SerpApi.Org
It's a term that few companies were paying attention to a year ago, but it has become central to the debate about AI stocks: tokens.
An AI token is the basic data unit processed by an AI model, which determines how much business customers must pay. The troubling trend is that they often end up paying a lot. This has led analysts to coin the term "token shock" to describe how companies find themselves grappling with the cost of embracing AI.
AI Tokens In The Spotlight
AI token expenses are "something that we're spending a bunch of time on, as probably pretty much everyone in corporate America is," JPMorgan Chief Financial Officer Jeremy Barnum told analysts on a July 14 conference call.
With AI costs rising, companies are trying to make sense of how and when they can best deploy top-performing models, such as those from Anthropic and OpenAI.
It's an important question. Enterprise adoption of AI is key to paying off the trillion-dollar data center buildout. To date, Anthropic, OpenAI and Google parent company AlphabetGOOGL have led the charge in getting businesses on board, as the leading makers of large-language models.
But concerns about costs for using those models has opened the door for new entrants. That includes open-source models from China.
Wall Street hopes to gain more insights into AI tokens this earnings season.
For JPMorganJPM, Barnum expects AI tokens to remain a trivial cost overall for the broader business. But the company is strategizing for how it can optimize its spending in future years.
"You really don't need the latest cutting-edge, incredibly expensive model to summarize an analyst report," Barnum told analysts on the company's second-quarter earnings. "The idea is use the right model for the right purpose."
Tokenmaxxing Maxes Out
The JPMorgan CFO's comments highlight how the conversation around AI spending has shifted.
Earlier this year, there were stories of so-called "tokenmaxxing," with companies racing to integrate AI into their operations as fast as possible. An internal dashboard at Meta PlatformsMETA became a viral story because it ranked employees by their individual token usage.
But costs from that approach clearly added up fast.
An Uber TechnologiesUBER executive told The Information in May that the company had nearly blown through its 2026 AI budget in four months. Earlier this month, PalantirPLTR Chief Executive Alex Karp said "something has gone completely wrong" for AI costs in a CNBC interview on July 2.
"As (AI) deployments expand, token consumption is growing faster than token prices are declining, creating what can increasingly be described as a form of 'token shock' for product team and corporate IT budgets," Bernstein analyst Mark Moerdler said in a recent client research note.
Despite those concerns, enterprise AI spending has continued to climb. That's according to data from the fintech startup Ramp, which offers software to help companies manage their spending, including on AI costs.
The Ramp AI Index found that overall AI spending by the median firm climbed 6.4% month-over-month to $10.66 per employee in June.
Many firms are still early in setting their AI strategies. So cutbacks in one part of the business could be offset by ramping adoption in another, Ara Kharazian, lead economist at the financial technology company Ramp, told IBD.
"The challenge for CFOs going forward will be to make these very ambiguous decisions on how to invest scarce AI resources and making sure the right teams and the right projects have access to the right budgets," Kharazian said.
Pricing AI Tokens
AI tokens are the key to that debate. They are the units AI labs use to measure overall consumption. But there is not an exact standard for what makes an AI token.
AI labs like Anthropic and OpenAI convert total units of text into AI tokens, and then charge per million tokens. This paragraph would cost about 50 tokens if "tokenized" by a large language model, according to a calculator offered by OpenAI.
Anthropic's top performing model, Fable 5, costs $10 per million input tokens, which refers to the information a user submits to a model. It costs $50 per million output tokens, or what the model produces.
There is a "flywheel" that can drive AI costs higher, according to Drew Armanino, a senior director and AI solutions practice lead at the accounting firm Armanino.
As AI models become more capable, employees turn to them for more tasks, which could then drive up token usage. The cost of that AI token can be a vague concept to an employee who is excited about using AI to comb their email or write code.
"But at some point, push comes to shove, and a CFO or a CTO will look at a series of bills and say, What am I getting for this?" Armanino told IBD. "That's the conversation that is hotly debated right now."
AI Stocks: China Open-Source
The focus on AI token costs has created a window for challengers to Anthropic and OpenAI — just as the frontier labs are pursuing initial public offerings.
"Low-cost, open-weight models (largely from Chinese labs) are catching up in performance to the proprietary frontier models and are becoming increasingly popular for AI use-cases given their significantly lower cost," William Blair analyst Arjun Bhatia said in a client note on July 20.
China's Moonshot AI recently launched a Kimi K3 model that has a total input-output cost of $18 per million tokens, according to Bhatia. Meanwhile, William Blair found that Claude Mythos 5 and GPT 5.6 Sol have input and output costs of $60 and $35 per million tokens, respectively.
That raises geopolitical concerns as China and the U.S. battle for AI leadership. The Trump administration is "showing signs" it could ban open-source models from Chinese companies, Axios reported.
Meta, SpaceX Target AI Costs
Of course, challengers to the frontier AI labs are not limited to China.
This month alone, Facebook parent Meta PlatformsMETA and xAI parent SpaceXSPCX have launched AI models they describe as cheaper for developers than Anthropic's leading offering.
The top AI labs have taken notice.
Anthropic in late June released Claude Sonnet 5. In a blog post, the AI startup touted how the model can complete agentic tasks at "an attractive price point."
OpenAI CEO Sam Altman recently told CNBC that the start-up's GPT-5.6 Sol model is 54% more token efficient on agentic coding tasks.
"Every enterprise now is thinking about spend and the value they're getting in exchange for AI, and this is what we really want to do," Altman told the news outlet.
Google, meanwhile, on Tuesday released an updated version of its lower-cost Gemini Flash model. The search giant said the new model reduces output token usage by 17%, compared to the prior Flash generation.
Managing AI Costs
Overall, analysts see an opportunity for software companies to benefit from the focus on AI costs.
Fears about AI competition have weighed on software stocks over the past 12 months. But some firms are positioning themselves as orchestrators that help enterprises use the proper model for each task.
"We believe large existing platforms such as MicrosoftMSFT, SalesforceCRM, and ServiceNowNOW are well positioned ... to take a lot of that complexity out of the conversation for organizations, as AI-outcomes on these platforms trend toward a more standardized cost," Evercore ISI analyst Kirk Materne wrote in a recent client note.
