Friday, 20 November 2026 Follow @AIDailyFive

The AI Daily Five · No. 30 · Thursday 19 November

What Does AI at Scale Really Cost?

From Anthropic’s record IPO and a cyberattack on European Commission staff systems to Google’s Gemini 4 Neon, Singapore’s tokenised central-bank bills and AI-generated criticism in peer review, today’s edition asks what AI at scale really costs.

The Marina Bay Sands skyline under a Singapore flag beside a glowing IPO screen with linked blockchain cubes.
Illustration: AI Story Front design team

1

Anthropic’s Record IPO Tests AI’s Economics

The Claude developer is heading to Nasdaq after pricing the largest initial public offering on record. Its financial disclosures reveal extraordinary demand for AI services alongside enormous infrastructure commitments.

Anthropic has priced its initial public offering at $83 per share, raising approximately $81.3 billion in what is expected to become the largest IPO in history.

The company announced the final pricing on 18 November, with its shares scheduled to begin trading on Nasdaq on Thursday under the ticker ANT.

The offering comprises 980 million shares, including approximately 956.4 million sold by Anthropic and 23.6 million sold by existing shareholders. The transaction values the developer of Claude at approximately $1.5 trillion on a fully diluted basis.

The listing tops the initial size of SpaceX’s record-setting offering earlier this year and represents one of the most significant public-market tests of investor confidence in artificial intelligence.

Yet Anthropic’s financial disclosures reveal a more complicated business than the valuation alone suggests.

The $900 Billion Infrastructure Commitment

Anthropic’s IPO prospectus describes approximately $900.9 billion in aggregate computing-related purchase commitments and lease arrangements.

These are contractual commitments extending over many years, not money already spent or a forecast of expenditure in a single financial year.

Approximately $70.7 billion is scheduled to fall due within 12 months of 30 September 2026, while the remaining commitments extend across subsequent years.

The figures reflect the enormous capital requirements involved in developing and serving advanced AI models.

Anthropic is also changing how it obtains computing resources. Previously dependent heavily on cloud providers, the company is developing a model under which it leases chips directly and uses dedicated facilities operated on its behalf.

The strategy may offer greater control over future infrastructure costs. It also commits the company to substantial expenditure before all the associated computing capacity becomes commercially productive.

Growth Does Not Eliminate Financial Risk

The IPO follows rapid revenue growth as enterprises adopt Claude for coding, research and business processes.

However, Anthropic remains unprofitable under standard accounting measures.

Its filing describes substantial operating losses and a business whose future financial performance depends partly on generating enough revenue to support continued investment.

The company also has commercial relationships with major technology providers that occupy several positions in the same market.

Amazon and Google’s parent company Alphabet have been investors, suppliers and distribution partners, while developing competing AI services.

As Reuters previously reported from Anthropic’s IPO disclosures, these overlapping relationships create both opportunities and commercial dependencies.

There is another important distinction in the company’s governance structure.

Anthropic is organised as a public-benefit corporation, with a long-term benefit trust designed to influence board composition. Its structure is intended to preserve commitments beyond shareholder returns, although its practical effectiveness will face greater scrutiny as the company becomes publicly traded.

2

EU Cyberattack Puts Staff Data at Risk

An attack on the European Commission has disrupted internal systems and raised concerns about the exposure of employees' personal information. Officials say the incident is largely contained, although the damage remains uncertain.

The European Commission is investigating a cyberattack that has disrupted internal services for around ten days and may have exposed sensitive personal information.

The incident affected systems used by employees to manage salaries, working hours, leave and medical reimbursement claims.

According to Reuters reporting on 19 November, the Commission believes it has contained most of the attack, but officials have not established the full extent of any resulting damage.

The attack was first reported in detail by Politico, which cited internal communications and EU officials familiar with the response.

Commission spokesperson Thomas Regnier confirmed that individual systems had been targeted and that affected parties had been informed.

What Systems Were Affected?

The disrupted applications included Sysper, the Commission’s internal human resources platform, and MyPMO, which handles services including medical expense reimbursements.

Both systems were reportedly restored earlier in the week, although Sysper experienced another interruption on Wednesday.

An emergency meeting of the Commission’s highest-level internal crisis-management body was convened to address the incidents. EU technology commissioner Henna Virkkunen subsequently briefed colleagues on the response.

