A GPU is not an AI factory.
Nvidia and AMD can design the compute. TSMC can manufacture the chips. Power companies can generate the electricity, and optical companies can move data at extraordinary speeds.
But someone still has to put those pieces together in a building that has power, cooling, networking, software, and a customer ready to use it.
That is the part of the AI story I want to examine here.
CoreWeave, Nebius, IREN, AXT, and Wolfspeed often get grouped together as “AI infrastructure” stocks. They do not do the same job. They operate at different handoffs between the parts of the supply chain we have already covered.
And the handoff matters. A component can be essential without the company supplying it becoming a great investment.
My question for each business is simple: What scarce resource does it control, and can it turn that scarcity into cash flow without leaving shareholders with the bill?
CoreWeave and Nebius: A GPU is not a cloud
Nvidia can deliver an accelerator. An AI customer needs a functioning cluster: thousands of accelerators connected by high-speed networks, supplied with power, kept cool, monitored, and made available through software.
That is what CoreWeave and Nebius are trying to sell.
Why doesn’t Nvidia simply do all of this itself? It could do more of it. But owning and operating data centers would mean financing buildings, carrying depreciating hardware, managing utilization, and potentially competing with the cloud companies that buy its products. The gap exists largely because companies have chosen different businesses to be in—not because Nvidia is incapable of building a data center.
The demand for specialized capacity is real. CoreWeave reported approximately $104 billion in revenue backlog at the end of June 2026. Nebius announced a five-year Meta agreement that includes $12 billion of dedicated capacity scheduled to begin delivery in early 2027.
Why do those contracts matter? Because they show that even very large technology companies are willing to buy outside AI capacity.
But a contract is not the same thing as cash in shareholders’ pockets. The infrastructure must be delivered, operated, and financed first. In the same quarter that CoreWeave reported roughly $2.6 billion in revenue, it also reported $640 million in net interest expense.
That is the tension in one set of numbers: enormous demand, enormous capital requirements.
The strongest objection to the neocloud thesis is also obvious. Why won’t AWS, Azure, and Google Cloud simply take this market?
They may take more of it. These companies already build and operate immense amounts of infrastructure. Specialized providers have an opening when they can bring up scarce capacity faster, configure it better for AI workloads, or give customers a useful alternative. That opening is not automatically a permanent moat.
If GPUs become easier to obtain and the hyperscalers match the service, renting out hardware could become a much less attractive business. CoreWeave and Nebius therefore need to prove that customers value their software, performance, reliability, and speed of deployment—not just temporary access to scarce chips.
GPU scarcity may win the first contract. A better cloud must win the next one.
IREN: A megawatt is not a data center
Our energy article dealt with where the electricity comes from. That is only half the problem.
Power must reach a site that has the right connection, buildings, cooling, fiber, and equipment. A company can have access to electricity and still be years away from delivering usable AI compute.
IREN sits at that conversion point. Its history in Bitcoin mining gave it experience developing large, power-intensive computing sites. Now it is converting and expanding that footprint for AI data centers and GPU cloud services.
The Microsoft agreement shows the ambition: GPU services across four facilities at IREN’s Childress campus, representing approximately 200 megawatts of IT load. But the distinction between planned and operating capacity matters. In its FY2026 results, IREN said it had delivered the first 50-megawatt phase to Microsoft; the remaining phases were still being commissioned or built.
That is why I watch delivery milestones more closely than a headline megawatt figure.
Why doesn’t Bloom Energy or a nuclear operator do this job? Generating power and operating a GPU cloud demand different equipment, capital, customers, and expertise. The power supplier may provide the electricity. IREN has to make that electricity useful to an AI customer.
Applied Digital offers another version of the handoff. It primarily develops and leases powered data-center infrastructure rather than selling the same full GPU-cloud service. At the end of June 2026, it reported 175 megawatts of live capacity at its Polaris Forge 1 campus. That is a different business model—and should be judged by different economics.
A useful way to keep the roles straight is this: the power company supplies the electricity; the developer makes a campus ready to use it; the cloud operator sells the resulting compute. Companies can move across those boundaries, as IREN has. But crossing a boundary also means taking on its costs and risks.
Power on a map is not power at a rack.
AXT: Before the optical module comes the material
In the Nvidia article, we discussed companies such as Lumentum, Coherent, Corning, and Marvell that help move data through AI systems.
AXT sits further upstream.
It makes compound-semiconductor substrates, including indium phosphide wafers used in certain optical devices. Those devices can become part of the high-speed links connecting AI equipment. AXT does not sell the finished transceiver or own the entire optical network. It supplies a starting material for some of the devices inside it.
Why does that matter? Because increasing demand for faster optical connections can reach all the way back to crystal growth. AXT said its second quarter of 2026 produced its highest quarterly indium phosphide revenue to date.
But I would not turn “more AI networking” into “every optical shipment benefits AXT.” Not every device uses its substrates. Customers can seek other suppliers or architectures. And a scarce product is of limited use if it cannot be shipped: fewer-than-expected Chinese export permits constrained AXT’s indium phosphide shipments in late 2025.
The question is whether AXT can expand qualified production, ship consistently, and retain attractive margins as supply grows.
Being upstream can create leverage. It can also put someone else between you and the customer.
Wolfspeed: Electricity still has to reach the chip
The energy story does not end when power enters the data center.
Electricity must be converted and controlled before a GPU can use it. As rack power rises, efficiency, heat, and the size of power equipment become more important.
That is where Wolfspeed’s silicon-carbide devices could play a role. In August 2026, Wolfspeed and LITEON announced that Wolfspeed’s technology had been qualified for LITEON’s 800-volt DC power platforms, which are designed for next-generation AI data centers.
Why does qualification matter? It puts the technology in a position to be used. Why isn’t it the end of the story? Because qualification is not the same as high-volume shipments or profitable revenue.
Wolfspeed also does not own silicon carbide outright. Other semiconductor companies compete in power devices, and different materials may make more sense in different parts of an electrical system. Treat this as an opportunity to evaluate, not a reason to assume that every AI rack will become Wolfspeed revenue.
Its financial history demands that discipline. Wolfspeed went through a Chapter 11 restructuring in 2025. That is a reminder that a useful technology and a healthy shareholder outcome are two separate things.
A design win is a possibility. Profitable volume is proof.
What could change the thesis
The bullish case is that these companies control hard-to-replicate pieces of a rapidly growing system: ready-to-use compute, powered campuses, qualified optical materials, and efficient power devices.
The opposing case is that today’s shortages fade. Hyperscalers build more capacity. Competing suppliers qualify. Projects arrive late or cost more than expected. Debt and dilution absorb the value before shareholders see it.
The thesis would strengthen as delivered capacity turns into sustained cash flow, customer relationships deepen beyond a temporary shortage, and growth requires less outside capital over time. The thesis would weaken if reported demand kept rising while returns on that demand did not.
Track different proof points for different businesses: utilization, financing costs, and customer retention for CoreWeave and Nebius; energized and delivered capacity for IREN and the data-center developers; permits, yields, and margins for AXT; and production revenue rather than announcements for Wolfspeed.
The companies in this article fill real gaps. That does not make them interchangeable, and it does not make every one of them a winner.
The practical principle is the same across all five:
A bottleneck is not a moat until it pays for itself.