We Have Seen This Compute Boom Before
AI is replaying the cloud era. Most companies will rent the frontier. The scrappy ones will learn how to own the economics.
Years ago, I worked with an engineer who loved to say that cloud technology was “just someone else’s computer.” He believed it strongly enough to make the phrase part of a shared team password, which said something about both his conviction and our security practices at the time.
I disagreed with him then, and I still do. Cloud was always much more than someone else’s computer. It gave us programmable infrastructure, elastic capacity, managed services, global reach, and a way for a small team to operate systems that would once have required a building full of hardware and a much larger staff.
Still, there was a useful truth hiding inside the joke.
The cloud boom convinced companies to stop buying all of their own compute and start renting it from service providers. We traded racks, power, cooling, hardware refreshes, and long capacity-planning cycles for APIs and a monthly bill. The servers did not disappear. They became somebody else’s capital expense.
The current AI boom is the next iteration of that same decision.
Enterprises are paying model providers and cloud platforms to own the accelerators, build the data centers, train the models, operate the inference stack, and keep the whole thing available. We send tokens into an API and receive intelligence-shaped output in return. It feels like software, but underneath it is an extraordinary amount of hardware.
We have seen this movie before. The cast is different, the computers are considerably more expensive, and this time the API occasionally gives you career advice. The business decision is still familiar: do we rent a capability from a provider, or do we build enough of it ourselves to control the economics?
The Frontier Will Be a Managed Service for Most Companies
Most enterprises will choose the provider.
That is not a failure of technical ambition. It is often the correct decision.
Frontier models require enormous investments in research, training, specialized hardware, networking, power, cooling, reliability, security, and capacity management. Alphabet reported $91.4 billion in 2025 capital expenditures, with the vast majority going to technical infrastructure. AWS has announced tens of billions of dollars in new AI infrastructure investment. Very few companies can participate in that race, and fewer still have a business reason to try.
An enterprise buying access to a frontier model is making the same basic choice it made when it adopted a managed database or a cloud object store. It is paying to consume the outcome without owning every layer required to produce it.
The advantages are real:
- Immediate access to the most capable models
- No GPU procurement or hardware lifecycle to manage
- Rapid improvements without a model migration project every quarter
- Elastic capacity for unpredictable workloads
- Mature security, compliance, and regional deployment options
- A smaller infrastructure team than an equivalent self-hosted platform would require
For many companies, model hosting will become another line in the cloud bill. Teams will choose a provider, connect it to their data and workflows, add governance, and focus their energy on the product they are building.
This is likely where most of the industry ends up. Cloud adoption taught us that convenience, speed, and access to managed capability usually win, even when owning the underlying hardware might be cheaper under the right conditions.
Reliance on service providers will grow because the providers are selling something more valuable than raw compute. They are selling relief from complexity.
The Other Group Will Learn What a Token Really Costs
There will also be a second group.
These companies will be more budget-conscious, more technically stubborn, or simply too scrappy to accept that every useful AI interaction must pass through a premium API forever. They will run open-weight models on hardware they own, colocate, reserve, or control more directly. Some will do it in a traditional data center. Others will use dedicated cloud capacity while keeping their inference stack portable.
Their first attempts will not always be cheaper.
Owning hardware does not magically create good unit economics. GPUs sitting idle are expensive furniture. Poor batching wastes capacity. Oversized models turn simple classification jobs into science projects. Weak observability makes it difficult to understand whether money is being spent on useful work, retries, bloated context, or an agent repeatedly investigating the same wrong assumption.
The companies that work through those problems will develop something valuable: a practical understanding of how AI value is created.
They will know which tasks need frontier reasoning and which can be handled by a smaller model. They will understand quantization, batching, caching, routing, context management, and hardware utilization. They will build evaluation systems that measure whether a cheaper model is actually good enough. They will learn that the best model for a workload is rarely determined by a single leaderboard.
Most importantly, they will understand their own demand.
The advantage is not simply owning GPUs. The advantage is knowing what the GPUs are doing.
Open-weight models make that possible. Model families such as Llama can be downloaded, adapted, evaluated, and operated without sending every request to the organization that produced them. Open-source inference servers, model routers, evaluation frameworks, observability tools, and agent harnesses are turning what once required a research lab into something a capable infrastructure team can operate.
This does not mean a small company will train the next frontier model in a server closet. It means the company can decide that a narrow, well-understood task does not need the frontier at all.
That distinction will matter more with each passing year.
The Cheap Experiment Becomes an Expensive System
Cloud bills have a familiar lifecycle.
The prototype is inexpensive. The first users barely move the graph. Adoption grows, the workload becomes important, and the monthly bill eventually reaches the meeting where someone asks why it increased by 400 percent.
AI systems are moving through the same lifecycle, but they contain a multiplier that ordinary applications do not.
A conventional application may make a database query when a user performs an action. An AI application may assemble context, call a model, call a tool, inspect the result, revise its plan, call another model, and then retry part of the process. An agent can consume more compute while appearing to perform the same user-visible task. Greater autonomy often means more inference.
