Is the AI giants' 'Token subsidy war' almost over?

Editor | Jingyu

Tokens are expensive, and burning through them hurts.

This isn't just the sentiment of people currently obsessed with Vibe Coding; even the Silicon Valley giants that were previously frantically advocating Tokenmaxxing have started imposing token limits on their own employees.

But a counter-intuitive fact is that for those of you currently using AI subscriptions, the tokens you are using have already been subsidized by the big AI companies, with the highest subsidy potentially being a staggering 70 times the subscription fee !

What's perhaps more worrying is that the two AI leaders, OpenAI and Anthropic, have both entered the sprint phase toward an IPO. Once these two companies go public,

will it be like the aftermath of the "subsidy wars" of the internet era, where the remaining companies start raising customer unit prices one after another, bringing token prices back to a rational level?

The good news is, this scenario might not happen. Recently, Google Ventures founder Bill Maris posed a question on the All-in podcast:

If Google decided to slash token prices by another 80%, how would OpenAI and Anthropic respond?

Coincidentally, not long ago, the startup team Agnes AI, in a live broadcast with GeekPark, explained in detail the potentially arriving "era of free tokens."

So, will the price of tokens rise or fall in the future? And what does this mean for people already addicted to AI ?

01, The Token Subsidy War is Fierce

Why is it said that the current price of tokens is actually not expensive?

Because, at least for AI subscriptions, the current prices from various AI companies are already "rock-bottom prices" after subsidies.

Recently, SemiAnalysis conducted a detailed evaluation comparing the actual value of tokens consumed against the subscription fees under OpenAI's and Anthropic's subscription models.

SemiAnalysis did something simple but effective——they actually used AI to complete various tasks under each AI platform's subscription plan, then used the public API pricing to back-calculate how much these tasks' tokens were worth. The results were as follows:

Note a pattern: the more expensive the plan, the higher the subsidy multiplier. This in itself indicates that these high-end plans are not designed to make money——they are a form of "inverse pricing," using the most aggressive losses to retain the heaviest users. Because heavy users are developers and enterprise decision-makers; once they are locked onto a platform, they bring entire teams and product lines with them.

Why continue operating at such a loss? The standard answer is: burn money first to gain scale, then raise prices to recoup losses once scale is achieved. This is how the mobile internet played out——Didi and Uber subsidized hundreds of billions of yuan in ride fees, and after the subsidies ended, ride prices went up; Meituan subsidized countless meal deliveries, and after the subsidies ended, delivery fees went up. This logic holds on one key premise: a lock-in effect was established during the subsidy period.

Didi could raise prices because drivers couldn't leave the order flow on the platform, and passengers couldn't leave the drivers on the platform. Meituan could raise prices because merchants couldn't leave its traffic and delivery network. When the subsidies ended, users were already "locked" into the ecosystem, with extremely high switching costs.

But the AI war has a fundamental difference from the internet era—— Tokens have almost no lock-in effect.

If Claude raises its prices, a developer can migrate API calls to GPT or Gemini within a day——interfaces are becoming increasingly standardized, and many development frameworks even have built-in multi-model switching capabilities. For ordinary users, it's even simpler: just change a URL. AI is not like ride-hailing, which has local driver networks; not like food delivery, which has logistics systems; not like social media, which has friend relationship chains. A token is a token, the same thing no matter who produces it.

This means that once subsidies stop, users can vanish instantly. Subsidies aren't "building a moat"; they're more like "maintaining a pulse"——as soon as someone offers a lower price, users are gone.

And this doesn't even account for a new variable that is making everyone's bills spiral out of control: AI Agents .

When you chat with ChatGPT, a single conversation might consume a few thousand tokens. But when you let an AI Agent execute a complex task——writing a piece of code and then auto-debugging it, analyzing a dozens-of-pages-long document and generating a report——in one round, the token consumption is 5 to 30 times that of a normal conversation. Some developers have tested and found that on the $100 Claude Max plan, a single Agent coding session can burn through nearly a hundred dollars' worth of tokens. Uber's CTO recently revealed that the company burned through its entire 2026 AI budget in just four months.

