The AI Industry’s Money In, Money Out and Money Made — A Cross-Show Cyber Dialogue
Summary
- Doubao’s decision to charge is not just about helping ByteDance recoup its investment; it is also testing the price point for China’s AI subscription market. 庄明浩 believes Doubao has built a clear lead in China’s chatbot market, and that ByteDance will not let lower margins or higher CapEx shake its commitment to AI; 阑夕 emphasizes that if Doubao stays free for the long term, Qwen, Yuanbao and other products will be even less willing to charge: “If you don’t charge and I don’t charge, how is the meme officer supposed to charge—and how does anyone make progress?”
- Subscriptions can cover inference, but for now they cannot cover the arms race of continuously upgrading models. Generative AI has to recompute every answer, while monthly fees are priced around average usage; the true cost of heavy users can far exceed their subscription fees, forcing products to keep adding quotas, throttling, and add-on packages. 阑夕 compares it to a gym: “Bet you won’t show up.” 庄明浩’s summary is that model-inference gross margins may reach 50%-60%, but training the next generation will still swallow those profits.
- Doubao’s RMB68, RMB200 and RMB500 tiers are probing both Chinese consumers’ willingness to pay and the value of AI as a productivity tool. RMB68 is clearly above the traditional “$20 overseas, RMB20 domestic” conversion, but still below RMB98 or RMB128; if the offering is limited to chat, the price may be hard to justify. More likely, it will bundle high-quality image, video and voice capabilities, or reserve the latest models for paid users. 庄明浩 and 高飞 both identify Qwen as the next player likely to start charging.
- AI investment has moved beyond corporate budgets to the physical limits of power, materials, production lines—and the planet’s resources. The program’s 2026 estimate is roughly $200B in Amazon CapEx, nearly $200B at Google, and close to $200B at Meta, with 4-5 giants together approaching $1T; 庄明浩 describes it as “how we got here without even noticing.” China may still have more than 10 model and multimodal players because market growth, financing liquidity and IPO windows temporarily make “you don’t have to leave the table as long as you don’t concede” a viable strategy.
- Profits are currently concentrated upstream, in the businesses that service machines; the closer a business gets to human users, the harder it becomes. Power, chips and Nvidia-style infrastructure are closer to “reliable in good times and bad,” while new cloud providers may have revenue but also carry heavy leverage. Application companies must pay both model and cloud bills while absorbing marketing, customer acquisition and payment friction. 高飞’s blunt formulation: “Everything that serves carbon-based life loses money; everything that serves silicon and machines makes money.”
- Subscriptions and APIs will not displace each other; they will settle into different mixes shaped by the US-China industry structures. China is stronger in To C, so subscriptions and telco bundles may gain traction first; the US already has mature To B, SaaS and API channels, and an application boom would amplify usage revenue. The 1-year forecasts from the 3 speakers range from a 40/60 subscription/API split in China to 50/50; 阑夕 cites Gemini data showing roughly 107T tokens per day in the main app versus 27T through the API, underscoring how a company’s capabilities shape its revenue mix.
- China’s real potential for differentiation may lie not just in cheaper models, but in reliable supply, emotional value and rapid hardware iteration. 庄明浩 sees the US as more focused on peak performance and returns, while China could fold Doubao’s personas, Qwen shopping and local services, toy companionship, learning devices and pet hardware into commercial loops, building products that are “beautiful and useless.” Combined with the power grid, supply chain and low-cost prototyping capabilities, AI hardware could follow a distinctly Chinese path.
Deep dive
1. Doubao’s Charging Decision Is Opening the Pricing Gate for the Entire Market
庄明浩 was not surprised: without DeepSeek’s breakout in early 2025, China’s chatbot war might already have been considered a Doubao victory by late 2024. At its current scale, Doubao is the player best positioned to test whether Chinese users will pay for general-purpose AI.
This is not about ByteDance being unable to afford free service. 庄明浩 believes AI remains a decision ByteDance can make “without any hesitation”; lower margins and higher CapEx will not change its strategic commitment. Charging is first and foremost a business-model experiment, not a cash crunch.
阑夕 takes the more aggressive view: Doubao “should charge.” As long as the leading product remains free, Qwen, Yuanbao and other followers risk losing users the moment they start charging, so Doubao has to set the precedent: “If you don’t charge and I don’t charge, how is the meme officer supposed to charge—and how does anyone make progress?”
