AI Picks and Shovels: 38 Public Companies in the Compute Supply Chain Beyond NVDA

The $1 Trillion Question Everyone Asks About NVDA

Microsoft, Google, Amazon, and Meta have collectively committed over $300 billion in AI infrastructure capex through 2027 — and that number keeps climbing with each earnings call. The reflexive trade has been NVDA, and for good reason: Nvidia's H100 and B200 GPUs sit at the center of nearly every major training cluster on earth. But the question serious investors are now asking is not whether Nvidia wins — it is who else gets paid when hyperscalers write those nine-figure datacenter checks. The answer is a supply chain 38 companies wide, spanning silicon design, memory stacks, fiber optics, power conversion, liquid cooling, and the real estate that houses it all. This piece maps that entire chain using live screener data so you can see exactly where the market's relative-strength leaders and laggards are sitting today.

The 7 Layers of the AI Compute Stack

Before looking at individual tickers, it helps to understand the architecture. A hyperscaler AI cluster is not just a rack of GPUs. It is a vertically integrated system where bottlenecks cascade from one layer to the next. Here are the seven layers and what breaks if any one of them is constrained.

Compute — The GPU and custom ASIC layer. Trains and runs the models. Every other layer exists to feed and cool these chips.

Memory and Storage — High-bandwidth memory (HBM) stacked on the GPU die, plus persistent flash and spinning disk for training data and checkpoints. Memory is frequently the throughput ceiling, not compute.

Networking and Optical — The fabric that connects thousands of GPUs inside a cluster. At scale, network latency and bandwidth become as important as raw compute.

Cooling and Thermal — AI clusters run at power densities that legacy air cooling cannot handle. Liquid cooling and thermal management hardware is now a first-class design constraint.

Power and Electrical — The datacenter draws more power than many small cities. Everything from switchgear to power distribution units to backup batteries is in shortage.

Substrate, Packaging, and Test — The equipment and materials that build the chips. Advanced packaging (CoWoS, HBM bonding) is one of the hardest bottlenecks to scale quickly.

Datacenter Real Estate — The hyperscalers do not own all their compute. Colocation REITs and tower companies provide the physical infrastructure and connectivity.

The Master Table — 38 AI Supply Chain Stocks (Sorted by Market Cap)

Compute — GPUs and Custom Silicon

Compute is the layer everyone sees, but the investment debate has matured beyond "buy Nvidia." NVDA sits at ELO rank #173 with a +6% six-month return — still positive, but the bulk of the easy appreciation is in the rearview mirror. The more interesting dynamic is at the custom silicon layer: AVGO at rank #142 is executing on hyperscaler ASIC programs at Google and Meta that could displace meaningful GPU share over the next two to three years. AMD at rank #49 is the closest thing to a GPU second-source, with MI300X winning meaningful inference deployments. Both SNPS and CDNS are the silent enablers — every custom ASIC tapeout runs through their EDA toolchains, and neither faces the fab capacity constraints that hardware vendors do. Marvell Technology (MRVL, not in our screener) is another active player in the custom ASIC space, designing networking and compute silicon for hyperscaler programs alongside Broadcom. The EDA pair have lagged the hardware names on a six-month basis, which makes them worth monitoring for a catch-up rotation if the broader theme continues.

Memory and Storage — The Throughput Ceiling

The memory layer is quietly running some of the best relative strength numbers in the entire table. WDC is ranked #2 out of the entire screener universe — a top 0.2% placement — with a +213% six-month return. STX is right behind at #3 with +160%. Both are riding the AI training dataset storage wave, where petabyte-scale demand is growing faster than general enterprise storage. MU at rank #17 is the HBM story: HBM3e stacked memory is the bandwidth interface between GPU compute and the data it processes, and Micron is one of only three global suppliers. The memory cycle has historically been violent in both directions; the current strength reflects genuine AI-driven demand but also carries mean-reversion risk if hyperscaler inventory builds. PSTG (Pure Storage) is in a rally trend after giving back -30% over six months — the product is well-positioned for AI workloads, but the stock is recovering rather than leading.

