There is a number floating around Wall Street right now that has the quality of a fever dream.
Seven hundred billion dollars. That is the combined capital expenditure that Amazon, Alphabet, Microsoft, Meta and Oracle have committed to spending in 2026 on AI infrastructure, from GPUs and data-center shells to the power that runs them. To put it in a frame that actually communicates the scale: $700 billion exceeds the entire annual GDP of Switzerland. It is more than the US government spends on Medicare in a calendar year. It is a sum of money so large that the companies writing those checks are, for the first time in decades, issuing debt at scale to fund it. Morgan Stanley estimated hyperscaler borrowing would top $400 billion in 2026, more than double the $165 billion borrowed in 2025.
Nobody disputes that the spending is happening. What matters to an investor is where along the supply chain the money lands, because “AI is a big trend” and “I know which stocks will benefit” are two different things, and the distance between them is where most retail investors leave money on the table.
Below I follow the money from the foundry to the software layer and check where pricing power really sits. My argument is that the suppliers selling into this spending, and the networking layer in particular, offer better economics than most of the operators writing the checks.
What the $700 billion means for shareholders
Before getting into individual stock categories, the aggregate spending picture deserves scrutiny rather than just admiration.
Amazon has committed to $200 billion in capital expenditure for 2026, a figure that, when it was announced, sent Amazon shares down roughly 9% in the sessions following the earnings call. Alphabet guided $175 to $185 billion. Microsoft indicated spending above its prior-year $88 billion base. Meta committed to a range that analysts interpreted as $115 to $135 billion. Oracle, the most aggressive spender relative to its size, targeted $50 billion.
The market’s negative reaction to some of these announcements is worth understanding rather than dismissing. When a company commits to spending $200 billion in a single year, the first question investors ask is “where does the cash come from and when does it generate a return.” Amazon and Microsoft are funding portions of their capex through debt issuance. Bank of America estimated that the five largest hyperscalers would spend approximately 90% of their combined operating cash flow on capital expenditure in 2026, which leaves little room for anything else the boards might want to fund.
For investors who own the hyperscalers themselves, this creates a specific tension. You are holding companies that are simultaneously generating extraordinary operating profits and consuming almost all of those profits in infrastructure investment. The bull case is that this investment generates revenue growth that justifies the spending in three to five years. The bear case is that the spending is partly competitive necessity, if you don’t build and your competitor does, you lose market position, which means some portion of it is maintenance capex dressed up as growth capex.
Neither case is obviously correct. What is clear is that the hyperscalers themselves are not the most elegant way to capture the AI infrastructure theme from an equity return standpoint.
Why suppliers often beat the operators
There is a durable historical analogy for the current situation: the California Gold Rush of 1849. The people who reliably made money during the gold rush were rarely the miners; most never struck it rich. Those who did well sold picks, shovels and jeans to them. Sellers of supplies captured a more predictable slice of the spending wave regardless of which individual mining operations succeeded.
The AI infrastructure buildout follows the same logic. Whichever lab ends up with the dominant AI platform, somebody still has to make the chips and build the data centers. Those suppliers collect a toll on the entire activity regardless of which AI application wins the consumer market.
The challenge is identifying which pick-and-shovel positions have genuine pricing power versus which ones are simply riding the wave and will face margin compression as competition arrives and customer sophistication increases.
The table below sorts the supply chain into six layers, from silicon to software, and puts a rough pricing-power and margin label on each.
| Layer | What It Includes | Pricing Power | Margin Profile |
|---|---|---|---|
| Silicon Manufacturing | TSMC, Samsung Foundry | Very High, near-monopoly at leading edge | About 60% gross margin (fiscal 2025) |
| GPU/AI Chip Design | NVIDIA, AMD, Broadcom custom ASICs | High, NVIDIA dominant, AMD credible alternative | NVIDIA about 71% gross margin (latest fiscal year) |
| Networking & Interconnect | Arista Networks, Marvell, Coherent | High, data center networking a bottleneck | Arista about 64% gross margin |
| Data Center Construction | Power companies, REITs, construction | Moderate, geographically constrained | 35-45% operating margins |
| Cloud Platform (Operators) | AWS, Azure, Google Cloud | Moderate, competitive market, switching exists | 25-35% operating margins |
| AI Software/Applications | Varied, massive range | Low to Very High, depends on defensibility | Highly variable |
The pattern that emerges from this table is consistent with what the pick-and-shovel framework would predict. Pricing power and margin profile are highest at the foundry and chip design layer, where physical scarcity and years of accumulated technical advantage create genuine moats, and generally lower at the application layer where competition is most intense.
