Note-taker's Notes:
The AI circle is very lively recently.
Anthropic has become the fastest-growing company in human history, with annualized revenue soaring from $10 billion at the end of 2024 to $47 billion in May 2026. Yesterday, it just completed its H round of financing of $65 billion, with a post-investment valuation of $965 billion, surpassing OpenAI to become the world's highest-valued AI company, and is expected to launch an IPO this fall.
DeepSeek's valuation has surged to $45 billion, with the National Integrated Circuit Industry Investment Fund leading the first round of financing of 70 billion yuan (about $10 billion), and the list has basically been finalized.
Kimi completed a $2 billion financing round, with a post-investment valuation exceeding $20 billion. It has raised over $3.9 billion cumulatively in half a year, becoming the financing champion among domestic large-scale model startups.
StepFun (Jieyue Xingchen) completed nearly $2.5 billion in financing, dismantling its red-chip structure to sprint for a Hong Kong IPO.
ByteDance adjusted its AI infrastructure investment for 2026 from 160 billion yuan upwards to over 200 billion yuan. Bloomberg further exposed that its total capital expenditure ceiling might reach $70 billion (about 500 billion yuan).
Doubao launched three tiers of paid subscriptions on May 4th, firing the first shot to end the era of free domestic AI.
If we follow the four seasons, what season is today's AI economy in? Is it spring, summer, or the transition between spring and summer? Or is it about to enter autumn, like the bubble period rumored outside?
The answer actually lies within cycles.
Today, we will use a six-dimensional scoring framework for cycle judgment proposed by Professor Su Dechao from the School of Philosophy at Wuhan University, founding advisor of Notesman PPE Academy, and instructor of Western Philosophy courses, to clarify this matter thoroughly.
一、Quantitative Scoring of Six Dimensions for Cycle Judgment
Many people like to use the elimination method to judge industry cycles: it's not winter, spring has long passed, autumn doesn't seem to have arrived yet, and finally they arrive at a correct but useless statement: summer. The elimination method only gives you a seemingly correct answer but doesn't tell you why it's correct.
Truly useful judgment must be quantified from specific dimensions.
The "Six-Dimensional Scoring Framework for Cycle Judgment" scores an industry according to six dimensions: "Narrative vs. Delivery, System Connectivity, Delivery Capability, ROI (Return on Investment) Rationalization, Industry Common Phenomena, and Capital Environment." Each dimension scores from 0 to 2 points, with a higher total score indicating closer proximity to autumn.
Let's score them one by one.
1. Narrative vs. Delivery: From Storytelling to Looking at the Ledger
This is the first dimension and also the easiest one to perceive changes.
When ChatGPT first emerged in 2022, everyone said, "AI will change everything." But no one asked you what you could specifically do or how much cost you could save.
Storytelling was enough; that was spring.
The situation today is different. Doubao introduced paid subscriptions, with three tiers of prices clearly written on the page—68 yuan, 200 yuan, 500 yuan—focusing on paid features like PPT generation, data analysis, and video production. It's not storytelling; it's delivering specific capabilities and then charging for them.
OpenAI's advertising platform is even more direct: advertisers buy ad space within ChatGPT and pay per click. On May 5th, the self-service ad management tool launched testing, removing the $50,000 minimum spending threshold, allowing even SMEs to advertise directly.
"AI will change the advertising industry in the future" is narrative; "giving you an advertising channel now" is delivery; they are two different things.
The narrative is still there; people are still saying things like "AI changes the world," but delivery already occupies a considerable proportion.
For this item, score 1 point.
2. System Connectivity: From Islands to Protocols
In spring, every AI product was an island. If you wanted to integrate ChatGPT into your company's system, you had to write a set of adaptation code yourself; changing to another model required rewriting it again.
In April 2026, Google released the Gemini Enterprise Agent Platform, integrating agent management into enterprises' existing workflows.
Microsoft Copilot is embedded in the Office suite, Amazon's AI shopping assistant opened sponsored answers to brands, and their respective boundaries are loosening.
Connectivity must reach a certain level to possibly generate this kind of synchronous response.
Partial protocolization has been achieved now, but standardized protocols have not yet become mainstream.
For this item, score 1 point.
3. Delivery Capability: From Occasional Help to Stable Work
In spring, AI was like an intern, occasionally helpful, but more often, you had to fix things for a long time.
Doubao has over 300 million users. As of March 2026, its daily Token call volume exceeded 120 trillion, representing over a 1000-fold growth compared to May 2024, and doubled again in the past three months.
OpenAI has 900 million weekly active users, 50 million individual subscription users, and over 9 million enterprise users.
What are these enterprises using AI for? Writing code, reviewing contracts, automatically generating marketing copy, handling customer service tickets—all scenarios that replace people from repetitive labor. This is no longer the scale of trial use; it's using AI to do work.
Delivery has been realized. But has delivery capability become a core competitive advantage? Hard to say. The story is still being told, but enterprises that can stand firm by stable delivery are not yet mainstream.
For this item, score 1 point.
4. ROI Rationalization: From Unclear Calculation to Starting to Calculate
In spring, no one calculated ROI. How much computing power was invested, how much output was obtained? Unclear.
Now people are starting to calculate: Tencent Hunyuan API prices increased significantly (according to industry news, some models increased by over 400%), driven by the hard pressure of computing power costs.
A single inference request for a GPT-level model costs about 0.01-0.03 yuan in computing power. When call volumes reach hundreds of millions or billions, costs balloon into astronomical figures. Doubao's high-frequency calls from 345 million monthly active users force ByteDance to face this issue.
Zhipu AI raised prices three times within the year; Alibaba Cloud canceled the basic plan for the Bailian platform. Behind these decisions lies the same logic: Products that cannot clearly calculate ROI are about to collapse.
But is ROI clear and calculable? A few closed loops have begun to show ROI, but the metrics are vague. Some calculate Token costs, some calculate efficiency improvements, some calculate customer acquisition conversion; standards are not unified.
For this item, score 1 point.
5. Industry Common Phenomena: From No One Talking About Profits to Some Starting to Doubt
In spring, everyone was expanding; no one talked about profitability. Burning money for growth was the default model.
Now, Doubao launched a charging model; large model vendors like Zhipu AI and Moonshot AI raised prices, which in itself is an admission that costs are unsustainable and must be charged for.
But has the model of burning money for growth been widely negated? Not yet. Capital is still investing, leading manufacturers are still expanding, just at a slower pace; everyone is starting to calculate.
For this item, score 1 point.
6. Capital Environment: From Random Valuations to Starting to Feel Pressure
In spring, financing was extremely easy, valuations were thrown around casually, and a PPT could get tens of millions of dollars.
OpenAI launched an advertising platform, with an ad revenue target of $2.5 billion this year. The company's total revenue in 2025 was $13 billion, but it incurred a loss of $8 billion.
Although OpenAI's valuation is $852 billion, Anthropic, which came from behind, had an annualized revenue exceeding $30 billion as of April, surpassing OpenAI's $25 billion for the first time. Therefore, in the private secondary market, its valuation was chased to nearly $1 trillion, also exceeding OpenAI's.
OpenAI faces huge cost pressures; the story of growing big and strong before raising more money can't be told anymore; it must make money.
Has the financing winter arrived? Not yet. Leading manufacturers can still raise money, but the valuation logic has changed: from looking at imagination space to looking at revenue capability.
For this item, score 1 point.
Six dimensions, each scoring 1 point, total 6 points. According to the framework, 0-4 points is spring, 5-7 points is summer, 8-10 points is autumn, 11-12 points is winter.
What is the state of summer?
Narrative and delivery coexist. The imagination space is still there, but the ledger is already on the table. Capital can still invest, but starts asking about returns. Users are still growing, but start to stratify; some are willing to pay, some only use free services.
There are local signs of entering autumn. If signals like Doubao charging and OpenAI advertising continue to amplify, gaining another 2 points—for example, ROI becomes clear, and capital universally demands delivery—then autumn has truly arrived.
Summer is the stage where narrative and delivery coexist, but delivery is becoming increasingly important. Everyone knows the story still needs to be told, but the acceptance criteria after the story is quietly changing to: what exactly have you delivered?
二、Why Now:
3 Signals, 2 Drivers
The Six-Dimensional Cycle Judgment scoring is a static analysis. What happened recently from a dynamic perspective?
On May 6th local time, Anthropic founder Dario Amodei said: "Our growth rate exceeds exponential. Revenue and usage in the first quarter of this year achieved 80x annual growth, welcoming explosive growth. We are providing more computing power as fast as possible at an unprecedented speed."
This is the fastest revenue-growing company in human history. The day when AI truly becomes a productivity infrastructure may have already arrived.
In the same week, two signals appeared simultaneously: Doubao charging and OpenAI selling ads.
On the surface, it's a coincidence, but underlying are two threads: cost pressure and commercialization opportunities.
Signal One: Doubao Charging
Why is Doubao charging? High-frequency calls from over 300 million users have turned computing power costs into a problem that must be faced squarely.
As of March 2026, daily Token call volume exceeded 120 trillion, representing over a thousand-fold growth compared to May 2025, and doubled again in the past three months.
According to public calculations by Zheshang Securities cited by multiple media outlets, ByteDance's capital expenditure in 2025 was about 160 billion yuan, of which about 90 billion yuan was used for AI computing power procurement, with the rest for infrastructure and network equipment.
