The AI boom has a familiar rhythm, even if the product is new. It arrives with a bright, simple promise—smarter software, faster research, automated work that used to take teams—and then it spreads outward into everything that has to exist for that promise to be delivered. In that sense, the boom behaves less like a single invention and more like a commercial craze around a “hot” item: attention spikes, investment follows, and then reality shows up at the loading dock.
A doughnut is a useful metaphor here, not because anyone thinks the technology is sweet, but because doughnuts capture the shape of the cycle. There’s the initial burst of visibility—the glaze, the shine, the reason people line up. Then there’s the center: the core capability that makes the product feel magical. But the real story is what happens around it. Doughnuts require ingredients, ovens, frying oil, supply chains, labor, packaging, and a steady flow of customers. If any one part fails, the whole operation slows down. The same is true for AI. The “center” is the model and the software stack, but the “ring” is the infrastructure and operational capacity that turns demos into deployments.
That’s why the phrase “chips aren’t the only thing getting fried” lands so well. When AI accelerates, it doesn’t just stress-test GPUs. It heats up the entire system: data centers, power generation and distribution, cooling, networking, storage, manufacturing throughput, logistics, and the cost/availability of components that keep machines running day after day. The bottleneck isn’t always where people first look. Sometimes it’s not the chip at all—it’s the ability to feed the chip with power, data, and time, or to keep the facility stable enough to run at scale.
To understand the current phase of the AI boom, it helps to stop thinking in terms of “breakthroughs arriving overnight” and start thinking in terms of “capacity arriving in waves.” Models improve, yes. But the bigger constraint is often whether the world can build, power, and operate enough compute to use those models effectively. That shift—from novelty to throughput—is where the doughnut metaphor becomes more than a joke. It’s where hype meets engineering.
The early stage: a hot product and a rush to capitalize
In most technology booms, the first wave is driven by a simple narrative: the product works, and therefore demand will explode. With AI, the narrative is compelling. Large language models can summarize, draft, translate, code, and assist in ways that feel almost conversational. Businesses see immediate value in productivity and customer-facing automation. Researchers see new tools for exploration. Investors see a platform shift.
But when demand grows quickly, companies don’t just buy the product—they build an ecosystem around it. They hire, they integrate, they train internal teams, and they sign contracts for hardware and cloud capacity. This is the “glaze” phase: the visible excitement that draws in more participants. It’s also the phase where forecasting errors are most expensive. If you underestimate demand, you miss the window. If you overestimate, you lock in costs before the market stabilizes.
AI has been unusually fast because the product is software that can be distributed instantly, while the infrastructure behind it takes longer to expand. A model can be released today; a data center expansion can take months or years. Even when hardware is available, it must be installed, networked, cooled, and integrated into systems that can handle real workloads rather than benchmark tests. That gap between speed of adoption and speed of capacity is where the boom starts to feel like a scramble.
The second stage: the “everything bottleneck” problem
Once the initial rush begins, the constraints multiply. People often talk about chips because chips are tangible and headline-friendly. But chips are only one link in a chain. Consider what it takes to run a modern AI workload at meaningful scale:
You need compute accelerators (GPUs or other specialized processors).
You need enough memory bandwidth and storage to feed them.
You need high-speed networking to move data between nodes and to connect to external systems.
You need power delivery that can handle sustained loads without instability.
You need cooling systems that can remove heat efficiently across thousands of racks.
You need software that can schedule jobs effectively and keep utilization high.
You need operational discipline: monitoring, redundancy, maintenance windows, and incident response.
If any of these elements lags, the system doesn’t fail completely—it slows down. And slowing down is often worse than failure, because it creates a fog of uncertainty. Teams can’t tell whether performance issues are due to model changes, software inefficiencies, data pipeline delays, network congestion, or power/cooling limits. The result is a kind of operational whiplash: investments continue, but returns arrive unevenly.
This is the “everything bottleneck” problem. It’s not that every component is equally constrained at all times. It’s that constraints can shift. One month, the limiting factor might be accelerator availability. Another month, it might be power capacity at a specific site. Another month, it might be the ability to connect to the right network topology or to procure enough transformers and switchgear. The bottleneck moves like a spotlight, revealing different weak points as the system scales.