The margin column is easier to trust when it comes from reported numbers, so the chart below uses the latest full-year gross margins in the StockVane database. NVIDIA leads at 71.1%, although that is down from 75.0% a year earlier. Broadcom follows at 67.8%, Arista at 64.1% and TSMC at 59.9%. AMD is the outlier at 49.5%, which is a reminder that the chip design layer is not one business with one set of economics.
Layer one: TSMC is the unavoidable toll road
Taiwan Semiconductor Manufacturing Company occupies a position in the AI supply chain that has no direct parallel elsewhere in global industry. For a more skeptical read on whether this spending pace is sustainable, see the AI Capex Bubble piece. The company currently holds approximately 68% of global foundry market revenue and an even larger share of leading-edge capacity, the advanced manufacturing nodes required to produce AI chips with the density and performance characteristics that modern training and inference workloads demand.
TSMC’s CoWoS packaging technology, the advanced process that allows multiple chiplets to communicate at extremely high bandwidth within a single package, is the enabling technology behind NVIDIA’s most powerful AI accelerators. Without CoWoS, the Blackwell architecture cannot be assembled in the configuration that produces its performance numbers. The demand for CoWoS has outpaced TSMC’s production capacity for two consecutive years, creating a queue that gives TSMC substantial pricing power over the chip designers who depend on it.
The financial profile reflects this structural position. TSMC’s gross margins have expanded from the low 50% range in 2022 to above 58% in recent quarters, driven by tight capacity and a product mix that has tilted toward the highest-value nodes. Revenue visibility extends well into 2027 because TSMC’s customers commit to capacity years in advance through long-term supply agreements, a level of forward booking more characteristic of regulated utilities than semiconductor manufacturers.
The geopolitical risk is real and worth sizing. Roughly 90% of TSMC’s manufacturing capacity sits in Taiwan, a geography that carries specific geopolitical sensitivity. This risk is not theoretical. It is the single most cited reason institutional investors who find TSMC’s fundamentals compelling still underweight the position. TSMC’s Arizona and Japan expansion programs address this risk slowly, but meaningful diversification of production away from Taiwan is a decade-long project, not a near-term story.
For investors who can hold the geopolitical risk in context, sizing the position relative to that specific risk rather than pretending it doesn’t exist, TSMC represents what is possibly the most defensible position in the entire AI supply chain.
Layer two: NVIDIA’s moat and the credible alternatives
NVIDIA’s position has been analyzed extensively, including in previous StockVane coverage, so this section focuses on what has changed and on the competitive dynamics that matter for investment decisions over the next 18 months.
The most significant development is NVIDIA’s explicit ambition to become what CEO Jensen Huang described as “the world’s leading CPU supplier”, a statement that signals the company is no longer content to dominate one category of semiconductor but intends to compete for the entire compute stack. The debut of Grace Blackwell systems that combine GPU and CPU capabilities in a single rack architecture, alongside the networking and software layers that tie them together, represents a fundamentally different competitive positioning than the “GPU company” framing that defined NVIDIA three years ago.
For investors, this matters because it changes the total addressable market calculation. GPUs address a large and growing market. The full compute stack, including CPUs and networking, addresses a market several times larger. Whether NVIDIA can capture meaningful CPU market share against AMD’s EPYC processors and Arm’s ecosystem is uncertain, but the direction of competition is now established.
The more important near-term question for the stock is the Vera Rubin transition, the next-generation architecture scheduled to ramp in the second half of 2026. Every major NVIDIA product transition has historically created a period of order uncertainty as hyperscalers decide whether to buy more of the outgoing generation or wait for the incoming one. This transition risk is not unique to NVIDIA, but at NVIDIA’s scale and valuation, a single weak quarter during the transition creates disproportionate stock price impact.