2. Generative Inference Invalidates the Old Internet Rule That More Users Mean Lower Costs
Traditional internet products can see marginal costs fall as scale grows. Generative models must recompute every request: “Every computation is a fresh computation, not copy-paste.” Each additional service interaction creates real Token and compute costs.
A monthly fee is essentially an average. Providers estimate how much a normal user consumes per month, then set a price of $10, $20, $100 or $200. Once heavy users run far above the average, costs exceed the price, leaving providers to recalibrate with quotas, throttling, add-on packages and reset cycles.
阑夕 explains subscriptions through the gym analogy: “Bet you won’t show up.” After OpenClaw took off, heavy users were able to burn through an entire subscription allowance, prompting Anthropic to restrict related usage. This was not simply a clash between competitors; it showed that subscription economics cannot withstand everyone operating at full load.
The key financial divide is training versus inference. Excluding pretraining, serving existing models can carry positive gross margins; but providers have no choice but to keep training the next generation. The program cited positive-cash-flow expectations around 2028 and 2030, respectively, meaning revenue must continue to outrun rising R&D spending.
3. RMB68 Could Be a Psychological Anchor
庄明浩’s map of overseas pricing is roughly as follows: about $10 gets users into a basic subscription, professionals may jump directly to $100, and $200 is a high-end personal tier. A direct conversion would put China’s monthly price at RMB1,500, which is unrealistic; the gaming industry’s RMB648 may be a more effective upper-bound reference.
The Doubao tiers discussed on the program are RMB68, RMB200 and RMB500 per month. An RMB998 tier also appeared to be listed on the page initially circulated, but seems to have been removed later. 阑夕 sees RMB68 as particularly calibrated: it is neither the customary RMB20 nor as high as RMB98 or RMB128.
If AI is treated as entertainment spending, 阑夕 cites QQ Music’s roughly 120M paying users as the ceiling for a single domestic product and argues that RMB68 would be difficult to push much beyond that base. If AI is positioned as a productivity tool, the price looks cheap. The real question is therefore: “What exactly is packed into those RMB68?”
The combination that might make the price work is high-quality image, video and voice capabilities, or delayed and quota-limited access to new models on the free tier. Creating a separate product is less effective than using version and quota differences to segment paid users naturally.
4. Qwen May Follow on Pricing, but Revenue Recapture Does Not Mean the Burn Stops
庄明浩 expects Qwen to start charging sooner; 高飞 says Qwen was the player he had intended to mention as well. Alibaba’s e-commerce margins are relatively thin, while the company is also funding cloud, infrastructure and model development. Its position as “the most aggressive” AI investor also means it has the strongest incentive to bring revenue in sooner.
If the market accepts Doubao’s pricing, competition will shift from “who dares to stay free” to “whose offering is worth the price.” 庄明浩 describes the move as recouping some of the money, not a strategic retreat: training, cloud and product investment will continue.
Tencent illustrates a different risk from relying on external models. 阑夕 relays Pony Ma’s metaphor: “We thought we had a ticket onto the ship, only to find that the ship was leaking.” Yuanbao relies heavily on DeepSeek, while the industry upgrades every 2-3 months; after DeepSeek moved from V3.0 to 3.1 without a major-version update, its user experience could begin to lag.
5. Paid Tiers Will Create an AI Divide, but Most People Do Not Need the Frontier
高飞 estimates that even if everyone eventually uses AI, only 10% or fewer may actually pay. People who use Claude Code and Opus 4.7 continuously already have a materially different understanding of the model frontier from users who only chat and search inside Doubao, Yuanbao or Qwen.
That gap will become a new “AI divide,” but demand is not uniform. Productivity users want to probe the limits; for everyday questions, last year’s GPT-4 may already be sufficient. For most people, the AI currently on their phone is already “the smartest thing they have encountered in their lives.”
A free version therefore does not have to be a crippled product, and a paid version cannot rely only on the words “smarter.” Providers need to identify which users, in which scenarios, require what level of quality and response speed before they will consider a recurring subscription worthwhile.
6. Advertising and Telco Bundles Offer Two Indirect Ways to Get Paid
庄明浩 believes AI advertising is difficult to plug directly into existing systems. Traditional ads are matched to pages or search keywords, while conversational ads must connect user inputs to products, bringing data de-identification and regulatory pressure. Once those obstacles are solved, China’s talent for making one party pay for another’s benefit could allow it to move faster than the US.
Telcos could turn Tokens into a third billing unit after voice and data. A mobile carrier could bundle Doubao’s RMB68 membership into a phone plan, lowering the psychological barrier because users pay through their bill. The carrier gains distribution rights over individual users, while the model provider must weigh incremental channel reach against the risk of being disintermediated.