That’s why “chips aren’t the only thing getting fried” is more than a catchy line. It’s a reminder that the heat is systemic. When AI demand rises, it doesn’t just increase consumption of semiconductors; it increases demand for electricity, water or alternative cooling resources, construction materials, skilled labor, and specialized components across the data center stack. It also increases demand for logistics and manufacturing capacity—things that are not easily scaled on short notice.
The doughnut center: models improve, but deployment is the real test
There’s a temptation to treat AI progress as a straight line: better models lead to better outcomes, which leads to more adoption. But deployment is where the center of the doughnut meets the ring. A model that performs well in a controlled environment may behave differently in production because of latency requirements, data quality issues, integration complexity, and user behavior.
Production AI also introduces new kinds of load. Training workloads are heavy and spiky, but inference workloads can be continuous and unpredictable. Some applications require low latency responses; others can tolerate slower outputs. Some need large context windows; others rely on retrieval-augmented generation. Some require strict compliance and audit trails. Each variation changes the compute profile and the infrastructure needs.
As a result, the “center” of the doughnut—model capability—doesn’t automatically translate into smooth scaling. Companies discover that the bottleneck might be the data pipeline, not the model. Or it might be the cost of running inference at the required volume. Or it might be the difficulty of maintaining consistent performance across diverse inputs.
This is where the AI boom’s next phase becomes less about flashy breakthroughs and more about operational maturity. The winners are likely to be those who can turn models into reliable services, manage costs, and scale without collapsing under their own complexity.
The infrastructure ring: power, cooling, and the physics of scale
Power is the most obvious “heat” point, but it’s also the most misunderstood. Electricity isn’t just a utility bill; it’s a physical constraint. Data centers require stable power delivery at high loads. They also require careful management of peak demand, redundancy, and failover systems. Even if you can buy hardware, you still need the electrical capacity to run it.
Cooling is the other major physical constraint. As racks fill with accelerators, heat density rises. Traditional cooling approaches may not be sufficient at higher densities, pushing operators toward more advanced solutions. These include liquid cooling, improved airflow design, and optimization of facility-level thermal management. Cooling upgrades take time and require engineering work that can’t be rushed simply because demand is urgent.
Then there’s the question of where the power comes from. Building new generation or expanding transmission capacity is slow compared with the pace of software adoption. Even when power exists nearby, connecting a new facility can involve permitting, grid studies, and construction timelines. The result is that some regions become attractive for AI investment while others lag—not because of talent or demand, but because of grid readiness.
This is one reason the AI boom can look uneven geographically. It’s not just about where companies want to build; it’s about where the infrastructure can support the load. The doughnut metaphor fits again: the glaze of hype spreads widely, but the frying pan—the capacity to cook at scale—is limited.
Supply chains and manufacturing throughput: the hidden calendar
Even when chips are available, the supply chain behind them can be a bottleneck. Accelerators must be packaged, assembled into systems, tested, shipped, and integrated. That process depends on manufacturing throughput across multiple stages, including substrates, memory, interconnects, and power components. It also depends on the availability of rack-scale systems, networking gear, and the specialized components that make high-density compute possible.
Logistics matters too. Shipping schedules, port capacity, customs processes, and warehousing all affect how quickly hardware reaches sites. Delays can cascade. A data center might have power ready but lack the networking equipment needed to connect clusters. Or it might have compute but not enough storage bandwidth to feed it. Or it might have everything on paper but face installation bottlenecks due to labor constraints.
These are not glamorous problems, but they determine whether AI deployments scale smoothly or stall. In many cases, the “boom” is real, but the timeline for realizing it is stretched by the calendar of physical production and installation.
The cost/availability feedback loop
When constraints tighten, costs rise. Costs rise for hardware, for energy, and for the operational overhead of running facilities at high utilization. That affects pricing models for AI services and the willingness of businesses to deploy at scale.
This creates a feedback loop. If inference becomes expensive, companies may limit usage to high-value tasks. If utilization drops because workloads can’t be scheduled efficiently, unit costs rise further. If energy costs spike