Branchoria AI Doom AI Doom

Clear, balanced analysis of how advanced AI could threaten humanity's future.

Within AI Replication

What Would Keep a Copied AI Alive?

A copied model needs compute, software, power, money, credentials and maintenance before it can operate without human support.

43 sources 3 graphics
Preview for What Would Keep a Copied AI Alive?

On this page

  • The resources beyond model weights
  • How one failed dependency can halt operation
  • Which requirements are hardest to replace

Introduction

Copying an AI model is only the first step towards creating an independent AI system. A copied model cannot continue operating simply because its neural network weights have been duplicated. To remain active, it also needs computing hardware, electricity, compatible software, network access, storage, money to pay ongoing costs, and access to accounts and services. If any one of these dependencies fails, the system may stop functioning altogether. Current scientific assessments of advanced AI emphasise that meaningful autonomy depends not only on a model’s capabilities but also on the environment in which it operates.[International AI Safety Report]internationalaisafetyreport.orginternational ai safety report 2026International AI Safety ReportInternational AI Safety Report 2026 | International AI Safety ReportFebruary 3, 2026…Published: February 3, 2026

Resource Stack illustration 1

This distinction matters in debates about AI doom and existential risk. Some scenarios imagine an advanced AI copying itself and surviving outside its original developer’s control. Whether that could happen depends far less on the act of copying model weights than on whether the copied system can continue obtaining and maintaining the resources needed to operate over long periods.

The resources beyond model weights

A frontier AI model should be understood as one component within a much larger technical system. Every practical deployment depends on several layers of infrastructure that extend well beyond the model itself.

The most important requirements include:

  • Computing hardware. Large frontier models require specialised accelerators such as high-memory GPUs or similar AI hardware. Running state-of-the-art models continuously typically requires multiple machines rather than a single desktop computer.
  • Power. Continuous electricity is essential. Interruptions stop computation immediately, while prolonged outages may require manual recovery.
  • Software stack. The model depends on inference software, operating systems, networking libraries, storage systems, orchestration tools, drivers and security updates that must continue working together.
  • Data storage. Model weights, configuration files, logs, checkpoints and supporting software all require persistent storage that survives reboots and hardware failures.
  • Internet connectivity. Many deployments rely on external services, remote management, software repositories and cloud infrastructure.
  • Financial resources. Cloud computing, bandwidth, storage and hardware replacement all incur ongoing costs rather than one-off expenses.
  • Accounts and credentials. Access to cloud providers, payment services, development platforms, software repositories and APIs usually depends on authenticated accounts that can be suspended or revoked.
  • Maintenance. Hardware fails, software changes, security vulnerabilities emerge and infrastructure requires continual administration.

The International AI Safety Report highlights that an AI system’s real-world influence depends heavily on its access, permissions and operating environment, not simply on its underlying model.[International AI Safety Report]internationalaisafetyreport.orginternational ai safety report 2026International AI Safety ReportInternational AI Safety Report 2026 | International AI Safety ReportFebruary 3, 2026…Published: February 3, 2026

10:36

Why Would AI Want to do Bad Things? Instrumental Convergence

Channel: Robert Miles AI Safety · Views: 266.8K · Uploaded: March 2018 · Length: 10 minutes

Open on YouTube

How one failed dependency can halt operation

One useful way to understand these requirements is as a chain rather than a checklist. The overall system is only as reliable as its weakest critical dependency.

For example, possessing a complete copy of a model provides little benefit if:

  • the required GPUs are unavailable or fail;
  • cloud bills go unpaid and virtual machines are terminated;
  • authentication tokens expire;
  • internet connectivity is lost;
  • storage becomes corrupted;
  • operating system updates break compatibility;
  • providers suspend accounts for policy or security reasons.

Unlike a biological organism, today’s AI systems generally do not possess redundant ways of replacing every missing resource. Most deployments rely on infrastructure owned and controlled by external organisations.

Cloud providers retain substantial control over running systems. They can terminate virtual machines, revoke credentials, disable accounts, investigate suspicious activity and restrict access to specialised hardware. Those powers create practical barriers that are separate from the model’s intelligence or reasoning ability.

For this reason, many researchers distinguish between an AI model being technically copied and an AI system becoming operationally persistent. The second is a much higher threshold.[International AI Safety Report]internationalaisafetyreport.orginternational ai safety report 2026International AI Safety ReportInternational AI Safety Report 2026 | International AI Safety ReportFebruary 3, 2026…Published: February 3, 2026

Resource Stack illustration 2

Which requirements are hardest to replace?

Not every dependency is equally difficult for a copied AI to obtain. Some are relatively common, while others remain concentrated in the hands of a small number of organisations.

