Another factor contributing to the weakening of the AI market: OpenAI also has to lower prices

TL;DR

· Uber executives directly said that the link between token consumption and real product improvements "does not yet exist"; OpenAI also admits that enterprise AI costs are becoming an increasingly tangible issue.

· AI demand has not disappeared, but companies are shifting from trial use to ROI audits, model providers are discussing price reductions, and the growth elasticity of cloud, GPU, and data centers needs to be re-verified.

· Related stocks: NVDA, MSFT, AMZN, GOOG, MU, AVGO, AMD, TSM, ARM, ORCL.

The AI stock market, which has been rising for two consecutive months, has recently started to weaken and pull back, and the market is beginning to look for reasons openly.

Interest rates, valuation congestion, and earnings disruptions can all explain this correction, but the market is also auditing a deeper assumption: whether increased enterprise token consumption necessarily leads to more revenue, efficiency, and profit.

Over the past two years, AI trading has followed a very smooth chain. Enterprises heavily use AI, token (the unit of measurement for model text processing) consumption rises, model vendors’ revenue grows, cloud providers sell more computing power, and demand for GPUs, HBM (high-bandwidth memory), servers, data centers, and electricity continues to expand. As long as token usage continues to grow, the market can interpret it as accelerated AI adoption and assign higher valuations to upstream hardware and capital expenditures.

But a recent change is that even model providers themselves are starting to discuss cost issues.

According to The Wall Street Journal, OpenAI is researching further price reductions for model calls to cope with enterprise budget pressures and competition from rivals like Anthropic. Meanwhile, OpenAI CEO Sam Altman recently publicly stated that more and more companies are viewing AI costs as an important issue, with some clients exhausting their planned AI budgets for the entire year as early as the first quarter.

This alone may not be enough to change the industry landscape, but it signals an important point: the market is shifting focus from model capabilities to costs, pricing, and return on investment.

The current looseness is not about "whether enterprises still use AI," but about "whether enterprises are willing to continue unconditionally paying for high-priced tokens."

Uber President and COO Andrew Macdonald said in a podcast that the link between token consumption growth and "useful consumer features" "does not yet exist." This statement comes from the buyers, not the sellers, investment banks, or model startups.

If the market previously believed "usage equals success," we are now entering a second phase: whether tokens can ultimately translate into revenue growth, reduced labor costs, or improved profit margins. Once this question is systematically raised by finance departments, valuation language in the AI industry chain will shift from "demand is unlimited" to "return verification."

Uber’s high adoption rate exposes budget pressures

Uber’s case is worth examining, not because it doesn’t understand AI, nor because it’s unwilling to use AI. On the contrary, Uber’s internal adoption rate of AI coding tools is very high. According to multiple media reports, among about 5,000 engineers, the monthly usage rate once reached 84% to 95%, with individual engineers’ monthly bills ranging from a few hundred dollars to $2,000.

The problem lies right here. When usage is high enough, bills are no longer just small experimental costs for innovation departments but become real costs that need to be explained at the operational level. As the company’s CTO previously disclosed, Uber’s annual Claude Code budget was exhausted within four months. Macdonald described this as a "mind-blowing" moment.

Within enterprises, AI tools initially entered budgets under the guise of "improving efficiency." Engineers generate code faster, customer service answers questions more quickly, operations teams write reports faster—all perceptible changes.

But as usage scales up, finance departments will look at more concrete questions: Does it bring in more revenue? Does it reduce actual human labor costs? Does it improve profit margins?

The "token maxxing" phenomenon Macdonald mentioned also indicates that high usage may be decoupled from high value. Token maxxing refers to teams or individuals consuming large amounts of tokens to maximize AI tool usage. The usage data looks good, but it may not correspond to better product results. For AI service providers, this is income; for companies, it might just be another runaway cloud bill.

Uber’s signal is more significant than the general "AI tools are too expensive."

