AI Debt Soars as Hyper-Scalers Issue Bonds to Fund Data Centers

This article looks at aspects of the feud between Anthropic and the US government.

• Two months ago, Anthropic developed the Claude Mythos model and then refused to release it to the public because the model was considered too good at finding code vulnerabilities and writing exploits for these. The model was only released to a selected group of software manufacturers and cybersecurity researchers.
• Anthropic released a sanitized version of Mythos earlier this month called Fable. The US government then told Anthropic that the Fable model was a threat to national security and placed export controls on it. Anthropic then revoked access to the model.
• There are doubts that the federal ban will survive legal scrutiny since allowing access might not be considered “exporting”.
• There is also the question that many key Chinese models are open-source, and bans on US model exports play into Chinese interests.
• Many cybersecurity researchers also believe that US interests are best served by making the Anthropic models available since they will help companies shore up defenses against cyberattacks.

Following on the non-release of Anthropic’s Claude Mythos model, a body made up of intelligence agencies from Australia, the US, the UK, New Zealand and Canada made a statement warning of the potential cyber-damage that emerging language models can cause.

• The statement writes: “Frontier AI models are anticipated to exceed current industry expectations, fundamentally transforming both offensive and defensive cyber capabilities. The timeline is not years, it is months.”.
• The statement also mentions: “We can only see what’s been released but there could be other models being developed by the likes of China, or other states and other actors and companies, that are just as advanced.”.
• The agencies are encouraging governments to act: “Cyber risk can no longer be treated as a purely technical issue. This is a core business risk and leadership responsibility”.

This InfoWorld article looks at an apparent contradiction in AI-based software development: AI is creating a new form of vendor lock-in, yet the history of software development is one of openness and non-proprietary systems.

• Software engineering is the most successful use case for AI. A JetBrains survey found that 74% of developers worldwide are using AI tools, with Claude Code now being the most used. 91% of these developers say their productivity has improved. Gartner claims that by 2027, company spending on AI tokens for code creation will bypass developer salaries.
• However, as one expert writes, “too many companies are building their entire AI strategy on top of proprietary platforms because the convenience is seductive”.
• There is currently a heavy dependence on proprietary AI systems from Anthropic, OpenAI, Google, Cursor and Microsoft. At the same time, there are many open-source initiatives with a large number of models available on HuggingFace. The community around the open-source OpenClaw AI agent is also very active.
• Open standards mitigates the risk of dependencies on proprietary systems. A recent example of such a risk is when Anthropic blocked access to models for some users of OpenClaw and OpenCode over what were seen as vague policy violations.
• Openness in the AI world can apply to models, representation formats, as well as the execution and orchestration layers. It also applies to the inference architecture. As one expert writes: “developers need visibility into how models run, how memory is used, and how performance scales”.
• There might be early signs that the AI community is moving towards openness, e.g., the Agentic AI Foundation and the Model Context Protocol (MCP) being open-source. One expert says: “Remember that it took many years before we saw the development and popularization of the open source cloud-native ecosystem… I think it would be a mistake to extrapolate from the current trajectory towards a closed, proprietary future.”

One Chinese and one Japanese firm have each announced the development of an AI model that they claim is close to Claude Mythos in performance.

• In Japan, Sakana AI – a company founded by ex-Google employees – announced the Fugu model which the company says “stands shoulder-to-shoulder with leading models like Anthropic’s Fable 5 and Mythos Preview”.
• Fagu is also designed for agent orchestration. This is a specific objective of the company which wrote “Orchestration Models are the next frontier, beyond bigger models”. Also, in an implied reference to the Anthropic models banned outside of the US: “Access to top models can disappear overnight. Collective intelligence is the practical hedge against this concentration of power.”.
• The Chinese cybersecurity firm 360 released two AI security tools: Tulongfeng is a model designed to discover cybersecurity vulnerabilities, Yitianzhen builds automated cyber defense and incident response. The founder is reported by Reuters as defining the need for such models “national strategic assets”.
• It remains to be seen what impact the ban on Anthropic giving access to their models abroad will have. The company’s run-rate revenue was 47 billion USD in May 2026, but it is not known what proportion of that revenue came from Asia. The Chinese and Japanese models could take an important market segment from Anthropic.

This Guardian article looks at efforts to implement a social media ban for children in the style of that enacted in Australia for children under 16 years of age.

• In Europe, the UK has announced a similar ban which it hopes to have in place by 2027. France intends to ban for under 15 year olds, Austria will ban for under 14 year olds, and Norway is extending its current ban for children under 13 years to children under 16 years.
• Spain is even considering a move that would make social media platform executives responsible for hate speech that appears on the platform.
• With many court cases appearing against the platforms, the current situation has been described as Big Tech’s “big tobacco” moment as more evidence appears about the harms and addictive effects of social media on children.
• The situation is a bit more complex in the US. Big Tech is influential among many federal and state politicians. Florida currently has the toughest laws with access by under 14 year-olds to social media forbidden; 15 year-olds require explicit parental consent.
• The effectiveness of these bans is debated. In Australia, the government claims that 5 million social media accounts have been closed. However, a survey of 900 parents suggests that two-thirds of children have managed to retain access to their accounts.
• Amnesty International described bans as an “ineffective quick fix” that is “out of step” with the realities of a digital generation. For the organization, “the most effective way to protect children and young people online is by protecting all social media users through better regulation, stronger data protection laws and better platform design.”.

