Corporate America Has Suddenly Decided To Stop Blowing Money On AI

It was fun while it lasted, but who realistically believed that corporate customers for AI were going to blow their entire budgets on a new technology whose efficacy remained largely aspirational? 

Many AI company executives and venture investors may have embraced the Silicon Valley hype - 'breathing one's own exhaust' - as they say, but reality began to intrude earlier this spring as the costs went up and returns...remained elusive. What has emerged is the realization that some high end AI are good for specific tasks, but not necessarily for making your back office in Omaha a tad more productive. Different AI strokes - at different price points - for different folks is now becoming the norm. The looming question is, how are all those data center investments and related costs going to produce returns under this more prudent corporate spending scenario? JL

Angel Au-Yeung and colleagues report in the Wall Street Journal:

Companies across the country are coming around to a radical idea upending the industry powering the global economy: They don’t have to blow their budgets on AI. Fed up with ballooning costs, companies are using lower-priced models, including some built in China. In many cases, they are adding the new, cheaper models alongside OpenAI and Anthropic’s products, shopping a la carte for AI. Investors and executives say the era of cheaper AI is here to stay. Companies in Silicon Valley were the first to treat AI like any other product and shop for the best deal. The approach began spreading across industries in recent weeks. “There’s zero loyalty.” Strategies to lower AI costs include limiting access to top models for new hires and using the most advanced AI systems to plan how tasks will be completed before turning to cheaper models for the execution. “We work with all of them, figuring out solutions to improve performance and get much better cost.”

Companies across the country are coming around to a radical idea with the potential to upend the industry powering the global economy: They don’t have to blow their budgets on AI. 

Fed up with ballooning costs, companies big and small are starting to use lower-priced models, including some built in China. In many cases, they are adding the new, cheaper models alongside OpenAI and Anthropic’s products, shopping a la carte for their artificial intelligence.

At Cursor, a startup that stands to benefit from this corporate epiphany because its software works with all models, Mike Saeks puts it simply: 

The most powerful and expensive AI models aren’t necessary for relatively mundane tasks. 

“It’s like driving a Lamborghini to go to the grocery store to pick up milk when that was designed to be raced around a track,” he said in an interview. 

Saeks is in a position to know. The former investment banker was hired two months ago by Cursor, a startup that Elon Musk’s SpaceX is buying for $60 billion, to be a field chief technology officer. That means he advises companies on how to measure and improve the returns on investment from their AI spending. The startup calls it tokenomics, a reference to tokens, the central unit used to measure AI usage.

Being economical—or tokenomical—is a dramatic reversal in mindset. Just a few months ago, it was a badge of honor to be using AI so much that you spent a lot on tokens. Companies rewarded employees for tokenmaxxing, flashing leaderboards that showed who had spent the most. Now they are thrift-maxxing.  

The shift in their budgeting isn’t just about how much U.S. companies are spending on AI. It’s also a geopolitical issue that pits the world’s economic superpowers against each other. Generally, the best-known U.S. models are closed, which means they are strictly controlled by the companies developing them. China is known for cheaper and open-weight models, which means they can be downloaded and customized.

Anthropic and OpenAI have accused the Chinese model-makers of ripping off their technology. The companies have called out Chinese AI startups that have made models in widespread U.S. use, including DeepSeek, MiniMax and Moonshot AI. 

Some companies have so far avoided using Chinese models because of security concerns, while executives at OpenAI and Anthropic have sounded the alarm about potential risks. Some executives and Trump administration officials have suggested a ban on such models; others argue that those favoring restrictions are trying to stifle competition.

On Friday, a group of tech companies including Nvidia, Microsoft and Palantir signed a letter in support of open models, urging U.S. policymakers to exercise caution on potential restrictions.

No matter how such fighting resolves, many investors and executives say the era of cheaper AI is here to stay. A number of American companies have released or are working on new open models as well.

Companies in Silicon Valley were the first to treat AI like any other product and shop for the best deal. The approach began spreading across industries in recent weeks. 

“There’s zero loyalty that I’m seeing,” says Marty Kausas, chief executive of Pylon, an AI customer support platform. “It really feels like a bloodbath right now.”

