Executive summary

In AI infrastructure investing, pricing power today sits with the companies that control scarce compute, but the binding constraint is shifting to electricity and power-ready data center capacity. For late-stage investors, the strongest opportunities combine control of a scarce input with an entry valuation that does not already price in perfection. Growth alone is not the test; durable profits at a sensible price are.

The AI gold rush: who actually profits?

The largest technology companies in the world are spending on AI infrastructure at a pace without precedent. With their second-quarter 2026 results, Amazon raised its 2026 capital expenditure outlook to about $220 billion, Alphabet to $195–205 billion and Meta to $130–145 billion, while Microsoft's calendar-2026 figure stood at roughly $175 billion after a lease-accounting change. Together that is about $730 billion, up from roughly $410 billion in 2025. Most of it goes to GPU clusters, custom silicon and data center construction.
That money flows somewhere. Some of it lands with model developers, some with chipmakers, some with the companies that build and power data centers. History suggests that during an infrastructure build-out, the suppliers of the scarcest input often earn the best margins: railroads needed steel, the internet needed fiber, cloud computing needed servers.
But controlling a bottleneck does not guarantee a good investment. Bottlenecks move, competitors arrive, and prices paid by investors can run ahead of the profits that eventually materialize.
So the question for a late-stage investor is not whether AI will grow. It is where the economic value of AI will accrue, and at what price that exposure can still be bought. The companies drawing the most attention are not necessarily the ones that will capture the most value.
Editor's note (remove before publishing): figures are company guidance from late July 2026 and use different definitions (Meta includes finance-lease principal; Amazon refers to cash capex). Refresh after Q3 2026 earnings, due the week of October 26.

Understanding the AI investment value chain

AI is not one investment category. It is a stack of five interdependent layers, each with its own economics, capital needs and competitive dynamics.

1. Applications and software — enterprise AI, copilots, vertical tools. Revenue comes from subscriptions and usage; capital needs are light, but switching costs are often low.
2. Foundation model labs — the companies training and serving frontier models. Revenue comes from API usage and subscriptions; capital needs are enormous, because frontier training runs have grown steadily more expensive, though efficiency gains offset part of that cost.
3. Semiconductors — GPUs, custom accelerators, networking and memory. Revenue comes from hardware sales; development costs are high, but the leaders earn exceptional margins.
4. Compute and data centers — GPU clouds, colocation operators and data center developers. Revenue comes from long-term capacity contracts; capital intensity is very high and often debt-financed.
5. Power and energy — generation, transmission, grid connection and cooling. Revenue comes from power purchase agreements and utility tariffs; projects take years and depend on permitting.

Visual for publication: five-layer AI value chain diagram. Note under it: the layers are interdependent, not a strict supply chain. Labs buy compute, cloud providers buy chips, and everyone needs power.

Money flows down the stack. A dollar spent on an AI application eventually pays for model inference, which pays for chips and data center capacity, which pays for electricity. The question is which layer keeps the most of that dollar as profit.

Who has pricing power across the AI stack?

Pricing power is the ability to raise prices, or hold them, without losing customers. Today it is unevenly spread, and it does not line up neatly with valuation.

AI model labs: high growth, high expectations

The leading labs are growing faster than almost any companies in history. Anthropic raised $65 billion at a $965 billion post-money valuation in May 2026, saying its run-rate revenue had crossed $47 billion; OpenAI was valued at $852 billion after its March 2026 round.
The economics are harder than the headlines. Labs pay for compute twice: once to train new models and again to serve every customer request. Performance gaps between frontier models tend to narrow within months, which limits how long any lab can charge a premium. Distribution, enterprise relationships and developer ecosystems are what turn a better model into durable revenue. For investors, the main question is not growth but how much growth is already in the price, and how much more capital each lab will need.

Semiconductors: the compute advantage

Chips are where AI profits are most visible today. NVIDIA reported $96.2 billion of revenue for the quarter ended July 26, 2026, of which $89.0 billion came from data centers, at a 75.0% gross margin. Margins like that come from combining hardware, software (CUDA) and networking into a system customers struggle to replace.
The threat is custom silicon. Every major hyperscaler designs its own accelerators, and Broadcom and others help them do it. The leaders keep pricing power while demand outruns supply; that gap is what investors must watch. Most of this layer is publicly listed and already priced accordingly.

GPU clouds and neoclouds: growth versus capital intensity

Neoclouds rent GPU capacity to labs and enterprises. CoreWeave, the public reference point, reported $2.6 billion of second-quarter 2026 revenue and a revenue backlog of about $104 billion, but carried roughly $35 billion of debt and posted a $626 million net loss, largely interest.
That is the model in one sentence: contracted demand, financed by borrowing, secured against hardware that depreciates. When utilization is high and customers are creditworthy, it works. When a large customer brings capacity in-house or a chip generation ages faster than expected, the debt remains.

Data centers: the value of available capacity

A data center is land, power, fiber and a building, but the scarce ingredient today is power. A site with energized grid capacity can be leased years before a greenfield site can be built. Long-term leases to hyperscalers give revenue visibility, at the cost of customer concentration and heavy upfront construction capital.

Power infrastructure: the emerging constraint

Power is the slowest layer to expand. New generation must be permitted, financed, built and connected to a grid that already has a long queue. That slowness is exactly what can give power owners pricing power: a gigawatt available in 2027 is worth more to a hyperscaler than a gigawatt available in 2032.  

Pricing power is only one part of returns. Capital efficiency, competition and the price paid matter just as much.

From GPUs to electricity: where is the AI bottleneck moving?

The first constraint was compute. In the early generative AI build-out, the scarce input was advanced GPUs, limited by a concentrated supply chain and advanced packaging capacity. That scarcity is why chip margins expanded, and why hyperscalers began designing their own accelerators.
The emerging constraint is electricity. The International Energy Agency projects that global data center electricity demand will more than double by 2030, to around 945 TWh, slightly more than Japan's entire consumption today, with demand from AI-optimized data centers more than quadrupling. In the United States, data centers are expected to account for almost half of electricity demand growth over the same period.
The US picture is sharper still. Lawrence Berkeley National Laboratory estimates that data centers used 176 TWh, or 4.4% of US electricity, in 2023, and could reach 325–580 TWh, or 6.7–12%, by 2028.
Supply is not keeping pace. At the end of 2025, more than 2,060 GW of generation and storage capacity was waiting in US interconnection queues, and only about 13% of the capacity that requested interconnection between 2000 and 2020 had reached commercial operation by the end of 2025. Where power can be delivered, and how fast, now shapes where data centers get built.
Nuclear is part of the answer, not all of it. Existing nuclear plants offer what AI data centers want most: around-the-clock, carbon-free power. Hyperscalers have signed long-term agreements with nuclear operators and backed next-generation reactor developers. But new reactors take years to license and build, and the IEA expects renewables and natural gas to meet most new data center demand this decade. For the private-market opportunity in small modular reactors, see nuclear energy investment opportunities.
The bottleneck is broadening rather than simply moving: chips still matter, but power availability now matters too, and its importance varies sharply by region and project.
AI infrastructure investing in private markets
Identifying the most profitable layer of the AI value chain is only the first step. For private-market investors, the more important question is whether that profitability can be accessed at a valuation that leaves room for attractive returns. A company with strong pricing power can still be a disappointing investment if its entry price assumes years of exceptional growth.
Much of the chip layer is already public and priced by the market every second. A meaningful part of the remaining exposure sits in private companies.
Private AI model developers. The leading labs remain private, with valuations near $1 trillion and revenue growing at extraordinary rates. Exposure offers participation in that growth, but entry prices are high, further dilution is likely as labs keep raising capital, and liquidity depends on an IPO or tender offer whose timing is uncertain.
Private compute and data center companies. Neoclouds, specialized compute operators and data center developers sit on contracted demand. The questions are the quality of their customers, how much of the capacity is financed with debt, and whether contracts outlast the hardware they pay for.
Private energy and nuclear companies. Advanced nuclear developers, fuel suppliers and power developers offer exposure to the slowest-moving and potentially most durable constraint. Returns depend on regulatory milestones and on executing first-of-a-kind projects on budget.
How investors access these companies. Late-stage investors typically use one of four routes: direct investment in a funding round, secondary purchases of existing shares, a venture or growth equity fund, or a special purpose vehicle (SPV) that pools investors into a single company. Each differs in what you actually own (shares, units in a vehicle, or a contractual economic interest), in total fees, in transfer restrictions imposed by the company, and in how and when liquidity arrives. For a detailed comparison, see how to invest in late-stage private companies.
Compliance note (remove before publishing): no ARCA offering is referenced here. Do not add specific available allocations until offering permissions are confirmed.
The risks behind the AI infrastructure thesis
1.    Overbuild. Capacity is being added on forecasts. If adoption slows or customers pause, the industry could hold more data centers and GPUs than it can fill at today's prices.
2.    Efficiency gains. Better models and chips cut the compute needed per task. Total usage may still grow, but the mix of winners can shift quickly.
3.    Customer concentration. Many infrastructure providers depend on a handful of hyperscalers and labs. One customer moving capacity in-house can reshape a business.
4.    Capital intensity. Debt, depreciation and refinancing can absorb most of the operating profit, as the neocloud model shows.
5.    Valuation and liquidity. A great company bought at the wrong price can still be a poor investment, especially if the exit takes longer than expected.

The ARCA Scarcity Test: five questions for AI infrastructure investments

Five questions are worth asking of any AI infrastructure opportunity. A company does not need a perfect score, but weakness on any one of them should be reflected in the price.

Visual for publication: render this as the branded Scarcity Test graphic (five numbered cards). Editor's note: if the investment team confirms this is ARCA's applied methodology, the intro can read "ARCA evaluates every opportunity against five questions."

The fastest-growing layer is not automatically the best investment. Durable economics bought at a disciplined price usually are.

Frequently asked questions

Is AI infrastructure a good investment?

It depends on the layer, the company and the price. AI infrastructure demand is large and growing, but returns vary widely: some layers carry high margins, others heavy debt and depreciation. Any investment should be judged on its competitive position, revenue quality and entry valuation, not on sector growth alone.

Which companies benefit most from AI infrastructure spending?

So far, chip designers have captured the most visible profits, followed by cloud providers with contracted capacity. As power becomes a constraint, owners of power-ready data center sites and reliable generation are gaining bargaining power. Which companies benefit most will continue to shift as bottlenecks move.

How can accredited investors invest in private AI infrastructure companies?

The usual routes are direct participation in a funding round, secondary purchases of existing shares, venture or growth equity funds, and special purpose vehicles (SPVs). Each differs in fees, what the investor actually owns, transfer restrictions and liquidity. Private investments are illiquid and carry a risk of total loss.

Conclusion: scarcity creates opportunity, valuation determines returns

AI is creating economic value across every layer of its stack. Today, chips capture the most visible profits, model labs carry the highest valuations, and power and data center capacity are becoming the constraint that decides how fast everything else can grow. Those positions will keep shifting as technology, supply chains and demand evolve.

For late-stage investors, the opportunity lies where control of a scarce input meets a price that leaves room for error, which is exactly what the five questions of the Scarcity Test are built to surface. The next great AI investment may not be the company building the most advanced model, but the one that sustainably monetizes the infrastructure the AI economy runs on.