The Daily Five · No. 8 · Tuesday 20 October

Today’s five stories ask who controls AI’s infrastructure, discoveries, wealth and oversight as it reaches chips, nuclear power, laboratories and the systems that watch over it.

By the AI Story Front editorial team · About our editors

Amazon challenges Nvidia with its own AI chips and nuclear power ambitions. Who will control AI’s infrastructure?

Amazon is developing an alternative to Nvidia’s computing ecosystem while backing a proposed nuclear power project. Together, the investments reveal how competition in AI increasingly depends on control over chips, cloud services and electricity.

Amazon is developing more of the technology and infrastructure required to operate artificial intelligence, announcing new software for its own AI chips alongside plans to support a nuclear energy project in the United States.

On 19 October, the company introduced TorchNeuron, software designed to allow developers to run existing PyTorch code on AWS Trainium processors without rewriting entire applications.

PyTorch is widely used to develop machine-learning models. Supporting it directly could make Amazon’s chips more accessible to researchers and businesses that currently depend on established hardware and software combinations.

Amazon says TorchNeuron requires as little as a single configuration change for some workloads. The software is scheduled for open-source release under the Apache 2.0 licence on 20 October.

The company argues that easier movement between different processors could reduce development costs and encourage competition in a market where Nvidia has a powerful position.

But developing and operating AI systems requires more than computing hardware.

In a separate announcement, Amazon revealed an agreement with X-energy and American Electric Power to explore building 12 small modular nuclear reactors in Indiana.

The proposed plant would provide up to 960 megawatts of electricity for the regional grid and Amazon’s data centres. Amazon would help finance development and become the facility’s first electricity customer if the project proceeds.

The reactors are not yet operational, and the agreement does not guarantee construction. The project will depend on further development, regulatory approval and commercial decisions.

Amazon estimates that the facility could support more than 1,000 construction jobs and over 100 permanent positions. These are projections rather than realised employment figures.

Training and operating large models requires substantial computing capacity, supported by reliable electricity supplies. Nuclear power offers the possibility of continuous low-carbon generation, although new facilities involve significant costs, long development periods and regulatory obligations.

Together, the two announcements show Amazon seeking greater influence over essential components of the AI economy: processors, software compatibility, cloud services and electricity.

Source: Amazon

The Storyfront Perspective

Amazon’s strategy illustrates how competition in AI is increasingly determined by control over infrastructure.

TorchNeuron could provide developers with a useful alternative to Nvidia’s ecosystem. Reducing the technical barriers to switching processors may encourage competition and lower costs.

Yet greater choice between chip suppliers does not necessarily mean greater independence from large technology companies.

Amazon is simultaneously expanding its position as a hardware developer, cloud provider and major energy customer. Businesses might become less dependent on one supplier while becoming more dependent on another.

The nuclear project introduces a different set of questions.

New low-carbon electricity could benefit the wider grid, support employment and contribute to reducing emissions. But the distribution of costs and benefits must be examined.

Who finances the infrastructure? How are electricity prices affected? What risks are carried by private investors, utilities and local communities?

Amazon says it intends to meet the costs associated with its data centres without transferring them to other electricity customers. Whether that commitment is achieved should be assessed through the project’s financing and eventual operation.

AI infrastructure is becoming economically and politically important. Its development cannot be judged only by the value it creates for technology companies.

The public also has an interest in how these resources are financed, controlled and shared.

Google and OpenAI accelerate scientific discovery. But who will control what AI discovers?

AI could transform medical research, from identifying new antimicrobial compounds to helping scientists investigate complex diseases. Its benefits, however, will depend on how discoveries are validated, controlled and used.

Google DeepMind and OpenAI announced separate scientific initiatives on 15 October, illustrating how artificial intelligence is moving beyond data analysis towards supporting the discovery process itself.

Google DeepMind described how researchers used its Antigravity platform to identify a potential antimicrobial peptide, a molecule that could help combat bacterial infections.

The system combined database searches, protein modelling and scientific literature to investigate previously uncharacterised biological sequences.

Using a protein language model called ProtNLM, researchers identified a 97-amino-acid candidate from the Lactococcin 972 family, with predicted activity against pathogens associated with respiratory infections.

The molecule is now undergoing laboratory validation. Its effectiveness and safety have not yet been demonstrated experimentally.

