Artificial Intelligence – Finance Debate

No one cares until it is too late

• As more is spent on AI infrastructure, the deals are becoming increasingly opaque and interrelated, such that the real size of liabilities becomes increasingly unclear, setting up the potential for some very nasty surprises.  
• However, while valuations are rising and demand is strong, nobody cares, but the minute there is a speed bump, the whole house of cards could very quickly collapse, as we have seen numerous times before.
• With Meta and xAI as the latest examples, the financial engineering cycle is ramping up once again and is being employed to allow companies to afford the massive cost of AI infrastructure while maintaining the façade of financial good health.
• Meta is raising $60bn to help cover its $100bn+ 2026 capex bill, half of which will be held off its balance sheet, so that unless one is really paying attention, one will never know that it is there.
• xAI is also entering into a creative transaction where it is raising $20bn into an SPV that will own the silicon chips and then rent them back to xAI.
• OpenAI’s creative financing strategies have also been well documented, but in my opinion, these are creative ways of getting other people (AMD shareholders and buyers) to pay its compute bill as opposed to obscuring the size of the potential liabilities.
• History is replete with examples of companies which looked great on paper but then collapsed under the weight of liabilities that suddenly materialised when things did not go as planned.
• Enron, Lehman Brothers, and WeWork are all household names due to their ignominious collapse when market conditions changed, and liabilities that were parked off-balance sheet came home to roost and exposed them as insolvent.
• The problem with this is that unless one is a forensic accountant, the liabilities are very difficult to spot, and with the optimism and news flow that surrounds AI, these deals will quickly be forgotten.
• As long as there is no disturbance in demand and the data centres fill up almost as soon as they open, there will not be a problem, but history shows us that this is very rarely the case.
• For example, OpenAI currently serves 800m users adequately with around 2GW of compute, but is planning on rolling out at least another 28GW of capacity.
• Furthermore, I suspect that the efficiency of data centres is going to improve substantially, meaning that by 2030, the number of tokens that 28GW produces will be many times greater than it is today.
• That means that to fill these data centres, OpenAI needs to be able to profitably sell many, many times more compute than it does today.
• The key here is profitably, as while OpenAI provides a lot of compute, the vast majority of it attracts no revenues, which is why OpenAI is estimated to be losing $10bn every three months.
• This is not a sustainable state of affairs, and the question of how OpenAI will make a positive return from the $1.4tn of capex it intends to spend over the next 5 years is a very valid one.
• The same question applies equally to Meta Platforms, which will spend over $100bn in 2026 but has no direct method of monetisation other than indirectly through the monetisation of its existing digital ecosystem.
• This is where the likes of Microsoft, Google, Nebius, Amazon and so on have an advantage as they intend to sell the tokens that they produce directly to the market, but whether the price remains high enough to earn the cost of capital remains to be seen.
• Consequently, all of these cross-holdings, investments and off-balance sheet structures are creating an increasingly unstable structure because it means that a problem in one area could easily affect everyone else due to the cross-holdings and inter-dependencies.
• This is precisely how collapses have happened in the past, and the way these deals are being structured makes me nervous that we are setting the scene for another one.
• To be fair, the off-balance sheet deals are being secured on the assets, but with the speed of obsolescence driven by TSMC and Nvidia, a demand wobble or valuation correction could easily cause the value of the assets on which the financing is secured to crater.
• The net result is that for as long as demand for compute remains robust and the market is willing to pour in cash, then nothing is going to happen.
• For the moment, this looks like how things are going to be well into 2026, but predicting the catalyst of such an event is virtually impossible.
• Hence, mitigation is the best strategy, which means owning the companies with the lowest valuations (less to correct), the strongest balance sheets and no exposure to off-balance sheet structures that could cause existential distress if the system decides to correct.
• Hence, I am looking further down the value chain to things like TSMC, Samsung, MediaTek and Qualcomm, who are exposed but will not be wiped out if there is a downward correction.
• I continue to hold Qualcomm and Samsung for exposure to this trend.