Major AI Events

2022: The Gateway Opens

November: ChatGPT Goes Public (NVDA +25.4%)

Market Notes: The Nasdaq continues to fluctuate amid interest rate hikes and recession expectations, while NVIDIA has rebounded from its lows, but AI has yet to form an independent market main theme; meme stocks: no clear AI meme stocks have emerged yet.

Monster Stocks Cumulative Table (As of 2022-11, Month-End Adjusted Closing Composite; NVIDIA as Leading Reference)

• On November 30, OpenAI released the ChatGPT research preview. The chat interface became the first mass-market product entry point for large models, transforming model capabilities from demonstrations at research institutions into software that ordinary people could use repeatedly. OpenAI: Introducing ChatGPT

Judgment of the month: What matters is not whether ChatGPT is the first to generate text, but that natural language has for the first time become a software interface with a sufficiently low barrier to entry.

December: From Demonstration to Diffusion (NVDA -13.6%)

Market Notes: U.S. tech stocks pulled back amid rate hike expectations, with Nvidia weakening along with the semiconductor sector; the spread of ChatGPT has not yet immediately translated into stock market action, and no clear “meme stock” has emerged.

Monster Stocks Cumulative Table (as of 2022-12, month-end adjusted closing composite; Nvidia as benchmark leader)

• ChatGPT enters a rapid diffusion phase. Subsequent OpenAI release notes show that users began requesting product features around context, web access, plugins, and code interpreter. The chat product was no longer just a one-off Q&A demo. ChatGPT Release Notes

Judgment for this month: The first industry increment comes from usage habits, not from new model names. Users first learn to delegate tasks to models, and only then do developers begin to look for workflows in which models can be embedded.

2023: Capabilities, Open Weights, and Capital Expenditures Accelerate Simultaneously

January: Cloud and Office Begin to Bind (BBAI +385.2%)

Market Records: US growth stocks rebounded from their 2022 lows, with Microsoft, NVIDIA, and AMD initially traded as cloud and chip recovery plays; C3.ai (AI) began to show thematic volatility, but a stable group of AI speculative stocks had not yet emerged.

Monster Stocks Cumulative Table (As of 2023-01, month-end adjusted close composite; NVIDIA as the bellwether benchmark)

• Microsoft announced the expansion of its partnership with OpenAI. The partnership places model capabilities, Azure computing power, and office software on the same commercial chain, with AI moving from a standalone product into cloud and productivity software. Microsoft: Extending the partnership with OpenAI

Verdict of the month: The advantage of closed-source models has expanded from parameter scale to cloud, distribution, enterprise contracts, and continuous services.

February: Bard and LLaMA Fork (688256.SS +121.4%)

Market Records: The competition between ChatGPT and Bard caused the AI narrative to be mapped onto stocks for the first time, with Nasdaq tech stocks strengthening; small-cap concept stocks such as C3.ai and BigBear.ai became the earliest batch of “meme stocks,” with their rises and falls driven more by sentiment and trading volume.

Monster Stock Cumulative Table (as of end of February 2023, month-end adjusted closing composite; NVIDIA as the benchmark leader)

• Google publicly introduces Bard. The search company begins to directly respond to ChatGPT, and generative AI enters competition for search entry points. Google: Bard and Search Updates
• Meta releases the LLaMA research model. A smaller model, open research, and community reproduction provide an important starting point for the open-model ecosystem. Meta: Introducing LLaMA

Judgment of the Month: The same competition splits into two tracks: one vying for stable cloud access, and the other competing for developers, weights, and local deployment rights.

March: GPT-4 Begins Connecting to Tools (300308.SZ +52.8%)

Market Records: Under the impact of banking risks, capital instead concentrated on large-cap tech stocks and computing power, with NVIDIA becoming the main theme around GTC; A-share stocks such as 360, iFlytek, Cambricon, and Inspur showed significant ChatGPT-concept volatility.

Monstrous Stocks Cumulative Table (as of 2023-03, month-end adjusted closing price compounding; NVIDIA included as the bellwether benchmark)

• March 14, GPT-4 launched for ChatGPT Plus. The discussion shifted from “can it write” to complex instructions, code, reasoning, and professional tasks. OpenAI: GPT-4 research
• March 23, ChatGPT began offering plugin experiments. Browsing, code interpreter, and third-party plugins pushed the model from responder to tool invoker. ChatGPT release notes

Judgment of the month: The value of a model’s capabilities increasingly depends on access to up-to-date information, computing environments, and external services.