The potential advantage lies in the speed and scale of investigation. AI can help researchers search millions of biological sequences, identify promising candidates and prioritise laboratory experiments.

Meanwhile, OpenAI introduced its Rosalind Academic Plan, offering eligible US university laboratories sponsored access to GPT-Rosalind, a model designed to support life sciences research.

The programme provides up to 30 free ChatGPT Business seats per laboratory for a maximum of 12 months, with access to GPT-Rosalind subject to approval.

GPT-Rosalind is intended to assist research involving drug discovery, proteins, genetic therapies and disease detection.

The two announcements represent different developments. Google has demonstrated a computational research workflow that produced a candidate for further investigation. OpenAI is expanding access to tools intended to support scientific work.

Both could reduce the time needed to examine complicated biological questions.

However, neither announcement establishes that AI has produced a new clinically effective treatment.

Turning a promising molecular candidate into medicine still requires experimental testing, safety assessments and clinical research.

There is also a broader issue concerning the governance of scientific discovery.

Advanced biological models may support beneficial research, but some capabilities could also be misused. More effective tools for analysing biological systems can introduce biosecurity concerns depending on what information they enable users to generate and how that information is applied.

Scientific institutions and AI developers must therefore consider not only what their systems can discover, but also which capabilities should be accessible, under what conditions and with what safeguards.

Source: Google DeepMind

The Storyfront Perspective

The potential contribution of AI to medicine is substantial. Faster research could help scientists investigate diseases, identify promising treatments and explore biological questions that have resisted conventional methods.

But scientific progress is not determined by speed alone.

Who owns the models, controls access to the research and holds the rights to commercialise the results?

The distinction matters because much biological knowledge emerges from publicly funded universities, international collaborations and shared scientific databases. Commercial AI companies increasingly provide the infrastructure used to analyse that knowledge.

This creates a potential tension between the collective production of scientific knowledge and the private ownership of the technologies that help transform it into commercial products.

Faster discovery also does not guarantee wider access to treatments. Patents, pricing decisions and unequal healthcare systems can prevent scientific advances from reaching the people who need them.

There are ethical questions beyond ownership. Tools capable of investigating biological mechanisms may support medical breakthroughs, but greater capability can also increase the consequences of misuse.

The answer is not to discourage scientific research. It is to ensure that access, safety and public benefit remain central to its development.

AI could make an important contribution to humanity. Whether it does so will depend on who controls its discoveries and what purposes they ultimately serve.

Anthropic pledges 5% of its shares to public benefit. But who exactly is the public?

The AI developer wants to dedicate part of its growing valuation to society. The proposal could provide substantial resources, but the company has yet to explain who will control the fund or decide who receives its benefits.

Anthropic has announced plans to allocate an initial 5% of its shares to a new public benefit fund, with further contributions possible as the company’s valuation increases.

According to its 18 October announcement, the initial allocation is subject to shareholder approval.

The company’s board may add another 2% of shares for every 50% increase in Anthropic’s value above a specified baseline.

The proposed arrangement would link the fund’s potential resources to the commercial success of the AI developer.

Anthropic says it could support established nonprofit organisations, distribute resources directly to selected populations or address the effects of AI on employment.

The company has already committed $200 million to research and initiatives concerning labour-market disruption.

It argues that future AI systems could generate substantial wealth while also affecting workers and employment opportunities.

The fund is intended to provide a more flexible mechanism for responding to those changes.

However, its legal structure deserves attention.

Anthropic says the fund will be established as a wholly owned limited liability company rather than a conventional charitable organisation.

The company argues that this will allow greater flexibility in how resources are distributed.

It also says equity held by the fund will not be used for lobbying or political donations.

But the announcement does not identify the eventual beneficiaries, establish how they will be selected or explain who will have decision-making authority over distributions.

Anthropic has promised further details about governance and valuation thresholds in the coming months.

The company also says its founders and employees have made substantial separate commitments to donate equity.

Combined with the new proposal, Anthropic estimates that approximately 20% of its shares will be dedicated to public benefit, with potential additional allocations as its valuation grows.

This figure should not be confused with an immediate distribution of 20% of the company’s wealth.

Some commitments relate to founders’ and employees’ equity, while others depend on shareholder approval or future decisions.

Nor does the announcement establish how much money the intended beneficiaries will eventually receive.

That will depend on the value of the shares, the fund’s operation and the arrangements through which resources are distributed.

Source: Anthropic

The Storyfront Perspective

Anthropic’s proposal raises a question that goes beyond corporate generosity: what does public benefit mean when the definition is established by a private company?

A fund supported by valuable shares could finance worthwhile initiatives, particularly for workers and communities affected by technological change.

But the announcement does not yet establish who will make those decisions or whether the people most affected by AI will have a voice in them.

The distinction is fundamental.

A company’s shareholders own financial interests in the business. A public benefit fund is supposed to serve a broader purpose. Yet the public does not automatically gain ownership, representation or decision-making rights because a company commits shares to social objectives.

Anthropic’s proposed fund remains wholly owned by Anthropic. That does not prove it will benefit wealthy investors rather than ordinary people. It does, however, make its future governance central to evaluating the proposal.

Who will sit on its governing body? Will workers, researchers and affected communities be represented? Can beneficiaries challenge decisions? And will the company retain the final authority over what constitutes public benefit?

These questions matter because AI’s economic effects are not distributed equally. As Dr Lorena Nessi argues in her analysis AI Digital Divide: Who Really Benefits from Artificial Intelligence?, the divide “is between those who can use, shape and benefit from these systems, and those who remain dependent on decisions made elsewhere.”

The financial gains may accrue to investors and companies that own models, infrastructure and intellectual property, while some of the costs of automation may fall on workers and communities with little influence over those technologies.

Redistributing part of the resulting wealth could help address those inequalities.

But the ability to distribute wealth is not the same as sharing power over how that wealth is created.

A meaningful public benefit initiative should therefore be assessed not simply by the value of the shares committed, but by its independence, transparency and accountability to the people it claims to serve.

Barcelona opens supercomputing resources to AI startups. Can public investment challenge Big Tech’s advantage?

Access to advanced computing remains one of the biggest obstacles for smaller AI developers. A European initiative aims to reduce that barrier, raising questions about whether publicly supported infrastructure can produce broader economic opportunities.

Barcelona’s Supercomputing Center is expanding support for artificial intelligence startups through an accelerator programme designed to improve access to computing infrastructure, technical expertise and specialised training.

The initiative is part of Europe’s developing network of AI Factories, which brings together supercomputing facilities and support services intended to help researchers, businesses and public organisations develop AI applications.

According to a 15 October announcement from Nvidia, Barcelona’s AI Factory has completed a pilot initiative and is rolling out its first full accelerator programme.

The programme is intended to support emerging AI businesses and strengthen the development of technological capabilities within Spain and the wider European market.

The collaboration forms part of broader efforts to connect publicly supported European computing facilities with startups that would otherwise struggle to obtain sufficient processing capacity.

Computing resources have become increasingly expensive and strategically important as AI models grow more capable.

Large technology companies can invest billions in infrastructure and secure long-term access to specialised processors. Smaller developers often face considerably greater constraints.

These barriers affect which organisations can experiment with advanced models, develop new products and compete commercially.

Supercomputing centres could help address that imbalance by providing access to infrastructure that individual startups could not reasonably build themselves.

Nvidia is supporting related European initiatives through its Inception programme, which offers eligible startups technical assistance and access to parts of its computing ecosystem.

The company also reports collaboration with AI Factories in France and Sweden, alongside European scientific computing institutions.

Barcelona’s programme therefore forms part of a broader attempt to strengthen the region’s AI capabilities.

However, access to infrastructure does not automatically establish commercial competitiveness.

Startups also require skilled employees, financing, customers and the ability to maintain their systems after initial support ends.

The long-term success of the programme will depend on which organisations receive assistance, what they can develop and whether publicly supported access results in sustainable businesses.

Source: Nvidia

The Storyfront Perspective

Public investment in computing infrastructure offers an important response to the concentration of economic power in AI.

Unlike industries where relatively modest resources can support experimentation, advanced AI increasingly requires expensive processors, specialised expertise and substantial operational capacity.

This creates structural barriers that market competition alone may not resolve.

Providing startups with access to public supercomputing facilities could widen participation and allow researchers and entrepreneurs to pursue ideas that would otherwise remain financially inaccessible.

But the distribution of those opportunities matters.

If public resources primarily support a small group of well-financed startups, the initiative could reinforce existing inequalities rather than reduce them.

There is also a distinction between public access and technological independence.

European computing centres may provide alternatives to commercial cloud services while continuing to depend on hardware and software developed by multinational companies.