The reported disruption also raised concerns about whether attackers could have accessed information contained within the affected systems.

That remains an important distinction.

The interruption of an application does not automatically mean that its underlying data has been stolen. Investigators have yet to establish publicly how much information, if any, was extracted.

Reports have also suggested that attackers used publicly available AI-assisted vulnerability-discovery technology. The precise role of such tools has not been independently established by a detailed technical account from the Commission.

Europe’s Digital-Security Responsibilities

The incident comes as the EU continues implementing an extensive cybersecurity and data-protection framework.

Its NIS2 Directive establishes requirements for cybersecurity risk management and incident reporting across critical sectors in member states.

Separately, EU institutions operate under their own cybersecurity framework, with CERT-EU supporting incident response and coordination.

The distinction matters because rules applying to private companies and national authorities are not necessarily identical to those governing EU institutions.

The Commission’s responsibilities extend beyond protecting its own information systems. It plays an influential role in developing the standards that European organisations are expected to follow.

The incident therefore invites examination of public-sector preparedness without assuming that any particular regulatory requirement was breached.

3

Google Gemini 4 Neon Targets Real-World AI Work

Google’s latest model aims to complete demanding professional tasks more efficiently, with improvements in coding and multi-step reasoning. Independent testing suggests the gains are meaningful but less dramatic than the launch rhetoric.

Google has launched Gemini 4 Neon, expanding its AI model family with a system designed for software engineering, scientific reasoning and complex professional workflows.

Announced on 18 November, Neon is intended to provide advanced reasoning capabilities while completing tasks more quickly and economically than some larger competing models.

It is rolling out through the Gemini API, Google AI Studio, business services and selected paid consumer subscriptions.

The release follows a rapid series of model announcements from Google and its competitors. Rather than simply introducing another larger model, Google is emphasising the relationship between capability, cost and the time required to complete a task.

How Does Neon Perform?

On the Artificial Analysis Intelligence Index, Gemini 4 Neon achieved a score of 61 at its highest reasoning setting.

That placed it among the leading models evaluated at launch, although not at the top of the overall ranking.

More revealing is the comparison with Google’s existing Gemini 4 Carbon.

According to independent analysis published on 18 November, Neon improved the intelligence score by just one point, while reducing the cost and time associated with completing benchmark tasks.

Artificial Analysis reported a cost of approximately $6.44 per task, compared with $7.76 for Carbon at its standard pricing.

The testing also found that Neon completed the measured workload in approximately 21 minutes, compared with around 36 minutes for Carbon.

These figures help distinguish improvements in measured efficiency from claims of substantially greater intelligence.

Google also highlights performance on professional evaluations involving legal research, coding and complex instructions.

Neon’s API pricing begins at $3 per million input tokens and $15 per million output tokens, with a context window supporting up to one million tokens.

However, the cost of an individual token does not reveal the total expense of completing a task. Models that generate longer responses, perform additional reasoning or require repeated attempts can become more expensive even when their listed prices appear competitive.

AI Competition Becomes More Practical

Google says Neon can process several types of information, including documents, spreadsheets, audio and video.

The model also supports multiple reasoning settings, allowing developers to choose different combinations of performance and resource consumption.

This matters for businesses developing applications that must complete work reliably within operating budgets.

An organisation handling thousands of customer enquiries may value predictable costs and fast responses more than a marginal improvement on an academic benchmark.

A research team solving a difficult technical problem may make a different choice.

Google has additionally highlighted cybersecurity applications and access through its Fairwind programme for eligible defenders.

These capabilities remain subject to deployment conditions and safety restrictions.

4

Singapore Tests Blockchain in Government Debt

The city-state has completed a live issuance of tokenised central-bank bills using digital central-bank money for settlement. The transaction offers concrete evidence of institutional blockchain adoption, although broader market use remains limited.

Singapore’s central bank has announced the successful issuance of S$100 million in tokenised Monetary Authority of Singapore bills, using wholesale central bank digital currency to settle the transaction.

The announcement on 19 November confirms that the four-week securities were issued on 6 November through the authority’s Singapore Dollar Test Network.

Five banks participated in the auction: DBS, HSBC Singapore, OCBC, Standard Chartered Singapore and UOB.

