The Arms Dealers of the AI Boom: Who Gets Paid No Matter Who Wins?
The AI investment story tends to focus on the companies building the models: OpenAI, Anthropic, Google, Meta, xAI and the other contenders competing to build increasingly capable artificial intelligence.
But there may be another way to invest in the AI build-out.
Instead of trying to predict which AI company ultimately wins, look at the companies supplying the infrastructure that all of them need.
During a gold rush, the more dependable business was not necessarily finding gold. It was selling picks, shovels and supplies to the people looking for it.
The AI equivalent is compute, networking, memory, electrical equipment, cooling, power generation and data-center construction.
And the scale of the build-out is extraordinary.
McKinsey estimates that approximately $5.2 trillion of capital investment may be required globally by 2030 to support AI-related data-center demand. Its base case assumes AI-related data-center capacity reaches approximately 156 GW by 2030, requiring about 125 GW of incremental capacity between 2025 and 2030.
McKinsey divides that $5.2 trillion approximately as follows:
| Infrastructure layer | Estimated investment | Share |
|---|---|---|
| Compute technology | $3.1 trillion | 60% |
| Energy, electrical and cooling | $1.3 trillion | 25% |
| Physical data centers | $0.8 trillion | 15% |
| Total | $5.2 trillion | 100% |
The actual opportunity may be larger. McKinsey notes that its estimate does not fully quantify spending by data-center operators or AI model developers themselves.
So who are the potential “arms dealers” supplying this enormous build-out?
Layer 1: Compute Technology — Approximately $3.1 Trillion
This is the most obvious layer and the one that has attracted the most investor attention.
But AI compute involves considerably more than GPUs.
NVIDIA — AI Accelerators and Systems
NVIDIA remains the dominant supplier of AI accelerators and has built an ecosystem extending well beyond individual GPUs into complete AI computing systems, networking and software.
It is the most obvious beneficiary of the AI infrastructure boom—but also one whose valuation and future performance are increasingly tied to maintaining its technological leadership.
TSMC — Advanced Semiconductor Manufacturing
TSMC represents an especially interesting arms-dealer proposition.
NVIDIA, AMD, Broadcom and hyperscaler-designed accelerators may compete with one another, but many leading-edge chips ultimately depend on TSMC’s manufacturing capability.
That makes TSMC less a bet on one accelerator architecture and more a bet on continued demand for advanced silicon.
Broadcom — Custom AI Silicon and Networking
Broadcom may benefit from an important evolution in AI infrastructure: hyperscalers increasingly designing custom accelerators for particular workloads.
If custom ASICs take some share from general-purpose GPUs, Broadcom can potentially benefit from the transition rather than being threatened by it.
Its networking technology provides another route into AI infrastructure spending.
SK hynix — High-Bandwidth Memory
AI accelerators require enormous memory bandwidth. High-bandwidth memory has therefore become a critical component of the AI compute stack.
SK hynix has emerged as one of the principal HBM suppliers.
Micron Technology — Memory
Micron provides another publicly traded route into the memory bottleneck.
The extraordinary memory requirements of AI servers mean the AI infrastructure cycle isn’t simply creating demand for processors—it is changing the economics of the broader memory industry.
Arista Networks — AI Networking
Thousands of accelerators are valuable only if they can communicate extraordinarily quickly.
That makes networking increasingly part of the AI compute architecture itself.
Arista is one of the important beneficiaries of hyperscale Ethernet networking and AI cluster construction.
Marvell Technology — Connectivity and Custom Silicon
Marvell provides data-center interconnect technology, networking silicon and custom compute products.
As AI clusters grow, moving data efficiently between processors, racks and data centers becomes increasingly important.
Compute watchlist: NVIDIA, TSMC, Broadcom, SK hynix, Micron, Arista Networks, Marvell Technology.
Layer 2: Energy, Electrical Infrastructure and Cooling — Approximately $1.3 Trillion
This may ultimately be the most interesting portion of the AI infrastructure story.
The first constraint was obtaining GPUs.
The emerging constraint is increasingly:
How do we power and cool them?
AI racks require enormous amounts of electricity, and next-generation systems can require radically different cooling and electrical infrastructure from traditional servers.
Vertiv — Power and Thermal Management
Vertiv supplies power management, UPS systems, thermal management and increasingly sophisticated liquid-cooling infrastructure.
As rack densities increase, cooling moves from being a building utility to becoming part of the compute architecture itself.
That makes Vertiv one of the purest infrastructure beneficiaries of AI.
Eaton — Electrical Distribution
Every data center needs to get electricity from the grid—or another generation source—to the computing equipment.
Eaton supplies switchgear, electrical distribution systems, UPS equipment and other critical electrical infrastructure.
Unlike the GPU market, there is relatively little uncertainty about whether the next generation of AI systems will require electricity.
Schneider Electric — Data-Center Electrical Infrastructure
Schneider combines electrical distribution, power management, automation and data-center infrastructure.