Such partnerships can be productive. They should nevertheless be assessed against the public objectives used to justify the investment.

A meaningful measure of success would be whether the programme supports a more diverse AI economy, creates sustainable opportunities and allows smaller organisations to retain control over the technologies they develop.

Microsoft wants AI agents to supervise other AI agents. Can automated oversight make them safer?

Microsoft is developing systems that monitor autonomous AI inside organisations and restrict their actions according to risk. The approach could improve security, provided automated supervision is supported by enforceable controls and independent assessment.

Microsoft has outlined a security approach that uses specialised AI agents to monitor other autonomous agents operating inside businesses and organisations.

In a technical announcement published on 19 October, Hayete Gallot, Microsoft’s executive vice president for security, described the need to treat autonomous AI systems as active participants within corporate environments.

These systems may access sensitive information, use software tools and carry out tasks without requiring human approval at every stage.

That autonomy creates security challenges that differ from those associated with conventional software.

An AI agent may be authorised to perform a task but take unexpected actions while attempting to complete it.

It could access information beyond what is necessary, interact with another system in an unintended way or continue operating after the conditions surrounding its task have changed.

Microsoft argues that security measures must therefore operate continuously, rather than relying solely on permissions established when an agent is deployed.

The company proposes using what the industry calls Guardian Agents, specialised systems intended to observe and respond to the actions of other agents.

These monitoring systems form part of the broader security architecture associated with Project Perception, which Microsoft originally announced in July.

Microsoft says the architecture can combine signals from identity systems, networks, data services, cloud environments and agent runtimes.

This information can be used to assess risk dynamically.

For example, an agent performing routine administrative work might initially have permission to access a limited set of documents.

If it subsequently attempted to reach sensitive information unrelated to its task, the security system could identify the change in behaviour and apply additional restrictions.

Microsoft describes this as combining AI-based monitoring with deterministic controls, meaning rules that can restrict actions regardless of how an AI model interprets a task.

The distinction is important.

An AI system assessing risk may make errors. Explicit permissions and technical boundaries can provide another layer of protection by preventing certain actions even when an automated assessment is inaccurate.

Microsoft also argues that organisations need better visibility into why agents behave in unexpected ways.

Its proposed approach includes tools intended to help security teams examine events, identify risks and adjust organisational policies.

However, the announcement does not establish that Guardian Agents can reliably identify and prevent every harmful action.

Nor does it provide an independent evaluation demonstrating that the approach is effective across different autonomous systems and organisational environments.

The technology represents a developing security architecture, not a guarantee that autonomous AI can operate without human oversight.

Source: Microsoft

The Storyfront Perspective

Using AI to supervise other AI systems is not inherently contradictory.

Autonomous agents may operate too quickly, and at too great a scale, for people to examine every individual action. Automated monitoring could help organisations recognise unusual behaviour and intervene before problems spread.

Microsoft’s emphasis on enforceable permissions is therefore important.

Security should not depend entirely on whether another AI model correctly interprets an agent’s intentions. Systems also need technical limits that cannot simply be overridden by persuasive instructions or unexpected behaviour.

The more difficult question concerns where authority lies.

When an organisation uses a platform to deploy AI agents, monitor their activities and enforce restrictions, a significant part of its operational security may depend on that platform provider.

This introduces a different form of technological dependence.

The same company could supply important parts of both the automation and the infrastructure governing its use.

That arrangement might improve consistency and simplify security management. It could also make independent scrutiny more difficult if organisations cannot adequately examine the rules, evidence and limitations behind automated decisions.

The answer is not necessarily to reject automated supervision.

It is to distinguish systems that assist organisations in exercising control from systems that require organisations to surrender control over their own security decisions.

The effectiveness of Guardian Agents should ultimately be measured through independent testing, transparent limitations and the ability of organisations to enforce their own policies.

Compiled with AI assistance. Written and checked by a human editor before publication.

Amazon challenges Nvidia with its own AI chips and nuclear power ambitions. Who will control AI’s infrastructure?

Google and OpenAI accelerate scientific discovery. But who will control what AI discovers?

Anthropic pledges 5% of its shares to public benefit. But who exactly is the public?

Barcelona opens supercomputing resources to AI startups. Can public investment challenge Big Tech’s advantage?

Microsoft wants AI agents to supervise other AI agents. Can automated oversight make them safer?

Five AI stories that matter, and what they mean. Every weekday.