From chatbots to chip racks: where AI profits may actually end up

Key takeaways

• Amazon’s Anthropic deal shows AI funding is increasingly about chips, cloud and power, not just software.

• ASML and TSMC suggest the AI build-out is still very much alive, and scarce capacity still holds pricing power.

• Nvidia still matters, but custom chips and long contracts are making the “picks and shovels” trade more specialised.

A chatbot writes the poem, answers the email, and gets all the applause. The bill, however, often lands somewhere else. That is what makes Amazon’s expanded Anthropic partnership so interesting. On 20 April 2026, Amazon announced it would invest up to USD 25 billion in Anthropic, while Anthropic committed to spend more than USD 100 billion over the next decade on Amazon Web Services technology. This is less a venture funding headline than a long-term capacity reservation dressed in startup clothing.

The funding round that is really a utility contract

The most revealing part of the Amazon-Anthropic announcement is not the cheque. It is the plumbing. Anthropic said it would secure up to 5 gigawatts of current and future Trainium chip capacity, expand inference, meaning the running of models after training, in Asia and Europe, and keep deepening its use of Amazon’s cloud stack. Amazon also said more than 100,000 customers already run Claude models on Amazon Web Services. That begins to sound less like backing a promising artificial intelligence lab and more like signing up a large industrial customer for the next decade.

That matters because AI is becoming capital-intensive in a very old-fashioned way. Software still matters, of course. But once demand scales, the winners are often the firms that own scarce inputs, not the ones with the flashiest demo. In AI, those scarce inputs are compute, advanced chips, networking, cooling, and the data-centre capacity needed to keep the whole system running. In other words, the chatbot may charm the user, but the rack gets paid.

The glamour is on the screen, but the cash may sit behind the curtain

The recent updates from ASML and TSMC make that harder to ignore. ASML lifted its 2026 revenue outlook as customers pressed ahead with expansion plans tied to artificial intelligence demand. A day later, TSMC raised its full-year growth outlook and pointed to capital spending landing at the high end of its range. That is not the tone of an industry stepping back for breath. It is the tone of builders asking for more concrete, more steel and a bigger site.

That is the more useful investor lens. Artificial intelligence is no longer only a software story. It is increasingly an industrial build-out, complete with bottlenecks, lead times and supply constraints. The visible part is still the chatbot, the assistant and the model demo. The less visible part is the factory, the foundry, the chip rack and the power bill. When demand runs ahead of supply, the companies that control key equipment and production capacity can end up with steadier economics than the businesses closer to the user.

Nvidia still sits at the centre of this world. Its graphics processing units, or GPUs, remain the benchmark for training advanced models. But the market is also shifting towards inference, which is the running of models after they have been trained. That part of the market places a bigger premium on speed, efficiency and cost. It opens the door to alternatives such as Google’s tensor processing units, or TPUs, Amazon’s Trainium chips and other custom designs built for narrower, more specific tasks. The artificial intelligence race is not moving away from infrastructure. It is moving deeper into it.

Not all shovels look the same anymore

This is where the old “just buy the shovel sellers” line starts to get a bit lazy. The shovels are no longer generic. Broadcom has signed a deal through 2031 to develop Google’s custom artificial intelligence chips and separately agreed to provide Anthropic access to about 3.5 gigawatts of artificial intelligence computing capacity from 2027. Google is also exploring additional chip designs with Marvell, including a memory processing unit and a new tensor processing unit aimed at running models more efficiently.

That changes the investor map. The earlier version of the AI trade was simple: Nvidia sells powerful chips, everyone queues up, end of story. The newer version is more crowded, more specialised and more strategic. Some customers want to reduce dependence on outside suppliers. Some want custom chips that lower costs for specific tasks. Some want tighter bundles that combine chips, software and cloud infrastructure into one sticky package. The profit pool may still sit with the picks and shovels, but now the tools are bespoke, the contracts are longer and switching costs may matter as much as the silicon itself. Slightly less poetic, perhaps, but much more useful.