Within Successor training
Could compute stop a runaway AI loop?
Even a brilliant automated researcher would still need chips, data centres, electricity, and capital to train a stronger successor.
On this page
- Why training successors requires physical infrastructure
- How chips, energy, and capital slow feedback loops
- Why compute limits reduce some risks but not all
Page outline Jump by section
Introduction
One objection to runaway AI self-improvement is simple: intelligence is not the same thing as compute. Even if an AI became an excellent researcher and could help design a stronger successor, it would still need vast amounts of computing hardware, electricity, data-centre capacity, and funding to train that successor. In the AI doom debate, this matters because many fast-takeoff scenarios assume that each generation of AI can rapidly create a more capable next generation. Compute constraints are one of the strongest reasons to think that such feedback loops could be slower, more expensive, and easier to observe than some classic intelligence-explosion stories suggest.
At the same time, compute limits are not necessarily permanent barriers. Recent decades have seen extraordinary growth in AI training compute, supported by expanding chip production, larger data centres, improving algorithms, and huge capital investment. The key question is therefore not whether compute matters—it clearly does—but whether physical infrastructure grows fast enough to keep pace with increasingly capable automated AI research. [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha…
Why training successors requires physical infrastructure
A common misunderstanding in discussions of recursive self-improvement is that a smarter AI could improve itself purely through software. In reality, training a frontier model is a large industrial process.
Creating a more capable successor typically requires:
- Tens of thousands to hundreds of thousands of specialised AI chips.
- Data-centre buildings and cooling systems.
- Reliable access to large amounts of electricity.
- High-speed networking equipment linking the chips together.
- Large financial investments to purchase, rent, or operate the infrastructure.
Even if an AI system generated brilliant new algorithms, those ideas would still need to be tested and validated through experiments that consume real computational resources. Frontier model training already operates at scales where hardware acquisition, cluster construction, and energy supply become major strategic concerns. [arXiv]arxiv.orgarXiv The rising costs of training frontier AI modelsarXivThe rising costs of training frontier AI modelsMay 31, 2024… [McKinsey]mckinsey.comthe next big shifts in ai workloads and hyperscaler strategiesMcKinsey & CompanyThe next big shifts in AI workloads and hyperscaler strategies17 Dec 2025 — Training workloads will demand up to one me… & Company
This creates a distinction that is often blurred in public discussions. An AI may be able to automate cognitive labour—writing code, designing experiments, analysing results—without automatically solving physical bottlenecks such as chip manufacturing or power generation.
How chips, energy, and capital slow feedback loops
Chips are not created at software speed
Advanced AI training depends heavily on specialised accelerators such as GPUs and related hardware. Manufacturing these chips requires complex global supply chains, advanced fabrication facilities, packaging technologies, and specialised equipment.
A hypothetical AI researcher might discover a better model architecture in days, but obtaining another hundred thousand cutting-edge chips could take months or years. Semiconductor production capacity expands slowly compared with software development cycles. This means that a self-improving AI system may encounter a hard resource constraint even if its research capability grows rapidly.
This is one reason why some researchers see compute governance—monitoring and controlling access to large-scale computing resources—as a potentially powerful safety lever. Unlike software, large training runs leave physical traces in supply chains, energy systems, and data-centre operations.
Electricity increasingly matters
Training and running frontier models requires substantial electrical power. Recent analyses suggest that the power demands of frontier AI training runs are growing rapidly and could reach gigawatt-scale levels during this decade. The broader data-centre sector already consumes hundreds of terawatt-hours of electricity annually worldwide. [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha… [axios]axios.comIt projects that by 2028, major AI model training operations could consume 1–2 gigawatts (GW) of power, and potentially escalate to 4–16… Industry participants increasingly describe power availability as a major bottleneck. New AI data centres can often be planned faster than they can obtain grid connections, with some regions facing multi-year waits for sufficient electrical capacity. [World Economic Forum]weforum.orgWorld Economic ForumIs power grid connectivity the strategic bottleneck for AI?May 18, 2026 — The underlying issue is that investment in…
For AI doom debates, this matters because recursive self-improvement models sometimes assume that more intelligence immediately translates into more capability. In practice, an AI that wanted to train a vastly larger successor might need access to infrastructure whose expansion is constrained by construction timelines, permitting, transmission networks, and energy production.
Capital is another constraint
Training frontier systems has become extraordinarily expensive. Analyses of training costs suggest that the largest models require investments ranging from tens of millions to hundreds of millions of dollars, with costs historically rising rapidly. Some forecasts suggest that frontier training runs could reach billion-dollar scales if current trends continue. [arXiv]arxiv.orgarXiv The rising costs of training frontier AI modelsarXivThe rising costs of training frontier AI modelsMay 31, 2024…
As a result, only a small number of organisations currently possess the resources needed to train the largest models. An autonomous AI would not merely need technical knowledge; it would also need continued access to enormous financial and industrial resources.