Like Eaton, it occupies an attractive position because it participates in the physical infrastructure required regardless of which AI accelerator or model architecture ultimately dominates.
GE Vernova — Generation and Grid Infrastructure
As data-center campuses move toward hundreds of megawatts and even gigawatt scale, the AI infrastructure story increasingly becomes an electricity-generation story.
GE Vernova participates in both generation and grid equipment.
This could give the company exposure to a later phase of the AI build-out that persists after the initial GPU deployment cycle.
Caterpillar — On-Site and Backup Generation
Large data centers require extremely high power reliability.
Caterpillar’s generator businesses provide backup and increasingly important distributed/on-site generation equipment.
Cummins — Power Systems
Cummins provides another route into large-scale standby and distributed generation.
The growing mismatch between data-center construction schedules and utility interconnection schedules could make on-site generation increasingly important.
Energy/infrastructure watchlist: Vertiv, Eaton, Schneider Electric, GE Vernova, Caterpillar, Cummins.
Layer 3: Physical Data Centers — Approximately $800 Billion
The final layer is the actual construction of the enormous facilities housing all this equipment.
This includes land, buildings, electrical installation, mechanical systems, HVAC, piping and construction engineering.
Comfort Systems USA — Mechanical Infrastructure
Comfort Systems is particularly interesting because modern data centers are extremely mechanically intensive.
Cooling, piping and HVAC become more important—not less—as rack density increases.
EMCOR — Electrical and Mechanical Construction
EMCOR provides electrical and mechanical construction services.
The enormous electrical requirements of hyperscale facilities create a potentially long-lived demand tailwind.
AECOM — Engineering and Infrastructure
AECOM participates in engineering, design and infrastructure development required to bring large-scale facilities into operation.
Turner Construction — Data-Center Construction
Turner is one of the major construction companies participating in the data-center build-out and is explicitly cited by McKinsey as an example of the “builder” category.
Turner, however, is privately held and therefore isn’t directly accessible through the public stock market.
DPR Construction — Hyperscale Construction
DPR is another major participant in advanced-technology and data-center construction.
Like Turner, it is privately held.
Physical infrastructure watchlist: Comfort Systems USA, EMCOR, AECOM, Turner Construction and DPR Construction.
AI Company Spending Commitments
The numbers have actually increased significantly during 2026.
The key is to distinguish annual corporate capex, contractual compute commitments, and multi-year infrastructure programs. They overlap, so we absolutely should not add all of these figures together.
Here are some of the most useful numbers as of August 2026:
| Company | Announced/planned spending | Time frame | What it represents |
|---|---|---|---|
| OpenAI / Stargate | $500B | ~2025–2029 | U.S. AI infrastructure |
| OpenAI broader cloud/compute plans | ~$750B | through 2030 | Planned cloud/data-center spending |
| Anthropic / AWS | >$100B | 10 years | AWS compute capacity |
| Alphabet/Google | $195–205B | 2026 alone | Corporate capex, overwhelmingly technical infrastructure |
| Meta | $130–145B | 2026 alone | Corporate capex, primarily AI/data-center infrastructure |
| Meta longer-term reported plan | ~$600B | through 2028 | Infrastructure investment |
| Big five hyperscalers | ~$750B | 2026 alone | Alphabet, Amazon, Meta, Microsoft, Oracle data-center capex |
The last figure is especially striking. Reuters currently estimates that Alphabet, Amazon, Meta, Microsoft and Oracle together are on course to spend roughly $750 billion on data centers during 2026 alone.
OpenAI — perhaps the most astonishing number
OpenAI’s Stargate announcement initially committed to $500 billion over four years, beginning with $100 billion.
That no longer appears merely aspirational. By October 2025, OpenAI said announced Stargate sites represented more than $450 billion of investment over the following three years and more than 8 GW of planned capacity, approaching its original 10-GW/$500B target.
More recently, reporting indicates that OpenAI’s broader planned cloud and data-center spending has risen to approximately $750 billion through 2030, including very large Microsoft and Amazon commitments.
That is remarkable for a company that does not have anything approaching the cash-generating capacity of Google, Microsoft or Meta—which is why financing these projects has become such an important part of the AI infrastructure story.
Anthropic — >$100B with Amazon alone
This one is particularly useful because Anthropic itself disclosed the commitment.
Anthropic’s Amazon infrastructure agreement
Anthropic says it has committed more than $100 billion over ten years to AWS technologies, securing as much as 5 GW of new compute capacity for training and running Claude. It already uses more than one million Amazon Trainium2 chips.
That’s a huge infrastructure commitment from a single AI laboratory.
Google — roughly $200B this year
Alphabet originally indicated $180–190 billion of 2026 capex, roughly six times its 2022 spending, with the overwhelming majority going to technical infrastructure.
It subsequently raised 2026 guidance to approximately $195–205 billion.
And Google has already indicated that spending should increase significantly again in 2027 relative to 2026.