7. Free Tokens Can Become Public Infrastructure, but Quantity Cannot Substitute for Quality
The program mentioned that Malta allegedly plans to provide free Tokens to “everyone,” though “everyone” needs to be taken with quotation marks and it may only be a benefit for one city. Tsinghua and other schools are also providing free Tokens to students, while companies can reimburse employee subscriptions. These are potential institutional experiments that could be replicated.
庄明浩 says the standard advice when companies discuss AI transformation is often to “reimburse all employee AI usage.” He cites 郑青’s analogy: “AI is basketball.” Watching 100 basketball tutorials will not teach someone to dribble; they first need a ball, a court and time to handle it themselves. If an organization truly treats AI as a core capability, it should first lower the barrier to actual use.
阑夕’s counterpoint is that “universal access, profitability and quality are an impossible trinity.” Without specifying which model sits behind the Tokens, talking about giving away 2B Tokens means nothing. Unlimited access to high-quality models has already become a recruiting perk for startups.
Universities face a particular contradiction. They may spend tens of millions of RMB each year on CNKI papers and other resources, while models could be at least as useful for research; at the same time, professors are sick of reading AI-generated papers. Eric Schmidt, Google’s former chairman, was booed by students after mentioning the AI race, but that cannot change the reality that the technology wave is impossible to stop.
8. AI CapEx Has Hit the Physical Constraints of Industry and the Planet
庄明浩’s 2026 figures are roughly $200B of CapEx at Amazon, nearly $200B at Google and close to $200B at Meta. The combined total for 4-5 giants is approaching $1T, a scale larger than the GDP of most countries.
That money ultimately has to become GPUs, storage, CPUs, data centers, power-generation equipment and raw materials. The question is no longer just whether companies are willing to invest, but whether the planet has enough resources, suppliers can ramp production lines in time, and the grid can carry the incremental load.
Once data centers are being envisioned in space, linear extrapolation is no longer reliable. 庄明浩’s central reaction is: “We got here without even noticing.” AI is pushing several industries at once toward limits humans have rarely approached before.
9. China’s Model Table Is Crowded Because the Funding Loop Has Not Broken
AI companies currently have 3 ways to survive: Google and others are funded by mature core businesses; OpenAI and Anthropic rely on massive financing; Kimi is trying to charge users while continuing to raise money from capital markets. Whichever path they take, they still have to answer the same question: where does the money come from?
阑夕 compares the cycle to a small-town debt loop: a traveler pays a hotel, the hotel pays the butcher, and the butcher pays someone else. The money does not increase, but the economy revives because circulation resumes. Certainty and liquidity remain high in AI: “As long as we spend a little more money, the circle can keep turning.”
Overseas, there is talk of “the 4 closed-source players.” 阑夕 lists Google DeepMind, Anthropic and OpenAI, then asks who the fourth is, arguing that US players are moving so fast that falling behind once may mean losing access to resources. Meta was particularly damaged by the Llama 4 episode. Competition in China is comparatively less intense; after the DeepSeek shock, Doubao still had time to review, regroup and catch up.
10. Back-Rank Players Can Still Raise Money Because the IPO Revenue Bar Is Lower Than Expected
Video models are the clearest example. After ByteDance released Seedance 2.0 and was seen as having opened up a decisive lead, Kuaishou’s Kling, Alibaba and MiniMax Hailuo still did not leave the field. Behind them, HiDream, Vidu, PixVerse and other players have raised financing in the billions of RMB.
One realistic reference point for investors is the IPO revenue of Zhipu and MiniMax: roughly RMB500M-RMB700M annually, or about $100M. If companies such as StepFun can reach around RMB500M, and other players reach around RMB300M, they may conclude that they too qualify for a listing.
In a rapidly expanding market, trailing the leader does not mean having no value—“there’s enough soup to go around.” AI companies with a meaningful user base may also exit through M&A. 高飞 sums up the possible outcome this way: “These companies dying just means they stay with us in another form.”
11. Subscriptions Monetize the Average, APIs Monetize Actual Calls—and User Behavior Adds Training Data
Heavy subscribers continue to be served because the many people who buy access but rarely use it help spread the cost. Providers dynamically adjust caps, throttling and quota resets; 阑夕 believes there must be a powerful financial model running in the backend: “They are not going to run a loss-making business.”
High-frequency users’ Prompts are not pure cost either. 高飞 argues that their distinctive image-making methods, writing approaches and workflows could become high-quality post-training data, allowing model companies to discover use cases they had not previously imagined.
