Our guides recommend boring index funds inside an ISA. That is genuinely where most UK beginners should start. But Zo has been curious about a question: if AI is going to keep growing, who actually benefits beyond NVIDIA? This is what we have been learning.
Why Everyone Thinks "AI = NVIDIA"
NVIDIA dominates AI headlines because its GPUs are the workhorses of AI model training. When companies like OpenAI, Google, or Meta want to train large language models, they buy thousands of NVIDIA H100 or B200 GPUs. NVIDIA's data centre revenue grew from $15 billion in 2022 to over $100 billion in 2025.
But here is the thing: NVIDIA does not operate in isolation. Every GPU needs memory. Every GPU rack needs networking. Every data centre needs power, cooling, and physical infrastructure. NVIDIA is the tip of an iceberg — and many investors are exploring what lies beneath.
The AI Infrastructure Stack
Think of AI infrastructure as a layer cake. Each layer is essential, and each has its own set of companies:
Layer 1: Chips (GPUs and Custom Silicon)
NVIDIA is the market leader, but it is not alone. AMD makes competing GPUs (MI300 series). Broadcom and Marvell design custom AI chips (ASICs) for hyperscalers like Google and Amazon who want alternatives to NVIDIA dependency.
The logic for custom silicon: if you are Google and you spend billions on NVIDIA GPUs, you have strong motivation to design your own chips (Google's TPUs) or commission custom ones from Broadcom to reduce your dependence on a single supplier.
Layer 2: Networking and Optics
This is the layer Zo has found most interesting. Modern AI training requires thousands of GPUs to communicate simultaneously — and they need to do it at extraordinary speeds (400Gbps and 800Gbps per connection, moving to 1.6Tbps).
The companies at this layer make the optical transceivers, coherent optics, and networking switches that enable this communication:
- Coherent Corp — optical transceivers and laser sources for data centre interconnects
- Lumentum — laser chips used inside high-speed transceivers
- Applied Optoelectronics (AAOI) — optical transceivers, focused on data centre customers
- Broadcom (networking division) — Tomahawk and Jericho switching ASICs that route traffic between GPUs
The thesis: as AI clusters grow from thousands to hundreds of thousands of GPUs, the networking and optical component demand grows proportionally — possibly even faster, because larger clusters need more interconnection bandwidth.
Layer 3: Data Centres and Neoclouds
GPUs need to live somewhere. The traditional hyperscalers (AWS, Azure, Google Cloud) are building massive data centres, but a new category has emerged — sometimes called "neoclouds" — that provide GPU compute specifically for AI workloads:
- Nebius — spun out of Yandex, building GPU-focused data centres in Europe and the US
- CoreWeave — GPU cloud provider backed by NVIDIA, recently IPO'd
- IREN (formerly Iris Energy) — data centre operator expanding into AI/HPC hosting
These are high-growth, high-capital-expenditure businesses. They are raising and spending billions to build out capacity as fast as possible. That creates both enormous upside potential and very real bankruptcy risk if AI demand slows.
Layer 4: Memory
AI models are memory-hungry. High-bandwidth memory (HBM) is stacked directly onto GPU packages to feed them data fast enough. The memory market is dominated by three players:
- Micron (US-listed)
- SK Hynix (Korean-listed)
- Samsung (Korean-listed)
HBM has become a major revenue driver — SK Hynix in particular has benefited enormously from being NVIDIA's primary HBM supplier. Micron is the most accessible for UK investors via US exchanges.
Layer 5: Cooling and Power
A modern AI data centre consumes staggering amounts of electricity and generates enormous heat. Each GPU rack can draw 60-100kW of power. Companies providing cooling and power infrastructure include:
- Vertiv — thermal management and power distribution for data centres
- Eaton — electrical systems, UPS units, power management
These are more mature, industrial businesses — less explosive growth than the chip companies, but also less likely to collapse if AI hype fades, because data centres need cooling regardless of whether the workload is AI or traditional cloud computing.
