Lots of Reasons to Worry about AI Infra, But Customer Concentration Cannot be Helped

There are lots of reasons for investors to worry about the artificial intelligence infrastructure business. After all, firms generally are committing capital to infrastructure faster than the economic value of AI is becoming visible. 

The investment question is therefore whether AI demand will grow fast enough, and profitably enough, to absorb the enormous quantity of infrastructure being financed today.

Concern	What investors worry about	Why AI infrastructure is particularly exposed	Potential consequence
Overbuilding / excess capacity	Infrastructure is being built on expectations of enormous future AI demand that may not materialize at the projected rate	Data centers and GPUs are highly capital-intensive and difficult to redeploy	Falling utilization, lower rental rates and impaired asset values
Return on invested capital	AI revenue may grow rapidly but still fail to produce returns commensurate with the enormous capital investment	A $1 billion data center requires substantial utilization for years to earn an attractive return	Lower ROIC and pressure on hyperscaler valuations
Rapid technology obsolescence	Today's expensive GPU cluster may become economically inferior before its expected useful life ends	AI accelerators are advancing unusually rapidly	Accelerated depreciation and residual-value losses
Falling AI compute prices	Competition and increasingly efficient models could cause the price of compute to fall faster than infrastructure costs	Compute is becoming increasingly commoditized	Infrastructure operators could experience margin compression
Financing leverage	The industry increasingly needs debt, leases, project finance, guarantees and SPVs to fund expansion	Capital requirements have grown faster than many companies' internally generated cash	Credit deterioration and refinancing risk
Off-balance-sheet exposure	Some economic obligations may be less obvious than conventional corporate debt	Guarantees, leases, joint ventures and residual-value guarantees complicate analysis	Investors discover greater effective leverage during a downturn
Customer concentration	A data center may effectively depend on one or two giant AI customers	New facilities are enormous and frequently built around anchor tenants	Default or spending cuts by one customer can undermine project economics
Circular financing / ecosystem dependency	Suppliers, customers and financiers may increasingly finance one another	Nvidia, hyperscalers, AI labs, data-center developers and private capital are becoming financially intertwined	Difficulty determining how much demand is truly independent
Power constraints	Infrastructure may be valuable but unusable because electricity cannot be delivered on schedule	AI facilities require enormous amounts of concentrated power	Delays, stranded land/capacity and higher costs
Permitting and community opposition	Local governments or residents may delay or block projects	AI data centers have unusually large power, water, land and noise footprints	Construction delays and unexpected costs
Grid / energy-price risk	Electricity may become substantially more expensive as AI competes for scarce generation	Power is becoming a major component of AI operating costs	Lower data-center margins and higher customer prices
Demand concentration in a few companies	Much of the spending ultimately depends on a handful of hyperscalers and frontier-model companies continuing to spend aggressively	Microsoft, Amazon, Google, Meta, Oracle and a small number of AI labs dominate demand	A capex slowdown could propagate rapidly through the supply chain
AI monetization uncertainty	AI usage may grow enormously without generating enough incremental profits to justify infrastructure spending	Consumers and businesses increasingly expect AI to become cheap or bundled into existing products	Revenue growth can lag infrastructure investment
Macro/interest-rate sensitivity	Higher rates make long-duration infrastructure investments less attractive	Data centers have large upfront costs and long payback periods	Lower valuations and more expensive project financing
Regulatory/geopolitical risk	Export controls, AI regulation, energy policy or restrictions on data centers could change demand	AI infrastructure is concentrated in a relatively small number of countries and suppliers	Assets may become stranded or economically less valuable

But one of the concerns is almost unavoidable. Standard business strategy is to diversify customer bases so that no change at any single customer account will imperil the business overall. But that seems virtually impossible in the AI infrastructure business, especially the “compute as a service” segment. 

The reason is that there simply are very few major buyers. 