April: Google DeepMind Merger (300476.SZ +52.8%)

Market Record: Leading US AI stocks are oscillating at high levels, while A-shares are shifting from broad-based gains to thematic rotation, with concept stocks such as iFlytek, Kunlun Tech, and 360 experiencing large volatility; “demon stocks” are starting to expand from model companies to the computing power and application chains.

Monster Stocks Cumulative Table (as of 2023-04, month-end adjusted close composite; NVIDIA as the leading benchmark)

• Google merges Google Brain with DeepMind to form Google DeepMind. Research teams, compute resources, and product roadmaps are now housed under a single organization, with Google clearly positioning the next-generation multimodal model as its core focus. Google: Formation of Google DeepMind

Judgment of the Month: Model competition is not only about research breakthroughs; it also requires organizing talent, computing power, and product distribution.

May: Long Context Meets GPU Orders (SMCI +112.4%)

Market Record: NVIDIA’s earnings report turned AI orders into a price signal for the entire market. On May 25, the stock price surged approximately 24% in a single day, and NVIDIA, AMD, Broadcom, as well as the A-share optical module chain were all追捧 together; this was the first batch of AI market movements with truly global diffusion effects.

Monster Stocks Cumulative Table (As of 2023-05, Month-end Adjusted Closing Price Compound; NVIDIA as Benchmark Leader)

• Anthropic releases a 100K context window. Long documents, contracts, and code repositories become model inputs, and long context begins to shift from a research metric to an enterprise product selling point. Anthropic: 100K Context Windows
• NVIDIA earnings turn AI expectations into GPU orders. In the first quarter of fiscal 2024, the data center business and next-quarter guidance significantly exceeded market expectations, and the market began to price AI along compute, networking, and data centers. NVIDIA: FY2024 Q1 Earnings

Judgment for this month: The upstream infrastructure can receive real orders first, while the revenue, renewal, and profit of the application layer will come much later; the longer the context, the more important the source, permissions, and verification become.

June: APIs Start Calling Functions (SOUN +49.2%)

Market Notes: US stocks continue to be priced around the “selling shovels” narrative, with Nvidia, Super Micro Computer, and networking equipment stocks strengthening; A-shares see rotation among computing power, optical modules, and servers, with pure-concept stocks like C3.ai showing noticeably higher volatility than the leading names.

Monster Stock Cumulative Table (as of 2023-06, month-end adjusted closing price compounded; Nvidia as the leading benchmark)

• OpenAI adds function calling, longer context, and lower pricing to the API. Developers can have the model output function parameters according to a JSON Schema, making the connection between the model and external tools more orchestrable. OpenAI: Function calling and other API updates

Judgment for this month: The key to tool calling is not whether the model says “I can help you,” but whether it can reliably produce structured results that are executable and verifiable.

July: Open Weights and Closed-Source Models Progress Together (SMCI +32.5%)

Market Record: US-listed AI leaders have shifted from one-sided rallies to high-level consolidation, while A-share and Hong Kong AI trading relies more on thematic rotation; NVIDIA, AMD, optical module and server stocks remain the core, and no new “monster stocks” capable of consistently decoupling from the sector have yet emerged.

Monster Stock Cumulative Table (as of 2023-07, month-end adjusted closing composite; NVIDIA included as benchmark leader)

• Meta and Microsoft release Llama 2. Model weights, starter code, and commercial usage terms are opened up more broadly, enabling developers to build their own applications either in the cloud or on-premises. Meta: Llama 2
• Anthropic releases Claude 2. Longer input, coding, and reasoning capabilities extend the closed-source model competition beyond chat experiences to long documents and developer APIs. Anthropic: Claude 2

This Month’s Takeaway: “Open source” is beginning to become an industry slogan, but what truly affects deployment is licensing, weights, VRAM, inference speed, and maintenance responsibility.

August: Code Llama Seizes the Code Entry Point (000158.SZ +26.2%)

Market Notes: Nvidia’s earnings report is expected to continue supporting compute-related stocks, but high-valuation AI small-caps are beginning to diverge; Nvidia and Super Micro Computer are outperforming most software concept stocks, as the market shifts from “anything AI-related goes up” to seeking real orders.

Monster Stocks Cumulative Table (as of 2023-08, month-end adjusted close composite; NVIDIA as the leading benchmark)

• Meta releases Code Llama. Open-source models begin to specifically target code generation, completion, and local development scenarios. Meta: Code Llama

This month’s verdict: coding formed its paid boundary first because coding tasks have testing, submission, and delivery of results, making the value much easier to verify than general chat.

