The AI Daily Five · No. 24 · Wednesday 11 November
Who Sets the Rules for AI?
From Nvidia’s chip deals in Asia and AI-verified Signal encryption to Anthropic’s IPO, tokenised European shares and the Paris AI safety declaration, today’s edition asks who sets the rules for AI.

1
Did Nvidia’s China Chip Deals Cross a Line?
New reporting connects a former Megaspeed executive at the centre of an alleged espionage investigation with substantial Nvidia chip transactions. The controversy has widened into questions about US export controls, corporate governance and political accountability.
The investigation surrounding Nvidia chief executive Jensen Huang has taken a more consequential turn. New reporting has shifted attention from allegations about his personal relationship with a Chinese businesswoman to the movement of advanced AI chips through companies operating between Singapore and China.
The Financial Times reported on 10 and 11 November that Alice Huang, a former executive and shareholder of Singapore-based Megaspeed, had helped develop business relationships involving Nvidia processors and Chinese technology companies.
Huang, who is not related to Jensen Huang, is alleged in earlier reporting to have had a relationship with the Nvidia chief executive and to have been connected to a Chinese intelligence operation. Those allegations remain contested, and there is no public finding that Jensen Huang knowingly provided intelligence or confidential information.
The Financial Times investigation examines how Megaspeed became a significant intermediary in the international AI computing market.
According to financial records reviewed by the newspaper, Megaspeed’s balance sheet grew from about $33 million to around $3 billion during 2024. Alibaba was reportedly a major customer, while related financing arrangements helped support the company’s rapid expansion.
The distinction between supplying computing services and exporting restricted hardware is central to the case. US rules restrict the shipment of certain advanced processors to China, but Chinese businesses may, under specified circumstances, lease computing capacity from facilities located elsewhere.
That arrangement creates a complicated enforcement problem. Authorities must establish not merely where a chip was purchased, but how computing infrastructure is owned, operated and accessed.
On 10 November, Reuters relayed a report by The Information alleging that Alice Huang and her associates helped move at least $7.5 billion in Nvidia chips to Chinese companies. Reuters said it could not independently verify the report.
Congressional Scrutiny Intensifies
The controversy is also becoming a political issue.
Republican congressional committee chairs previously requested a closed-door briefing from Jensen Huang on 18 November concerning potential foreign intelligence activity and Nvidia’s security procedures.
More recently, Democratic Senate leader Chuck Schumer accused President Trump of ignoring an intelligence warning about Jensen Huang more than a year ago, calling it a “dereliction of duty”, and criticised the decision to give China access to advanced AI chips.
The company’s independent directors have retained outside counsel to investigate. Nvidia says it has no reason to believe its trade secrets or US national-security information were compromised.
Megaspeed has denied wrongdoing, while China’s foreign ministry has rejected allegations of espionage.
These denials and the unresolved investigations are significant. The available reporting establishes that serious questions have been raised about Nvidia’s commercial relationships and the effectiveness of export controls. It does not establish that every disputed shipment was illegal or that the company knowingly participated in espionage.
Nvidia occupies an unusually important position in global AI infrastructure. Its processors are central to the computing systems used by leading AI developers, while governments increasingly treat access to advanced chips as a national-security concern.
The Megaspeed controversy illustrates how difficult those controls may become when hardware, financing and computing services move through international networks of companies.
2
Can AI Prove Encrypted Messages Are Secure?
Google DeepMind researchers have used AI agents to produce mathematical proofs for important parts of Signal’s encryption software. The work offers a new approach to cybersecurity, although its guarantees are limited to the properties and code examined.
Most people trust encrypted messaging applications without ever seeing the software that protects their conversations. Even security specialists can struggle to establish whether complex encryption code implements its mathematical design correctly.
Google DeepMind has demonstrated a new approach to that problem using artificial intelligence.
On 10 November, the company introduced AlphaProof Code, a system of AI agents designed to help researchers construct mathematical proofs about software.
The team used the system to verify important security properties of libsignal, the reference implementation of the Signal messaging protocol.
Unlike conventional software testing, which checks how a programme behaves in selected situations, formal verification uses mathematical reasoning to establish that specified properties hold under defined assumptions.
In its technical research, DeepMind reports that AlphaProof Code generated approximately 6.5 million lines of machine-checked proof code over four weeks.
The researchers examined parts of Signal’s post-quantum PQXDH handshake and Triple Ratchet protocol, technologies designed to protect communication even when attackers obtain certain information or when future computing capabilities become more powerful.
Their proofs address properties including authentication, forward secrecy and post-compromise security.
Forward secrecy helps protect previously exchanged messages if a key is later compromised. Post-compromise security concerns the ability of a communication system to recover protection after certain forms of compromise.
What Exactly Did the AI Prove?
The researchers combined human expertise with AI agents capable of proposing specifications and constructing proofs.
Those proofs were checked using Lean, a mathematical proof assistant whose verification kernel determines whether the supplied reasoning follows the accepted rules.
This distinction matters. The AI generated the proof material, but the mathematical verification did not depend on simply trusting the AI’s conclusions.
The system also identified problems when researchers deliberately reintroduced two previously corrected vulnerabilities into test code.
However, the result is not a guarantee that Signal can never be compromised.
