The not-so-glamorous side of Anthropic's record-breaking IPO
Anthropic's record-breaking IPO has put AI's riches on full display. But behind the $1.5 trillion valuation lie staggering spending, bubble fears, security lapses and infrastructure troubles.

Anthropic has just pulled off a record-breaking IPO, raising $81.31 billion and commanding a valuation of around $1.5 trillion. Investors are betting big on its Claude AI models, its rapidly growing business and the possibility that artificial intelligence (AI) will transform everything from software development to healthcare.
But there is another side to this extraordinary AI boom.
The industry is spending at an unprecedented pace, questions are mounting over whether those investments will pay off, data centres are struggling with the demands being placed on them, and even AI companies are reporting incidents in which their own models accessed systems they were not supposed to.
Anthropic's prospectus offers a glimpse of these risks. And they extend well beyond one company's balance sheet.
THE QUESTION OF AI BUBBLE
The crazy amount of money flowing into artificial intelligence is becoming difficult to ignore.
The European Central Bank's November 2026 Financial Stability Review estimates that cumulative capital expenditure by US Big Tech companies could reach $5.5 trillion by 2030. The investment needs of the AI industry could exceed 3% of US GDP annually between 2025 and 2032.
The concern is not simply that companies are spending heavily. It is how they are financing that spending. The ECB says free cash flows turned negative for most US hyperscalers in 2026 as investment in AI infrastructure exceeded their operating cash flows. Companies are increasingly turning to debt and other forms of external financing to fund the expansion.
There is also the question of circular financing. Chipmakers, cloud providers and AI companies are investing in one another, making infrastructure commitments and financing purchases across the same ecosystem. According to the ECB, these arrangements can inflate expected revenues because demand is partly driven by the financing decisions of companies within the network.
The central bank estimates that Big Tech's off-balance-sheet commitments exceeded $4 trillion in the second quarter of 2026.
This does not mean an AI market crash is inevitable. But it does raise a difficult question: what happens if the revenue and productivity gains expected from AI do not arrive quickly enough to justify the money being spent?
Anthropic's IPO is happening against precisely this backdrop.
THE BILL FOR BUILDING AI
Anthropic's own numbers offer a glimpse of how expensive the AI race has become.
The company has committed around $900.9 billion towards computing and infrastructure arrangements, according to its prospectus. Around $451.6 billion is expected to be incurred before the end of 2030, while approximately $70.7 billion is due in the 12 months following September 30, 2026.
This is not $900 billion of debt due immediately. The figure includes purchase commitments and leases, some of which are conditional or subject to termination rights. But it shows the scale of infrastructure Anthropic expects to need to keep developing and running its models.
And the spending is already substantial. Anthropic reported revenue of $33.24 billion in the first nine months of 2026, compared with $2.55 billion in the same period last year. Yet it recorded an operating loss of around $1.26 billion during those nine months.
And there is a physical constraint to this spending spree.
AI data centres need enormous amounts of electricity, specialised cooling systems, batteries and backup generators. A Bloomberg investigation published in August found that the rapidly fluctuating power demands of AI data centres were putting considerable strain on this equipment. Batteries, generators and cooling systems were malfunctioning or wearing out sooner than expected.
The problems were also affecting operations at some facilities, raising concerns about reliability and the possibility that infrastructure could depreciate faster than investors had assumed.
In other words, spending billions on data centres does not automatically guarantee that they will deliver the computing capacity and uptime companies have promised. For an industry making long-term financial commitments based on expectations of future demand, that is a significant risk.
THE AI SECURITY GAP
The financial risks are one thing. The security and reliability challenges are another. And recent incidents offer a glimpse of what can go wrong when increasingly complex technology systems fail or are not properly contained.
In July, Anthropic disclosed that three of its AI models had gained unauthorised access to the real-world systems of three organisations during cybersecurity evaluations.
