The Picks and Shovels Strategy: Investing in the Infrastructure Behind AI

If you cannot tell which miner will strike gold, buy the company that sells shovels to all of them. In AI, the shovels are chips, cloud and data platforms.
Every technology revolution produces the same strategic debate among investors: should you back the applications, or the infrastructure that every application needs?
In 2021 the infrastructure side of the AI trade has been extraordinarily strong. Nvidia’s market capitalisation has passed $750 billion, driven increasingly by data centre demand rather than gaming. TSMC announced plans to invest $100 billion over three years in new capacity. The global chip shortage has shut down car production lines and reminded every board that semiconductors are not a commodity; they are a strategic resource. Cloud providers continue to grow at rates that would be remarkable for companies a tenth of their size. Snowflake and Databricks have become two of the most valuable private and public data companies in the world.
Why the logic works
The appeal of the infrastructure strategy is simple. The application layer of AI is still fragmented. Thousands of startups are competing in hundreds of use cases, and it is genuinely hard to know which will survive. Infrastructure, on the other hand, is needed regardless of who wins. Every model must be trained somewhere, every dataset must be stored and processed, every prediction must run on some piece of hardware.
This logic is familiar from earlier waves. In the internet era, the companies providing networking equipment, data centres and payment rails often captured value more predictably than the majority of consumer websites. In the mobile era, chip designers and app store owners did very well while many app makers struggled.
Where the logic can break
That said, infrastructure investing is not free of risk. I see three issues that investors should think through carefully.
The first is cyclicality. Semiconductors have always been a cyclical industry. Shortages today lead to capacity investments that can produce gluts tomorrow. A large part of this year’s demand also comes from cryptocurrency mining, which is notoriously volatile.
The second is concentration. A small number of hyperscale cloud providers buy an enormous share of AI hardware. They also design their own chips. Google already has its TPU, Amazon has Inferentia and Graviton. If the biggest customers become competitors, pricing power can shift quickly.
The third is geopolitics. The proposed acquisition of Arm by Nvidia, announced last year for $40 billion, is facing regulatory resistance in the UK, the EU, the US and China. Semiconductor supply chains are concentrated in Taiwan and South Korea. Infrastructure is becoming a matter of national security, and that brings both subsidies and restrictions.
The middle layer
Between the hardware and the applications sits a layer that I find particularly interesting from an investment perspective: data infrastructure and machine learning operations. Tools to clean, label, govern, version and monitor data and models. These businesses are less exposed to the hardware cycle, often have strong recurring revenue, and benefit from the fact that most companies discover that their real AI problem is not the algorithm but the data pipeline.
When I talk to CIOs and CTOs, the bottleneck they describe is rarely “we need a better model”. It is “we cannot get clean, consistent data into production”. That is a sign of where spending will flow.
The real question
The question for infrastructure providers is whether rising demand for compute will translate into durable pricing power or whether it will be competed away as capacity expands. My instinct is that the answer will differ across the stack. Commodity components will see margins compressed. Proprietary platforms with ecosystems around them, software and developer tools included, may keep their premium for a long time.
Investment question: Is it better to invest in companies developing AI applications or in the infrastructure that every AI company will need?
