A layered map of the AI compute supply chain, from the companies that pay to the chips, memory, packaging, networks, power and buildings, with the latest dated number for each layer and where the bottlenecks and risks sit.
Which stocks do Wall Street’s best‑ranked analysts agree on? See them in Exent Gems →
It is the chain of companies that turns AI budgets into running GPU clusters: buyers sign contracts, clouds and neoclouds buy chips and rent buildings, and suppliers deliver processors, memory, servers, networks, cooling and electricity, funded by lenders and shareholders. The map below reads top to bottom, following the money. Scope: listed companies in the US and Asia that sell into AI data centers, with dated figures from their own releases. Excluded: software and model companies, and private suppliers.
Every layer grew in 2026, at very different speeds. The latest figure each company reported, with its date:
| Layer | Company | Latest figure (reported) | Period | Bottleneck or risk |
|---|---|---|---|---|
| GPUs | NVIDIA | Data Center revenue $89.0B (+117%); Q3 outlook $108.0B total | Quarter to Jul 26, 2026 | A few large buyers; fast product cycles |
| Custom chips, networking | Broadcom | AI semiconductor revenue $16.7B (+221%); Q4 outlook $21.7B | Quarter to Aug 2, 2026 | Concentrated in a handful of hyperscaler designs |
| Networking | Arista | Revenue $3.036B (+37.7%) | Quarter to Jun 30, 2026 | Competes with NVIDIA’s own networking |
| HBM memory | SK hynix | Revenue 79.3T won, operating profit 60.5T won | Q2 2026 | Memory has a long history of boom-and-bust pricing |
| HBM memory | Micron | Revenue $54.23B (a year earlier $11.32B) | Quarter to Sep 3, 2026 | Same cycle risk; capacity takes years to add |
| Foundry, packaging | TSMC | Revenue US$46.72B (+41.2%); HPC 68% of revenue | Q3 2026 | Advanced packaging capacity; geographic concentration |
| Power, cooling | Vertiv | Net sales $3,274M (+24%) | Q2 2026 | Order timing tied to data-center build schedules |
| Turbines | GE Vernova | Gas backlog + slot reservations 116 GW (83 GW at end-2025) | Jun 30, 2026 | Manufacturing capacity; years of lead time |
| Neoclouds | CoreWeave | Revenue backlog ~$104B; Q2 net loss $626M | Jun 30, 2026 | Debt, customer concentration, delivery |
The buyers at the top are few: what Microsoft, Alphabet, Amazon and Meta plan to spend is tracked in big-tech AI spending in 2026. The neocloud layer, where the financial stress showed in October 2026, is mapped in what is a neocloud and compared company by company in neocloud stocks compared.
In 2026 the three tightest points are high-bandwidth memory, advanced packaging and electrical power. HBM is stacked next to each GPU, and only three companies make it at scale; Counterpoint Research put their Q2 2026 revenue shares at 50% (SK hynix), 33% (Samsung) and 18% (Micron). Packaging that joins GPUs and memory is concentrated at TSMC, whose high-performance computing business made up 68% of its Q3 2026 revenue. Power is slowest to add: GE Vernova’s gas-turbine backlog plus slot reservations rose from 83 GW at the end of 2025 to 116 GW by June 30, 2026, and the company expects at least 125 GW by year-end. On the demand side, Goldman Sachs Research projected in May 2026 that US data-center power demand would rise from 31 GW in 2025 to 66 GW in 2027 (a projection, not a measured figure). A bottleneck usually means pricing power for the supplier that owns it, until new capacity arrives.
Three ratios from the table tell you more than the headline: growth, concentration and margin. Each uses only the reported numbers.
| Question | Arithmetic | Result | What it tells you |
|---|---|---|---|
| How fast did Micron grow? | $54.23B ÷ $11.32B | 4.8× (+379%) | Memory prices and volumes both rose; cyclical sectors can reverse |
| How dependent is NVIDIA on data centers? | $89.0B ÷ $96.2B | 92.5% of revenue | One end market drives almost everything |
| How profitable is the HBM leader? | 60.5T won ÷ 79.3T won | 76% operating margin | Shortage-level margin; it falls if supply catches up |
Interpretation (ours): very high growth plus a very high margin is what a bottleneck looks like at its peak. It is also what a peak looks like in a cyclical industry, which is why the next step is always to ask what happens to price when capacity is added.
