Who could have imagined that in the AI era, the first thing to explode wouldn't be employee productivity, but the boss's bill?

Last year, Token consumption was a badge of honor for enterprises embracing AI, a sign of innovation. This year, the winds have shifted entirely.

Tencent has begun tightening its AI budget, while Palantir's CEO publicly criticizes model vendors for exorbitant pricing. Before AI replaces employees, bosses are already figuring out how to save on Tokens.

The reason is simple: Tokens are expensive!

Deloitte estimates that a company with annual revenues of around $13 billion could spend up to $700 million annually on AI. And this is just the beginning.

According to Jellyfish, a foreign R&D management platform, AI-related spending per developer has surged 18.6-fold over the past nine months.

But bosses don't spend money without expecting results. So, where's the return?

Jellyfish data reveals that engineers who use the most Tokens are roughly twice as productive as low users but consume 10 times as many Tokens.

In other words, Token consumption is outpacing value creation.

Perhaps, before chasing the AI dream, enterprises should tally the ROI of Tokens.

/ 01 / From 'Token Maximization' to 'Token Discipline'

Last year, 'Tokenmaxxing' was all the rage in Silicon Valley.

The idea was simple: encourage employees to use AI as much as possible. Many companies even used Token consumption as a proxy for AI adoption and innovation.

This year, the narrative has changed. More companies are crunching the numbers on Tokens.

Recently, Nikesh Arora, CEO of Palo Alto Networks, told CNBC that Token prices need to drop by 90% for true large-scale enterprise deployment.

Palantir's Alex Karp was even more blunt. He criticized OpenAI, Anthropic, and others for their Token pricing model, calling it 'fundamentally flawed': companies pay hefty fees but may end up with Tokens that create no value.

According to Karp, if AI helps a company earn an extra $1 billion, model companies should take 30% of that—not charge per Token.

Similar stories abound this year.

A Priceline employee revealed that a routine Cursor contract renewal cost 4-5 times more than before. Uber exhausted its annual AI coding budget by April.

AI spending is spiraling out of control.

Earlier reports from Jiupaicaijing revealed that before June this year, some Tencent departments allocated up to $2,000 per employee per month for AI usage—equivalent to a first-tier city employee's monthly salary. Even after adjustments in June, many employees still had Token budgets ranging from $140 to $700 monthly.

Tencent's '2025 R&D Big Data Report' shows that by the end of 2025, the company and its wholly-owned subsidiaries employed about 87,400 people, with 76% (66,400) in R&D.

Assuming $430 per R&D employee per month, the nominal annual Token budget alone approaches $2.4 billion.

Overall corporate spending is even more staggering.

Deloitte surveyed 548 large enterprises and found that 74% had invested in AI in the past year, with over half allocating 36% of their digital budgets to AI.

Deloitte estimates that a company with $13 billion in annual revenue could spend up to $700 million annually on AI.

But as spending rises, where are the returns?

Jellyfish found that engineers who use the most Tokens are about twice as productive as low users but consume 10 times as many Tokens.

In short, AI investments are far outpacing returns.

Why is AI so expensive?

Agents operate entirely differently from traditional software.

Traditional software follows fixed execution paths with predictable costs. Agents, however, autonomously search, call tools, reflect, retry, and carry context—making their execution highly uncertain.

Many companies only realize how many times a workflow has looped—and how many Tokens it burned—when they receive their cloud bill at month's end.

Last year, a study on Agent programming tasks found that an Agent completing a similar task could consume 1,000 times more Tokens than a typical code Q&A session. Even for the same model and task, Token consumption could vary by up to 30 times across runs.

More importantly, more Tokens don't necessarily mean better results. Studies show that model accuracy typically peaks at moderate Token consumption before entering a stage of diminishing returns.

Thus, Tokens—once an innovation metric—have become a cost center.

/ 02 / Token Efficiency Determines the Winners in Large Models

With Token ROI in the spotlight, model tiering has become inevitable.

While SOTA models like Claude Fable 5 remain irreplaceable for high-value tasks, their 'performance excess' for most ordinary tasks results in low ROI. Users are flocking to cheaper open-source models for these tasks.

This trend is already evident.

Earlier this year, a16z surveyed 100 large enterprises and found that 37% were already using five or more models in 2025. By 2026, 81% of respondents were testing or using three or more model families in production—up from 68% less than a year earlier.

a16z summarized enterprise model usage patterns: leading models handle external, high-visibility, performance-sensitive tasks, while internal, standardized, repetitive tasks are driven by cost.

Currently, closed-source frontier models still dominate enterprise spending.

According to The Information in May this year, among 34 leading AI startups it tracked, OpenAI and Anthropic captured 89% of revenues.

The logic is simple: for cutting-edge applications like programming, users are price-insensitive and willing to pay more for performance.

But not everyone believes this pattern (pattern) will last. AI investor Gavin Baker recently argued:

While most economic value still resides in the smartest models, market share may gradually shift toward cost-effective alternatives.

Others disagree. Gavin Baker, a prominent AI investor, noted:

Although open-source and low-cost models handle most Tokens, most economic value still lies in the smartest models. This may change, with large models' market share shifting from leaders to cost-effective options.

The logic is straightforward: lower model prices mean lower costs per unit of intelligence. The same budget can process more Tokens or enable more business units to adopt AI, boosting ROI.

Ultimately, the winners in the model layer will be those with the highest Token efficiency.

Freda Duan, a partner at Altimeter Capital, put it bluntly: only two types of models will survive—those that are exceptionally good or exceptionally cheap.

The former represents task efficiency; the latter, cost efficiency.

The most likely future scenario is that enterprises will use intelligent routing to precisely split tasks: frontier models handle high-value complex reasoning, while cost-effective models manage massive basic tasks, maximizing output per unit cost.

If the U.S.'s greatest strength today lies in its frontier models, China's advantage increasingly resides in cost efficiency.

To some extent, DeepSeek represents the 'death line' for large models.

A model only slightly better than DeepSeek but much pricier has no long-term justification. For enterprises and developers, paying disproportionate costs for marginal performance gains defies basic business sense.

This tiered mechanism underpins the U.S.-China AI competition. China's AI ecosystem is squeezing overseas counterparts not by directly defeating U.S. giants but through a stealthier 'bottom-up replacement.'

More U.S. AI app companies are quietly switching their underlying infrastructure to Kimi, Qwen, or DeepSeek to cut costs, while maintaining their product experience and workflows upfront. For example, Perplexity's new computer application system uses China's Z.ai-developed GLM 5.2 as one of its underlying models.

Many believe the ultimate goal of large models is to create the smartest AI.

But automotive history suggests another outcome: Ferrari defined speed; Toyota defined industrial efficiency.

The same may apply to AI. While U.S. models push intelligence boundaries, Chinese models are redefining cost floors. As more enterprises calculate Token ROI, the industry's future may hinge not on who is strongest, but who can make intelligence an affordable industrial commodity.

That's precisely where Chinese models have their greatest opportunity.

By Qi