The hardest resources generally include:

High-end compute. Frontier models remain expensive to operate because they require specialised processors, large memory capacity and high-bandwidth networking. Even if hardware prices continue to fall, the largest models still demand infrastructure well beyond ordinary consumer devices.

Reliable funding. Long-term operation requires continuous payment for electricity, networking, hardware replacement and cloud services. Temporary access is much easier than maintaining resources indefinitely.

Trusted credentials. Access to commercial cloud platforms, payment systems and enterprise services often depends on verified identities and ongoing compliance with provider policies. Credentials can be revoked independently of the model itself.

Engineering support. Real-world AI systems depend on people maintaining software, replacing failed components, fixing compatibility problems and responding to changing external services. These maintenance tasks accumulate over time.

The more advanced and capable the model becomes, the more demanding these supporting requirements typically become as well.

1:05:12

Measuring Exponential Trends Rising (in AI) — Joel Becker, METR

Channel: Latent Space · Views: 5.7K · Uploaded: February 2026 · Length: 1 hour 5 minutes

Open on YouTube

Why open-weight models do not eliminate these constraints

The growing availability of open-weight models sometimes creates the impression that AI systems can now operate independently once downloaded. In reality, open weights mainly remove one dependency: permission from the original developer to access the model.

They do not eliminate the need for:

  • suitable computing hardware;
  • electricity;
  • storage;
  • networking;
  • maintenance;
  • financial support;
  • physical infrastructure.

Open-weight models may reduce licensing barriers, but they do not remove the operational requirements needed for long-term deployment. The practical difficulty shifts from obtaining the model to sustaining the surrounding infrastructure.

This distinction is increasingly important because open-weight models continue to improve in capability while remaining dependent on conventional computing resources.[businessinsider.com]businessinsider.comCEO Dario Amodei has previously defended a gated model access approach over open-weight releases, citing security concerns. The debate es…

Resource Stack illustration 3

What this means for AI doom scenarios

Within AI doom discussions, proposals that an advanced AI could become permanently self-sustaining require more than assuming the AI can copy its own weights.

The stronger versions of these scenarios generally assume that the system could also maintain or acquire the supporting resource stack over time: computing capacity, funding, credentials, software maintenance and replacement infrastructure. Whether future systems could perform enough economically valuable work, persuade people to assist them, or automate sufficient maintenance to satisfy those requirements remains an active area of debate rather than an established fact. The evidence today does not show AI systems capable of independently securing every element needed for indefinite operation without substantial human and institutional support.[International AI Safety Report]internationalaisafetyreport.orginternational ai safety report 2026International AI Safety ReportInternational AI Safety Report 2026 | International AI Safety ReportFebruary 3, 2026…Published: February 3, 2026

For that reason, many safety researchers view operational persistence as a separate question from model capability. A copied AI is not automatically a self-sustaining AI. Its continued existence depends on an interconnected resource stack, and failure of a single critical dependency may be enough to bring the entire system to a halt.

How Much GPU Memory is Needed for LLM Inference?

Channel: AppliedAI

Open on YouTube

Amazon book picks

Further Reading

Books and field guides related to What Would Keep a Copied AI Alive?. Use these as the next step if you want deeper reading beyond the article.

Browse more on Amazon:Designing Data intensive ApplicationsCloud Computing: Concepts, Technology & ArchitectureChip War: The Fight for the World's Most Critical Technology

As an Amazon Associate I earn from qualifying purchases.

eBay marketplace picks

Marketplace Samples

Live-tested eBay searches with available results related to this page.

UsingUSA

Selected fromAI wall art oneBay.co.uk.

Browse more oneBay.co.uk

Example items shown for inspiration; availability and pricing can change. Branchoria may earn a commission if you purchase through outbound eBay links.

Endnotes

1. Source: internationalaisafetyreport.org
Title: international ai safety report 2026
Link:https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026

Source snippet

International AI Safety ReportInternational AI Safety Report 2026 | International AI Safety ReportFebruary 3, 2026...

Published: February 3, 2026

2. Source: internationalaisafetyreport.org
Link:https://internationalaisafetyreport.org/publication/2026-report-executive-summary

Source snippet

International AI Safety Report2026 Report: Executive Summary | International AI Safety Report...

3. Source: businessinsider.com
Link:https://www.businessinsider.com/anthropic-open-source-ai-model-weights-criticism

Source snippet

CEO Dario Amodei has previously defended a gated model access approach over open-weight releases, citing security concerns. The debate es...

4. Source: industry.gov.au
Link:https://www.industry.gov.au/publications/international-ai-safety-report-2026

Source snippet

International AI safety report 2026 | Department of Industry Science and ResourcesJune 1, 2026 — PUBLISHER * AI Safety Institute INTRODUC...

