当企业无力支付他们的AI时

中文版正在随流水线自动整理，稍后刷新即可；可先阅读下方英文原文。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Thought Leaders

What Happens When Businesses Can’t Afford Their AI Anymore?

Add Unite.AI to your preferred sources on Google

In its early years, Uber made a cross-town ride cheaper than a taxi. Investors covered the difference until the company needed to make money. Data from Rakuten Intelligence showed US ride-hailing prices rising 92% between 2018 and 2021, and by 2023 even Uber’s CEO was surprised to hear what a three-mile trip in New York now cost.

Artificial intelligence is approaching a similar moment. Anthropic confirmed in June that it had confidentially filed for a US public listing, and OpenAI is reportedly preparing to follow. Public companies answer to shareholders every quarter, and that scrutiny tends to change pricing.

That raises a question many businesses have yet to ask: what happens if the AI their products rely on becomes more expensive, more restricted, or both? The answer depends less on what the AI labs decide than on how each business has built around them, and there is still time to prepare.

AI Costs Have Moved from the IT Budget to the Business Model

The first AI bills were small enough to file under experimentation. Today AI sits inside products and workflows, so its price feeds directly into margins. Current prices reflect a period of heavy investment.

In a test by research firm SemiAnalysis, a $200-a-month Claude Max plan delivered around $8,000 of usage a month at standard API prices. That is generous for users, and it shows how far some prices sit from the underlying cost of the service.

Some adjustments have already begun. In mid-2025, the AI coding tool Cursor moved its $20-a-month plan to usage billed at API rates, and Anthropic introduced weekly usage limits for its heaviest Claude Code users. Developer Simon Willison suggested the era of investor-subsidised AI usage may be ending.

Analysts are giving similar advice. Gartner’s August 2026 guidance urges leaders to prepare for likely AI price increases. Its analysts have also warned that organisations that don’t understand how AI costs scale can misjudge them by 500% to 1,000%.

Falling unit prices help, but only up to a point. Epoch AI shows the price of a given level of AI performance falling sharply each year, yet MIT researchers found newer reasoning models use far more processing to reach their answers. The bill grows with usage, and usage keeps growing.

What Public AI Labs Could Mean for the Startups Building on Them

A listing brings discipline. Public AI labs will need to show investors a path to sustainable margins, and pricing is one of the most direct levers available. That may mean higher list prices, tighter usage limits, features moved into premium tiers or older models retired sooner.

Startups also share a market with their suppliers. AI labs increasingly offer their own applications, as Anthropic has with Claude Code, so it pays to be clear about the value a business adds beyond the model itself.

There is a constructive side. Public companies disclose more, giving customers clearer insight into how AI is priced. Competition between providers, along with fast-improving open models, should keep prices in check for those ready to switch.

How Companies Can Reduce Their Dependency

The business risk comes down to two questions. Would customers value the service enough to pay more if costs rose? And is there enough margin to absorb a price increase? For any business unsure of the answers, five practices help, and none require starting from scratch.

Know the cost of each task. Many teams know their monthly AI bill but not what a single customer action costs. Measuring cost per task shows which features, clients or use cases would become unprofitable if prices rose by 30%, 50% or even 100%.

Use AI where it adds the most value. Large models are strongest at handling ambiguity, such as approaching a problem for the first time. Repetitive work often suits conventional software. In software testing, for example, AI can work out how to check a user journey once, and that check then runs thousands of times as ordinary code without calling a model.

Match the model to the job. Epoch AI found that the best open models able to run on a single consumer graphics card trail the leading models by about a year or less. For routine work, a smaller or self-hosted model can deliver acceptable results at a fraction of the price.

Keep the option to switch. Products built around one provider’s specific features are the hardest to move. A simple layer between the product and the model, plus regular testing of alternatives, turns a future price rise into a negotiation.

Price by outcome. Billing customers for tokens passes every change in AI prices straight through to them. A flat fee per outcome, such as a completed task or a resolved request, keeps pricing stable for customers and gives the business room to absorb changes in its own costs. The same logic helps when buying: tools priced per outcome make future spending easier to predict.

Affordability Is a Design Decision

Uber’s price rises did not end ride-hailing. They ended business models that assumed rides would always be subsidised. AI is likely to follow a similar path. As the major labs move into public markets, AI will be priced closer to what it costs to provide, and businesses will be judged on whether their products still work at that price.

The companies best placed for that shift treat AI as a dependency to be measured and managed. For businesses that haven’t started, it is not too late. Each of these steps can begin this quarter, and each one makes the next price change easier to absorb.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.