AI spending hits a plateau as price wars and efficiency gains take hold

The spending plateau

Data from Ramp indicates a significant shift in how companies are paying for artificial intelligence. While the volume of tokens—the basic units of text processed by AI—continues to rise, the actual money spent on those tokens has largely flatlined. This trend became particularly visible over the late summer months of August and September.

In July, token spending saw a massive jump of 42% compared to the previous month. However, that growth plummeted to just 1.5% in the following two months. This divergence suggests that while people are using AI more frequently, the cost of that usage is no longer scaling at the same pace as the demand.

Ramp's lead economist, Ara Kharazian, noted that the most intense AI spending comes from venture-backed, high-growth tech companies. He warned that if these specific firms start to slow down due to market factors, it could create cascading effects across the entire industry. The data shows that the most dedicated firms, often called 'AI-pilled,' have seen their spending stutter slightly below their summer peaks.

A buyer's market emerges

The primary driver behind this spending slowdown is a fierce price war between the major AI providers. OpenAI recently cut the prices for its GPT-6 Sol and GPT-6 Luna models by half. In response, Anthropic significantly reduced pricing for its Fable 5.1 and Opus 5.5 models to stay competitive in the marketplace.

Anthropic admitted in a recent filing that these price cuts actually slashed its revenue growth by about $600 million. Both companies are also offering enterprise discounts to secure long-term contracts. While these tactics are designed to lock in customers, they are simultaneously making it harder for the companies to hit high revenue targets quickly.

Kharazian describes the current environment as a buyer's market where competition is high. He points out that Meta and Nvidia are also offering cheap or open-source models. Additionally, new tools like Jev allow users to choose between predetermined outputs, making simple tasks like 'Yes/No' answers even cheaper to execute than they were before.

Efficiency and optimization

Beyond the price cuts from the providers, companies are becoming much more sophisticated at managing their own AI budgets. As teams get more experienced with the technology, they learn how to be more thrifty. This involves moving away from using the most expensive models for every single task and instead finding the right balance of power and cost.

One common method for this is using 'routers.' These are systems that automatically direct a user's request to the cheapest model capable of handling it. For example, a company might use a very expensive model for complex coding but route simple internal queries to a much cheaper, lightweight model to save money.

This shift is reflected in the spending gap between different types of companies. While the top 10% of firms spend between $600 and $1,000 per employee, the median firm spends only $12 to $15. It is unlikely that every industry, such as car dealerships or doctor's offices, will ever reach the high spending levels seen by the most tech-heavy firms.

Valuations face pressure

The cooling of AI spending is already beginning to reflect in the financial outlook for the industry's biggest players. Bloomberg reported that Anthropic's potential valuation dropped from an estimated $2 trillion in early October to between $1.2 trillion and $1.3 trillion. The company is also facing a high 'junk' rating from credit agencies.

OpenAI has followed a similar path by lowering its proposed valuation from $1.4 trillion to $1.25 trillion. They also reduced their latest fundraising goals by approximately $10 billion. These adjustments suggest that the market is beginning to price in the reality of slower revenue growth as the initial hype settles.

Investors are now looking for the next 'revolution' to drive growth, similar to how AI coding agents fueled the current boom. Some hope that personal agents will be the next big thing, but Kharazian remains skeptical. He argues that personal agents may not be as economically productive as coding, and based on current products, the era of hypergrowth spending may be slowing down.