This VentureBeat article suggests that software developers need to cultivate more project management capabilities as AI is increasingly used to write their code. The software development bottleneck has moved from writing code to the decisions about what to build.

• The article identifies five software development eras distinguishing the work of software engineers. The first era is the Stack Overflow era (2014 to late 2022) where engineers’ issues lived on a single web portal. Content on Stack Overflow has dropped 77% since the arrival of ChatGPT.
• The second era is the browser-tab era (late 2022 to 2024) where first generation language models were accessed from outside of the IDE, with code being pasted from the console into the code editor.
• Next came the IDE-native era (2024 to 2025) as tools like Cursor and Claude Code moved the language model into the text editor and gave the models access to the full repository. Escalation to “senior” software engineers began to disappear and Claude prompts exceeded the number of bash commands typed by developers.
• The spec-driven era (2025 to 2026) saw larger context windows increase the amount of development that could be done in a single session. A single session would accomplish what previously required many Jira tickets. For instance, an AWS engineering team recently completed a project with six people in 76 days; the project had initially been planned for 18 months with 30 engineers.
• Finally, the routines era (2026) relies on agent technologies where agents are given tasks and told to work on these – the software developer can give these instructions in the evening and inspect the work done on arrival in the office in the morning.
• For engineers, technical competence is still important. Agents may write 70% of a code base but do not understand issues such as memory leaks or thread safety. Technical debt will manifest itself when the first major bug is discovered. It is also important for engineers to adopt a project manager mindset to help create ideas.

A startup called Unconventional AI has developed an image generation model that they claim performs as well as Stable Diffusion or OpenAI’s GPT Image 1 but which uses up to 1000 times less energy.

• The Un-0 model is built on an oscillator-based architecture – a complete departure from the architectures used in generative AI.
• The Un-0 model currently runs on a software simulation of Unconventional’s oscillator chips but the start-up, which currently has only 50 employees, hopes to build the chip soon.
• For the start-up owner, formerly of Databricks, “AI scaling is hard because of energy. It’s going to be the fundamental limit in the next few years. You just can’t go past it”.

This TechCrunch article looks at the sharp rise of Micron over the past year. The company supplies memory cards for PCs and games consoles.

• Micron was trading last week with a market cap of 1.27 trillion USD. This compares with a market cap of 1.39 trillion USD for Meta and 1.42 trillion USD for Tesla.
• Micron’s revenue has quadrupled year-over-year to 41.45 billion USD. Profits over the last year have gone from 1.88 billion USD to 28.2 billion USD.
• This rise in the company’s valuation is due to a current shortage of memory (DRAM, NAND and High-Bandwidth Memory) required by data centers. The data center demand is also placing a strain on the supply chain of PC manufacturers like Dell and HP. This has all led to an increase in the price of PCs, Apple devices and game consoles.
• The fall in supply of memory has been termed “RAMageddon” and is expected to last into 2027 at least.
• The largest challenge for memory chip makers such as Micron and Samsung is scaling since this requires building large new manufacturing facilities.

Reuters examines the increase in corporate borrowing needed to finance AI. Hyper-scaler companies like Amazon and Alphabet (Google’s parent company) have issued 600 billion USD in bonds over the past 12 months.

• The companies are selling bonds on global markets to raise funds. Amazon raised 14.5 billion USD in May in the Euro Zone, and Alphabet has issued bonds in Canadian dollars, Japanese yen and Swiss Francs.
• The surge in bonds highlights the huge costs around data center development. BNP Paribas expects the hyper-scaler companies to spend 725 billion USD on data centers this year. This is double the spending level seen mid-2025.
• Analysts believe that the volume of AI debt could push the investment-grade market above 2 trillion USD for the first time ever in 2026. For Morgan Stanley, “if we start to see companies coming to the bond market over and over again, then I think it starts to be a concern”.

The well-known environmental activist, Erin Brockovich, has taken up the cause of residents worried of the impacts of new data centers in their communities.

• Erin Brockovich became known in 1993 when she took a major role in a lawsuit against Pacific Gas and Electric Company (PG&E) on behalf of residents of Hinkley, California, whose groundwater had been contaminated. The settlement for 333 million USD was at the time the largest ever.
• Brockovich has received thousands of emails from residents in communities where data centers are being built. Their complaints vary from the noise and pollution levels, the damage to animals and the environment, and the impact on utility bill costs.
• Water is a particular concern. A large data center requires 23 million liters of water each day for cooling. This is the equivalent to what 50’000 people drink.
• Based on data in the emails, Brockovich built an open-source map showing where data centers are operating or are currently being built. The map shows that as of June this year, 33 AI data centers have been completed and are operational, 68 are under construction and 41 are proposed.
• Brockovich expressed surprise at the apparent secrecy around construction. Many of the residents only learned about the data centers after the project had been signed on. In Hill County, Texas, county commissioners voted a one year moratorium to halt building after public opposition. The commissions were then sued for 100 million USD in damages by the construction companies.
• The Guardian writes: “It feels like a step into post-democracy, which is a tech bro fantasy, a world in which laws and regulations have been obviated. The big tech companies seem to have blueprinted their fantasy and started building it.”.