Today, much of that consumption feels unusually affordable. Providers offer free access, flat-rate subscriptions, credits, aggressive token pricing, and generous limits because they are competing for adoption. Investors and platform companies are funding the infrastructure buildout while customers are still learning where AI belongs.
That is normal for a new platform market. It is also temporary.
The infrastructure eventually has to produce a return. Data centers, accelerators, power contracts, research teams, and global inference capacity cannot remain abstract financial concepts forever. Even as the cost per token falls, usage can grow faster. Cheaper inference enables larger contexts, more background agents, more generated media, more retries, and AI features in places where nobody previously thought to put them.
Cloud compute became cheaper and more efficient over time. Cloud bills did not disappear.
The same will be true for AI.
The real cost may arrive through higher prices, tighter limits, premium service tiers, data residency charges, reserved capacity, or simply an enormous increase in how often applications call models. A token can become cheaper while the system becomes more expensive.
The companies that treat current pricing as a permanent law of nature will be surprised. The companies measuring cost per successful business outcome will be ready.
This Is Where the Scrappy Companies Gain Ground
Large enterprises can absorb a surprising amount of inefficiency. They can sign broad agreements, negotiate volume discounts, and place AI spending inside a transformation budget while the use cases mature.
Up-and-coming companies have less room for theater.
That constraint can become an advantage. A team that cannot afford to send every problem to the largest model is forced to understand the problem more precisely. It has to ask:
- Does this task require generation, or would a deterministic system be better?
- Does the model need the full document, or can retrieval provide a smaller relevant section?
- Does every step need a frontier model?
- Can a smaller open-weight model handle the common path?
- Can results be cached or reused?
- How much capacity is idle?
- What does a successful outcome cost, not merely a million tokens?
These are infrastructure questions, but they are also product questions. Answering them produces better systems.
The likely winning architecture is not religiously “open” or “closed.” It is selective.
A company may use a frontier provider for difficult reasoning, a smaller hosted model for routine work, and an open-weight model on its own hardware for high-volume predictable tasks. A routing layer can make that choice based on quality, latency, privacy, and cost. Evaluation determines whether the substitution is safe. Observability shows whether the expected savings are real.
This is the AI equivalent of a mature cloud strategy. Good cloud engineering was never about putting every workload in the same place. It was about understanding the workload well enough to make intentional tradeoffs.
The scrappy companies that build this capability early will have options later. They will be able to move a workload when provider pricing changes. They will be able to take advantage of improving open-weight models. They will know whether dedicated hardware makes sense because they will have actual utilization data. They can keep frontier providers as a powerful tool without allowing one provider’s price sheet to become their business model.
That optionality is a competitive advantage.
Owning the Economics Requires Owning the Understanding
There is a broader lesson here that applies beyond infrastructure.
In “Extending the Reach of Every Engineer,” I argued that agentic coding should expand an engineer’s capability rather than replace the understanding beneath it. The same principle applies to AI platforms.
Abstraction is valuable. Ignorance is not.
A managed service can remove operational work, but it does not remove responsibility for architecture, cost, security, quality, or business outcomes. An API can hide the inference server, but it cannot tell an organization whether the model call was necessary. A model can return an answer, but it cannot decide how much that answer was worth.
Organizations should be developing this understanding now, while provider access is plentiful and experimentation is relatively inexpensive.
That does not require buying a data center. It requires building the muscles to evaluate models, measure workloads, route requests, control context, and understand cost. It means avoiding application designs that are inseparable from one model’s behavior. It means maintaining an escape path before a pricing change turns portability into an emergency project.
Most companies will still choose managed frontier models, just as most companies chose cloud infrastructure. The point is not that everyone should self-host. The point is that the companies capable of self-hosting, substituting, or negotiating from a position of real knowledge will make better decisions.
Nothing New, and Nothing to Fear
Every major abstraction creates anxiety among the people who understand the layer being hidden.
Virtual machines were going to make hardware knowledge irrelevant. Cloud services were going to eliminate infrastructure engineering. Containers were going to make operating systems disappear. Serverless was going to end server management.
None of those things happened.
The work moved. The useful level of understanding changed. Some responsibilities became centralized, some became automated, and new ones appeared. The engineers who understood both the abstraction and what lived beneath it became more valuable, not less.
AI infrastructure will follow the same pattern.
Frontier providers will make extraordinary capabilities available to almost everyone. That will create more dependence on a small number of service providers, and for most organizations the trade will be worthwhile. At the same time, open-weight models and open-source tooling will give determined teams a path to understand and control more of the stack.
The real divide will not be between companies that use closed models and companies that use open ones.
It will be between companies that understand their AI economics and companies that do not.
Cloud taught us that not every company should own a data center. It also taught us that moving infrastructure to a provider does not make cost, architecture, or operational discipline disappear.
AI is teaching the same lesson again.
The frontier is convenient today, and in many cases remarkably cheap. We should use it. We should learn from it. We should build valuable things with it.
We should also remember that the bill is coming.
When it does, the winners will not necessarily be the companies with the largest model available in a dropdown. They will be the ones that know when the frontier is worth paying for, when an open-weight model is enough, and how to turn hardware into a durable business advantage.
They will not be afraid of the infrastructure beneath the API.
They will understand it.