September: ChatGPT Goes Multimodal (000158.SZ +21.9%)

Market Records: Long-end U.S. Treasury yields rose, and AI leaders pulled back from their highs, though Nvidia remained the relatively strong core name; the A-share application sector cooled off, with no new sustained “demon stocks” emerging.

Monster Stocks Cumulative Table (as of 2023-09, month-end adjusted close compounded; NVIDIA included as benchmark leader)

• OpenAI brings image understanding, voice, and image-based conversation to ChatGPT. Text models are beginning to become the entry point for multimodal work. OpenAI: ChatGPT image and voice updates

This Month’s Takeaway: The objects models handle have expanded from text to screenshots, meeting notes, and documents, but understanding materials does not mean knowing whether they are real.

October: Image Generation Enters the Dialogue (300502.SZ +47.5%)

Market Notes: U.S. tech stocks continue to be weighed down by interest rate expectations and export control concerns, with AI leaders trading sideways to weaker. In the A-share market, computing power and application stocks are taking turns rebounding; the rally looks more like a short-term thematic play, and no new global AI “dark horse” has emerged yet.

Monster Stocks Cumulative Table (as of 2023-10, month-end closing price with adjusted composite; NVIDIA as the leading reference)

• DALL·E 3 enters ChatGPT Plus and Enterprise. Image generation moves from a standalone prompting tool into a conversational editing workflow, and OpenAI begins experimenting with a provenance classifier to identify generated content. OpenAI: DALL·E 3

Judgment of this month: The competition in AIGC is shifting from one-shot generation to iteration, copyright, provenance, and credibility.

November: DevDay and Governance Crisis (PLTR +35.5%)

Market Records: The leading U.S. AI stocks have regained strength amid earnings reports and the OpenAI governance crisis, with Microsoft, Nvidia, and AMD offering higher certainty than small-cap concept stocks; A-share AI trading has warmed up, but no new monster stocks have emerged that can deviate from fundamentals over the long term.

Monster Stocks Cumulative Table (as of 2023-11, month-end adjusted closing price composite; NVIDIA as the leading benchmark)

• OpenAI DevDay releases GPT-4 Turbo, Assistants API, vision capabilities, and lower prices. API, retrieval, function calling, and the assistants framework further embed models into applications. OpenAI: New Models and Developer Products Announced at DevDay
• OpenAI’s board abruptly fires and later reinstates Sam Altman. A governance crisis within a week makes model-company governance, shareholder relations, and talent dependency part of the industry’s reality. OpenAI: Sam Altman Returns as CEO, OpenAI Has a New Initial Board

Judgment for this month: Beyond the risks associated with the model company’s products, there are also risks related to the board of directors, core talent, capital, and partners.

December: Gemini Brings Multimodality into Products (BBAI +25.9%)

Market Notes: Leading US AI stocks are approaching their yearly highs, with NVIDIA emerging as the clearest “selling picks and shovels” mainline for the year; the A-share and Hong Kong stock markets remain structurally driven, where servers and optical modules more easily attract capital than pure application concepts.

Monster Stock Cumulative Table (as of 2023-12, month-end adjusted closing compound; NVIDIA as the leading reference)

• Google releases Gemini 1.0. Gemini is tiered from the start into Ultra, Pro, and Nano, covering cloud, applications, and on-device, with a strong emphasis on native multimodality. Google: Introducing Gemini

This Month’s Assessment: Model competition is no longer limited to a single web chat entry point, but is simultaneously fighting for search, mobile, cloud, and developer tools.

2024: Multimodality, Reasoning, and Agents Enter Products

January: GPT Store Attempts to Distribute Assistants (SMCI +86.3%)

Market Records: The beginning of 2024 was still driven by US semiconductors, with NVIDIA, AMD, Broadcom, Arm, and Super Micro Computer becoming high-beta targets; A-shares AI first rallied strongly then diverged, with no unified monster stock across the entire market yet.

Monster Stocks Cumulative Table (as of 2024-01, month-end adjusted closing composite; Nvidia as the bellwether benchmark)

• OpenAI launches GPT Store. Users can publish and discover custom GPTs, as model platforms begin experimenting with combining prompts, knowledge bases, and tools into distributable software units. OpenAI: Introducing the GPT Store

Judgment of the Month: The challenges of platformization have shifted from “can we build an assistant” to discovery, distribution, permissions, and sustained usage.