Think about that trajectory:
2022: ~$31B
2026: ~$200B
2027: expected to be significantly higher again.
That is industrial-scale capital deployment.
Meta — $130–145B this year
Meta’s current 2026 capital-expenditure expectation is approximately $130–145 billion, driven substantially by AI infrastructure.
More importantly for our arms-dealer thesis, current reporting indicates roughly $600 billion of spending through 2028.
So this isn’t simply a one-year GPU buying spree.
Something even bigger is appearing off the balance sheet
This is where I think the story becomes particularly interesting.
The hyperscalers now have enormous purchase commitments and lease obligations for future chips, compute capacity, energy and infrastructure.
Recent Financial Times analysis puts purchase commitments across major hyperscalers at nearly $1.5 trillion, plus roughly another $1.5 trillion of lease obligations identified by Goldman Sachs. Alphabet alone reportedly had $811 billion of purchase commitments at Q2 2026, much of it associated with long-term technical infrastructure and energy contracts. Meta had approximately $349 billion.
Again, we cannot add these numbers to capex—they represent different accounting categories and some spending will ultimately flow through capex.
But they tell us something important:
The AI infrastructure build-out is increasingly contractual rather than merely aspirational.
That materially strengthens the arms-dealer thesis.
And it changes how I would think about our 18-stock universe.
We shouldn’t merely ask:
“Which companies benefit from AI?”
We should ask:
Where have hundreds of billions of dollars of already-announced capital commitments created capacity bottlenecks?
That would lead me particularly toward TSMC/advanced packaging → HBM → networking → cooling → switchgear/transformers → generation → electrical/mechanical construction.
Those bottlenecks are potentially where the suppliers acquire pricing power, which is much more interesting than simply having rising revenue.
There is also a fascinating empirical-dashboard opportunity here: add the AI Infrastructure dataset we discussed, but classify each company by infrastructure layer + bottleneck + spending exposure, and then overlay our regime-aware Bollinger signals. That would combine the fundamental $5T capital-flow thesis with an empirical entry/exit mechanism.
The 18-Company AI Infrastructure Watchlist
Putting the layers together produces a very different AI portfolio from simply owning the largest technology companies.
| Layer | Company | Primary exposure |
|---|---|---|
| Compute | NVIDIA | Accelerators/systems |
| Compute | TSMC | Semiconductor fabrication |
| Compute | Broadcom | Custom silicon/networking |
| Compute | SK hynix | HBM |
| Compute | Micron | HBM/memory |
| Compute | Arista Networks | Networking |
| Compute | Marvell | Interconnect/custom silicon |
| Energy | Vertiv | Cooling/power |
| Energy | Eaton | Electrical infrastructure |
| Energy | Schneider Electric | Electrical/power management |
| Energy | GE Vernova | Generation/grid |
| Energy | Caterpillar | Generation |
| Energy | Cummins | Power systems |
| Physical | Comfort Systems USA | Mechanical/cooling construction |
| Physical | EMCOR | Electrical/mechanical construction |
| Physical | AECOM | Engineering |
| Physical | Turner Construction* | Construction |
| Physical | DPR Construction* | Construction |
*Privately held.
This is not a recommendation to buy all 18 companies. It is a map of the supply chain from which an investor can identify opportunities.
Why the Arms-Dealer Thesis Matters
The attractiveness of this approach is that we don’t necessarily have to answer the hardest question in AI:
Who wins?
Suppose one AI model company spends billions building infrastructure and subsequently fails.
The GPUs were still purchased.
The semiconductor was still fabricated.
The memory was still installed.
The network was still constructed.
The switchgear was still purchased.
The cooling equipment was still installed.
The electricity was still generated.
The data center was still built.
And if another company acquires that infrastructure and uses it for a different model, the infrastructure suppliers may get paid again during the next upgrade cycle.
This doesn’t eliminate investment risk. An AI infrastructure overbuild could hurt virtually every company in this ecosystem. Technological change could also strand particular types of equipment.
But it changes the question from:
Which AI company will win?
to:
Which scarce resources will almost every AI company need?
That may be a much easier question to answer.
But How Do You Buy Stocks That Have Already Soared?
There is another problem.
Many of the companies supplying the AI boom have already experienced extraordinary price appreciation.
Buying an excellent company at almost any price is not necessarily an excellent investment.
That is where an empirical trading system can complement the fundamental thesis.
Instead of concluding:
“AI infrastructure will grow, therefore buy the stock.”
we can separate selection from entry.
Fundamental analysis determines:
What companies are eligible to own?
The trading system determines:
Under what empirical conditions is capital authorized to enter?
That distinction is particularly valuable during a market boom.
The objective isn’t to predict when NVIDIA, Vertiv, Eaton or another AI infrastructure company will correct.
It is to maintain a ranked universe of desirable companies and wait until market behavior provides an acceptable entry condition.
In other words:
Fundamentals create the backlog. Price behavior authorizes the pull.
Read Next: How to Trade Parabolic AI Stocks