Hyperscalers plan to spend more than $700 billion on capex in 2026. Here is the supply chain that money flows through, from chips and memory to power plants, and who sits where in it.
The four largest cloud companies plan to spend roughly $720 billion to $745 billion on capital expenditures in 2026, most of it on AI data centers (TMT Finance). Every dollar of that flows down a supply chain of chips, memory, cables, servers, buildings, and power plants. Understanding who sits where in that chain is the key to reading AI news as a trader.
“AI infrastructure” is the physical and financial plumbing that lets companies train and run AI models. It is not the chatbots or apps themselves. It is the hardware and facilities underneath them, and the companies that sell into that build-out.
Think of it as a stack with eight layers:
The layers move together, but not in lockstep. When a hyperscaler raises its capex budget, chip and memory stocks usually react first. Power and construction names follow as projects break ground. When the market worries about AI returns, the most leveraged layers, like the neoclouds, tend to fall hardest.
Read it top to bottom: buyers’ budgets flow down to system builders, then to chip designers, then to the factories. Power and cooling sit beside the stack because every data center needs them.
Compute is the largest and most-watched layer. AI models run on accelerators: chips built to do enormous numbers of simple calculations in parallel. Two approaches compete here. Merchant GPUs are sold to anyone. Custom chips (ASICs) are designed for one customer’s workloads.
Nvidia (NVDA) is the dominant supplier. Its data center segment brought in $89.0 billion in fiscal Q2 2027 (May–Jul 2026), up 117% from a year earlier, and the company guided the next quarter to about $108 billion in total revenue (Nvidia). Nvidia sells far more than chips. It sells full rack-scale systems, networking gear, and the CUDA software platform that most AI developers are trained on. That software lock-in is a big part of why its lead has lasted. Its newest platform, Vera Rubin, is ramping into full production at CoreWeave, Google Cloud, Microsoft Azure, Oracle, and Nebius.
Advanced Micro Devices (AMD) is the main merchant alternative. Its Instinct GPUs compete on price and memory capacity, and it is building rack-scale systems of its own. For AMD, the key question is how quickly large customers commit to it as a true second source.
Broadcom (AVGO) leads the custom-chip business. It co-designs accelerators, which it calls XPUs, for companies that want to cut their dependence on Nvidia. Its AI semiconductor revenue hit $16.7 billion in fiscal Q3 2026 (May–Jul 2026), up 221% year over year, with about $21.7 billion guided for the next quarter. Named customers include Google (TPUs), OpenAI, and Meta (Converge Digest).
Marvell Technology (MRVL) is the other large custom-silicon designer. It also sells the electrical-to-optical chips inside data-center links.
Cerebras Systems (CBRS) builds wafer-sized AI chips and rents them out as a cloud service. It went public on Nasdaq at $185 a share in May 2026 (Cerebras).
CPUs still matter. They manage the accelerators and handle general tasks. Arm Holdings (ARM) licenses the designs behind most custom data-center CPUs, including Nvidia’s own Grace and Vera chips. Intel (INTC) and AMD sell traditional x86 server CPUs.
What to watch: Nvidia’s data-center guidance, Broadcom’s custom-chip commentary, and any sign that hyperscalers are shifting spend from GPUs toward in-house chips.
Nvidia, AMD, Broadcom, and Marvell design chips but do not manufacture them. That work goes to foundries, which rely on a small group of equipment makers. This layer is less visible than compute, but it is where many of the true bottlenecks sit.
Taiwan Semiconductor Manufacturing (TSM) manufactures nearly all leading-edge AI accelerators, for Nvidia, AMD, Broadcom’s customers, and Apple alike. It also owns most of the capacity for CoWoS, the advanced packaging method that joins a processor to its memory stacks on one module. When AI chip supply is tight, packaging capacity is often the reason. Intel (INTC) is trying to build an outside foundry business to compete, though it remains far behind TSMC in AI volume. Samsung also offers foundry services but trades in Korea, not on a US exchange.
The foundries buy their tools from a handful of specialists:
In packaging and test, Amkor Technology (AMKR) is the largest US-listed outsourced packaging company, and Teradyne (TER) makes the test equipment that checks AI chips and memory before they ship.
What to watch: TSMC’s monthly sales reports and capex plans, and equipment makers’ orders from memory producers. Equipment names often move ahead of the chip cycle, because factories order tools a year or more before output arrives.
An accelerator is only as fast as the data it can reach. That makes memory one of the tightest parts of the AI supply chain in 2026. Research cited by Benzinga estimates that memory can make up more than half the cost of an AI accelerator. Major suppliers have already committed their 2026 high-bandwidth memory output. Some industry voices expect relief only when new factories reach full production in 2028 or 2029 (Benzinga).
There are three main products:
The key players:
Memory has historically been a boom-and-bust business. Prices soar when supply is short, then collapse when new factories come online. The AI-era question is whether long-term HBM contracts make this cycle smoother than past ones.
What to watch: Micron’s earnings and pricing commentary, Korean export data, and any news of HBM supply falling short of accelerator demand.