Why compute limits reduce some risks but not all
Compute constraints are one of the strongest arguments against extremely fast, invisible, software-only intelligence explosions.
If training a more capable successor requires:
- Purchasing or controlling scarce chips,
- Securing large energy supplies, [axios.com]axios.comIt projects that by 2028, major AI model training operations could consume 1–2 gigawatts (GW) of power, and potentially escalate to 4–16…
- Building additional infrastructure,
- Spending vast sums of money,
then capability growth may remain tied to observable economic and industrial processes rather than occurring entirely inside computers.
From this perspective, runaway AI development could resemble the growth of a major industrial sector rather than a sudden overnight event. Governments, companies, and outside observers would have more opportunities to detect warning signs and intervene.
However, this argument has important limitations.
First, compute has historically expanded very quickly. Frontier training compute has grown by roughly 4–5 times per year over long periods, while leading-model training compute since 2020 has increased by several orders of magnitude. [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha…
Second, algorithmic improvements can partially substitute for hardware. Researchers routinely discover methods that achieve similar capabilities using less compute. Improvements in efficiency have historically been substantial, meaning that raw hardware shortages do not translate directly into proportional slowdowns in capability growth. [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha… [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha…
Third, a runaway process does not necessarily require each generation to be vastly larger than the previous one. If an AI became dramatically better at research while operating on roughly similar compute budgets, significant capability gains might occur before infrastructure limits became binding.
The central disagreement
The deepest disagreement is not whether compute matters. Almost everyone agrees that it does.
Instead, the debate concerns the relative importance of intelligence and infrastructure.
Those who see compute constraints as a major brake argue that physical resources ultimately govern how quickly capabilities can grow. Intelligence alone cannot manufacture chips, build power stations, or construct data centres overnight. Therefore, they expect self-improvement loops to be slower, more visible, and more manageable than classic intelligence-explosion scenarios suggest. World Economic Forum [reuters]reuters.comKevin Zhang, TSMC’s Senior VP of Business Development, noted that customers — including those in mobile, IoT, and high-performance AI dat… Those more concerned about AI doom respond that even a partially constrained feedback loop could still be dangerous. If AI systems automate most AI research and development, they may dramatically accelerate the rate at which available compute is converted into capability gains. In that world, infrastructure remains a constraint, but a much less effective one than human research labour is today. Rapid capability growth could still occur even without unlimited hardware. Stanford HAI [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha…
What compute limits imply for p(doom)
For people estimating p(doom)—the probability that advanced AI leads to existential catastrophe—compute constraints usually act as a moderating factor rather than a complete solution.
They weaken the strongest versions of the claim that a single AI could instantly and invisibly race far beyond human capabilities. Physical infrastructure introduces delays, costs, dependencies, and opportunities for monitoring.
But compute limits do not eliminate concerns about loss of control. A system may become strategically dangerous long before it reaches the maximum scale allowed by global chip production. Moreover, history shows that when powerful technologies generate economic and geopolitical advantages, societies often invest heavily in expanding the underlying infrastructure.
The practical takeaway is that compute constraints are best understood as a speed limiter, not necessarily a safety guarantee. They reduce some pathways to runaway self-improvement, but they do not by themselves rule out the possibility that increasingly autonomous AI systems could drive capability growth faster than human institutions can reliably manage. [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha… [Epoch AI]epoch.aiThe amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha…
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: epoch.ai
Link: https://epoch.ai/Source snippet
The amount of compute used to train frontier language models has grown exponentially. Since 2020, the trend among top-5 models ha...
-
Source: epoch.ai
Title: AITraining compute of frontier AI models grows by 4-5x per year
Link: https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-yearSource snippet
Training compute of frontier AI models grows by 4-5x per yearMay 28, 2024 — Our expanded AI model database shows that the compute used to...
Published: May 28, 2024
-
Source: arxiv.org
Title: arXiv The rising costs of training frontier AI models
Link: https://arxiv.org/abs/2405.21015Source snippet
arXivThe rising costs of training frontier AI modelsMay 31, 2024...
Published: May 31, 2024
-
Source: mckinsey.com
Title: the next big shifts in ai workloads and hyperscaler strategies
Link: https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-next-big-shifts-in-ai-workloads-and-hyperscaler-strategiesSource snippet
McKinsey & CompanyThe next big shifts in AI workloads and hyperscaler strategies17 Dec 2025 — Training workloads will demand up to one me...
-
Source: epoch.ai
Title: power demands of frontier ai training
Link: https://epoch.ai/publications/power-demands-of-frontier-ai-trainingSource snippet
How much power will frontier AI training demand in 2030?11 Aug 2025 — The power required to train the largest frontier models is growing...
-
Source: axios.com
Link: https://www.axios.com/2025/08/12/ai-training-power-needsSource snippet
It projects that by 2028, major AI model training operations could consume 1–2 gigawatts (GW) of power, and potentially escalate to 4–16...
-
Source: reuters.com
Link: https://www.reuters.com/business/retail-consumer/energy-use-forcing-rethink-ai-chip-design-tsmc-says-2026-05-28/Source snippet
Kevin Zhang, TSMC’s Senior VP of Business Development, noted that customers — including those in mobile, IoT, and high-performance AI dat...
-
Source: arxiv.org
Link: https://arxiv.org/html/2405.21015v2Source snippet
arXivThe rising costs of training frontier AI modelsFeb 7, 2025 — This paper develops a detailed cost model to address this gap, estimati...
-
Source: epoch.ai
Title: what will ai look like in 2030
Link: https://epoch.ai/blog/what-will-ai-look-like-in-2030Source snippet
?Sep 16, 2025 — But algorithmic efficiency has already been improving within the existing compute growth. There is no particular reason t...
-
Source: arxiv.org
Title: arXiv Compute Requirements for Algorithmic Innovation in Frontier AI Models
Link: https://arxiv.org/abs/2507.10618 -
Source: arxiv.org
Title: arXiv Will Compute Bottlenecks Prevent an Intelligence Explosion?
Link: https://arxiv.org/abs/2507.23181 -
Source: hai.stanford.edu
Title: 2026 ai index report
Link: https://hai.stanford.edu/ai-index/2026-ai-index-reportSource snippet
Stanford HAIThe 2026 AI Index Report | Stanford HAIIndustry produced over 90% of notable frontier models in 2025, and several of those mo...
-
Source: epoch.ai
Link: https://epoch.ai/trendsSource snippet
in [Artificial]({{ 'artificial-goals/' | relative_url }}) IntelligenceFeb 5, 2026 — Frontier AI systems are advancing rapidly from increases in compute, hardware performance, softwa...
-
Source: epoch.ai
Title: open models threshold
Link: https://epoch.ai/data-insights/open-models-thresholdSource snippet
Frontier open models may surpass 1e26 FLOP of training...15 Jan 2025 — Historical trends suggest that the largest open model will surpas...
-
Source: epoch.ai
Link: https://epoch.ai/data/ai-modelsSource snippet
shold that grows over time as...Read more...
-
Source: epoch.ai
Title: power demands of frontier ai training
Link: https://epoch.ai/blog/power-demands-of-frontier-ai-trainingSource snippet
How much power will frontier AI training demand in 2030?Aug 11, 2025 — The power required to train the largest frontier models is growing...
-
Source: epoch.ai
Title: reports the models’
Link: https://epoch.ai/gradient-updates/r-and-d-vs-training-computeSource snippet
Final training runs account for a minority of R&D compute...23 Mar 2026 — Most of the final training run compute spending comes from our...
-
Source: epoch.ai
Title: open models
Link: https://epoch.ai/topics/open-modelsSource snippet
Open-Weight Models: Data & ResearchFrontier open models may surpass 1e26 FLOP of training compute before 2026. By Luke Emberson. Models w...
-
Source: epoch.ai
Link: https://epoch.ai/topics/data-centersSource snippet
AI Data Centers: Data & ResearchUsing satellite imagery and permit data, Epoch tracks the scale and growth of AI data centers and superco...
-
Source: epoch.ai
Link: https://epoch.ai/aboutSource snippet
About UsWe investigate the drivers and bottlenecks of AI progress and scaling. We were among the first to systematically monitor trends i...
-
Source: epoch.ai
Title: compute for robotic manipulation
Link: https://epoch.ai/data-insights/compute-for-robotic-manipulationSource snippet
Compute is not a bottleneck for robotic manipulationAug 8, 2025 — Compute is not a bottleneck for robotics, while training data is. Front...
-
Source: epoch.ai
Title: how much does it cost to train frontier ai models
Link: https://epoch.ai/blog/how-much-does-it-cost-to-train-frontier-ai-modelsSource snippet
?3 Jun 2024 — The cost of training top AI models has grown 2-3x annually for the past eight years. By 2027, the largest models could cost...
-
Source: arxiv.org
Link: https://arxiv.org/html/2512.04142v1Source snippet
From FLOPs to Footprints: The Resource Cost of Artificial...3 Dec 2025 — These findings demonstrate that the environmental impact of lar...
-
Source: arxiv.org
Link: https://arxiv.org/pdf/2504.16138Source snippet
Trends in Frontier AI Model Count: A Forecast to 2028by I Kumar · 2025 · Cited by 5 — The model predicts that there were 23 AI models exc...
-
Source: weforum.org
Link: https://www.weforum.org/stories/2026/05/electricity-data-grid-connectivity-strategic-bottleneck-ai-transformation/Source snippet
World Economic ForumIs power grid connectivity the strategic bottleneck for AI?May 18, 2026 — The underlying issue is that investment in...
Published: May 18, 2026
-
Source: linkedin.com
Link: https://www.linkedin.com/posts/epochai_todays-largest-ai-models-are-trained-on-activity-7336059172972556288-C_emSource snippet
Epoch AI's PostToday's largest AI models are trained on over 10^26 FLOP. How many will surpass this threshold in the coming years? In a n...
-
Source: merriam-webster.com
Link: https://www.merriam-webster.com/dictionary/epochSource snippet
Definition & Meaning5 May 2026 — epoch applies to a period begun or set off by some significant or striking quality, change, or series of...
Published: May 2026
-
Source: epochai.substack.com
Title: power demands of frontier ai training
Link: https://epochai.substack.com/p/power-demands-of-frontier-ai-trainingSource snippet
AI Training Power DemandPower demands for frontier AI training have been growing at 2.2x per year, with frontier runs now exceeding 100 M...
Additional References
-
Source: medium.com
Link: https://medium.com/technology-media-telecom/frontier-ai-peaked-heres-what-comes-next-8b9fc65eaa6cSource snippet
Frontier AI Peaked. Here's What Comes NextAccording to the latest market study by Omdia, parameter growth in frontier AI models has slowe...
-
Source: linkedin.com
Link: https://www.linkedin.com/posts/epochai_training-frontier-ai-models-requires-a-lot-activity-7326637083240947714-98PGSource snippet
Frontier AI models use 1.5x more energy than thoughtTraining frontier AI models requires a lot of power — but how much? We find that fron...
-
Source: greenmediography.nl
Link: https://greenmediography.nl/reference/the-rising-costs-of-training-frontier-ai-models/Source snippet
The rising costs of training frontier AI modelsThe costs of training frontier AI models have grown dramatically in recent years, but ther...
-
Source: ourworldindata.org
Link: https://ourworldindata.org/grapher/hardware-and-energy-cost-to-train-notable-ai-systems -
Source: linkedin.com
Link: https://www.linkedin.com/posts/prithpal-khajuria-9044251_frontierai-aiinfrastructure-energy-activity-7437888342869880832-xunpSource snippet
Integrate energy strategy early. 2. Map infrastructure risks openly. 3. Prioritize flexible site...Read more...
-
Source: linkedin.com
Title: hidden cost ai energy data centers hardware race shreyas shah f8ksc
Link: https://www.linkedin.com/pulse/hidden-cost-ai-energy-data-centers-hardware-race-shreyas-shah-f8kscSource snippet
The Hidden Cost of AI: Energy, Data Centers, and...Global [AI compute]({{ 'compute-kyc/' | relative_url }}) capacity has grown at an estimated 3.3x per year since 2022, reachi...
-
Source: flyfrontier.com
Link: https://www.flyfrontier.com/Source snippet
Frontier Airlines: Low Fares Done RightAs Home of Low Fares Done Right, find great deals and cheap flights to destinations all over North...
-
Source: researchgate.net
Link: https://www.researchgate.net/publication/391150569_AI%27s_Power_Requirements_Under_Exponential_Growth_Extrapolating_AI_Data_Center_Power_Demand_and_Assessing_Its_Potential_Impact_on_US_CompetitivenessSource snippet
ure power needs, summarize current bottlenecks for rapid data center construction...Read more...
-
Source: sparkco.ai
Title: ai data center build out capacity prediction markets
Link: https://sparkco.ai/blog/ai-data-center-build-out-capacity-prediction-marketsSource snippet
AI Data Center Build-out Capacity Prediction MarketsA comprehensive 2025 industry analysis of how prediction markets price AI infrastruct...
-
Source: ourworldindata.org
Title: artificial intelligence training computation
Link: https://ourworldindata.org/grapher/artificial-intelligence-training-computationSource snippet
Computation used to train notable artificial intelligence...Mar 12, 2025 — This data is based on the following sources. Epoch AI – Param...
Topic Tree