Formal verification establishes particular properties of specified software under stated assumptions. It cannot rule out every implementation error, compromised device, malicious application or weakness outside the verified components.
Nor does generating millions of lines of proof establish that the process is inexpensive or easily reproduced by other organisations.
The researchers have released verification artefacts and supporting resources, allowing other specialists to examine the methods.
The development suggests that AI could help make rigorous software verification more practical, particularly for systems where failures could expose sensitive information.
3
Who Stands to Gain From Anthropic’s IPO?
The Claude developer plans one of the largest stock-market flotations in history, with major technology suppliers also committing $26 billion in private investment. The proposed deal highlights the financial relationships underpinning the AI industry.
Anthropic has formally launched the investor roadshow for its proposed initial public offering, seeking to sell 825 million shares at between $70 and $78 each.
The Claude developer announced the price range on 10 November, marking a new stage in its planned listing on the Nasdaq stock exchange under the ticker ANT.
At the midpoint of $74 per share, the offering would be worth approximately $61 billion. However, not all those proceeds would go to Anthropic.
According to its stock-market announcement, the company plans to issue approximately 801 million new shares, while existing shareholders would sell almost 24 million. Anthropic would receive proceeds from its newly issued shares, after expenses, but not from shares sold by existing investors.
The proposed flotation comes as investors question whether the extraordinary spending on AI infrastructure can generate sufficiently large returns.
Anthropic’s revenue has grown rapidly as businesses adopt its models for programming and professional work. Nevertheless, the company remains unprofitable while investing heavily in computing capacity and development.
The IPO would expose those economics to greater public-market scrutiny.
Its Suppliers Are Also Becoming Investors
One of the most significant details appears in Anthropic’s SEC filing.
Seven major technology companies have agreed to purchase a combined $26 billion of Anthropic shares through a separate private placement associated with the IPO.
They include Google, Microsoft, Nvidia, AMD, Broadcom, Micron and TSMC.
These companies occupy different positions across the computing infrastructure that AI developers need, from cloud services to advanced processors and semiconductor manufacturing.
Their investment commitments could strengthen Anthropic’s financial position, but they also illustrate how closely connected the largest AI companies have become.
Some organisations developing AI models receive financing from businesses that also supply the equipment and infrastructure needed to train and operate those models.
The arrangements are not necessarily improper. Technology companies routinely invest in customers and commercial partners.
Nevertheless, such relationships can complicate assessments of market competition and demand, particularly when investment capital eventually supports purchases from the same corporate ecosystem.
Anthropic is incorporated as a public benefit corporation, a structure that requires it to consider its stated public-benefit purpose alongside shareholder interests.
That governance model will face a significant test as the company enters public markets, where investors may place considerable pressure on growth and financial performance.
The offering is not yet complete. Its final price, proceeds and timing remain subject to market conditions and regulatory requirements.
5
Can Countries Agree on AI Safety Standards?
France has brought together 16 governments and ten technology companies around a new declaration on advanced AI safety. The agreement supports independent evaluation and incident reporting, but contains no binding enforcement mechanism.
Sixteen governments and ten AI companies have endorsed an international declaration calling for stronger safety standards, more transparent evaluations and better cooperation when advanced AI systems create serious risks.
The Paris Call for AI Safety and Security was signed on 9 November and published by the French presidency.
The government signatories include France, Canada, Ireland, the Netherlands, Norway and Kenya, alongside ten other countries.
Among the corporate signatories are OpenAI, Anthropic, Microsoft, Mistral AI, Cohere and Hugging Face.
According to the official declaration, participants agreed on three broad areas for cooperation: evaluating frontier AI systems, clarifying responsibilities for safety and developing systems for reporting serious incidents.
One objective is to establish more comparable methods for assessing advanced models.
AI developers currently use different benchmarks and testing procedures, making it difficult for regulators and researchers to compare the capabilities and risks of competing systems.
The declaration proposes common technical standards and a public repository where non-sensitive evaluation findings could be shared.
It also addresses the growing autonomy of AI systems, including the possibility that models may act in unexpected ways or become difficult to control.
Independent Evaluation and Emergency Safeguards
Signatory companies commit to disclosing relevant model capabilities, limitations and known risks.
They also support evaluation by qualified independent specialists where appropriate, alongside internal monitoring and testing procedures.
Another provision calls for developers to maintain mechanisms that can restrict, pause or shut down model operations under their control.
These measures are intended to address situations where a model demonstrates dangerous capabilities or behaves in ways that existing safeguards cannot adequately manage.
The declaration also recognises the importance of whistleblower protections and confidential channels for reporting concerns.
Governments are encouraged to establish arrangements that allow employees and other individuals to raise safety concerns without facing retaliation.
However, the agreement is voluntary.
It does not establish a new international regulator, create enforceable penalties or set firm deadlines for every commitment.
The declaration also leaves several implementation details unresolved, including how independent assessments would be organised, which findings must become public and what consequences would follow if a participant failed to comply.
The United Kingdom and United States are not included in the initial published list of government signatories, although the agreement remains open to further endorsements.
The significance of the initiative will therefore depend on whether participating governments and companies translate the declaration into practical requirements and transparent reporting.
The AI Daily Five