庄明浩 sees China as stronger in To C. If individual payments work, subscriptions could account for a larger share. The US has mature To B, SaaS and API infrastructure; once enterprise applications take off, every application will run on top of a model through an API, potentially producing more usage revenue.
阑夕 cites Google I/O data: Gemini has roughly 900M monthly active users, with about 107T Tokens per day in the main app versus 27T through the API. The 1-year forecasts from the 3 speakers range from a 40/60 subscription/API split in China to 50/50, with 高飞 and 庄明浩 both favoring the latter.
12. Profits in the Five-Layer Cake Are Systematically Moving Upstream
The 5 layers identified by Jensen Huang are power, chips, infrastructure, models and applications. 庄明浩 believes that over the past 3-plus years, the only consistently profitable layer has been the “shovels” layer represented by Nvidia, while the upstream layers remain “reliable in good times and bad.”
Data centers could make money in principle, but new cloud providers borrowed heavily to expand for a demand surge. Current revenue may not even cover future interest expense, leaving both their financial statements and leverage looking ugly. They may still be profitable once the business matures, but gross margins will likely fall toward the average range of traditional cloud providers.
Model inference gross margins may reach 50%-60%, but providers must keep funding training. Application companies pay both model and cloud providers, then absorb their own operating, marketing, traffic-buying and customer-acquisition costs. 阑夕’s summary: the first few layers are “if you have inventory, someone will buy it,” while the application layer faces the problem that “there aren’t enough people to foot the bill.”
13. Falling Token Prices Have Not Automatically Rescued Applications Because Everyone Chases the Latest Model
Application founders often project future high margins by assuming “Token costs are falling at 10x speed.” 庄明浩 points out that they may still lose money next year: nobody wants to use an old model, while the price of the latest model will not fall in tandem and may even continue rising.
高飞 translates the profit distribution into a “machine economy.” AI needs power, silicon, servers and optical communications, and every one of those links makes money. Applications ultimately serve humans, yet are the hardest to charge for: “Everything that serves carbon-based life loses money; everything that serves silicon and machines makes money.”
阑夕 remains “cautiously optimistic.” Upstream providers cannot pass pressure down to applications forever. The real signal is a clear payer and a clear ROI, not a large call count; otherwise, the industry’s profit pool will never close.
14. Model Profitability Is the Industry Thermometer; the Next Challenge Is Customer ROI
The program’s estimates are that OpenAI could reach break-even around 2029, Anthropic sooner, and Midjourney could approach break-even the following year. The dates may shift, but the common condition is that revenue growth must remain faster than training investment.
Once the Claude Code battlefield opened, model revenue began accelerating. Zhipu and MiniMax generated roughly RMB500M-RMB700M in revenue last year, then reached RMB200M-RMB300M or more in a single quarter. That made “exponential growth” and a calculable path to profitability seem less distant.
The market’s questions are moving downstream. It used to ask whether the models could make money; now it asks what enterprises are actually producing after consuming so many Tokens and writing so much code, and whether the ROI works. 高飞 believes the new question itself shows that people are no longer as worried about the earlier question of model profitability.
阑夕 offers a “politically incorrect” test: “Just look at layoffs. Layoffs mean there is value,” because they indicate that AI has replaced human labor. He acknowledges that the formulation is somewhat politically incorrect.
15. China May Find a Different Commercial Path Through Emotional Value and Hardware Iteration
庄明浩 believes US commercialization will pursue peak returns more pragmatically, while China may have more room for imagination in products that are “beautiful and useless.” Doubao could enter through personas and companionship, letting users customize characters such as an older-sister archetype rather than selling efficiency alone.
庄明浩 also points to Qwen’s integration with shopping and ride-hailing. Even if the commercial value remains debatable, it demonstrates a different kind of malleability; when users encounter certain “IQ tax” products, Qwen may even tell them not to buy. A single transaction may not be profitable, but it could strengthen overall trust and ecosystem value.
庄明浩 argues that China’s grid prioritizes stable supply, while the US private-power system could see new data centers drive up city electricity prices and trigger voter backlash. Chinese chips may lag by a generation, but cost advantages, resource allocation and administrative coordination could ease the constraints.
高飞 is more bullish on the hardware variable. A mature supply chain can turn almost any idea into a prototype and then a product quickly and cheaply, while AI gives “dumb hardware” a smart brain. Camera earbuds, companion toys, learning devices and pet AI may mostly fail, but repeated iteration could still take China down a new hardware path.