# The AI boom's picks and shovels

_Before the software wins, someone has to build the infrastructure. Understanding where the money actually flows in an AI boom, and why compute, power and physical infrastructure come first, is the prerequisite for reading the investment theses that follow._

Neil Woodford · 23 June 2026 · 7 min read

![The AI boom's picks and shovels](https://cdn.sanity.io/images/v3acfbvo/production/305d09f383396c5fa1a8dce4c2e125f0adbc968d-2240x1260.png?w=1600&fit=max&auto=format)

---

In my [AI thesis](https://www.noisecancelling.co/topics/ai-industrial-revolution), the revolution is industrial before it is intelligent. Every large-language model, every AI agent, every productivity tool runs on physical infrastructure that has to be built, powered and cooled before any of the software value can be realised. The money, in my view, goes to the infrastructure layer first.

_[Embedded media](https://en.wikipedia.org/wiki/Railway_Mania)_

This is not a new pattern. It played out in the railway boom of the 1840s, in the electrification of industry in the late nineteenth century, and in the internet build-out of the 1990s. In each case, the picks-and-shovels suppliers, the companies selling the tools of the revolution rather than trying to win it, captured durable economic value even when the final applications were hard to predict and many early software and content businesses failed.

The AI iteration of this pattern appears to be underway. Understanding it requires understanding three things: what the infrastructure actually consists of, who is paying for it and why, and where the genuine bottlenecks lie.

## What the infrastructure consists of

Running an AI model is extraordinarily compute-intensive. Training a large model, the process of adjusting billions of numerical parameters until the model can usefully predict the next word, classify an image or reason through a problem, requires weeks or months of continuous computation across thousands of specialised chips. Inference, the moment the trained model answers your question, is cheaper per query but runs at enormous scale, billions of times a day.

The chips that do this work are overwhelmingly graphics processing units (GPUs), alongside large numbers of advanced memory chips called high-bandwidth memory (HBM) chips. GPUs were originally designed to render video game graphics, where millions of similar calculations have to run in parallel. Training and running AI models is structurally the same kind of problem: massive parallel arithmetic rather than the sequential logic that conventional central processing units (CPUs) were designed for. _(Neil in the margin: HBM stacks memory chips vertically and sits right next to the GPU, feeding it data fast enough to keep all those parallel cores busy. It's a genuine bottleneck — only a handful of firms can make it, which is why he names them later.)_

Those chips have to sit somewhere: in data centres, the large warehouse-scale buildings that house the chips and servers. A modern AI data centre requires extraordinary amounts of electricity. A large facility might draw 100 megawatts or more, enough to power a city of 80,000 homes. The chips also generate heat, which requires substantial cooling infrastructure, itself another engineering and energy constraint. _(Neil in the margin: For scale, that's roughly a tenth of a typical nuclear reactor's output running continuously, for a single building. The frontier sites now being planned are measured in gigawatts — power-station territory.)_

The pipeline from raw silicon to a working AI data centre is therefore: chip design, chip fabrication, server assembly, data centre construction, power connection and cooling. Every link in that chain is a bottleneck at different points in the build cycle.

## Who is paying, and why does it matter?

In 2024 and 2025, the four largest technology companies in the world, Microsoft, Alphabet (Google's parent), Amazon and Meta, each announced capital expenditure programmes running into tens of billions of dollars annually, a significant portion of which was directed at AI infrastructure. In 2026, Microsoft alone guided to over $80 billion in capital spending, much of it on data centres.

These are not speculative dot-com promises. They are real capital deployments by businesses with genuine, substantial earnings. Microsoft, Meta, Alphabet and Amazon each generate free cash flow of between $50 billion and $100 billion a year. These companies, among others, are now spending a large share of that cash on AI infrastructure because they are competing with one another and with several Chinese companies for a globally dominant position in what each believes will be the next platform shift in computing technology. _(Neil in the margin: The cash left after running costs and the capital spending needed to keep the lights on — the genuinely discretionary money. Neil's distinction from the dot-coms turns on this: real cash being redeployed, not freshly raised equity being burned.)_

This is the distinction between the current AI build-out and the internet bubble of 1999 and 2000. In that earlier cycle, much of the spending was financed by newly issued equity in what were typically loss-making companies with entirely speculative revenue projections. Most of the money they raised was spent on building brand awareness, aggressive marketing and TV advertising. The eventual, inevitable crash was a reckoning between the capital deployed and the cash generated.

The current cycle has many different characteristics and different financing structures: for example, some of the biggest companies building AI infrastructure, although clearly not all, are among the most profitable businesses ever to have existed, and are funding the build predominantly from cash earnings. Others like OpenAI, Anthropic and SpaceX are funding their investments from equity and debt issuance, but most of the investment is being deployed in physical assets. Whether these will deliver the returns the companies are anticipating and their valuations demand is not yet clear, indeed far from it, but the assets themselves have a future utility that never applied to businesses like boo.com and Pets.com _(Neil in the margin: Two emblematic dot-com flameouts: boo.com, a UK fashion retailer, burned roughly $135m in 18 months; Pets.com famously spent lavishly on advertising before collapsing in 2000. The contrast is that they left behind no durable assets.)_

## The picks and shovels

The phrase "picks and shovels" comes from the California Gold Rush, where the surest profit was not in panning for gold but in selling the tools to those who were. The equivalent suppliers in the AI build-out are the companies whose revenues are secured by the infrastructure spending itself, rather than by whether any particular AI application succeeds.

