Opinion: Why we plan for AI to get more expensive
AI capacity is short and prices are already moving. The two risks this creates for teams that build on AI, and how we plan for higher prices.
· 5 min read
By Track49
We use AI assistants every day, and so do most of the teams we work with. Many of them have rebuilt their development process around tools that feel cheap and always on. We think neither can be relied on, and that a business depending on them should plan for the day they change.
What is going right
Some prices are falling fast. Epoch AI measured how quickly the price of a given level of performance drops. The price to reach GPT-4’s score on a set of PhD-level science questions fell by about 40 times a year. Depending on the benchmark and the level chosen, the decline ranged from 9 to 900 times a year. Capability is moving just as quickly. At the moment, the models we use improve every few months, and they take on work that would have been unrealistic a year ago. It is a remarkable period to be building software.
Our concern is narrower. Cheaper tokens and better models still need chips, power and buildings to run on, and those are growing far more slowly than demand. The companies that grew fastest, Anthropic and OpenAI, have hit those limits first. Anthropic’s annualized revenue went from $9 billion at the end of 2025 to more than $30 billion by April. OpenAI’s API traffic rose from 6 billion to 15 billion tokens a minute between October and March. The chips, power and data centers to serve growth like that take years to build, not months.
What the supply side looks like
The evidence has been building all year. In April, The Wall Street Journal reported on the state of AI capacity, and the details are worth reading closely.
- Reliability. Over the 90 days to April 8, Anthropic’s API was available 98.95% of the time. Established cloud providers typically maintain 99.99%. Over a quarter, that is the difference between about 13 minutes of downtime and about 22 hours. Some enterprise customers switched providers because of it.
- Rationing. GitHub added new limits to Copilot on April 10, citing rapid growth and high concurrency. Windsurf replaced its credits with daily and weekly quotas. Anthropic adjusted its session limits and offered extra usage off-peak to spread the load.
- Products cut, partly for capacity. OpenAI closed the Sora app. Its CFO, Sarah Friar, said she spends much of her time looking for near-term compute capacity.
- Prices already moving. An hour on a current Blackwell GPU rose from $2.75 to $4.08 in two months, an increase of 48%, according to the Ornn Compute Price Index. CoreWeave raised prices by more than 20% and now asks smaller customers for three-year contracts. So far the increases are in wholesale compute and in how subscriptions are metered, not in list prices per token. We expect list prices to follow, and agentic work to keep raising the number of tokens each task consumes.
- Power is the hard limit. Vultr’s chief executive, J.J. Kardwell, pointed to long hardware lead times and slow data center construction, and said that available power through 2026 is already spoken for. Bank of America analysts expect demand to exceed supply through at least 2029.
- Communities are pushing back. Data Center Watch counted at least 75 US data center projects, worth about $130 billion, blocked or delayed by local opposition in the first three months of 2026 alone.
New chips arrive on long lead times. Grid connections and substations take years, and each new site now has to win over the people who live next to it. When a resource is that constrained and demand keeps rising, the price eventually reflects it. Today’s AI pricing was set while vendors competed for customers, and we don’t expect it to hold once capacity, not market share, is what they are rationing.
Two risks for teams that depend on it
Operational dependency. If your engineers stop when the model is slow, rate-limited, repriced or down, you have added a single point of failure to your delivery process, and you don’t control it. The same is true of production features that call a hosted model. A 429 response, a deprecated model version or a quiet change in model behavior can all break a workflow that has no fallback.
Comprehension debt. Code that an assistant wrote and nobody on staff understands still has to be maintained. While the assistant is cheap, that cost is hidden: you ask it again. If it becomes expensive or unavailable, the code is suddenly hard to change. That is how a legacy system starts, and it can start in a codebase that is only a year old.
What we do about it
None of this is a reason to stop using AI. It is a reason to use it in a way that still works if the price triples or the service is rationed for a week.
- Keep people who understand the systems you run. Review AI output to the point where someone could have written it, and could change it without help.
- Let the toolchain catch mistakes, whoever made them. A strict compiler, a real test suite and enforced project rules make the origin of a change matter less.
- Write decisions down. Design rationale belongs in the repository, in specifications and decision records, not in a chat history that expires with a subscription.
- Model the spend at other prices. Take your current token usage and price it at three and ten times today’s rates. If the answer is alarming, find out now which work actually needs a frontier model.
- Keep the provider replaceable. Put model calls behind one interface, keep a small evaluation set for each task, and you can measure whether a cheaper or self-hosted model is good enough before you need to switch.
That last point is why we self-host open-weight models alongside the frontier models we use, and help some clients do the same. Routine work runs on models we control, on hardware we choose, at a cost we can forecast. Frontier models handle the work where they make a measurable difference. If a provider raises prices or limits capacity, the routine work continues, and the decision about what goes where stays with us.
Sources
- Epoch AI, LLM inference prices have fallen rapidly but unequally across tasks, March 12, 2025.
- Data Center Watch, Q1 2026: Data Center Watch Report, covering January to March 2026.
- The Decoder, The AI industry is running out of compute, with outages, rationing, and rising GPU prices, April 13, 2026, reporting on The Wall Street Journal and the Ornn Compute Price Index.
- ai
- operations
- risk
- engineering