Applications exceeded twice the issuance amount, according to the Monetary Authority of Singapore (MAS).

Unlike many digital-asset initiatives announced as future projects, this transaction has already taken place.

What Makes the Transaction Different?

MAS bills are short-term debt securities issued by Singapore’s central bank as part of its monetary operations.

In the new transaction, the securities were created and recorded directly on distributed-ledger infrastructure.

This differs from issuing a conventional security and subsequently creating a digital representation of it.

The settlement process used wholesale central bank digital currency, a form of digital central-bank money intended for transactions between authorised financial institutions.

MAS says this approach can reduce the need for reconciliation and certain intermediary processes.

The distinction is important because transferring tokenised securities does not automatically guarantee that the corresponding payment is completed.

Combining the securities and settlement infrastructure can potentially reduce the operational complexity of transferring financial assets and money between institutions.

The bills are scheduled to mature on 4 December 2026.

What Happens Next?

MAS plans to expand the number of participating institutions and explore the use of tokenised government securities in collateral management.

Collateral plays an important role in financial markets, allowing institutions to secure borrowing and other transactions using assets whose value can be recognised by counterparties.

Making those assets easier to transfer and verify could improve some financial operations, particularly when several institutions need to coordinate.

However, a successful issuance among five participating banks does not establish that the technology is ready for universal adoption.

Legal certainty, interoperability and operational resilience remain significant questions.

Speaking at the Singapore FinTech Festival on 19 November, MAS managing director Chia Der Jiun acknowledged that tokenised financial assets remain small relative to conventional markets.

MAS also published additional frameworks covering tokenised collateral and digital-money settlement arrangements.

The regulator intends to expand live trials during 2027.

5

Is AI Overloading Scientific Peer Review?

New commentary in Nature Machine Intelligence argues that generative AI can produce scientific criticism faster than researchers can evaluate it. The concern is about the quality of academic judgement, rather than the quantity of feedback.

Artificial intelligence is changing the economics of scientific peer review, potentially making it easier to generate objections to research than to determine whether those objections matter.

That is the argument advanced in a correspondence published in Nature Machine Intelligence on 18 November.

The authors, Yiteng Sun, Fan Li, Ching-Hung Lee and Chun-Hsien Chen, argue that AI-assisted tools can generate substantial volumes of manuscript-specific feedback in seconds.

Determining whether each criticism exposes a genuine problem, however, still requires scientific expertise and time.

This creates a potential imbalance in an academic system where reviewers’ comments carry considerable influence over publication decisions.

When Criticism Becomes Inexpensive

Peer review is designed to test the quality of research before publication.

Independent specialists examine a manuscript’s arguments and methods, identify possible errors and recommend improvements.

Traditionally, formulating detailed criticism required substantial intellectual effort. That effort naturally limited the volume of questions a reviewer might reasonably raise.

Generative AI changes the calculation.

A reviewer can use AI to identify possible methodological weaknesses, challenge interpretations and produce extensive comments quickly.

Some of those observations may be valuable. Others may be irrelevant, inaccurate or disproportionate to the paper’s actual claims.

Authors must then devote time to assessing the criticisms and deciding whether revisions are necessary.

The researchers suggest that peer-review systems need an obligation threshold: a clearer distinction between comments that genuinely require authors to respond and observations that do not materially affect the study.

This is a conceptual proposal, rather than an experimentally demonstrated solution.

The Debate Over AI in Academic Publishing

The concern is part of a wider discussion about the role of generative AI in scientific publishing.

Nature Portfolio’s guidance on artificial intelligence states that reviewers remain responsible for their assessments and must protect confidential manuscript information.

The publisher permits certain forms of AI assistance while insisting that expert judgement cannot be delegated to automated systems.

Earlier this year, a Nature Nanotechnology editorial argued that AI can support peer review, provided its output is validated by human experts.

Another contribution to Nature in August argued that blanket bans on AI assistance risk ignoring the practical challenges facing researchers and journal editors.

These perspectives are not necessarily contradictory.

AI may help reviewers identify genuine problems, while simultaneously increasing the number of comments that authors must investigate.

The important distinction concerns how the technology is used and whether its output improves the evaluation of scientific evidence.

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