The "Second Derivative" Thesis
Some investors describe this as investing in the "second derivative" of AI:
- First derivative: companies that sell AI products (OpenAI, Google, Meta)
- Second derivative: companies that sell to the first derivative (NVIDIA sells GPUs to AI companies)
- Third derivative: companies that sell to the second derivative (optical companies that sell to NVIDIA's customers)
The theory is that third-derivative companies are less efficiently priced because fewer analysts cover them, they have smaller market caps, and the connection to the AI theme is less obvious. Whether that theory holds in practice is debatable — many of these stocks have already risen 200-500% since early 2023 as the market caught on.
How UK Investors Can Access These Stocks
The vast majority of AI infrastructure companies are listed on US exchanges. UK investors can access them through several brokers:
- Trading 212 — commission-free US share dealing, fractional shares available
- Hargreaves Lansdown — wide selection of US stocks, £11.95 per trade
- AJ Bell — from £5 per US trade, good research tools
- Interactive Investor — flat monthly fee includes US trading
All of these let you hold US shares inside a Stocks & Shares ISA, so any capital gains are tax-free. You will need to complete a W-8BEN form to reduce US dividend withholding tax from 30% to 15%.
The Simpler Alternative: Just Buy the Index
Here is something worth understanding: a single global index fund like Vanguard FTSE All-World (VWRL) or iShares MSCI ACWI already gives you meaningful exposure to this theme. NVIDIA, AMD, Broadcom, Micron, and Vertiv are all in the index, weighted by market cap.
As these companies grow and their revenues increase, their weighting in the index automatically increases too. You do not need to pick the individual winners — the index does it for you, retrospectively, at zero research cost.
The trade-off is that an index fund will never give you the outsized returns of picking a single stock that triples. But it also protects you from the scenario where a single stock you picked drops 80% because a technology shift made it obsolete.
The Risks You Must Understand
Before anyone considers individual AI infrastructure stocks, understand these risks:
- Cyclicality: Semiconductor and data centre spending is cyclical. Booms are followed by busts. The AI spending cycle could slow just as sharply as it ramped up.
- Valuation: Many AI infrastructure stocks already trade at 30-60x earnings. That prices in years of continued growth. If growth disappoints even slightly, the stocks can fall 30-50% rapidly.
- Technology risk: Today's bottleneck can become tomorrow's commodity. If NVIDIA moves optical networking on-chip, separate optical companies lose their market.
- Concentration risk: Buying individual stocks means your portfolio's fate depends on a small number of companies rather than the broad market.
- Liquidity risk: Smaller companies can have thin trading volumes, meaning you might not be able to sell at a good price when you want to.
FAQ
What are AI infrastructure stocks?
AI infrastructure stocks are companies that provide the physical hardware, networking, power, cooling, and data centre capacity that AI systems need to operate. This includes GPU makers, optical networking firms, memory manufacturers, data centre operators, and power/cooling companies.
Can UK investors buy AI infrastructure stocks?
Yes. Most AI infrastructure companies are listed on US exchanges (NASDAQ or NYSE). UK investors can buy them through brokers that offer US share dealing — including Trading 212, Hargreaves Lansdown, AJ Bell, and Interactive Investor. You can hold them inside a Stocks and Shares ISA.
Is it better to buy individual AI stocks or an index fund?
For most investors, a global index fund is the safer and simpler approach. A fund like Vanguard FTSE All-World already includes NVIDIA, AMD, Broadcom, and other AI beneficiaries weighted by market cap. Individual stock picking concentrates risk and requires constant research — most professionals underperform index funds over time.
What is the second derivative thesis in AI investing?
The second derivative thesis means investing not in the AI companies themselves, but in the companies that AI companies buy from. If NVIDIA sells GPUs, the second derivative is whoever sells critical components to NVIDIA — optical transceivers, advanced packaging, memory chips, and so on.
What are the risks of investing in AI infrastructure stocks?
Key risks include: AI spending cycles could slow, creating sharp drawdowns; many companies trade at very high valuations already pricing in years of growth; technology shifts could make current infrastructure obsolete; smaller companies face liquidity risk and higher volatility; and concentration in one theme leaves you exposed if AI hype fades.
Related reading: