AI is becoming increasingly powerful, yet large language model companies are still not profitable — who exactly is reaping the benefits of this wave of dividends?

In the summer of 2026, more sweltering than the weather is the capital market's complex sentiment toward AI.

Just last week, an internal email from ByteDance's Volcano Engine sent shockwaves through the industry: certain advanced features of the Doubao large model enterprise edition have concluded their free trial period, officially launching commercial billing. As the news broke, some exclaimed "the free lunch of AI is over," while others sneered "finally, it's time for retail users to foot the bill."

Almost simultaneously, the mid-year earnings season kicked off. Flipping through the dozens of published financial reports from major internet giants and listed companies, a highly contrasting picture emerges: on one hand, buzzwords like "AI empowerment," "large model implementation," and "intelligent transformation" have become standard fare in earnings calls, with AI-related Capex (capital expenditure) in corporate IT spending often growing 30%, 50%, or even doubling year-over-year; on the other hand, apart from a handful of established cloud vendors, almost all listed companies focused on AI business are seeing their net profit margins decline, and their loss margins are expanding.

This is rather peculiar. If AI truly lives up to its hype, capable of writing code, designing, handling customer service, and analyzing reports, helping enterprises save substantial labor costs, then those saved funds, combined with the fees enterprises pay for AI services, should have made large model companies rolling in profits. Yet the reality is, whether it's OpenAI across the ocean, domestic players like DeepSeek and Moonshot AI, or even ByteDance's newly monetized Doubao, they all seem to be "losing money to make a splash."

This inevitably prompts a question: in this vigorous AI revolution, since everyone acknowledges that AI is useful, who exactly has pocketed this massive "efficiency dividend"?

To solve this puzzle, we need to step outside the emotional narrative of "AI replacing workers" and dive deep into the books of the entire industry chain.

01

Let's start with a set of data. A leading e-commerce giant mentioned in its Q2 2026 earnings report that its technology investment increased by 42% year-over-year, mainly used to build intelligent customer service and supply chain optimization systems. The CFO's explanation was straightforward: "This is to cope with rising labor costs, and we expect to recoup the investment within the next 18 months."

This "expected return" captures the prevailing mindset of most enterprises launching AI projects today. However, from an accounting perspective, such investments are recorded as costs in the current period, while the returns are deferred. This creates a notable feature in current financial reports: the growth rate of Capex (capital expenditure) far outpaces the short-term revenue growth rate.

This phenomenon is not new in tech history. The widespread adoption of cloud computing in the past and the explosion of mobile internet both went through phases where infrastructure was built ahead of schedule, while application monetization lagged behind. But what makes this AI wave unique is the astonishing speed of "cash burning."

Training a large model easily starts at tens of millions of dollars; maintaining daily inference services, the electricity bills for GPU clusters are already astronomical. When enterprises shout "All-in AI" in their earnings reports, they are essentially placing a high-stakes bet: wagering that the long-term efficiency gains brought by AI will cover the massive current investments. But for vendors providing AI capabilities, the uncertainty of these "future returns" forces them to face an awkward reality—the more they sell, the more they might lose.

ByteDance's Doubao introducing paid services is essentially not about achieving immediate profitability, but about setting up a "rest stop" in this long marathon to ease cash flow pressure. But whether this sip of water can quench the thirst depends entirely on the profit distribution logic of the entire industry chain.

02

To figure out where the money has gone, we can simplify the AI industry chain into the classic "gold rush" model.

The upstream sector consists of the "shovel sellers" who are guaranteed to profit. Whether in the US or China, the most certain winners in this wave of AI are undoubtedly computing power providers. The biggest winner in the global market is undoubtedly NVIDIA. Despite facing various restrictions and competition, its Blackwell architecture chips remain hard currency. As long as large models keep iterating, the demand for computing power will never cease. NVIDIA doesn't just sell chips—it sells "tickets to the game." Its high gross margins are built on extremely high technological barriers. Domestically, Huawei's Ascend series has become the backbone of domestic computing power. As the domestic substitution process accelerates, Ascend chips are seeing growing penetration in key sectors such as government affairs and finance. While the profit margin per individual chip may not match NVIDIA's, it benefits from steady volume growth and policy moats.

For upstream players, regardless of whether downstream miners strike gold, as long as mining continues, shovels must be purchased continuously. They are the first group in this AI dividend to truly put money in their pockets.

The midstream sector is home to the "tool sellers" struggling to survive in a tight spot—these are the large model vendors we care most about, including OpenAI, Anthropic, as well as domestic players like DeepSeek, MiniMax, and Moonshot AI.

Their situation is the most awkward. On one hand, they have to pay exorbitant computing power rents or procurement fees to upstream suppliers; on the other hand, to capture market share, they are forced into brutal price wars.

ByteDance's Doubao offered free services for a long time before, DeepSeek positioned itself with "extreme cost-effectiveness," Kimi competed through differentiated long-text capabilities... Behind all these moves is a precipitous decline in API call prices. In certain vertical sectors, the price per million tokens has even dropped below 1 RMB.

This leads to a result: midstream vendors have turned into "transporters," packaging expensive computing power into cheap tokens to sell downstream. The essence of this business model is using high-cost resources to generate low-gross-margin revenue. Without achieving absolute market dominance, it's extremely difficult to see light at the end of the profitability tunnel.

The downstream sector is made up of mixed-blessing "gold miners"—enterprise users leveraging AI across all industries. Large corporations like ByteDance, Baidu, and Alibaba are both AI users and providers. For these giants, AI's value lies in optimizing internal processes and reducing operational costs, and they have truly reaped the rewards of "cost reduction and efficiency improvement."

Small and medium-sized enterprises (SMEs) are the biggest beneficiaries of this wave of AI inclusivity. Previously unable to afford programmers for development, they can now build simple applications by spending a few hundred yuan on API calls. However, most SMEs haven't secured excess profits from this, because the efficiency gains brought by AI are quickly eroded by market competition—their competitors are also using AI.

As a result, downstream users do save money, but this saving is more of a "defensive gain" rather than an "offensive gain."

03

This brings us to a core business paradox:

If AI truly helps an enterprise save 1 million yuan, why is the enterprise unwilling to pay 100,000 yuan to large model companies?

The answer lies in the extreme imbalance of value distribution.

The exorbitant "shared overhead costs" have become the last straw weighing down downstream enterprises. The cost structure of large models is characterized by "high fixed costs and low marginal costs." Training a model once can burn tens of millions, which is a sunk cost. To make the model perform better, continuous RLHF (Reinforcement Learning from Human Feedback) is required, which translates to endless ongoing expenses. Yet the API fees paid by downstream enterprises often only cover the electricity and operation costs during the inference phase, and are nowhere near enough to amortize the massive upfront training and R&D costs. It's like staying in a hotel where your room rate only covers utility bills, not the mortgage.

As open-source models advance, the moat for closed-source models is narrowing. When the capability gap between models shrinks to a certain point, they turn into utility products similar to "tap water" or "electricity." For utilities, users will only accept extremely low prices. Midstream vendors are trapped in the "innovator's dilemma": if they don't innovate, they get eliminated; if they innovate too fast, the cost of educating the market becomes too high, and their innovations are easily replicated.

The cost savings AI brings to enterprises are primarily reflected in labor. But on financial statements, these savings show up as reductions in "administrative expenses" or "sales expenses," rather than increases in "operating revenue." Business owners save money, but that money never flows to large model vendors—it becomes part of the enterprise's net profit. Large model vendors are only helping enterprises "cut costs," but not participating in sharing the gains from "revenue growth."

ByteDance's Doubao launching paid services is precisely an attempt to break this paradox. It wants to tell the market that high-quality tools have their price. But whether this can reverse the overall loss situation in the midstream depends on whether users are willing to pay a premium for "great performance," instead of only paying the rock-bottom price for "basic usability."

04

Looking ahead over the next 12 months, I believe the large model industry will bid farewell to "unrestrained growth" and enter the second half of "refined operation." Three trends will become increasingly prominent:

The first shift is from "competing on parameter size" to "competing on user retention." In the past, vendors loved to show off rankings and parameters, competing over who had more parameters and higher benchmark scores. But enterprise users have grown numb to this. Going forward, the competition will focus on user retention rates and renewal rates. Can the model operate stably? Does it produce many hallucinations (fabricated content)? Can it seamlessly integrate into existing workflows? These are the critical factors that determine survival. Models that only perform well on benchmarks but cannot be practically implemented will be eliminated.

The second shift is from "public cloud" to "private deployment." For industries with extremely high data security sensitivity such as finance, government affairs, and healthcare, public cloud APIs cannot meet their requirements. Over the next year, we will see a flood of private deployment orders. For midstream vendors, this presents an opportunity to raise average order values. Although implementation is resource-heavy and has long cycles, once a contract is signed, it guarantees long-term cash flow. This is also a competitive stronghold for major players like Huawei Cloud and Tencent Cloud.

The third trend is that "Agents" will become a breakthrough point for monetization. Pure conversational AI (Chatbots) is hard to charge for, as it's far too easy to be replaced. But if AI evolves into an Agent that can automatically execute complex task chains (such as automatically booking tickets, generating financial reports and sending emails, maintaining code repositories), its value will no longer be limited to being a "Q&A tool"—it will become a "digital employee." At that point, Pay-for-Results will become feasible. This is also the key direction where ByteDance's Doubao will focus its efforts after launching paid services.

Currently, this wave of dividends has mostly settled in the hands of computing power infrastructure providers (upstream) and a subset of giants that have dramatically optimized internal efficiency through AI (downstream).

Large model vendors (midstream) are currently in the "darkness before dawn." ByteDance's Doubao launching paid services is a signal that the industry is attempting to rebuild a healthy business closed loop. But this is destined to be a long-term battle.

For ordinary people like us, instead of worrying about whether AI will take our jobs, it's better to think about how to leverage AI to become an "irreplaceable digital artisan." After all, in a gold rush, the people who ultimately get rich aren't necessarily miners, nor necessarily shovel sellers—but they are definitely the ones who know how to use tools to create value.

This dividend ultimately belongs to those who are more efficient.

This article is from the WeChat public account "Xin Xiang", authored by Chao Hai, and published by 36Kr with authorization.