<p>There is an old lesson from gold rushes worth remembering. The people who went looking for gold often came home poorer. The people who sold them food, tents, shovels and entertainment did better. The Klondike rush of 1897 is a particularly good example. Around 100,000 hopeful prospectors set off towards the Yukon. Seattle became their supply depot. About 70,000 passed through the city and in just nine months, Seattle merchants sold as much as USD 25 million of goods, against barely USD 325,000 in all of 1896. Up in Skagway, the economics were equally instructive. Prostitutes earned more than cooks, dressmakers or nurses. There was gold in the hills, but there was also plenty of money to be made from people looking for it. Something similar may now be happening with artificial intelligence. </p><p>The obvious winners so far are the people selling the shovels. Nvidia sells the chips, whilst others provide cooling systems, electrical equipment, cables, transformers and power. They get paid as the infrastructure is built. But the people building the data centres are making a much larger bet. They are spending vast sums today on the assumption that demand for AI computing will turn up tomorrow. And the sums are remarkable. Large technology companies are expected to spend trillions of dollars on AI related capital expenditure over the second half of this decade. Increasingly, this is not being financed entirely from the cash flows of Microsoft, Amazon or Google. Debt, private credit and special financing vehicles are entering the picture. Some of the borrowing is sufficiently complicated to sit far away from the corporate balance sheet. </p><p>For the moment, there is no sign of empty data centres. Vacancy in major markets is extremely low and demand remains ferocious. That is precisely why everyone wants to build one. The danger appears if the future turns out to be merely good rather than spectacular. Suppose AI demand continues to grow, but at 15 per cent a year instead of the 30 or 40 per cent that most spreadsheets assume. Nothing has gone wrong with AI and ChatGPT still works. Companies continue to automate and we continue asking machines to write our emails. There is simply too much computing capacity chasing the demand that actually arrives and that is when the financial arithmetic changes. A data centre has enormous fixed costs with land, building, power and interest on borrowed money. Much of the equipment inside it ages remarkably quickly. A warehouse built ten years ago is still a warehouse but a top end AI chip bought today could look elderly a few years from now. </p><p>So a fall in utilisation can hurt. Revenues weaken while depreciation, interest and operating costs remain. If too much capacity comes on stream, prices for computing can fall. Then refinancing gets harder and highly leveraged developers discover that long term infrastructure financed against short term technological assumptions is an uncomfortable combination. India should watch this carefully, as data centre investment here is accelerating rapidly and some extraordinarily ambitious capacity targets are being discussed. The question is whether every project being planned today will earn the return its investors expect. That is why your columnist would not call data centres a bubble. Not yet, anyway. The present market is genuinely tight and AI may indeed transform the world. But history offers a useful warning. A technology can change everything and still leave some of the people who financed its infrastructure nursing dreadful losses. </p><p>During the Klondike rush, the clever fellow was often the chap selling the shovel. The same may prove true of AI. The gold miners now wear suits, carry spreadsheets and build data centres. </p>
<p>There is an old lesson from gold rushes worth remembering. The people who went looking for gold often came home poorer. The people who sold them food, tents, shovels and entertainment did better. The Klondike rush of 1897 is a particularly good example. Around 100,000 hopeful prospectors set off towards the Yukon. Seattle became their supply depot. About 70,000 passed through the city and in just nine months, Seattle merchants sold as much as USD 25 million of goods, against barely USD 325,000 in all of 1896. Up in Skagway, the economics were equally instructive. Prostitutes earned more than cooks, dressmakers or nurses. There was gold in the hills, but there was also plenty of money to be made from people looking for it. Something similar may now be happening with artificial intelligence. </p><p>The obvious winners so far are the people selling the shovels. Nvidia sells the chips, whilst others provide cooling systems, electrical equipment, cables, transformers and power. They get paid as the infrastructure is built. But the people building the data centres are making a much larger bet. They are spending vast sums today on the assumption that demand for AI computing will turn up tomorrow. And the sums are remarkable. Large technology companies are expected to spend trillions of dollars on AI related capital expenditure over the second half of this decade. Increasingly, this is not being financed entirely from the cash flows of Microsoft, Amazon or Google. Debt, private credit and special financing vehicles are entering the picture. Some of the borrowing is sufficiently complicated to sit far away from the corporate balance sheet. </p><p>For the moment, there is no sign of empty data centres. Vacancy in major markets is extremely low and demand remains ferocious. That is precisely why everyone wants to build one. The danger appears if the future turns out to be merely good rather than spectacular. Suppose AI demand continues to grow, but at 15 per cent a year instead of the 30 or 40 per cent that most spreadsheets assume. Nothing has gone wrong with AI and ChatGPT still works. Companies continue to automate and we continue asking machines to write our emails. There is simply too much computing capacity chasing the demand that actually arrives and that is when the financial arithmetic changes. A data centre has enormous fixed costs with land, building, power and interest on borrowed money. Much of the equipment inside it ages remarkably quickly. A warehouse built ten years ago is still a warehouse but a top end AI chip bought today could look elderly a few years from now. </p><p>So a fall in utilisation can hurt. Revenues weaken while depreciation, interest and operating costs remain. If too much capacity comes on stream, prices for computing can fall. Then refinancing gets harder and highly leveraged developers discover that long term infrastructure financed against short term technological assumptions is an uncomfortable combination. India should watch this carefully, as data centre investment here is accelerating rapidly and some extraordinarily ambitious capacity targets are being discussed. The question is whether every project being planned today will earn the return its investors expect. That is why your columnist would not call data centres a bubble. Not yet, anyway. The present market is genuinely tight and AI may indeed transform the world. But history offers a useful warning. A technology can change everything and still leave some of the people who financed its infrastructure nursing dreadful losses. </p><p>During the Klondike rush, the clever fellow was often the chap selling the shovel. The same may prove true of AI. The gold miners now wear suits, carry spreadsheets and build data centres. </p>
