A primer on AI

Charles Stross’ 2005 novel ‘Accelerando’ imagines a world tipping into ‘The Singularity.’ The pace of technological change quickens, then quickens again, until powerful artificial minds dismantle the Earth itself, converting the planet into solar-powered chips on which to run their own thinking. Physical humans end up as refugees, scratching a living on freezing asteroids at the edge of the solar system. It sounds wild and a little silly. Yet this kind of post-human future is the source of genuine reflection among some very powerful people in technology circles.

I don’t intend to litigate whether they are right to worry. My point is more modest. Artificial intelligence is not a bubble in the sense of being nothing; it is plainly going to be an impactful technology that drives real social change. But the economic relationships that underpin it are new, and they are being worked out very quickly and at an unprecedented scale. For those of us on the outside, the blended worlds of finance and technology that produce Large Language Models (LLMs) can be difficult to keep track of.

It strikes me that the more of us paying attention, the better. So here is a primer on the main links in the emerging AI supply chain: who makes what, who rents from whom, and where the money flows.

Chips

The biggest beneficiary of the AI boom in share-price terms has been Nvidia. The company traditionally made graphics cards for gaming, and it turns out the calculations required to train an AI model are very similar to the ones a game demands: enormous numbers of simple sums, all running in parallel. Nvidia’s current benchmark ‘Blackwell’ chips (soon to be succeeded by its ‘Rubin’ range) have become a geopolitical hot potato; both the Biden and Trump administrations have maintained bans on exporting the most capable versions to China. Plenty of other firms make chips, and several companies further down the chain, including Google, Amazon and OpenAI, are racing to develop them too. But for now, no one quite replicates what Nvidia produces.

As an aside, almost every chip mentioned here, whoever designs it, is physically manufactured by the Taiwanese company TSMC, which specialises solely in making chips for others and can therefore be trusted not to steal the cutting-edge designs it is handed.

Datacentres

The chips need power to run, and shelter from the elements, and the quantities involved are vast. A set of specialist firms has grown up to provide ‘compute’ to everyone else: they find the land, secure the power, and build the infrastructure to keep racks of chips running, often on cheap ex-agricultural land far from cities. CoreWeave and Oracle are two prominent examples. They then rent out access to their chips. This is, in effect, the AI-specific version of the cloud businesses that Microsoft and Amazon have run for years, and both of those giants are major players at this level too.

Model development

Anthropic, OpenAI and Google DeepMind sit at the vanguard of training the most capable models. ‘Training’ is the computationally heavy process that establishes a set of mathematical ‘weights’ inside a model; these weights are what make it behave the way it does. It is worth understanding one counterintuitive point here: although a model can appear to learn from your conversation, those weights do not change as you talk to it. The model is fixed until the lab trains and releases a new one.

The labs doing this work need two expensive things: the most brilliant minds available, and a great deal of computing power. They build their own datacentres and also rent time on other people’s; they will take all the compute they can get.

Inference

‘Inference’ is the term for actually using a trained model to do work. It is a different computational task from training. It leans more heavily on memory bandwidth than raw processing, and so it draws on a slightly different profile of datacentre service. When you type a question into Claude or ChatGPT, you are using an inference service: a consumer-facing interface, backed by a model and the necessary computing power. Many of us now pay for these services. Claude and OpenAI both train models and then make them available through an interface. But there are also businesses that offer an inference service without training any models of their own; they often bundle several behind a single interface of their own design.

Model users

This is you, sitting and typing into the chat window with your favourite model. But the largest users are often the big technology companies themselves, whose staff are heavily encouraged to lean on AI in their daily work. A less visible set of heavy users sits in financial services. Firms such as Jane Street and Hudson River Trading (a hybrid of market-maker and quantitative hedge fund) have built their own internal models on the same underlying ‘transformer’ architecture that Google pioneered in 2017, and that powers the language models the rest of us use. Much as weather forecasting has crept further into the future over recent decades, these firms use their models to extend how far ahead they can anticipate the movement of market prices. That foresight is a large part of how they make money.

What characterises this whole supply chain is change. Google designs chips, builds datacentres, trains models and sells inference. Microsoft currently relies on OpenAI for its model, but would like to train its own. OpenAI has a nascent chip it is building. Nvidia is developing datacentres and a model. The lines between supplier, customer and competitor are blurred in a way that makes the economics hard to predict. This will solidify over time.  

Why does any of this matter to your financial plan? Because a remarkable share of the value in global stock markets is now concentrated in this supply chain, which means a big share of most ordinary investors’ portfolios is too. There are also some heady claims about the impact these technologies are going to have on work and society. We could regret our decisions about AI long before its children start to devour the planet; my instinct is that the more we try to understand what is going on from a technical and economic standpoint, the better placed we all are to hold our leaders accountable.

If you would like to talk through what sits inside your own portfolio, you can get in touch here.