Currency trading is our business. But that does not mean we ignore what is happening elsewhere in financial markets. Quite the opposite: a serious disruption in equities or credit will quickly find its way into currencies.
And few developments fascinate us more — occasionally even astonish us — than the current AI infrastructure race.
Global data-centre investment is expected to reach roughly $3 trillion by 2030, with hyperscalers — giant cloud operators such as Amazon, Microsoft, Alphabet, Meta and Oracle — among the main drivers.
The arithmetic is uncomfortable. If capital of that magnitude ultimately has to clear, say, a 15% hurdle rate, it points to roughly $450 billion of annual economic return at full deployment. And even that may understate the challenge: GPUs and other AI hardware become obsolete far faster than traditional infrastructure. Unlike a bridge or an office building, much of this asset base will require substantial and recurring reinvestment.
So where will the return come from?
Certainly not from a few ChatGPT subscriptions.
The monetisation case must be much broader: cloud-compute rental, API consumption, enterprise copilots and agents, workflow automation, security and data services, advertising uplift, model hosting — and perhaps entirely new software categories that do not yet exist at scale.
That may happen. But it is not guaranteed.
Meanwhile, the way this extraordinary build-out is being financed is already changing. Hyperscalers remain immensely cash-generative, but borrowing has risen sharply. More intriguing is the growth of off-balance-sheet financing. Dedicated vehicles or joint ventures own data centres and borrow against them. The hyperscaler may hold only a minority stake while committing to long-term leases, capacity purchases or guarantees. Economically, the obligation remains — but much of the debt sits elsewhere. The BIS has gone so far as to describe such structures as “shadow borrowing.”
Does that sound faintly familiar?
Don’t get us wrong, it is not 2008. There are no subprime mortgages at the centre of this cycle, the underlying assets are productive, demand is real, and the largest technology companies remain extraordinarily profitable. But the rhyme is difficult to ignore: leverage migrating into special vehicles, increasingly complex financing chains, investors assuming continued demand, and obligations becoming less visible precisely when confidence is highest.
And there is another part of the equation that gives us pause: pricing.
A token is a basic unit of input or output processed by an AI model, and inference services are commonly priced per million tokens. Silicon Data’s usage-weighted benchmark fell to $0.97 per million tokens at the end of August — down 29% in that month alone and more than 50% from its May peak.
The longer-term decline is even more dramatic. OpenAI has documented a 99% reduction in per-token pricing between a model introduced in 2022 and a newer, more capable model introduced less than two years later.
Optimists have a perfectly reasonable answer: if price(p) × volume(q) keeps rising, falling prices do not matter.
True — but that only protects revenue. It says nothing about margins. If prices keep falling while compute, energy and infrastructure costs remain substantial, rapidly growing usage can coexist with margin compression. It is not unthinkable that volume grows but profits stagnate.
It may all work out. But the combination of enormous capital commitments, falling prices and uncertain margins makes that outcome far from certain.
We will keep watching closely. Because if the emperor turns out to have no clothes after all, “volatility” may prove an inadequate description of what follows. A major market disruption would become a very real possibility.
For us, that would be a reason to reduce risk — quickly and decisively.