Computing Product Category	Dominant Buyer Segment	Market Dynamics & Demand Concentration Drivers
High-Performance AI Accelerators (GPUs/TPUs)	Microsoft, Meta, Google, Amazon (AWS)	Hyperscalers routinely absorb upwards of 40-50% of total advanced enterprise chip allocations, deploying hundreds of thousands of specialized accelerators per cluster to train and serve frontier models.
Custom AI Silicon & ASICs	Google (TPUs), Amazon (Trainium/Inferentia), Meta (MTIA), Microsoft (Maia)	Demand is entirely insourced and consolidated among the top cloud operators designing proprietary chips to reduce dependency on merchant silicon and optimize workload economics.
Enterprise Server Racks & Motherboards	Hyperscale Data Center Operators & Large Cloud Providers	Traditional enterprise server buyers are bypassed by custom Open Compute Project (OCP) specs ordered at massive scale by a handful of operators, leaving original design manufacturers (ODMs) highly concentrated.
High-Bandwidth Memory (HBM)	Major GPU Makers fulfilling Hyperscaler Orders	Production capacity for advanced multi-layer memory stacks is heavily pre-committed to fulfill multi-billion dollar buildouts driven by the major cloud titans.
Advanced Data Center Networking (InfiniBand / High-Speed Ethernet Switches)	Microsoft, Meta, Google, Amazon, Oracle	Ultra-low latency fabric requirements limit early-stage adoption of cutting-edge networking gear to the major cloud providers scaling distributed GPU training fabrics.

But that is not unique to the AI infrastructure business. Some infrastructure markets are inherently oligopsonistic (there may be many suppliers, but only a handful of economically viable buyers).

TSMC doesn't have the option of saying, "If Apple doesn't order this next-generation process, we'll simply find 100 other customers." The universe of customers capable of economically using a 2-nanometer-class process is tiny: Apple, Nvidia, AMD, Qualcomm, Broadcom, MediaTek and a handful of others.

Some other industries also have very-concentrated buyers. In the commercial aerospace business, tier-one avionics and structural component suppliers sell almost exclusively to a duopoly of global aircraft manufacturers: Boeing and Airbus.

In the defense contracting industry, specialized aerospace, radar, and cybersecurity firms rely almost entirely on a single primary buyer: the U.S. Department of Defense or allied national governments.

In the advanced semiconductor equipment industry, companies producing critical lithography systems sell to a handful of chip fabrication giants such as TSMC, Samsung, and Intel.

In railway rolling stock, manufacturers of heavy locomotives and specialized railcars interface with heavily consolidated markets dominated by national freight networks or state-run transit authorities.

Market	Typical number of economically meaningful buyers	Why buyers are so few	Examples of suppliers
AI hyperscale data centers	~5–10	Enormous power, capital and computing requirements mean only hyperscalers and a few AI companies can consume large campuses	Nvidia, Credo, Arista Networks, Cisco, HPE
Advanced semiconductor fabs	~5–10 major customers	Leading-edge chips require enormous volumes and only a few companies design them	TSMC, Samsung, Intel Foundry
Extreme-ultraviolet lithography	Essentially 2–3	Only a handful of semiconductor manufacturers need leading-edge EUV equipment	ASML
Commercial aircraft	~50–100 significant airlines globally, but much smaller number of major customers	Aircraft cost hundreds of millions of dollars and fleet procurement is concentrated	Boeing, Airbus
Large commercial jet engines	~10–20 major airline/airframe customers	Few aircraft platforms and three major engine manufacturers	GE Aerospace, RTX/Pratt & Whitney, Rolls-Royce
Cruise ships	~10–20 major cruise operators	Extremely expensive specialized vessels; only major cruise companies can order them	Meyer Werft, Fincantieri, Chantiers
Nuclear reactors	Very few utilities/developers	Gigawatt-scale projects cost billions and require specialized sites, regulation and financing	GE Vernova, Westinghouse, EDF, KHNP
Military aircraft	A handful of governments	National-security requirements restrict buyers; export markets are politically constrained	Lockheed Martin, Boeing, RTX, Northrop Grumman
Military satellites / launch systems	Very few governments	Classified requirements and national-security restrictions sharply constrain customers	Lockheed Martin, Northrop, Boeing, SpaceX
Large LNG projects / LNG carriers	Relatively few global buyers	Huge projects require long-term contracts and specialized infrastructure	Cheniere, QatarEnergy, shipyards
Electricity generation equipment	Hundreds of utilities, but few very large buyers	Large turbines and grid equipment are bought in relatively infrequent, lumpy projects	GE Vernova, Siemens Energy, Mitsubishi Heavy
High-end semiconductor packaging	Relatively few hyperscalers/chip designers	Advanced packaging is expensive and tied to particular chip architectures	TSMC, ASE, Amkor
Mining equipment for ultra-large mines	A few dozen major mining companies	Huge trucks, shovels and autonomous systems are economically viable mainly at enormous mines	Caterpillar, Komatsu, Epiroc
Offshore oil platforms / FPSOs	A few dozen oil companies	Projects cost billions and are geographically and technically specialized	SBM Offshore, MODEC, Samsung Heavy
High-speed rail equipment	Few national/state buyers	Large projects are government-controlled and geographically constrained	Alstom, Siemens Mobility, CRRC
Container ships	Many shipping companies, but highly concentrated among largest carriers	Very large vessels require substantial capital and are ordered in batches	Hyundai, Hanwha Ocean, CSSC

There are many reasons investors can worry about the health of the AI infrastructure business. But customer concentration does not seem a concern that can realistically be avoided. Some industries are just like that: there are few potential buyers.