The question is, can such a token subsidy war be sustained? Who is likely to be left standing at the end of the melee?

Bill Maris believes the answer is clearly the traditional giants.

02, Token as a weapon

To understand the true brutality of this subsidy war, one must first see a structural asymmetry——the ammunition sources for the warring parties are completely different.

Google's annual advertising revenue exceeds $300 billion. This isn't money from investors, nor capital burned from fundraising, but a money-printing machine that runs automatically every day. Billions of people worldwide open search engines, watch YouTube, and use Gmail daily, and advertising fees flow automatically into the account. It doesn't need roadshows, doesn't need to appease analysts, and doesn't need to explain to anyone why it's spending this money.

Google subsidizing AI tokens with its advertising profits is like someone sitting on an oil well fighting a price war at gas stations——his oil bubbles up from his own land, while his competitors' oil is bought with bank loans.

OpenAI and Anthropic are those buying oil with loans.

OpenAI has cumulatively raised over $180 billion, with its latest valuation exceeding $850 billion. Anthropic has raised over $130 billion. This money comes from venture capital and strategic investors——they aren't giving money for charity; they expect these companies to go public and deliver substantial returns upon exit.

And after going public, the real trouble begins. Being public means financial statements are open to the world. Every quarter, Wall Street analysts will scrutinize revenue, profits, customer acquisition costs, and marginal costs. When they calculate that for every $1 in subscription fees you receive, you're actually losing $70——even the most brilliant growth story won't support the stock price.

Bill Maris laid out this logic very bluntly on the podcast. His exact words were: "If I were Google, and I decided to arbitrarily cut token prices by 80%, what would happen to OpenAI's and Anthropic's business models?"

The host pressed him on how likely that was. Maris didn't hesitate: " 100%. Capital as a weapon, tokens as a weapon. "

This isn't an analyst's speculation. Bill Maris is the founder and CEO of Google Ventures and was also a Special Projects VP at Google, having incubated Waymo and Google X. Everyone present understood: this wasn't a hypothesis; this was him having seen firsthand how Google fights a war.

The scenario he painted is simple: Google announces an 80% price cut for the Gemini API. What would enterprise customers do? If the product quality is comparable——and in many benchmarks, Gemini is already on par with Claude and GPT——but the price is four-fifths cheaper, would you continue using the expensive one?

Maris gave the answer himself: "If you're a company, and you can go to Google and Gemini and pay 80% less for essentially the same product, why wouldn't you? And then the pressure on those companies becomes very severe."

And OpenAI and Anthropic have almost no symmetrical countermeasures. They can't follow suit with price cuts——they have no money-printing machine; every dollar is investor money. They also can't maintain a premium based on a technological gap——the gap between large models is rapidly narrowing. Today you might lead by three months, and in three months you'll be caught up. This isn't like the generational technological gap of the iPhone versus Nokia. The moat between AI models is more like a dam built of sand, easily overtopped when the tide rises.

Under Bill's narrative, Google has a strong chance of winning, but in the world of AI, can Google really monopolize? Meta can open-source a free model at any time; China has DeepSeek and ByteDance; Amazon is pushing its own models. When you drive token prices down to dirt cheap, competitors don't disappear——they also cut prices.

In the AI war, there may be no winners.

03, The "Infinite Game" of Tokens?

Even those with the least knowledge of history can somewhat make the following judgments about the endgame of the current AI war:

The first is the "Internet Service" script ——the story of Didi, the story of Amazon: subsidize first, then monopolize, then raise prices to reap profits. In this script, today's price war is just the prologue; eventually, one or two winners will occupy the vast majority of the market and gain pricing power. If so, the current huge losses are a worthwhile investment——just like Amazon lost money for twenty years before ultimately becoming a dual champion in e-commerce and cloud computing.

