI (Sevilla et al., 20 August 2024) found that data centre campuses of 1 to 5 GW are likely possible by 2030, estimated a 2e29 FLOP run at about 6 GW, and concluded that power is likely the first bottleneck, followed by chip manufacturing capacity. It also estimated that training spread over several sites could reach 2 to 45 GW. This post makes no claim about transformer or turbine lead times, or about cooling requirements, because no source with figures was checked. Cooling is a real engineering problem at this density. **Unknown:** how much power ASI would need, or whether it would be trained in one place at all. ## 2. Interconnected compute clusters Frontier training runs on tens of thousands of accelerators wired together, and at that size hardware failure is routine. - Meta's Llama 3 paper ([Grattafiori et al., 2024](https://arxiv.org/abs/2407.21783)) describes training on a cluster of 16,384 Nvidia H100 GPUs. Trade press (Tom's Hardware, 2024) headlines the paper's reliability findings as about one failure every three hours, with roughly half attributed to GPU and HBM3 memory faults; the paper itself is the authority for exact counts. - Epoch AI's 2024 analysis lists latency as one limit on how large a single synchronous run can grow, estimating that cumulative latency on modern GPU setups would cap training runs at 3e30 to 1e32 FLOP. No source was found for claims that failures at larger scale come "every few minutes", or that a specific interconnect latency such as "sub-millisecond" is required. Those are extrapolations, not published findings, so they are not made here. **Unknown:** how large a cluster ASI would need, or how its failure rates would scale. ## 3. Synthetic data and formal verifiers Two separate questions are often blurred together: whether human-written text runs out, and whether other signals can substitute. - **Public text.** Villalobos et al., [*Will we run out of data?*](https://arxiv.org/abs/2211.04325) (Epoch AI; v1 October 2022, revised June 2024), project that if trends continue, models will be trained on datasets roughly equal in size to the available stock of public human text between 2026 and 2032, or earlier if models are overtrained. This is a projection about public text, not a statement that data is "effectively exhausted". Epoch's 2024 scaling report also estimated the equivalent of 400 trillion to 20 quadrillion tokens available for training by 2030, once other sources are counted. - **Checkable tasks.** DeepSeek-AI's [DeepSeek-R1 paper](https://arxiv.org/abs/2501.12948) (arXiv January 2025; *Nature* 645, 2025) reports improving reasoning through reinforcement learning without human-labelled reasoning examples, with strong results on verifiable tasks such as mathematics and coding. Tasks with automatically checkable answers are the natural place for machine-generated training signal. Whether that carries over to open-ended domains with no checker is an open research question. Formal verifiers are not an established requirement for ASI. They are a hypothesis some researchers explore, for example in [Automated AI Research](../articles/vol06_asi/ASI-AIRD-001_automated_ai_research.md), not an established requirement. **Unknown:** what data, real or synthetic, ASI would need. ## 4. Capital - Stargate's announced figures above ($500 billion target, over $400 billion planned over three years) are the clearest public marker of intent at the top end. No source gives a cost for ASI, and there is no verifiable public data on researcher pay, so none is claimed. Policy responses to this scale of spending, such as compute thresholds and export rules, are covered in [Export Controls and AI Geopolitics](../articles/vol08_safety_governance/GOV-EXPORT-001_export_controls_and_ai_geopolitics.md) and [International AI Governance and Compute Governance](../articles/vol08_safety_governance/GOV-INTL-001_international_ai_governance_and_compute_governance.md). **Unknown:** the cost of ASI, or whether cost is the binding limit. ## 5. Isolation and containment Two different ideas sit under this heading: protecting model weights from theft, and containing a system that might itself be dangerous. - **Protecting weights.** RAND's *Securing AI Model Weights* (Nevo et al., 2024) defines five security levels, SL1 to SL5, scaled to increasingly capable attackers up to the most capable nation-states, and catalogues attack vectors. Read the report itself for the specific measures at each level. - **Containment of the system itself.** Whether isolation can safely contain a system more capable than its overseers is an unsolved problem, discussed in [AI Alignment and Control](../articles/vol06_asi/ASI-CTRL-001_ai_alignment_and_control.md). No published standard for isolating a hypothetical ASI was found, and nothing here claims that air-gapped clusters are the answer or are in use. ## At a glance | Resource | Sourced, dated fact (frontier AI today or projected) | What nobody knows for ASI | |---|---|---| | Power | Data centres about 415 TWh in 2024, projected about 945 TWh in 2030 (IEA); 2e29 FLOP run about 6 GW by 2030 (Epoch, Aug 2024) | Whether ASI needs more or less | | Clusters | Llama 3 trained on 16,384 H100 GPUs; frequent hardware failures reported | Size and reliability needs | | Data | Public text stock may be matched between 2026 and 2032 (Epoch, revised Jun 2024) | Whether synthetic data suffices | | Capital | Stargate: nearly 7 GW and over $400 billion planned (Sep 2025) | Total cost | | Security | RAND defines SL1 to SL5 for weight protection (2024) | How to contain ASI | ## Frequently asked questions **How many chips and how much power would ASI need?** No one knows, because no ASI exists and no public source gives a figure for it. The nearest sourced projection is Epoch AI's August 2024 estimate that a 2e29 FLOP training run, possible by 2030, would need about 6 GW. That is a projection for scaled-up frontier models, not for ASI. **Is human-written data running out?** A projection says the stock of public human text could be matched by training datasets between 2026 and 2032 (Epoch AI, revised June 2024). That does not mean training stops: Epoch also estimates much larger token totals when other sources are counted, and reinforcement learning on checkable tasks reduces dependence on human examples. **Can open-source or smaller labs build ASI?** Nobody can say, because it is not known what ASI requires or whether it is reachable by scaling. What is sourced is that the largest announced projects involve hundreds of billions of dollars and gigawatts of power, which few organisations can fund. That says nothing definitive about methods that need less. ## Sources - International Energy Agency, [Energy and AI: Energy demand from AI](https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai) (2025). - Lawrence Berkeley National Laboratory, [2024 United States Data Center Energy Usage Report](https://datacenters.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report) (19 December 2024); figures as summarised in the US Department of Energy's [release](https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers) (20 December 2024). - Sevilla et al., Epoch AI, [Can AI scaling continue through 2030?](https://epoch.ai/blog/can-ai-scaling-continue-through-2030) (20 August 2024). - Villalobos et al., [Will we run out of data? Limits of LLM scaling based on human-generated data](https://arxiv.org/abs/2211.04325) (arXiv, revised 4 June 2024). - SoftBank Group, [OpenAI, Oracle, and SoftBank expand Stargate with five new AI data center sites](https://group.softbank/en/news/press/20250924) (23 to 24 September 2025). - Grattafiori et al., [The Llama 3 Herd of Models](https://arxiv.org/abs/2407.21783) (arXiv, v3 November 2024). - Tom's Hardware, [Faulty Nvidia H100 GPUs and HBM3 memory caused half of the failures during Llama 3 training](https://www.tomshardware.com/tech-industry/artificial-intelligence/faulty-nvidia-h100-gpus-and-hbm3-memory-caused-half-of-the-failures-during-llama-3-training-one-failure-every-three-hours-for-metas-16384-gpu-training-cluster) (2024); trade-press report. - DeepSeek-AI, [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://arxiv.org/abs/2501.12948) (arXiv, January 2025). - Nevo et al., RAND, [Securing AI Model Weights](https://www.rand.org/pubs/research_reports/RRA2849-1.html) (2024).