_[Watch: Neil's explanation of the picks and shovels suppliers — Are we in an AI bubble? And will the UK need an IMF bailout?](https://www.noisecancelling.co/the-show)_

Nvidia is the most obvious example. Nvidia designs the GPU chips that run most large AI models. It does not fabricate them. That work goes to TSMC (Taiwan Semiconductor Manufacturing Company), which manufactures to Nvidia's designs using the most advanced lithography processes in the world. The combination of Nvidia's chip architecture and TSMC's fabrication capability represents a concentration of the AI supply chain in a very small number of hands. Other picks and shovels suppliers would also include the companies that manufacture the machines that make semiconductors like ASML, and the small number of globally significant manufactures of memory chips and particularly, HBM chips including SK Hynix, Micron and Samsung. _(Neil in the margin: Lithography is the printing of circuit patterns onto silicon; the cutting edge uses extreme-ultraviolet machines made solely by ASML — which is why he flags ASML next. The whole leading edge funnels through a startlingly small number of suppliers.)_

_These companies appear here as worked examples of the infrastructure layer, not as current positions or recommendations. The point is the pattern, not the stock._

Below the chip layer sit several others. Advanced packaging, the process of connecting multiple chips together into a single package that performs better than any individual chip, has become a critical constraint. Power semiconductors, which convert and regulate electricity at the voltages these facilities need, are in short supply. The power utilities that will eventually connect new data centres to the grid face a pipeline of connection requests that, in some regions of the United States and the United Kingdom, stretches years into the future.

## The power constraint

The power question has become one of the most discussed bottlenecks in the AI build-out. The International Energy Agency estimated in 2024 that data centres already consumed about 1 to 1.5% of global electricity demand, and that AI-specific growth could push that figure substantially higher through the rest of this decade.

That creates pressure at two points. First, the energy companies that supply the power: utilities with grid infrastructure near large data centre clusters, gas peakers that can provide reliable power when renewable generation varies, and nuclear operators whose output is both reliable and low-carbon. The large technology companies have signed power purchase agreements, long-term contracts to buy electricity directly from generators, at a pace not seen since the early years of renewable energy deployment. _(Neil in the margin: "Peaker" plants are gas turbines kept on standby to fire up quickly when demand spikes or wind and solar fall away. They're expensive per unit of power but valued for being dispatchable on demand — the opposite of intermittent renewables.)_

Second, the companies that build and equip the data centres themselves: the specialist construction businesses, the providers of cooling systems, the makers of high-voltage switchgear and transformers. Transformer manufacturing became a public bottleneck in 2023 and 2024. Lead times for large power transformers, which had historically been 12 to 18 months, now extend to three years or more in some cases, as the grid simultaneously needs to absorb new renewable generation and supply new AI load.

## The bottleneck cycle

These constraints are not permanent. They are a feature of any rapid industrial build-out: the moment demand accelerates faster than supply chains can respond, shortages develop and prices rise. Suppliers invest in new capacity. Over two to five years, the shortage typically eases.

This cycle has implications for how to read the investment opportunity. Early in a build-out, with demand growing rapidly and supply constrained, the pricing power tends to sit with the infrastructure suppliers. Later, as supply catches up, the competitive advantage may shift to the applications layer, the businesses whose AI-powered products actually win customers and generate revenue.

My judgment is that the productivity gains from AI will be real and significant. They are potentially transformational for the service sector in particular, but the precise shape and timing remain genuinely uncertain. That uncertainty, not about whether AI matters, but about which applications win and on what timescale, is arguably why the infrastructure layer is the more attractive in the early stages of a build-out. The chips and the power have to be paid for regardless of which software wins.

## The 1999 comparison, revisited

The analogy to the late 1990s is both useful and potentially misleading. The internet bubble saw genuine infrastructure overbuild in fibre-optic cable: so much was installed that it took nearly a decade for demand to catch up with the capacity. They were, however real productive assets. They were just overbuilt relative to near-term demand, and the companies that built them, sometimes with debt, were often destroyed in the process, while the infrastructure itself proved durable and was eventually used. _(Neil in the margin: Firms like Global Crossing and WorldCom laid vast quantities of fibre on optimistic traffic forecasts; most sat unused as "dark fibre" for years. The cautionary parallel is precise — sound assets, ruinously mistimed financing.)_

_[Embedded media](https://www.youtube.com/watch?v=zPWbNPHHy8o)_

The current AI infrastructure build might follow a similar path: real physical assets, potentially overbuilt for near-term demand, with capital at risk even if the underlying long-term thesis is correct. That is a distinct risk and one we will be monitoring as closely as we can, but the lessons from history (canals, railways, Japanese real estate, internet infrastructure, US real estate and housing and infrastructure in China) show that the aftermath of infrastructure overbuild can be painful both for the industry participants and for those banks and investors that lent money to and invested in these projects.

## What this changes

Understanding what drives the infrastructure layer does not tell an investor which companies to back but it should help them to better understand the AI investment landscape as a whole: which revenues are secured by infrastructure spending that is already committed, which depend on applications that are still competing for customers, and where the physical constraints sit that govern the pace of the whole build-out.