It’s not saying AI is useless, but that when AI moves from trial budgets into operational budgets, companies need to prove that every dollar spent on tokens can generate measurable business results. High adoption no longer automatically equals success; it will instead reveal cost structures first.

Cost pressures are beginning to propagate up the industry chain

Enterprise buyers are starting to do the math, and platforms are also changing their charging methods.

GitHub announced that starting June 1, 2026, Copilot will switch to a usage-based billing model, introducing monthly AI Credits (AI usage quotas). For light users, this may just be a change in billing structure; for frequent users of intelligent agent features, some heavy users report that single-session costs can reach dozens of dollars, sparking community discussions.

This indicates that platforms are no longer willing to cover unlimited token costs in fixed subscription fees.

In the past, users paid a monthly fee, and the platform bore the underlying model call costs. Now, as the number of intelligent agent calls, context length, and multi-turn tasks increase, cost pressures are becoming explicit. The more you use, the more you pay—this is a correction to the "unlimited AI" narrative.

More importantly, this pressure has already propagated from the application layer to the model layer.

Over the past two years, the mainstream narrative in the large model industry has been cost reduction, efficiency improvement, and scale expansion. But as enterprise procurement departments begin to audit ROI, model providers face new questions: if clients are unwilling to continue paying high prices for tokens, how will growth be sustained?

OpenAI’s recent signals are quite typical. On one hand, Sam Altman admits that enterprise budgets are under pressure; on the other hand, market reports suggest OpenAI is researching further price cuts. This indicates that industry focus is shifting from "whether the model is leading" to "whether the unit cost of intelligence is low enough."

For enterprise clients, the most important question is no longer which model is the strongest, but which model can generate more business results within the same budget.

Microsoft’s internal cutback on Claude Code licensing also points in the same direction. According to reports from The Verge, Axios, TechRadar, Microsoft’s Experiences & Devices division has canceled most internal Claude Code licenses and shifted to its own Copilot tools. The scale and reasons are still to be fully disclosed, so it cannot be directly concluded that Microsoft has confirmed cost-driven cuts in external tool procurement.

But this move at least shows that large tech companies are reallocating external model call costs.

The impact on the AI industry chain is not about how much revenue a particular tool loses, but about the buyer discipline starting to propagate upward. Companies can limit quotas, choose cheaper models, shift some tasks to open source or self-built solutions, or ask vendors for discounts. Model vendors and application-layer companies will still have demand, but pricing power is no longer solely determined by "more powerful models" but also by "whether customers can justify the costs."

Cloud providers will also be affected. In the past, cloud revenue from AI was strongly associated with model training, inference, and enterprise applications—more token usage meant more demand for cloud services. But if companies start lowering the unit token cost or shifting high-frequency, low-value tasks to cheaper inference paths, cloud revenue elasticity may fall short of previous expectations.

High usage must prove high value

At this point, companies are starting to audit partly because AI usage has entered a sufficiently large sample stage, and inefficiencies are no longer easy to ignore.

Entelligence.AI’s May 2026 study analyzed 2,444 organizations and over 1 million pull requests. According to their estimates, only $0.18 of every $1 spent on AI tokens actually generated real value reaching users; $0.44 was spent fixing bugs introduced by AI; $0.27 was rework; and $0.11 was spent on review friction.

This data cannot be generalized to the entire industry. It comes from the vendor’s own research page and mainly reflects software engineering scenarios, not independent audits or academic papers. But it is enough to illustrate a key point: enterprises are indeed under ROI audit pressure, especially in scenarios where AI-generated content still requires human review, correction, and integration.

The most obvious benefit of AI tools is faster generation, but what companies really pay for is deliverable results. If AI-generated code causes more bugs, requiring more review, rework, and testing, the time saved on the front end will reappear on the back end. For individual users, this may just be an experience issue; for large enterprises, it becomes a financial and organizational management issue.

This also explains why growth in token usage can no longer be simply equated with AI success.