Training a frontier model takes tens of thousands of accelerators working as one machine. Moving data between them is now almost as important as the chips themselves. This layer has three parts: scale-up (linking chips within a rack), scale-out (linking racks within a building), and scale-across (linking whole data centers).
Over short distances, copper cables carry signals. Beyond a few meters, data travels as light through optical transceivers built around lasers. In March 2026, Nvidia invested $2 billion each in Coherent and Lumentum, paired with multibillion-dollar purchase commitments, to secure laser capacity it sees as a coming bottleneck (Futurum).
A shift to watch is co-packaged optics, which moves the optical engine right onto the switch chip to save power. It could reshape who captures value in this layer.
What to watch: Arista’s and Broadcom’s networking commentary, optics makers’ guidance, and supply signs for indium phosphide lasers.
Chips arrive at the data center as finished servers and full racks. A modern AI rack can hold dozens of accelerators, with liquid cooling, power shelves, and networking already wired in. The companies that build these racks handle huge dollar volumes. Their margins are thin, though, because most of the value belongs to the chips inside.
Much hyperscaler hardware is also built by Taiwanese original design manufacturers such as Foxconn, Quanta, and Wistron. These trade in Taipei, not on US exchanges.
What to watch: order backlogs and gross margins. Rising AI server revenue alongside shrinking margins is a common pattern in this layer. It tells you whether a company is winning business profitably or just shipping boxes.
This layer writes the checks that fund everything above. It has three groups: hyperscalers, neoclouds, and data-center landlords.
The hyperscalers are the giant cloud and internet companies. They buy most of the world’s AI hardware, both to run their own AI products and to rent it to others. Their 2026 capex guidance, as compiled by TMT Finance in August 2026:
| Company | Ticker | 2026 capex guidance |
|---|---|---|
| Amazon | AMZN | about $220B |
| Alphabet | GOOGL | $195B–$205B |
| Microsoft | MSFT | about $175B (new lease-accounting basis) |
| Meta Platforms | META | $130B–$145B |
| Oracle | ORCL | about $40B (fiscal 2027) |
These budgets are the single most important driver for the whole stack. Investors now scrutinize them closely, because spending is running ahead of cash flow and some of it is being financed with debt. Oracle has become a major AI cloud provider largely through large, multi-year contracts with AI labs.
Neoclouds are newer companies that do one thing: rent out GPU capacity. They grow fast, borrow heavily, and depend on a few big customers, which makes their stocks the most volatile part of the stack.
Equinix (EQIX) and Digital Realty (DLR) are real estate investment trusts that own and lease data-center buildings. They are steadier than neoclouds, but less directly tied to AI chip demand.
The binding constraint on AI growth is increasingly electricity, not chips. Morgan Stanley estimates US data centers will need about 68 gigawatts of power from 2026 through 2028. After counting projects under construction and available grid capacity, that leaves a potential 38-gigawatt gap (Motley Fool). For scale, one gigawatt is roughly the output of a large nuclear reactor.
What to watch: utility interconnection queues, turbine and transformer lead times, and power-purchase deals between tech companies and generators. The power shortage is likely to outlast any single chip generation.
| Layer | What it provides | Major players (US tickers unless noted) |
|---|---|---|
| Compute | GPUs, custom accelerators, CPUs | NVDA, AMD, AVGO, MRVL, CBRS, ARM, INTC |
| Foundry and equipment | Chip manufacturing, tools, packaging, test | TSM, ASML, AMAT, LRCX, KLAC, AMKR, TER |
| Memory and storage | HBM, DRAM, NAND, hard drives | MU, SK Hynix (000660.KS), Samsung Electronics (005930.KS), SNDK, WDC, STX |
| Networking and optics | Switches, lasers, transceivers, cables | ANET, AVGO, CSCO, COHR, LITE, FN, CIEN, CRDO, ALAB, APH, GLW |
| Servers and systems | Rack-scale AI servers | DELL, SMCI, HPE, CLS |
| Hyperscalers | Cloud platforms and AI capex budgets | AMZN, MSFT, GOOGL, META, ORCL |
| Neoclouds | Rented GPU capacity | CRWV, NBIS, IREN, CORZ, APLD |
| Data-center REITs | Buildings and leases | EQIX, DLR |
| Power and cooling | Generation, electrical gear, liquid cooling | GEV, CEG, VST, TLN, BE, ETN, VRT |
| Construction | Grid and facility build-out | PWR, EME, FIX, FLR |
SK Hynix and Samsung trade on the Korea Exchange, not in the US. Schneider Electric (Paris) and the Taiwanese server makers are major players too, but also trade only on non-US exchanges.
Educational only. Not advice. Kapnotes provides research and commentary for informational and educational purposes only. Nothing here is a recommendation to buy or sell any security. Company names are used to illustrate how the AI supply chain is structured, not to endorse any investment.
Kapnotes is unbiased, catalyst-driven research for active traders. We screen the market's biggest movers and hand you a focused list of trading opportunities, updated in real time throughout the day.