The second is the "Utility" script . Tokens become a standardized basic resource, like electricity, bandwidth, or cloud storage. No one can maintain pricing power long-term because product differentiation is too small and switching costs are too low. Competition pushes prices infinitely toward the cost line, with profit margins approaching zero. Eventually, the government might step in to regulate——just as it did with electricity and telecommunications a hundred years ago.

The divergence between these two scripts hinges on one word:

Lock-in .

Didi could raise prices because passengers were locked into the driver network, and drivers were locked into the order flow. Amazon could raise prices because merchants were locked into its logistics and traffic ecosystem.

The lock-in effect is the cornerstone of the "lose money first, profit later" model .

But AI tokens——as has been repeatedly argued above——have almost no lock-in. APIs are standardized, and switching costs are practically zero. The core condition for the first script to hold true simply does not exist for the product that is the token.

If the second scenario—the "utility" infrastructure endgame—is closer to reality, what we are witnessing is not a war that will eventually produce a winner, but a war of attrition with no end.

Meituan founder Wang Xing once described this state of competition. His insight was: in some competitions, the concept of "winning" does not exist. The goal of participants is not to defeat their opponents, but to ensure they always remain at the table. Because as long as you are still at the table, you can continue to raise funds, hire people, and iterate. Leaving the table is the only way to lose.

Using this framework to re-examine today's AI landscape, many seemingly contradictory things suddenly become clear.

OpenAI's latest valuation exceeding $800 billion is not because training models requires that much money. It needs that much money to continue fighting the price war. Fundraising is not to win; it is to "qualify to keep fighting."

Google's plan to slash token prices by 80% is not to eliminate OpenAI and Anthropic. It is to ensure it remains a core player in the AI era—just as it once used free Android to ensure it was not thrown off the table by the mobile era.

And Anthropic raising the API pricing for its latest flagship model, Fable 5, to double that of its predecessor—$10 per million input tokens and $50 per million output tokens—appears to be a "price hike," but is actually a proactive screening for enterprise clients willing to pay for high-end capabilities, because it knows deep down: it cannot out-burn Google in a consumer-side subsidy war.

Every round of price wars expands the scale of AI usage. Greater scale means more data, more scenarios, and more developers flooding into the ecosystem. This, in turn, makes all participants' models stronger. The combatants use the war itself to attract resources and upgrade themselves—this is not a zero-sum game of life and death, but a process where everyone grows stronger through competition, yet none are likely to earn windfall profits.

Doesn't this sound like the eventual shape of the electric power industry?

140 years ago, both Edison and Westinghouse thought they were fighting for a winner-take-all market. They bet their entire fortunes on the idea that "whoever defines the standard for electricity owns electricity." But the fate of electricity tells us a simple truth:

When a technology is important enough, universal enough, and standardized enough, it no longer belongs to any single company. It belongs to infrastructure .

On the surface, the AI competition looks like Google versus OpenAI versus Anthropic, a contest of model capabilities and a battle of fundraising scale. But zoom out, and the real function of this competition is: it is accelerating the push of AI into an infrastructure layer that no single company can monopolize.

When Bill Maris says it is "100% going to happen," he may not just be predicting that Google will cut prices. He may be unconsciously predicting a larger trend—that in the world of AI, tokens will ultimately belong to no one. Just as no one "owns" electricity today.

For OpenAI and Anthropic, this implies something unsettling: even with technological leadership, even after raising astronomical amounts of capital, the future they are chasing—"making big money from AI"—may never have existed in the first place. What they face is not a temporary price war, but a structural destiny—what they are striving to build may, in essence, be the next generation of water, electricity, and highways.

For users, to some extent, this may be good news. Because as long as the token subsidy war continues, people can still enjoy the "good deal" of $20 in cost for $400 worth of computing power.