Artificial Intelligence – Panic Debate

Demand keeps the story intact.

• A brutal sell-off has hammered the AI trade, but the key question is whether this is the bursting of the bubble or the market doing what it frequently does, which is to put high-quality assets on sale at a huge discount.
• Economist and value investor Benjamin Graham coined the saying “In the short run, the market is a voting machine, but in the long run, it is a weighing machine” which was later made famous by Warren Buffett.
• In practice, this means that the fundamentals of a company will always prevail over public sentiment, meaning that one needs to figure out whether anything fundamental has changed that would warrant the large repricing of AI assets we have seen in the last 2 weeks.
• The reality is that very little has changed other than the political landscape in the USA, which is likely to have some impact on the speed at which data centres can be built.
• Starting in 2027, I expect to see more delays, more objections and slower permitting as harder questions are asked of the data centre operators when they want to build a new data centre.
• I still think that this is a reasonably straightforward problem to solve, as Meta demonstrated in Richland Parish, Louisiana, where its data centre has revitalised the local economy and increased tax revenues so much that teachers received a large, one-time bonus.
• Offers such as providing cheap electricity for residents from data centres with their own power supply and contributions to local projects should overcome resistance, but it is now going to take longer than before to bring these data centres to completion.
• The problem is that sentiment has just swung against data centres and the AI rollout in the midterm elections.
• Data centres were an issue, and the side that favours slowing or curtailing the build-out won a landslide.
• Hence, the view is that there is now a mandate against data centres, although there is strong evidence to suggest that the landslide was much more a repudiation of the current administration as opposed to a policy debate.
• Either way, the net result is that the roll-out in USA will very likely be slower for at least the next two years
• Outside of this, nothing has changed, raising the question of whether this justifies taking Nvidia down by 25%, cutting the neo-clouds in half or more and the memory sector by 30% to 50%.
• Other bellwethers such as AMD are down 40% in a month, while Arm has almost halved and the list goes on and on.
• There remain two places in the AI industry where I see a bubble, which are in the spot price for AI compute and the valuation of Anthropic and potentially OpenAI.
• It is becoming clear that there is no real defensible position in AI models, meaning that anyone who creates and sells them for a living needs to both maintain an edge and have a touchpoint with the user or customer.
• This makes picking the winner in this space notoriously difficult, meaning that the lowest-risk part of the industry remains the picks and shovels of AI.
• Large chipmakers, cloud service providers and technology suppliers like Arm, TSMC, ASML and GlobalFoundries will be unaffected by who wins, as a fall-off in revenues from one will be offset by growth in another.
• The silver lining to the political change in the USA is that AI compute is now going to become even more scarce and for a longer period of time, while at the same time those that monetise it have lost between 30% and 50% of their value.
• Furthermore, there have been no signs of a fall in demand for data centre equipment, as everyone remains sold out at least 12 months in advance, as demonstrated by AMD just a couple of days ago.
• The comparison to the US Railroad roll-out of 1870-1890, which was followed by mass bankruptcies and a restructuring of the industry, is often made, and in many ways the comparison is appropriate with the exception of demand.
• In the 19^th century, the tracks were laid (often in parallel by competing companies), but there were no trains to run on them for quite some time.
• The same happened in 1999 – 2000 when massive amounts of dark fibre were installed even though there was no traffic, as it was thought that the fibre would trigger traffic.
• This was correct, but it took a long time to materialise, meaning that those who had borrowed money ran out of time and went bankrupt.
• However, with AI, there is so much demand for AI services that none of the providers can keep up, and the customers of the picks and shovels are selling everything they can build and in many cases at fantastic prices.
• This is the equivalent of running the train tracks at 100% capacity as soon as they are laid, or the dark fibre being full the minute that it has been completed.
• This is why I have been fine with as much spending as these companies feel they need, but only for as long as they have done so out of operating cash flow.
• The minute they started borrowing money, they put interest rates and bond investors in the driver’s seat, which I suspect is as much to blame for the current panic as the political change in the USA.
• Hence, I think that the biggest risk to the AI trade is a liquidity crunch where cash dries up, start-ups have more difficulty in raising money and growth in AI usage materially slows down.
• Hence, the way to look at the sector at the moment is to take the current company share price and work out how much of a slowdown the shares are pricing in and use that as a guide as to whether one should go fishing for the bottom.
• Against this backdrop, the memory players are now on sale, as is Nvidia, Qualcomm, and GlobalFoundries.
• Other assets such as Arm and Nebius are getting to the point where I would take a closer look at them.
• I can’t see the catalyst that will switch the market from voting to weighing, but as long as demand for AI remains robust, the catalyst will come, meaning that there is a rare opportunity on offer.