Networking and Optical — The Invisible Backbone

Networking is the layer most retail investors underestimate. At the scale of a hyperscaler training cluster, moving data between ten thousand GPUs fast enough to keep them all busy requires switching bandwidth measured in petabits per second. ANET at rank #86 is the established Ethernet AI fabric player, though its six-month return of +5.4% is more modest than the optical component names. The optical side has been explosive: CIEN +159%, COHR +126%, AAOI +272%, and FN +45%. These companies make the transceivers and coherent optics modules that carry data at 400G and 800G inside and between campuses. CRDO is the less-followed name here — its active electrical cable technology is a power-efficient alternative to optical transceivers for very short hops between GPU racks inside a pod. Lumentum (LITE, not in our screener) is another optical component name worth researching as an adjacent player in the 800G transceiver ramp.

Cooling and Thermal — The Physics Problem Nobody Can Ignore

Nvidia's H100 GPU dissipates roughly 700 watts. The B200 pushes past 1,000 watts per chip. Multiply that by tens of thousands of chips per cluster and the resulting heat density makes traditional air cooling physically inadequate. VRT (Vertiv) at +60% six months is the consensus pick here — it designs the precision liquid cooling and power distribution systems that go directly into AI racks. MOD (Modine Manufacturing) at +46% is the smaller, higher-torque alternative — a company that has pivoted aggressively into datacenter thermal management from its industrial HVAC roots. JCI and TT are larger industrials with datacenter HVAC exposure, both in uptrend. AAON (not in our screener) is another notable HVAC manufacturer with datacenter-specific product lines worth researching as a smaller adjacent name. The cooling layer tends to be less cyclically volatile than the semiconductor equipment names because thermal infrastructure has a replacement cycle, not just a build cycle.

Power and Electrical — The Constraint Everyone Just Noticed

Power availability has become the binding constraint on datacenter expansion in several markets. GEV (GE Vernova) at ELO rank #6 — top 1.6% of the screener universe — has the most direct exposure: gas turbines and grid equipment are on multi-year backlogs as utilities scramble to provide dedicated power feeds to hyperscaler campuses. Its six-month return of +91% reflects that scarcity premium. ETN (Eaton) and HUBB (Hubbell) are the power distribution plays — switchgear, UPS, and electrical components that go into every new datacenter build. ENS (Enersys) is the battery backup angle: as datacenters shift from lead-acid to lithium backup systems, Enersys is capturing that upgrade cycle with +69% six-month performance. GE (GE Aerospace) technically appears in this layer due to the power generation overlap, but its primary business is aviation and it sits in a downtrend — the AI infrastructure thesis is secondary at best. AYI (Acuity) is also in a downtrend and is the weakest link in this layer on a trend-and-momentum basis.

Substrate, Packaging, and Test — The Manufacturing Bottleneck

Advanced packaging — particularly CoWoS (chip-on-wafer-on-substrate) used to bond HBM dies to GPU dies — is one of the hardest bottlenecks in the entire AI supply chain to scale quickly. AMAT at rank #13 (top 3%) and LRCX at rank #36 (top 8.2%) are the two largest semiconductor equipment companies by market cap, and both are direct beneficiaries of memory and logic fab expansions. KLAC at rank #26 provides the inspection and process control equipment that becomes more critical — not less — as process nodes shrink. TER at rank #4 (top 0.7%) rounds out the top tier: semiconductor test is a mandatory step after every wafer is fabricated, and HBM test in particular is technically demanding. AEIS (Advanced Energy Industries) is the power-conversion play within this layer, supplying the precise RF and DC power systems that semiconductor fabs need to run their etch and deposition tools. Entegris (ENTG, not in our screener) and MKS Instruments (MKSI, not in our screener) are adjacent specialty materials and instrumentation names worth researching if you want deeper exposure to the consumables side of the fab equipment cycle.

Datacenter Real Estate — The Land Underneath the Cloud

The hyperscalers cannot build fast enough on their own. Land permitting, power interconnect queues, and construction timelines mean that leasing space in an existing campus is often faster than building. EQIX at rank #7 (top 1.9%) is the global interconnection leader — its campuses are where the internet's major networks meet, and AI traffic is accelerating interconnection revenue. DLR at rank #35 provides wholesale colocation capacity. IRM (Iron Mountain) at rank #100 is the surprise — a company best known for records storage that has been quietly building out AI-targeted colocation capacity with strong momentum. AMT and CCI are both in downtrends, weighed down by rate sensitivity in their tower REIT structures — the AI infrastructure story is real but the rate environment has dominated price action.

