David McWilliams: The economics of the AI boom are actually quite old-fashioned

David McWilliams: The economics of the AI boom are actually quite old-fashioned

This week saw the largest share offering the world has ever seen. Anthropic, the AI company that makes Claude, floated on the US stock market with an initial value of just over $1 trillion – a thousand billion dollars. Founded only six years ago, it is today the 10th most valuable corporation in the US. With the share offering coming a week after the Jensen Huang scandal, it seems that the world is captured by some AI story or another. Even global politics is in thrall to this new technology. On Thursday, when Donald Trump, Xi Jinping and Vladimir Putin sat together, it was clear that the new cold war was as much about tech as arms. We are looking at an industrial transformation on the scale of the move to oil and gas in the early 20th century. It will change the world, but that doesn’t mean everyone can profit from it. Despite all the money flooding into what is known as “the AI trade”, there seems to be precious little understanding of the economics underpinning the AI boom. Where is the money going, why is everything so expensive and what are the risks that it collapses following one big speculative mania? When so few people have the answers, it makes sense to ask the questions. READ MOREThe AI story is portrayed as a high-end technology story, but when you drill down a bit, the economics of AI seem close to old-fashioned heavy industry. It involves pouring concrete, connecting power grids and buying manufactured microchips at a large scale. The lion’s share of the money goes on imported hardware. Very little of the spending is creating new tech jobs. And a growing share of the AI buildout – estimated to be $1 trillion this year – is being financed with borrowings that will need to be paid back sooner or later.Who is profiting at the moment, who is paying, and how long can it last? Most of the profits are being made by chipmakers largely based in Asia. Most of the cost is borne by the companies in the US doing all the borrowing, those building or renting the data centres. In addition, as almost everything uses chips – your phone, your car and your fridge – we are all paying through higher costs in our data-driven economy. The amount of new debt means the AI boom is dependent on the rate of interest, the patience of lenders and the tolerance of local communities for data centres, which were a big issue in the US midterms, as was a general dislike of the tech titans who dominate the AI world. A data centre needs four things: chips, land, water and electricity. Building a new facility can take about two years. There’s a lot of concrete to pour. The amount of money, water and electricity required is enormous. Consider a 1 gigawatt (GW) campus of data centres. Granted, this is a huge data centre but it gives us an idea of the scale of the industrial transformation in front of us. One GW of electricity roughly corresponds to the output of a nuclear power plant or the annual electricity for some 650,000 houses – bigger than Dublin. About 10.5 billion gallons of water are needed to cool the machines, about enough for 100,000 households, and it needs around 700-1,100 acres of space, about half the total size of Phoenix Park. Are you beginning to grasp the scale? [ There is little indication we have a plan for what to do if the AI bubble burstsOpens in new window ]Most data centres are not this big, running on the megawatt rather than gigawatt scale, closer to 100-250MW of energy, but the demands of AI mean this figure is trending relentlessly upwards, which makes it worthwhile to look at the jumbo data centre. A 1GW data centre costs around $50 billion – roughly $15 billion for the facility and $35bn for the compute to power it. Chips make up roughly 70 per cent of the total cost. Because chips dominate the cost, the profits flow to whoever makes them. Nvidia and the amorous Jensen Huang may well design these chips, but they do not manufacture them. Rather, the likes of SK Hynix, TSMC and others do, with Korea, Taiwan and Japan controlling manufacturing. Others will want to get in on the game. [ AI fever is producing some strange symptomsOpens in new window ]The hyperscalers, companies such as Google and Meta, are designing their own chips, elbowing in on Nvidia’s turf. Geo-strategically, rival suppliers, particularly in China, will emerge. Greater supply of chips will bring down the cost and as the cost of chips falls, the price of using AI will fall. The dilemma for companies such as Anthropic is that if the price of using it falls, its revenues will fall, undermining its new position as the 10th most valuable company in the US. Not only is the industrial model changing rapidly, with new competitors emerging at every stage of the supply chain, but maybe – more critically – the financing of the AI transformation has changed from cash flow to debt and equity. The hyperscalers, such as Google, used to finance their AI businesses from their large cash flow but, as the costs have risen, even a big company like Google has had to start borrowing money. Around 10 per cent of all global investment-grade bonds are already related to AI, possibly rising to 16–19 per cent by next year. The AI borrower is in a highly competitive market. For example, the US government bond yields roughly 5 per cent with essentially no risk, so every data centre bond competes with sovereign debt.Moreover, central banks that are trying to bear down on inflation are raising interest rates, pushing up rates everywhere. Markets are already getting nervous. Hyperscalers with other subscription businesses can borrow cheaply, but others such as Oracle are paying the widest spread. There is also another group of AI-related businesses called neo-clouds that offer cheaper data centre services than Google, Meta, Amazon or Microsoft. You might have heard of names like CoreWeave, Nebius, Lambda and Crusoe. They are paying much higher interest to get money and are most exposed to any change in interest rates or market sentiment. Over the past few weeks, the huge amounts of money involved have put the wind up financial markets. Can the AI companies like Anthropic that have yet to make a profit generate money to validate their huge valuations? Will the hyperscalers continue to bankroll our datacentres? Will banks start to call in loans? And can all this credit continue to flow into a sector that is spending so much but – at least so far – is earning so little? We’ve seen boom and bust before. Will this be the biggest one?

Original Source

Read the full article at Irishtimes →

KhanList aggregates and links to publicly available news content. We do not host full articles from third-party sources. Always verify important information with original sources.