February: Sora and the Long Context Race (SOUN +347.0%)

Market Notes: Nvidia’s earnings reports and AI server orders have pushed the market into an acceleration phase. High-volatility targets such as Super Micro Computer, Arm, and SoundHound have taken turns becoming “meme stocks”; A-share optical modules, servers, and computing power leasing have also shown notable activity.

Monster Stocks Cumulative Table (as of 2024-02, month-end adjusted closing composite; NVIDIA as the leading benchmark)

• OpenAI releases research preview of Sora. Text-to-video brings the generative model competition to temporal consistency, physical world simulation, and video source verification. OpenAI: Sora
• Google releases Gemini 1.5, emphasizing long context and MoE architecture. The model begins handling longer videos, codebases, and multi-document inputs. Google: Gemini 1.5

Judgment of the Month: The next threshold for multimodality is not generating a single beautiful frame, but maintaining coherence over long durations and within long contexts.

March: Models Begin to Stratify by Capability (300502.SZ +29.1%)

Market Records: The GTC and NVIDIA’s new-generation computing power narrative continue to push semiconductors higher, with NVIDIA, Super Micro Computer, Arm, and the server chain seeing high-level position rotation; highly elastic A-share targets such as Cambricon, Industrial Fulian, and Zhongji Xudong experienced significant volatility.

Monster Stocks Cumulative Table (As of 2024-03, Month-End Adjusted Close Compound; NVIDIA as Benchmark Leader)

• Anthropic releases the Claude 3 family. Opus, Sonnet, and Haiku cover the enterprise API with different capabilities, speeds, and costs; model selection begins to turn into a routing problem. Anthropic: Claude 3 family

Verdict for this month: “The strongest model” is no longer the only answer; latency, price, and task difficulty are equally decisive in determining whether a product is usable.

April: Llama 3 Expands the Open Ecosystem (300308.SZ -39.7%)

Market Recap: Rising interest rates and earnings expectations triggered a pullback in US AI stocks, with Super Micro Computer and some high-valuation small caps seeing steeper declines; the A-share AI theme cooled in sync, with the speculative rally temporarily fading.

Meme Stock Cumulative Table (as of 2024-04, month-end closing price with adjusted composite; NVIDIA as the leading benchmark)

• Meta Releases Llama 3. The 8B and 70B models are open-sourced to the community, with expanded training data scale, tooling, and cloud partners. Meta: Llama 3

This month’s verdict: Open-weight models are shifting from research alternatives to a complete ecosystem, where deployers can trade cost and data boundaries for control.

May: GPT-4o Enables Real-Time Interaction (NVDA +26.9%)

Market Records: NVIDIA’s earnings report once again confirms AI capital expenditures, with related stocks such as NVIDIA, Dell, Super Micro Computer, and TSMC strengthening; A-share optical module and server chains followed, with funds shifting back toward “hardware monster stocks” where orders can be seen.

Monster Stocks Cumulative Table (as of 2024-05, month-end adjusted close composite; NVIDIA included as bellwether reference)

• OpenAI released GPT-4o. Text, vision, and audio capabilities were unified into a single faster model, with real-time voice interaction becoming a product focus. OpenAI: GPT-4o

Judgment for this month: Multimodal is no longer just about uploading files; real-time interaction is beginning to compete for mobile, customer service, and office entry points.

June: AI Enters Devices and Workflows (ARM +35.8%)

Market Records: Nvidia briefly became the world’s most valuable company by market capitalization in June, with stock splits and index funds further amplifying the attention; in A-shares, computing power chain stocks such as Zhongji Innolight, Eoptolink, and Industrial Fulian outperformed most application stocks, and Nvidia was the most typical AI leader during this phase.

Monster Stocks Cumulative Table (As of 2024-06, Month-End Adjusted Composite Close; NVIDIA as Benchmark Leader)

• Apple launches Apple Intelligence. On-device models, personal context, Private Cloud Compute, and ChatGPT integration are woven into the system experience of iPhone, iPad, and Mac. Apple: Apple Intelligence
• Anthropic releases Claude 3.5 Sonnet. Improvements in code, vision, and tool use push coding and Agent tasks to the center of the closed-source model competition. Anthropic: Claude 3.5 Sonnet

Judgment of the Month: AI is moving beyond “opening a website” and becoming embedded into devices and workflows, with privacy, permissions, and latency now part of the experience.

July: Large Models Shift Toward Low-Cost Routing (SOUN +28.9%)

Market Record: The market is rotating from large-cap tech stocks to small-caps and non-AI sectors, with NVIDIA and Super Micro Computer rallying first before pulling back; A-share computing power stocks are oscillating at high levels, and the characteristics of speculative stocks have shifted from “continuous rallies” to “high volatility and rapid rotation.”