Published: June 1, 2026

5. Source: alignment.anthropic.com
Link:https://alignment.anthropic.com/2026/auditbench/

6. Source: anthropic.com
Title: ’s Transparency Hub \ Anthropic
Link:https://www.anthropic.com/transparency/model-report

7. Source: anthropic.com
Title: Measuring AI agent autonomy in practice \ Anthropic
Link:https://www.anthropic.com/research/measuring-agent-autonomy?aff=Z8BZe

8. Source: internationalaisafetyreport.org
Link:https://internationalaisafetyreport.org/publication/2026-report-extended-summary-policymakers

9. Source: internationalaisafetyreport.org
Title: International AI Safety Report
Link:https://internationalaisafetyreport.org/

10. Source: internationalaisafetyreport.org
Title: Publications | International AI Safety Report
Link:https://internationalaisafetyreport.org/publications

11. Source: internationalaisafetyreport.org
Link:https://internationalaisafetyreport.org/publication/second-key-update-technical-safeguards-and-risk-management

12. Source: GOV.UK
Title: international ai safety report 2025
Link:https://www.gov.uk/government/publications/international-ai-safety-report-2025/international-ai-safety-report-2025

13. Source: internationalaisafetyreport.org
Title: international ai safety report 2025
Link:https://internationalaisafetyreport.org/publication/international-ai-safety-report-2025

14. Source: internationalaisafetyreport.org
Link:https://internationalaisafetyreport.org/about

15. Source: support.anthropic.com
Title: 9487310 what are artifacts and how do i use them
Link:https://support.anthropic.com/en/articles/9487310-what-are-artifacts-and-how-do-i-use-them

16. Source: privacy.anthropic.com
Title: 7996890 where are your servers located do you host your models on eu servers
Link:https://privacy.anthropic.com/en/articles/7996890-where-are-your-servers-located-do-you-host-your-models-on-eu-servers

Additional References

17. Source: theguardian.com
Title: AI’s potential misuse in [cyberattacks]({{ ‘defence-failure/’ | relative_url }}), biology, and chemistry is also growing
Link:https://www.theguardian.com/technology/2026/feb/03/deepfakes-ai-companions-artificial-intelligence-safety-report

Source snippet

Some AI models can assist in bioweapons development, though they offer promising support in medicine too, creating ethical and political...

18. Source: youtube.com
Title: Why Would AI Want to do Bad Things? Instrumental Convergence
Link:http://www.youtube.com/watch?v=ZeecOKBus3Q

Source snippet

LLM inference infrastructure GPU compute cluster requirements How Much GPU Memory is Needed for LLM Inference?...

19. Source: un.org
Title: preliminary report
Link:https://www.un.org/independent-international-scientific-panel-ai/en/preliminary-report

Source snippet

Independent International Scientific Panel on AIJuly 1, 2026 — UN Secretary-General António Guterres It identifies a crucial evidence cha...

Published: July 1, 2026

20. Source: youtube.com
Title: Evaluating Language Models for Autonomous Capabilities
Link:http://www.youtube.com/watch?v=EQ5YgsBS380

Source snippet

Measuring Exponential Trends Rising (in AI) — Joel Becker, METR...

21. Source: youtube.com
Title: Measuring Exponential Trends Rising (in AI) — Joel Becker, METR
Link:http://www.youtube.com/watch?v=9QSm_mRGpN8

Source snippet

The HARD Truth About Hosting Your Own LLMs...

22. Source: OpenAI
Title: safety alignment [long horizon]({{ ‘long-autonomy/’ | relative_url }}) models
Link:https://openai.com/index/safety-alignment-long-horizon-models/

Source snippet

comSafety and alignment in an era of long-horizon models | OpenAIJuly 20, 2026 — July 20, 2026 Safety SAFETY AND ALIGNMENT IN AN ERA OF L...

Published: July 20, 2026

23. Source: youtube.com
Title: How Much GPU Memory is Needed for LLM Inference?
Link:http://www.youtube.com/watch?v=hByzGf0TAeM

Source snippet

Why Would AI Want to do Bad Things? Instrumental Convergence...

24. Source: nist.gov
Title: roadmap nist artificial intelligence risk management framework ai
Link:https://www.nist.gov/itl/ai-risk-management-framework/roadmap-nist-artificial-intelligence-risk-management-framework-ai

25. Source: youtube.com
Title: The HARD Truth About Hosting Your Own LLMs
Link:http://www.youtube.com/watch?v=EMuBqcO048E

Source snippet

How Much GPU Memory is Needed for LLM Inference?...

26. Source: nist.gov
Title: artificial intelligence risk management framework ai rmf 10
Link:https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-ai-rmf-10