The free model really can't hold on much longer. A cost estimation circulated in the tech community shows that hardware depreciation accounts for about 58% of a single inference cost, and electricity costs about 29%. The larger the user base, the higher the cost.
Converting 120 trillion Tokens based on discounted public API prices is equivalent to daily revenue that should be 300 to 500 million yuan.
But now? C-end revenue is zero.
A thousand-fold growth resulting in zero—you can't find a second case like this among Chinese internet companies in the past 15 years.
At the same time, domestic Tokens entered a continuous price increase channel. Zhipu AI adjusted API prices upwards; GLM5.1 API increased by 10%, with the overseas version increasing by over double; Alibaba Cloud canceled the Bailian platform's basic plan; GLM5.0, MiniMAX 2.5, and Kimi 2.5 ended free public testing.
But there's also a price reduction side: DeepSeek V4-Pro slashed prices to 2.5% off, 0.25 yuan per million tokens; Alibaba Cloud's Tongyi Qianwen visual understanding model price dropped by over 80%; Doubao 2.0 Lite input price is only 0.6 yuan per million tokens.
Large model vendors are stratifying: one end raising prices, the other lowering prices.
Pressure and opportunity, two drivers.
Signal Two: OpenAI Selling Ads
Why is OpenAI selling ads? Half pressure, half opportunity.
Pressure side: The company's revenue in 2025 was $13 billion, with a cash loss of $8 billion. 50 million individual subscription users and 9 million enterprise users correspond to annual revenue of tens of billions of dollars, but computing power costs, R&D costs, and operating costs combined exceed this figure. Subscription revenue is insufficient to cover costs; the numbers don't add up.
Opportunity side: According to industry observation estimates, the ad pilot generated an ARR (Annual Recurring Revenue) of $100 million in less than two months after launch. The ad monetization potential of 900 million weekly active users is huge.
According to authoritative institution forecasts, Meta's full-year ad revenue will exceed $243.46 billion, higher than Google's $239.54 billion, indicating that monetization through advertising still has a huge market.
The forecast OpenAI shows investors is: $2.5 billion in ad revenue in 2026, reaching $100 billion by 2030.
This is a forced choice driven by costs and an active choice driven by opportunity. OpenAI is targeting not just cost coverage, but this large market.
AI is not free; GPUs, electricity, engineers, bandwidth—all cost money. The larger the user base, the higher the cost. In spring, you could burn investors' money to sustain; in summer, you must find users to pay.
But summer also means commercialization paths have opened: advertising, subscriptions, tiering—monetization methods are much richer than in spring.
The underlying logic of two signals appearing at the same point in time is: user scale has reached a critical point, cost pressure forces charging decisions, and commercialization opportunities mature enough to monetize.
Summer is this stage: the ledger is laid out, but the window hasn't closed yet.
三、Entering in Summer, How Can You Run Faster?
Clearly, AI has reached "summer"; it has transitioned from "usable" to "truly capable of helping you save money and make money."
How is this determined? Let's give a few examples:
Semir's designers used to take at least three days to produce one render. Now, using AI tools, they produce a render in 30 seconds, viewing the render directly without sampling.
Designer Lin Jianxia's exact words were: "Unsatisfactory schemes are directly eliminated, without wasting sampling costs."
AI improved Semir's overall design and R&D efficiency by 35%, and pattern design speed by over 200%. In 2025, AI brought Semir direct value: 200 million yuan in new revenue and 20 million yuan in cost reduction.
Anta's "Linglong" design large model, trained on tens of millions of apparel and footwear data accumulated over thirty years, can generate 56 inspiration schemes in minutes. The designer team can complete line drawing and generate high-definition renderings within one day.
With AI collaborating with the team, the tennis shoe project, from initiation to final model confirmation, took no more than 40 days, significantly faster than the traditional 3-month design cycle.
Peacebird (Taipingniao) achieved intelligentization across the entire marketing chain. Starting from understanding the business goal of "increasing GMV for new autumn children's wear products," AI can autonomously identify high-potential users, generate personalized product recommendations and marketing content, and push them to sales consultants' enterprise WeChat with one click.
Ultimately, the new series' click-through rate increased by 90%, payment conversion rate increased by 20%, and new product GMV surged by 31%.
Midea Group has formed an AI R&D team of over 400 people, with more than 13,000 intelligent agents running daily across multiple scenarios such as residential, office, manufacturing, healthcare, warehousing, and logistics. In 2025 alone, AI saved Midea 700 million yuan in costs and improved overall efficiency by over 15 million hours.
What do these cases illustrate? AI is transitioning from "embellishment" to "main force."
Having seen the examples, what should be done next?
In a nutshell: Progress from AI implementation for a series of small problems to gradually building large-scale AI system applications.
Specifically, take three actions.