Broadcom occupies a distinct position that merits explicit attention. The company designs custom AI accelerators, specifically custom ASICs for Google, Meta, ByteDance and Apple, that serve as alternatives to NVIDIA’s general-purpose GPUs for inference workloads. Broadcom’s AI revenue grew at a rate that surprised even optimistic analysts in 2025, and the company’s guidance for fiscal 2026 implies AI-related revenue approaching $25 billion. The custom silicon model captures a specific market: hyperscalers willing to invest in chip design to reduce unit economics at massive scale. As AI inference volumes grow into the hundreds of billions of queries per day, the cost per inference becomes a central competitive variable, and custom silicon wins on unit economics versus general-purpose alternatives.
Layer three: the networking bottleneck
When a data center installs 100,000 GPUs, the chips themselves are only part of the capital requirement. Every one of those GPUs needs to communicate with every other GPU at extremely high bandwidth and extremely low latency during training runs. The networking infrastructure that enables this communication has become one of the genuine supply chain bottlenecks in the AI buildout, and the companies that make the switches and optical components are operating in a demand environment with characteristics similar to what GPU manufacturers experienced two years ago.
Arista Networks has become the networking vendor of choice for AI-scale data center builds, having captured meaningful share of the hyperscaler switching market through a combination of technical performance and software-defined networking capabilities. Revenue growth in the most recent fiscal year exceeded 20%, and the company’s backlog provides visibility into demand that is not purely dependent on quarterly purchase order timing.
The optical transceiver market, the components that convert electrical signals to optical signals for transmission across fiber, is experiencing similar dynamics. As data center distances grow and bandwidth requirements increase, coherent optical technology has become a critical enabling layer for AI interconnect. Companies focused on this space have seen demand acceleration that significantly outpaced their production capacity planning from 18 months ago.
The networking layer is less widely owned by retail investors than the GPU layer, which is partly why it remains interesting. The coverage ratio of sell-side analyst attention relative to fundamental opportunity is lower than for the more prominent names, and the valuation premiums are correspondingly less stretched in some cases.
Layer four: power will separate winners from losers
Every GPU cluster consumes power. A modern AI training cluster running tens of thousands of GPUs draws electricity at a rate that challenges the supply capacity of regional power grids. Data center power consumption has moved from an operational detail to a strategic variable, hyperscalers are now actively negotiating with utilities, investing in dedicated generation and, in some cases, acquiring nuclear power assets specifically to ensure the energy supply their planned infrastructure requires.
This dynamic has created a specific investment theme that sits partially outside the traditional technology sector. Utilities with data center exposure and independent power producers that can contract directly with hyperscalers have seen demand for their products and services accelerate in ways that most of them were not fully planning for 18 months ago.
The financial profile of power-focused AI beneficiaries is different from semiconductor companies: lower gross margins, higher capital intensity and longer time horizons between investment and return. But the demand visibility is exceptional, a hyperscaler that commits to a data center site commits to the power contract simultaneously, creating multi-decade revenue visibility for the power supplier that contrasts with the quarterly order variability that affects GPU and networking vendors.
For investors who want AI exposure with lower volatility than pure semiconductor plays, the power infrastructure layer offers a structurally different risk profile while still capturing meaningful upside from the spending wave.
Cloud operators are spenders, not suppliers
One of the most common mistakes in AI sector investing is treating AWS, Azure, and Google Cloud as beneficiaries of AI infrastructure spending in the same category as TSMC or Arista. They are not the same.
TSMC and Arista are selling into the spending wave. the three big clouds are the ones doing the spending. Their margins reflect that distinction. The cloud platforms are simultaneously among the most competitively valuable franchises in global technology and among the most capital-intensive businesses in history. AWS generates extraordinary operating profit, above $105 billion annualized at recent run rates, but that profit is being recycled into data center construction at a rate that leaves limited free cash flow relative to market capitalization.