Very unevenly: suppliers in the same layer moved in opposite directions, and the highest volatility was in the neocloud layer. Exent market data, year to September 30, 2026:
| Stock | Layer | 1-year total return | Largest drop in the year (daily) | Annualized volatility |
|---|---|---|---|---|
| NVIDIA (NVDA) | GPUs | +22.7% | −20.2% | 33.6% |
| AMD | GPUs | +278.1% | −27.8% | 67.8% |
| Broadcom (AVGO) | Custom chips, networking | +7.2% | −33.0% | 46.2% |
| Micron (MU) | HBM memory | +537.6% | −30.3% | 65.0% |
| Dell (DELL) | Servers | +284.3% | −32.3% | 73.5% |
| Supermicro (SMCI) | Servers | −14.3% | −66.2% | 81.7% |
| Arista (ANET) | Networking | +39.7% | −28.3% | 54.4% |
| Vertiv (VRT) | Power, cooling | +60.1% | −24.8% | 56.6% |
| GE Vernova (GEV) | Turbines | +54.9% | −30.6% | 53.4% |
| Oracle (ORCL) | Hyperscaler cloud | −50.6% | −58.3% | 60.7% |
| CoreWeave (CRWV) | Neocloud | −36.3% | −61.2% | 95.7% |
Past returns are history, not a forecast, and these figures end before the October 2026 selloff in neocloud shares.
By asking which layer is short of capacity, who owns that shortage, and how long it will last, rather than by buying the theme. The “picks and shovels” idea is explained in picks-and-shovels investing, and what past big winners had in common in how to find the next Nvidia. A checklist for any supplier:
It goes wrong when spending slows or capacity catches up, and the layers with the most borrowing and the most cyclical pricing feel it first. Gartner said on October 14, 2026 that over $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030 and that over 80% of neocloud providers would cease independent operations by then, which includes being acquired. A pullback by one of the few large buyers would travel down every arrow in the map. Who should not rely on this page: anyone looking for a list of stocks to buy; the map shows where to look, not what to own. Limitations: company figures use different fiscal quarters, private suppliers are missing, and market-share data come from one research firm. For sizing a position in a theme this volatile, see speculative sleeve size.
If you look for growth companies in a theme like this, the hard part is the shortlist; Exent tracks Wall Street analysts and ranks them by their past accuracy.
Chipmakers (NVIDIA, AMD, Broadcom), memory makers (SK hynix, Samsung, Micron), TSMC for manufacturing and packaging, server makers (Dell, Supermicro), networking (NVIDIA, Arista, Broadcom), power and cooling (Vertiv, GE Vernova) and data-center landlords.
Three companies: SK hynix, Samsung and Micron. Counterpoint Research put their Q2 2026 HBM revenue shares at 50%, 33% and 18%.
Manufacturing of advanced AI processors, and the packaging that joins them to HBM memory, is concentrated at TSMC. High-performance computing was 68% of TSMC’s Q3 2026 revenue of US$46.72B; packaging capacity takes years to expand.
Because generation and grid equipment take years to build. GE Vernova’s gas-turbine backlog plus slot reservations reached 116 GW at Jun 30, 2026, and Goldman Sachs projects US data-center power demand of 66 GW in 2027, up from 31 GW in 2025.
NVIDIA supplies its GPUs and is also an investor and a customer; server, networking and power-equipment makers supply the racks; landlords such as Applied Digital, Galaxy and Core Scientific lease it data-center space.
Partly. Neoclouds are among NVIDIA’s customers, but data-center revenue of $89.0B in one quarter comes mostly from much larger buyers, so NVIDIA’s results depend more on Big Tech spending than on any single neocloud.
There is no general answer. In the year to Sep 30, 2026, results ranged from +537.6% for Micron to −36.3% for CoreWeave in Exent market data, so the layer and the company matter more than the theme. Past returns are not a forecast.
Both connect GPUs into one cluster. InfiniBand is NVIDIA’s specialized networking; Ethernet is the open standard used by Arista, Broadcom and others, and AI builders use both.
Exent tracks 4,000+ analysts, keeps only the elite few with proven records, and surfaces the Gems where they agree.
Educational content, not investment advice. Examples are illustrative; past performance does not guarantee future results.