AI Value Migration Should Resemble Prior Computing Trends

To the extent that the value of generative artificial intelligence models is based on computational power or speed, it is virtually inevitable that raw processing power will cease to be the driver of customer value as the differences in performance between models diminishes and as open source alternatives proliferate. 

We have seen that shift in many types of computing products. Inference costs, for example, dropped 600 times between 2020 and 2026, for example. The price of the cheapest available output tokens fell from roughly $0.13 per million tokens in mid-2024 into the $0.01-$0.03 range in 2025-2026, according to one study. 

In the personal computer industry, that meant marketing eventually shifted away from processor speed to other attributes, while the overall value shifted to applications. 

So we might well predict that the cost of using models will continue to drop, while model value also shifts. The likely outcome is that, as cheaper computation expanded the addressable markets for computation, so cheaper inference will grow the addressable use cases for inference. 

 

PC era	AI era
CPU cycles	Tokens
MHz/GHz	Model intelligence
RAM/storage	Context/knowledge
Faster processor	Better model
Cheaper computing	Cheaper inference
PC hardware commoditization	Model/token commoditization
Software captures value	Applications/agents capture value

If typical computing product models also apply, then value will migrate “up the stack.” Instead of evaluating inputs (processor speed; model power), we shift to evaluating outputs “what does it do for me?” or “what are the economic results?”). 

Eventually, we stop evaluating value in discrete ways, as capabilities are simply integrated into many other products. The analogy perhaps is electricity, an input used by many products, but not itself a user-relevant output. 

The implications for value in the AI value chain would seem to be clear as well. Over time, value gets produced beyond workflows or even outcomes. At some point, AI becomes invisible, as electricity supply is invisible. 

We assume its existence, as we assume networking exists, or computation exists. 

At that point, AI becomes infrastructure for other products, the way electricity, computation and networking are available for use by many types of products. 

It might take some time, but the PC analogy also suggests the evolution path for AI. When computation was scarce, computation itself was valuable.

When computation became abundant, software became valuable. When software became abundant, data, networks, platforms and workflows became increasingly valuable.

When intelligence becomes abundant, the scarce resource may become the ability to direct intelligence toward economically valuable outcomes, as arguably was true not only of PCs but also transistors and optical fiber networks. 

Scarcity is the driver. Early on, inference capability is scarce, so that drives the value metrics. Later, when inference is plentiful, scarcity shifts elsewhere: “what are the outcomes?”

On the other hand, the value of some frontier models should remain, as commodity PCs coexist and embedded processors coexist with graphics processing units and servers. One popular example might be smartphones. 

Smartphones illustrate that the physical device can remain the value-bearing product even after its underlying computing capabilities become commoditized. The reason is that the smartphone bundles computing with several other scarce things.

PCs remain “place based.” They sit on desks. We use them for work, learning or play. Smartphones are used ambiently and personally, with sensors, cameras and location awareness that make them a platform “for life.” 

Component	Historically scarce	Today
CPU	Computing power	Increasingly commoditized
Storage	Capacity	Increasingly cheap
Display	Resolution/size	Mature technology
Camera	Image quality	Still highly differentiated
Battery	Energy density	Still constrained
Radio	Connectivity	Increasingly standardized
Sensors	Capabilities	Cheap but useful
Software	Basic functionality	Ecosystem differentiator
Industrial design	Physical experience	Still differentiated
Network	Connectivity	Major source of utility
Ecosystem	Applications/services	Extremely valuable
Convenience	Always-available computing	Very valuable

The point is that value migrates toward whatever remains scarce. For PCs, scarcity migrated toward software and applications.

For smartphones, it migrated toward ecosystems, connectivity, design, cameras, convenience and network effects.

For AI, the scarce things might be context, proprietary data, customer relationships, trust, workflow integration, distribution and the ability to turn intelligence into economically valuable action.