Tokens are both revenue billing units and cost measurement units. For model vendors, more tokens mean more revenue; for enterprises, more tokens only justify continued budgets if they lead to more revenue, lower costs, or higher profit margins.

If the market previously regarded token growth as a leading indicator of hardware demand, it now needs to add another half: the conversion rate of token value. Only when token consumption can reliably translate into business results will cloud AI revenue, GPU orders, HBM expansion, and data center construction have more solid end-user support.

Willingness to pay propagates up the industry chain

Macro strategist Andreas Steno Larsen recently pointed out that the Silicon Data-related LLM Token Expenditure Index is one of the key charts to watch in the current market. Reports show that this index tracks enterprise spending or pricing levels for every million tokens, and after rising sharply in early 2026, it showed signs of decline near the end of May.

It’s important to keep boundaries here. Silicon Data’s public page mainly provides product descriptions; the index methodology and full historical data are not fully disclosed. It cannot be taken as a hard conclusion but as a signal to observe changes in enterprise willingness to pay.

A decline in token expenditure does not necessarily mean a decrease in AI usage.

In fact, the current market looks more like witnessing a shift from "computing power competition" to "unit intelligence cost competition." Enterprises still need AI but may not be willing to continue purchasing AI at previous prices.

If OpenAI finally initiates a new round of price adjustments, it will ease enterprise budget pressures and also mark the industry’s entry into a price competition phase. The market will then need to reassess: will future growth come from new demand or from usage expansion after price reductions?

AI demand may still grow, but the revenue quality and upstream elasticity of transmission may change.

The impact varies across different segments. Application and model layers will face initial price pressures: companies will demand clearer ROI, reduce low-value calls, or switch costs between models.

Cloud providers face revenue elasticity issues: with the same usage, if unit prices fall, cache and batch processing improve, or self-built solutions increase, cloud AI revenue growth may not be as strong as token volume growth.

Upstream, GPU, HBM, advanced packaging, servers, and data center investments are about future capital expenditures. If enterprise payment discipline makes model vendors and cloud providers more cautious about future revenue, hardware orders and data center construction will be re-evaluated.

Larsen’s warning is not that hardware demand will immediately disappear, but that if token pricing continues to weaken, the market will start to doubt the slope of this AI infrastructure investment cycle.

The decline in AI stocks and token billing audits are not simply cause-and-effect. You can’t say chip stocks fell because Uber burned through its budget, but they are on the same chain: when valuations already reflect long-term high growth, any signals about terminal willingness to pay and ROI will be amplified into reassessments of upstream capital spending.

Next, look at revenue elasticity and order momentum in earnings reports

Current evidence does not support the idea that "the AI bubble has burst." Enterprises are not stopping AI use, and developers will not revert to a state without Copilot, Claude, or other intelligent tools. A more reasonable judgment is that AI adoption is transitioning from early enthusiasm into a budget discipline phase, where the market begins to distinguish which use cases can prove returns and which are just billings.

The most important validation moving forward is not just finding another company saying AI is too expensive, but whether the language in earnings reports from cloud and software companies is changing. Can Microsoft, Amazon, and Google’s AI cloud revenue growth continue to be highly elastic? How do renewal, downgrade, and complaint rates for enterprise tools like Copilot and Claude Code change after usage-based billing? These will be more telling than daily stock prices about whether buyer discipline is systematically strengthening.

On the hardware side, watch for signs of downward revisions in GPU, HBM, and data center orders. As long as cloud providers continue to increase capital spending and advanced chip orders remain tight, the willingness to pay for tokens will look more like a healthy adjustment. If cloud AI revenue elasticity weakens and upstream orders and data center construction slow down, the market will price it as a deeper cycle turning point.

The AI trading cycle is not over, but its pricing language is changing. Previously, the market asked "how many tokens were used," now it asks "how much profit did these tokens ultimately generate." This difference will determine the direction of valuation divergence across the AI industry chain in the future.

Click to learn about the job openings at Rhythm BlockBeats

Join the Rhythm BlockBeats official community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia