Michael Burry Sounds Alarm on AI Investment Bubble

This article from RTE looks at how the Russian government is attempting to control access to information on the Internet.

• The Kremlin blocked access to Western websites and news agencies in 2022. The intent was to control the narrative of high Russian casualties in the Ukraine war. Facebook and Instagram were also banned.
• Three major messaging applications have now been banned in Russia: WhatsApp, Signal, and now Telegram. The Telegram App was reportedly used by 100 million Russians and its ban even brought criticism from well-known pro-war bloggers.
• The Kremlin has supported the development of a Russian messaging App called Max. To force users to adopt Max, the strategy is to turn Max into a super-app that allows users to do many administrative tasks and pay bills. The approach is similar to the Chinese app WeChat.
• For one foreign policy expert, “The Kremlin has always wanted to create a sovereign information space modeled on the Chinese example”. China has a locked-in information space where certain historical and political information such as the 1989 Tiananmen Square massacre is absent.
• It is reported that Putin himself does not use the Internet. He is thought to spend most of his time in a bunker and only communicates using Internet-less dedicated phone lines.

This article summarizes TechCrunch’s Equity podcast discussion on a deal struck between xAI and Anthropic which gives Anthropic all of the computing capacity at xAI’s Colossus 1 data center in Tennessee (USA).

• For one journalist, the move to allow another firm develop its AI models in xAI premises questions xAI’s ambition to make Grok a leading AI model.
• The business model for xAI is also questioned, and whether the selling of data center compute capacity could be a key part of that. There is also the problem that xAI is facing an environmental lawsuit around Colossus.
• It seems that Elon Musk’s current concern is the upcoming IPO of SpaceX – which could be valued at 1.75 trillion USD. The xAI entity has been integrated into SpaceX, though could be made independent again under the name SpaceXAI.

In an MIT Technology Review article, the 2024 Nobel Prize winner in economics, Daron Acemoglu, shares some of his concerns around current AI development.

• For one thing, Acemoglu is not yet convinced of an AI job apocalypse and is not calling for an AI corporate tax to compensate AI-driven layoffs.
• His first concern centers on AI agents. The power of agents is that they can work independently on tasks rather than just answering questions. However, they are not yet able to swivel between tasks or data formats. He cites the example of an x-ray technician who juggles between tasks like verifying images, archiving, and checking with colleagues. The inability of agents to juggle as effectively as people do is currently saving many jobs from AI takeover.
• A second issue is the hiring of top economists by AI firms. OpenAI hired Ronnie Chatterji from Duke University in 2024 and Google DeepMind hired Alex Imas, an economist at the University of Chicago as its “director of AGI economics”. Acemoglu is worried that AI firms are “interested in economists just to further their viewpoints or further the hype”.
• The third issue is the absence of apps that make AI easier to use – the way the MS PowerPoint and MS Word made the production of slide decks and documents easier. He notes that anyone can chat with an AI model, but it takes effort for a worker to get practical and productive use from the AI.

This VentureBeat article looks at the evolution of priorities by AI decision-makers as AI costs rise. It follows a Gartner report that reveals that AI infrastructure spending will increase by 401 billion USD this year. At the same time, data-economy reports that average GPU utilization by organizations is around 5% - meaning that companies are over-provisioned and over-spending by a factor of 20.

• This under-utilization can be explained by an initial fear of scarcity on the part of large companies. They bought compute capacity in anticipation of requiring it at a later time.
• This mindset is changing according to a VentureBeat survey, as companies become more interested in metrics like the ratio of inference cost to total cost of ownership. In other words, the measure is no longer GPU activity (the number of powered on GPU chips) but GPU productivity (how many useful tokens are generated per dollar spent).
• Another change concerns the technical challenges that data center providers encounter. First, they need to measure GPU productivity and not just GPU power consumption. Second, they need to customize their architectures to changing volumes of training and inference jobs. Currently, training jobs account for 70% of business volume, and inference jobs 30%.
• Several technical advances that help data centers are mentioned. RDMA (Remote Direct Memory Access) removes significant latency by allowing data copies between RAM and GPUs without passing by the CPU. Another improvement is on GPU caching where data can be prefilled for AI jobs. A final technique mentioned is GPU cache compression; Google’s TurboQuant for instance supports a lossless compression rate of 6 to 1.

This article looks at various aspects of the problem of deep-fake sexualized content posted on the Internet. The term used today for this content is nonconsensual intimate imagery, or NCII.

• NCII can result from AI being used to modify pornographic content with the faces of other people. Another approach is the use of “nudify” apps which are quite common today including Grok and Crushmate (an app for which advertisements appeared on the Meta platform).
• There is a general consensus that NCII constitutes “a new form of sexual violence”. Victims have been known to experience depression and even suicidal ideation.
• Another aspect of the problem is the NCII AI tools are trained on the estimated 10’000 terabytes of porn available on the Internet. When porn is used in NCII deep-fakes, the actors of the original porn films can lose revenue.
• This can also lead to new business models. In one case, porn actors signed a deal with Spicey AI to have AI twins, even though the company soon went out of business. Another approach is to have AI videos with porn content taken down. Takedown Piracy has digitally fingerprinted more than half a billion videos and the organization had managed to get 130 million copyrighted videos taken down from Google.

The International Data Center Authority (IDCA) has said that the amount of electricity used by data centers for AI has risen by 15% over the last two years.

• In the US and the UK, data centers are now consuming 6% of electricity from the grid. The global average is 2%. The consumption rate is 19% in Singapore and 11% in Lithuania.
• Annual investment in data centers is now 1 trillion USD – nearly 1% of the global economy. The number of data centers worldwide is now estimated to be 10’000.
• The IDCA is warning that data center development is “sparking societal and political concerns”. It writes that “significant community and political pushback starts to occur in nations once their data center footprints have reached the 5% consumption level of national grids.”.
• Greenpeace UK is warning that an “unchecked AI boom” will lead to higher energy bills, water supply problems and “a new lifeline for fossil fuels”.
• The International Energy Agency says that data center energy use increased by 17% in 2025, compared to a 3% increase in global electricity demand. It also says that 13% of data center energy consumption in the US are unused “zombie” services: unused services that are not switched off. The wasted consumption exceeds 3 GW.
• The IDCA also warns of physical attacks on data centers as they become part of countries’ critical infrastructures.

MIT Technology Review reported on the main developments in the court case brought by Elon Musk against Sam Altman and Greg Brockman whom he accuses of abandoning the non-profit mission of developing AI for the betterment of humanity.

• In particular, Musk claimed that he invested in OpenAI precisely for this mission.
• OpenAI’s lawyer claimed that OpenAI remains a nonprofit dedicated to developing AI safely, and that Musk is suing simply to handicap a competitor of his xAI company. The lawyer claimed that the “OpenAI nonprofit is the best-resourced nonprofit in the world”.
• For one law professor interviewed in the article, the non-profit “doesn’t have any voice”. In particularly, “it doesn’t have much money, and OpenAI doesn’t think it has any obligation to fund it… It’s unclear how on earth the nonprofit is supposed to exercise its duties and control the entire company.”.
• Musk’s lawyer grilled Sam Altman on his alleged history of lying. Former OpenAI board executives testified that Altman had lied to them – which led to Altman being temporarily sacked from the company. In a separate development, the US House oversight committee is launching an investigation into potential conflicts of interest involving Altman.

In the court case between Elon Musk and OpenAI, a federal jury found that OpenAI is not liable for Musk’s claims that CEO Sam Altman and president Greg Brockman unjustly enriched themselves from Musk’s startup contribution.

• The jury’s finding is a non-binding, advisory verdict, which the judge accepted.
• Microsoft was also found not liable in the verdict.
• The jury decided that Musk’s lawsuit, filed in 2024, did not fall within the statute of limitations to bring his case. In a tweet afterwards, Musk wrote: “Regarding the OpenAI case, the judge & jury never actually ruled on the merits of the case, just on a calendar technicality. There is no question to anyone following the case in detail that Altman & Brockman did in fact enrich themselves by stealing a charity. The only question is WHEN they did it!”.
• OpenAI argued that Musk was aware of the for-profit plans from 2017 – which means the lawsuit would have to have been submitted by 2020.
• The verdict is a relief for OpenAI which can now keep planning for an IPO which could see a valuation of 1 trillion USD.

This InfoWorld article investigates the pros and cons of a business model where a company can sell its excess compute capacity to others. This model challenges the dominance of hyper-scaler companies.

• One advantage of this model for consumers is pricing as companies do not have the same cost structures and margin expectations as cloud providers.
• Another advantage is market efficiency, as companies can get access to resources without the traditional cloud intermediaries. It also gives consumers alternatives to hyper-scaler lock-in, which could be important for specialized AI loads.
• The model does have cons however. The main problem is that few companies have the infrastructure to provide excess compute capacity to others. An infrastructure would need to handle billing, identity control, service-level agreements, automation and workload observability.
• Another challenge is ensuring that a compute provider respects compliance conditions specified by consumers regarding data management.
• The article argues that this new model may emerge gradually. The first step will be the appearance of compute brokers who find and classify available compute capacity, and then locate interested consumers.

Michael Burry, the investor famous for betting against the housing market (book and movie called “The Big Short”) is warning of parallels between the current AI economy and the Dot-com bubble of the late 1990s.

• In a post on X, Burry wrote that capital tied to AI is more concentrated than Internet related investment at the time of the Dot-com bubble.
• He noted that 87% of venture capital funding is directed at AI, compared to below 40% VC investment in Internet in the 1990s. He also mentions that 38% of high-yield bond issuance is AI-related, as is 49% of investment-grade debt issuance.
• The article mentions that 100 billion USD of investment-grade bonds issued during the Internet boom had turned to junk by 2002.
• Burry underlines that a bubble is not about the over-pricing of stocks, but rather the dependence of too many parts of the financial system on a single narrative – AI in this case – holding up. The current narrative is that massive AI investment in data centers will eventually pay off which has become a risky assumption.
• He also warns against the line of thought that says even if one over-spends in the first wave of a technology, the infrastructure becomes essential, which is more or less what happened in the post Dot-com era. However, this argument might not hold for AI because GPUs and CPUs can become obsolete faster than the network infrastructure built in the 1990s.