3 Underowned Picks Worth Watching

Credo Technology — CRDO

Credo makes active electrical cables and SerDes IP that reduce power consumption for short-reach interconnects between GPU racks. The stock is at roughly $30 billion market cap with a modest +2.3% six-month return, which means it has not yet moved with the broader optical and networking group. The technology addresses a real problem — as GPU clusters get denser, the per-port power budget for optical transceivers becomes prohibitive, and AECs offer a cheaper, lower-power alternative for distances under a few meters. The valuation is not cheap, but the relative underperformance versus peers like COHR (+126%) and CIEN (+159%) makes it worth watching for rotation.

Advanced Energy Industries — AEIS

Advanced Energy is one of the least-discussed names in the AI supply chain, sitting at roughly $14 billion market cap with +80% six-month performance. It supplies precision power conversion systems to semiconductor fab equipment makers — when AMAT or LRCX sell a new etch tool, it often contains AEIS power components. The company also sells DC power infrastructure into datacenters directly. Two separate AI capex tailwinds — chip fab buildout and datacenter construction — feed into the same product lines. ELO rank data is not available in the screener for AEIS, but the price action has been strong and the business has a clear, defensible position in both supply chains.

Enersys — ENS

Enersys makes industrial batteries and UPS systems. The AI angle is straightforward: every new datacenter needs backup power, and the industry is replacing legacy lead-acid backup systems with lithium battery alternatives that Enersys makes. At roughly $7 billion market cap and +69% over six months, the stock is in a clear uptrend with a thesis that does not depend on AI model performance — it depends only on datacenters being built and upgraded. The smaller market cap means it flies under most institutional screens, and the backup power story has received less attention than cooling or networking despite being equally non-optional infrastructure.

What Would Break This Trade

The Capex Peak Problem

Hyperscaler AI capex has grown so fast that the primary risk is not whether AI is real — it is whether datacenter buildout spending has overshot near-term demand. Capital spending cycles in technology historically overshoot on the way up and undershoot on the way down. If Microsoft, Google, Amazon, or Meta cuts their datacenter capex guidance — even while reaffirming AI strategy — the downstream supply chain will feel it immediately. Semiconductor equipment names are particularly exposed because their order books are forward-looking, and any pause in fab capacity expansion would flow through to order cancellations before it appears in revenue.

The ROI Gap

The hyperscalers are spending hundreds of billions of dollars building AI infrastructure faster than AI-driven revenue is growing. Microsoft's AI revenue is scaling, but not yet at a rate that justifies the pace of infrastructure spending on a standalone basis. Google, Amazon, and Meta are in similar positions. This creates a window where investor patience is the key variable. If any major hyperscaler publicly signals that AI ROI is taking longer than expected — or if AI pricing deteriorates as competition drives inference costs down — the risk-off reaction would pressure the entire supply chain, even names with solid fundamentals.

Power and Grid as the New Bottleneck

The irony of the AI supply chain is that the constraint shifts. Twelve months ago it was GPU allocation. Six months ago it was advanced packaging capacity. Today the emerging constraint is power interconnect queues — the time it takes utilities to connect a new datacenter to the grid is now measured in years in many markets. If grid constraints become the binding limit on AI cluster deployment, it compresses demand for everything upstream: less GPU demand, less memory demand, less networking equipment. The companies that benefit most from the constraint shift are the power infrastructure names — GEV, ETN, HUBB — but a grid bottleneck is negative for the compute and memory layers that are currently leading the table on relative strength.

Track These Stocks on Stock Alarm

The full universe of 38 names above is filterable by sector, market cap, trend state, and ELO relative strength rank in the full screener. Use the Technology and Industrials sector filters as a starting point, then sort by ELO rank to isolate the names with the strongest current momentum.

For a ranked view of relative strength across the entire screener universe, the ELO power rankings show every stock sorted by its 26-week relative strength score. The top five from this table currently sitting in that leaderboard: WDC at #2, STX at #3, TER at #4, GEV at #6, and EQIX at #7. Those placements reflect genuine momentum breadth across four different layers — memory, test, power, and real estate — which is a useful signal that the supply chain trade is not narrowly concentrated.

To see which institutional investors and superinvestors own positions in these names, the institutional 13F view shows fund ownership data by ticker so you can see which names are held by the investors whose process you respect.

For more context on the AI investment landscape and the semiconductor industry that underpins it, these related pieces are worth reading:

• Are we in an AI bubble?
• Top AI technology stocks 2026
• Semiconductor industry explained

Set alerts on any of these tickers — free to start.