Want to know how comfortable your retirement will be? Baffled by the jargon thrown around by A.I. moguls? Well, you’re in luck, because deciphering four concepts will go a long way toward answering that crucial question about your personal finances.
The math that lies behind this Silicon Valley shorthand will help determine if America’s largest A.I. companies can earn enough revenue to justify historic levels of capital spending. Next year, Amazon, Google, Meta and Microsoft are expected to spend nearly $1 trillion, according to estimates by Bloomberg News. That’s roughly what the entire world is expected to spend in 2026 on fossil fuel supplies.
Since 10 dominant tech companies represent more than 35 percent of the value of the entire S&P 500, as of late July, how they fare will meaningfully shape what happens to the more than $47 trillion Americans have invested in retirement plans.
So let’s start with …
A.I. agents
Agents are computer systems that analyze data, consider scenarios and choose strategies to solve what can be complicated problems. They can be applied across a wide variety of issues with minimal human involvement.
Siri or Alexa — both reintroduced this year as A.I.-powered chatbots — can answer your weather forecast question. A more sophisticated agent could automate processes for the National Hurricane Center to predict extreme weather events far into the future, ensure vulnerable areas receive regular updates and alert officials on steps they should take. For companies pressured by antsy shareholders and boards to increase productivity, A.I. agents have been welcome news.
While that all sounds great, it comes at a cost — one that’s often significantly higher than anticipated. In an April interview in The Information, Uber’s chief technology officer, Praveen Neppalli Naga, said the first few months of 2026 had already exhausted his company’s entire annual A.I. budget.
Such surprises led OpenAI’s Sam Altman to acknowledge things had changed He said that in January, “people were totally happy with the amount they were spending” on A.I. Five months later, cost had emerged as a “huge issue” for his customers.
This uncomfortable expense comes down to the purchase of …
Tokens
While my middle-aged mind first goes to subways and county fairs, in A.I.-speak, tokens refer both to the basic units of text that are processed by computer models and to the unit of billing for A.I. The cost of A.I. depends on the number of tokens used and how they are deployed. Asking a bot like ChatGPT a simple question is pretty inexpensive. The cost of a single token has declined, but the sheer number of tokens needed to run complicated A.I. agents has risen, and that means the overall cost has ballooned.
In June, Walmart’s chief technology officer said the company was limiting tokens per employee to control costs. Employees hitting the cap would have to wait for a reset to continue. While that could potentially slow innovation, it also pushed employees to consider other ways to approach a task, including a traditional internet search or phoning an experienced colleague. Old school, FTW.
Other users have switched to less expensive A.I. tech — often what are called …
Open-weight models
These A.I. models, a specialty of Chinese tech companies, are publicly accessible and transparent (think “open for inspection”). The word “weight” refers to the way the model was trained. Open models tend to be significantly less expensive than so-called closed models, like those produced by OpenAI, Anthropic and most leading U.S. companies, and are often favored by tech start-ups with limited budgets.
Closed models are easier to use but cost more and feel like proprietary black boxes. I might like my A.I. model’s response, but I have no idea how it generated that answer. The big worry is that workers will switch from American A.I. companies to open Chinese models.
China, like the United States, sees A.I. as a way to support economic growth through faster productivity. It has made A.I. development a national priority, with an emphasis on open models and A.I. applications such as robots. It used a July A.I. conference hosted in Shanghai, and attended by President Xi Jinping and representatives from dozens of countries, to unveil a new open A.I. model — Kimi K3.
In the last week of July, the five most used A.I.s tracked by OpenRouter, a company that helps users switch between different models, were Chinese, representing 43 percent of all token usage that week. U.S. competitors barely registered on the leaderboard.
China’s gains are manageable for U.S. companies, so far, since spending on open models is a small fraction of what closed-model companies receive. But top U.S. executives and Wall Street analysts worry the shift could go further, which would hurt future revenue and sky-high stock market valuations.
Washington is asking an additional question: What are the geopolitical and security risks if China’s A.I. keeps improving and takes market share away from U.S. companies? The country and its capabilities will be front and center for A.I. policy this September, when President Trump and President Xi are set to meet in person. The summit is likely to touch not just on the two countries’ approaches to A.I. but also more granularly on the costs of key A.I. inputs — what’s often referred to as …
Compute
Once a verb, now a noun, it refers to the computational power needed to train and run A.I., and the hardware that allows A.I. models to work. Think semiconductor chips, power and data centers. Despite record spending that has weakened the balance sheets of A.I. companies, their access to compute is constrained. This includes steady access to critical minerals used to make chips, which today are largely processed in China.
These constraints, coupled with strong demand, have pushed up costs for A.I. companies. Apple’s chief executive, Tim Cook, during a quarterly earnings call in July, described recent input price increases as a “100-year flood” that is pushing the company to raise prices this year.
Generating more compute requires huge investment and a bet on A.I.’s success. To put together all the jargon in one place (deep breath): Revenue from users deploying A.I. agents is one way to pay for compute, but the costs of token usage, along with the existence of inexpensive open-weight alternatives, suggest there may be limits to how much money these tech giants can make. And that is a central conundrum that will help determine the future of U.S. and global financial markets — and your savings.
Rebecca Patterson is an economist and senior fellow at the Council on Foreign Relations who has held senior positions at JPMorgan Chase and Bridgewater Associates. She is a co-host of the podcast “The Spillover.”
The Times is committed to publishing a diversity of letters to the editor. We’d like to hear what you think about this or any of our articles. Here are some tips. And here’s our email: [email protected].
Follow the New York Times Opinion section on Facebook, Instagram, TikTok, Bluesky, WhatsApp and Threads.