Pylon CEO Marty Kausas says his company has received more than $1 million in free tokens from AI companies. Eli Imadali for WSJ

AI companies are offering customers like him deals “like crazy” to keep them, he said. His company has received months of unlimited free usage. So far this year, Kausas estimates Pylon has received around $1.6 million in free tokens from one vendor, $65,000 from another and $10,000 from another.

Microsoft has weighed adding Chinese models on its platforms including DeepSeek, and executives at a host of startups say they have had conversations about using open models with top executives at financial institutions, healthcare companies, insurance providers and others.

Cursor can easily embrace model mixing because its coding software is “model agnostic,” meaning that users can toggle between advanced models from OpenAI, Anthropic, SpaceX, Google and Chinese model-makers. 

Strategies to lower AI costs include limiting access to top models for new hires and using the most advanced AI systems to plan how tasks will be completed before turning to cheaper models for the execution, Cursor’s Saeks says. 

Cursor recently ran an experiment to evaluate the cost of building a web browser from scratch. Doing the entire task on OpenAI’s GPT-5.5 cost a little more than $10,000. Using Cursor’s Composer coding model in combination with Anthropic’s Opus 4.8, cost $1,339. 

“The best model for a task used to change every few months,” Saeks says. “Now it feels like it’s happening multiple times per week.”

An OpenAI spokeswoman said a newer model—GPT 5.6 Sol—has been trained to be much more token efficient, meaning it can complete advanced tasks for less money. Anthropic released a powerful, lower-cost model on Friday. A company executive said its users can choose higher intelligence or lower cost within its ecosystem. Both companies say they support open-weight models.

Anthropic and OpenAI have spent years in a race for AI supremacy—trying to develop evermore advanced, premium models only to find now that companies want basic ones, too. The increasing popularity of such models has turned the AI race on its head, threatening the heady valuations of Anthropic and OpenAI as they prepare for public listings. To fight back, they are trying to lock in customers, offering partnerships, tens of thousands of dollars in incentives and heavily subsidized AI usage.

Executives who favor using Anthropic and OpenAI products say they want the highest level of intelligence available and are willing to pay a premium for it. 

Zoom, known for its videoconferencing platform, has been using Llama, an open-weight model released by Meta Platforms, for three years now, and by fine-tuning it, Zoom has saved substantially, Zoom’s chief technology officer Xuedong Huang says. The company currently uses a combination of Anthropic, OpenAI and open models.

Huang referenced an ancient Chinese myth where three ordinary people combine their wits to equal one genius. He believes this method is the “secret sauce” for companies. 

Barry McCardel, co-founder and CEO of Hex, an AI data analytics platform, said using open models has worked better for many of the startup’s customers, because it’s well-equipped to customize the models using company data or for specific tasks. 

“We are continually assessing how much we want to commit to any one lab given how dynamic a moment this is,” he said. In the last two weeks, about 50% of Hex’s customers have adopted Kimi, the model produced by China-based Moonshot into their workflow. “Any day, the labs can drop a model that hits the frontier,” and McCardel said his company wants to stay flexible. 

In the spring, the startup Telnyx, which produces infrastructure for real-time AI agents, or autonomous bots, was running 1,000 agents with a top Anthropic model and an open operating system. The Claude usage, under a top-tier subscription level, cost $200 per employee a month.

A screenshot of an engineer's control deck for AI agent supervision at Telnyx. Telnyx

Then Anthropic stopped allowing third-party operating systems to run on its subscriptions, which it saw as a violation of its terms of service. To continue, Telnyx would have to pay per use.

“We did the math and it was going to be like 100 grand per day,” says David Casem, the company’s CEO.

So he turned to open models.

“They worked,” he says. “It’s not like we don’t use OpenAI or Anthropic models, we still do,” he says. “They just don’t do everything anymore.”

Today, a family of models made by Chinese startup Z.AI is powering the company’s 1,400 agents, which costs around $100 per agent per day. Anthropic’s most powerful model, Fable, acts as a conductor that plans out work while open-weight models do the implementation. OpenAI’s Sol handles a review of what the open-weight models produce.  

Gabe Pereyra, president of the legal AI startup Harvey, said the company trained GLM-5.2 and provided it with a tool that allows it to “call” Anthropic’s Fable 5 model if it determines that “this is a really hard task.”

“We work with all of them,” he said. “We’re figuring out a bunch of solutions like this to maintain performance or improve performance and get much better cost.”