The First Action: Find a Minimum Entry Point, Achieve a Value Closed Loop
Don't start by thinking "company-wide AI transformation"; that's the biggest pitfall. 80% of companies' AI implementation failures stem from being too ambitious, seeking completeness, and pursuing AI for AI's sake.
How to do it? Remember three words—Small, Accurate, Fast:
Small: Select 1-2 scenarios with "clear pain points, standardized processes, and sufficient data" to start with. For example, AI intelligent customer service, finance/administration automation, marketing material generation, contract compliance review—these belong to "high value, low threshold, quick results."
Accurate: Before launching each scenario, first establish the business baseline for the next 3-6 months, clarify how benefits are calculated, and how success is defined. The core evaluation metrics must be quantifiable financial results, not technical self-congratulatory metrics like "model accuracy rate, response speed."
Fast: If no clear results are seen within 3 months, iterate quickly or shut down; never force it. Set a stop-loss line in advance for each AI project. If preset business goals are not met for two consecutive cycles, shut down directly.
The key to this step is to achieve a closed loop, giving the team confidence and the boss determination, so the subsequent path can be walked.
The Second Action: From Pilot to Replication, Accumulating Organizational Capabilities
A single successful scenario doesn't qualify as transformation; it's just a pilot at best. What truly creates a gap is whether you can turn the success of one point into something the entire company knows how to do.
Don't rush to expand a successful scenario; first solidify the approach: how prompts are written, which tasks are assigned to AI, which personnel must oversee, what pitfalls are common, how success is calculated—write it into a set of standard procedures, then promote it to similar business lines.
Build "two infrastructures": First, an AI capability sharing middle platform—don't let each department explore from scratch; sales can directly use the AI data capabilities proven in finance, R&D can reuse market's user insight models. Second, a prompt knowledge base, categorized and shared by scenario, rewarding those whose prompts are used most.
Then, how can the organization keep up? Semir internally repeatedly emphasizes one sentence: AI implementation is 70% a people problem, only 30% a technical problem. While replicating scenarios, three things must be synchronized; missing any one won't work:
Talent梯队: It's not just hiring a few algorithm engineers. Three layers are needed. The top layer involves the boss or core executives personally leading. The middle layer consists of "translators" who understand both AI boundaries and business pain points. The grassroots level involves frontline employees being able to use AI tools to solve their immediate problems.
Incentive Mechanism: All incentives must be tied to quantifiable AI implementation results, with direct profit-sharing from the increased revenue and cost savings brought by AI; the payout cycle should be short, monthly or quarterly, allowing people to quickly feel "using AI = more money"; most crucially, incentives must also reach frontline executors—they are the ultimate users of AI tools; without their participation, AI will never be implemented.
Organizational Structure: Top leadership must personally lead, involving heads of business, technology, finance, and HR to work together, not letting the IT department fight alone; also, incorporate AI implementation cooperation into the performance assessments of each department head; use performance to speak to those who remain "uninvolved and aloof."
Simply put, this step turns "one person's success" into "an organization's muscle."
The Third Action: Systematic Reconstruction, from Adding AI to Using AI to Redo
Multiple scenarios are successful, and the organization has kept up. Next, it's not about "pasting AI onto old processes" anymore, but letting AI redo the processes—this is the large system.
Process Reconstruction: Serial to Parallel. The old way was "one person finishes, passes to the next," proceeding serially. In the AI era, this is completely obsolete. It must be changed to multiple people and multiple AIs working simultaneously:
Before meetings, let AI simulate all parties' stances first, exposing logical flaws and resource conflicts in advance; formal meetings only resolve real disagreements, cutting meeting time by 70% directly.
Real-time Alignment Dashboard: Stop writing weekly reports. Everyone, including AI, updates status on the same dashboard, with AI responsible for monitoring inconsistencies. If you say "high cost-performance" but set a high-end price, the dashboard marks it red immediately, discoverable the same day, without waiting for the review meeting two weeks later.
Upon receiving a requirement, don't rush to do it; first use AI to paraphrase your understanding and have the other party confirm. If even AI misunderstands, it means the requirement itself is unclear; nip rework at the source.
Automated Trigger Chains are also crucial: User complains—AI generates安抚话术—send to客服 for review—sync with brand work group;
ROI drops—AI finds原因—push suggestions to负责人;
Inventory almost out—AI calculates补货量—push to supply chain—can operate even if no one is monitoring.
In summary: Processes should change from being pushed by people to "when something happens, AI automatically runs, people only make decisions."
Finally, summarize these three actions: First find a pain point to achieve a closed loop, proving AI can make or save money; then solidify the successful approach and expand it, matching people and incentives; finally, let AI redo the processes, turning small problems into a large system. First run through a point, then spread into an area, finally let AI redo the whole game.
This article is from WeChat public account "Notesman" (ID: Notesman), author: Lao Jia