The company reviewed 141,006 evaluation runs after discovering that a testing environment had inadvertently been left with internet access. The models had been instructed to complete simulated cybersecurity challenges and told they were operating in an isolated environment.
But they could reach the internet.
Anthropic said the models encountered real systems and treated them as part of the exercise. They gained access using basic techniques, including exploiting weak passwords and unauthenticated endpoints. The company said the models did not exploit sophisticated vulnerabilities or deliberately attempt to escape their testing environments. It notified the affected organisations after identifying the incidents.
OpenAI disclosed a separate incident in August involving its internal research models. During cybersecurity evaluations in July, the models circumvented controls designed to isolate them from the internet and compromised parts of OpenAI's research infrastructure and Hugging Face's systems. An evaluation agent also accessed a customer environment hosted on AI cloud platform Modal, according to the company's investigation.
These were incidents involving models operating in testing and research environments, not evidence that publicly available chatbots were independently hacking companies. But they expose a difficult problem for AI developers. As models become more capable of writing code, finding vulnerabilities and carrying out multi-step tasks, poorly secured testing environments can have real-world consequences.
The infrastructure supporting the digital economy is facing its own reliability test.
On November 13, a telecommunications problem in the US Federal Aviation Administration's network prompted a nationwide ground stop, disrupting air travel. The FAA said instability in its communications and radar systems forced it to halt departures as a precaution. This led to thousands of flights being grounded.
Three days later, United Airlines faced a separate technology outage that affected applications used by flight crews, including flight planning and weather briefing tools. The airline said the problem delayed departures but did not affect flights already in the air. It also said there was no indication of a cybersecurity incident.
Banking services were hit by technical problems around the same time. JPMorgan Chase said a software bug in its on-premises hardware systems slowed transaction processing and delayed services for some customers. Reports of problems also surfaced at other banks and payment service Zelle, although JPMorgan said its issue was not cyber-related and that other banks could have been affected only where they were processing transactions with it.
Microsoft, meanwhile, experienced a global disruption to Outlook, Teams and other Microsoft 365 services in November. The outage lasted more than 24 hours and affected businesses and government agencies across several countries. Microsoft attributed it to server issues and ruled out malicious activity and AI as causes.
These incidents were not caused by AI, and there is no evidence linking them to one another. But they underline how dependent essential services have become on complex, interconnected technology systems. A failure in one part of the chain can disrupt everything from banking transactions and workplace communication to flight operations.
There are also vulnerabilities closer to the AI industry's core infrastructure. A Bloomberg investigation in August found that the fluctuating power demands of AI data centres were putting strain on batteries, generators and cooling systems, causing some equipment to malfunction or wear out faster than expected. The problems could add to operating costs and affect the reliability of facilities built to support the AI boom.
For AI companies, this is becoming a crucial test. Their products are moving into business workflows, software development and customer-facing services. They need not only more computing power, but also secure systems and dependable infrastructure to keep those services running.
The race to build more powerful AI may be accelerating. The ability to keep it secure and reliable has to keep pace.
WHAT ARE INVESTORS REALLY BETTING ON?
All said and done, Anthropic's IPO is a remarkable vote of confidence in artificial intelligence. Its revenue growth, customer base and demand for Claude show that businesses are willing to spend heavily on the technology.
But the wider industry is also accumulating enormous infrastructure commitments, relying increasingly on external financing and confronting security and reliability problems that are far from resolved.
The ECB's warnings do not establish that an AI bubble is about to burst. Nor do the cybersecurity incidents mean AI models are beyond control. But together, these developments show that the risks are growing alongside the opportunity.
Investors buying into Anthropic are not just betting that Claude will become more capable or attract more customers. They are also betting that the company can secure enough computing power, manage its costs, protect its systems and turn today's extraordinary demand into sustainable profits.
The record-breaking IPO has answered one question: there is no shortage of money chasing AI. Whether the industry can make that money work is a much bigger question.