This doesn’t make the hyperscalers bad investments. It makes them different investments with a different thesis. You are betting on their ability to convert AI infrastructure investment into AI-related cloud revenue at a rate that grows faster than their cost of capital. Evidence for this thesis exists, Microsoft’s Azure AI revenue and Google Cloud’s backlog both showed strong growth in recent quarters, but the path from infrastructure investment to revenue recognition spans multiple years and involves customer adoption cycles that are not yet complete.
The suppliers collect their revenue today, from the spending that is happening now, with gross margins that expand as volumes grow. The operators collect their revenue later, after infrastructure is built and customers have ramped their AI usage, with margin profiles that depend on utilization rates that are still developing.
How to spread exposure across the theme
Given the analysis above, a rational approach to AI infrastructure exposure involves owning different parts of the supply chain in proportions that reflect both conviction level and risk tolerance.
| Category | Key Names | Bull Case | Primary Risk | Suggested Role |
|---|---|---|---|---|
| Foundry | TSMC | Irreplaceable capacity, pricing power, margin expansion | Geopolitical, Taiwan concentration | Core position, size to risk tolerance |
| GPU Design | NVIDIA, AMD | Dominant platform, expanding TAM with CPU ambitions | Valuation, product transition execution | Core with disciplined entry |
| Custom Silicon | Broadcom | Secular shift to custom ASIC for inference at scale | Customer concentration, design cycles | Satellite position |
| Networking | Arista, Marvell | Demand acceleration, under-owned compared with GPU names | Competition from white-box alternatives | Tactical position |
| Power Infrastructure | Selected utilities, IPPs | Multi-decade demand visibility, defensive revenue | Regulatory risk, slow capital deployment | Defensive AI allocation |
| Cloud Operators | AWS, Azure, Google Cloud | AI revenue monetization, strongest long-term moats | High capex consuming margins, timing of ROI | Hold if owned, selective on new entry |
The most contrarian position in this table, the one with the highest gap between institutional attention and fundamental quality relative to that attention, is networking. Sell-side research has picked over the GPU supply chain to exhaustion by sell-side research. Networking, which faces the same demand acceleration and similar capacity constraints, receives less coverage and correspondingly less valuation premium. This asymmetry tends to close over time as the investment community catches up to the fundamental picture.
The question I ask before buying
Here is the specific intellectual test I apply before adding AI infrastructure exposure at current prices.
The spending is real. Seven hundred billion dollars in committed capex does not reverse in a quarter. But the stock prices of the suppliers also reflect expectations about what comes after 2026, whether the capex cycle lasts into 2027 and 2028 or is a concentrated front-loading of investment that creates a subsequent air pocket.
Goldman Sachs pointed out that analyst consensus estimates for hyperscaler capex were roughly 20% growth at the start of both 2024 and 2025, and actual growth exceeded 50% in both years. The pattern of consistent underestimation suggests that the momentum has more runway than consensus acknowledges, but momentum that has been underestimated for two years also has a specific risk: when it finally decelerates, the surprise goes in the opposite direction.
For companies whose valuations assume continued 50%+ capex growth from hyperscalers, and there are names in the AI infrastructure complex where that assumption is clearly embedded in forward multiples, the risk is asymmetric. If capex growth comes in at 25% instead of 50%, those stocks reprice materially even though 25% growth is an extraordinary outcome by any historical standard.
The suppliers with the most defensible positions are those whose revenue is diversified enough across the supply chain that a single customer’s moderation in spending doesn’t create a dramatic earnings miss. TSMC, serving every chip designer simultaneously, has that diversification. NVIDIA, whose revenue is highly concentrated in a handful of hyperscaler customers, has somewhat less.
Concentration is a reason to size a position carefully; it is no reason to skip the name.
My test is growth in hyperscaler capex. If it lands near 50% again in 2027, the suppliers’ forward multiples look earned. If it slows to 25%, expect the most concentrated names to reprice sharply even though 25% would still be an extraordinary year.
Financial disclaimer: The content on StockVane is for educational and informational purposes only and should not be construed as professional financial advice. Stock market investing involves risk of loss.
Pingback: Which Tech Stocks Are Worth Buying Cautiously in Q4 2026 – StockVane