Monsters Stock Cumulative Table (as of end of July 2024, month-end adjusted closing composite; NVIDIA as the bellwether benchmark)

• Meta releases Llama 3.1 405B. Open-weight models begin directly competing with frontier closed-source models, and provide long-context and agent reference systems. Meta: Llama 3.1
• OpenAI releases GPT-4o mini. Smaller and cheaper models drive localization or batching of simple tasks, and the economics of model routing become clearer. OpenAI: GPT-4o mini

Judgment for this month: The unit economics of AI products don’t depend solely on the strongest model. Low-cost models that handle large volumes of simple tasks are what give API calls a chance to become long-term budget items.

August: AI Act Enters into Implementation (000158.SZ +69.3%)

Market Notes: The unwinding of the yen carry trade triggered a sharp selloff in global tech stocks followed by a rapid recovery, with the price swings of NVIDIA and AMD widening once again; AI stocks in the A-share market are still primarily driven by rotation, without the emergence of any new single monster stock.

Monster Stock Cumulative Table (As of 2024-08, Month-end adjusted closing composite; NVIDIA as the bellwether benchmark)

• EU AI Act takes effect. Risk classification, transparency, and general-purpose AI model obligations begin entering the formal implementation phase. EUR-Lex: Regulation (EU) 2024/1689

This month’s verdict: Regulation is no longer an abstract discussion before launch; it will gradually make its way into model documentation, deployment processes, and enterprise procurement checklists.

September: A new curve emerges for reasoning models (000158.SZ +180.8%)

Market Notes: Rate cut expectations support a rebound in U.S. tech stocks, but capital is starting to diversify from high-valuation chips to software and platforms; Nvidia, Arm, and Super Micro Computer remain representative of high volatility, while A-share AI is in repeated bottom-building.

Monster Stocks Cumulative Table (as of 2024-09, month-end adjusted closing composite; NVIDIA as benchmark leader)

• OpenAI released o1-preview and o1-mini. Reasoning models shift more computation into the inference stage, where math, code, and complex tasks begin to be priced based on “how much they think.” OpenAI: o1-preview and o1-mini

This month’s verdict: Beyond “bigger,” a capability curve of “thinking longer” has emerged. Waiting, token, and review costs all need to be calculated together.

October: Agents begin operating computers and searching (SMCI -30.1%)

Market Recap: Leading U.S. AI stocks diverged around earnings reports, with Nvidia remaining relatively resilient, while Super Micro Computer and small-cap AI stocks saw larger volatility. In A-shares, capital rotated toward new themes such as computing power, satellite communications, and robotics, with the performance of speculative “demon stocks” clearly becoming shorter in cycle.

Monster Stocks Cumulative Table (as of 2024-10, month-end adjusted close compounded; NVIDIA as the leading stock benchmark)

• Anthropic releases computer use capability. Claude can now view the screen, move the mouse, click, and type—agents evolve from calling APIs to operating general-purpose software. Anthropic: computer use
• OpenAI releases ChatGPT Search. The chat entry point now directly returns real-time web answers with links, further narrowing the boundary between search and Q&A. OpenAI: ChatGPT search

Judgment for this month: The value of an Agent comes from execution, but the stronger the execution capability, the less the permissions, prompt injection, error rollback, and responsibility boundaries can be omitted.

November: MCP Connects Tools and Data (SOUN +85.1%)

Market Record: After the U.S. election, AI themes heated up alongside software, autonomous driving, and robotics sectors. Targets such as NVIDIA, Palantir, and SoundHound strengthened but experienced sharp divergence; Hong Kong tech stocks rebounded in tandem, though a stable basket of “AI monster stocks” has yet to form.

Monster Stocks Cumulative Table (as of 2024-11, month-end adjusted closing composite; NVIDIA as the leading benchmark)

• Anthropic open-sources Model Context Protocol. MCP attempts to connect models with external tools and data sources through a universal protocol, signaling that tool integration is evolving from single-company implementations toward an ecosystem-wide protocol. Anthropic: Model Context Protocol

Judgment of the Month: The bottleneck for Agents is gradually shifting from the model itself to tool protocols, context management, and permission systems.

December: Sora Goes Live, Gemini Enters the Agent Era (SOUN +113.1%)

Market Records: By year-end, AI trading expanded from chips to software, data, and robotics. Nvidia experienced volatile trading at high levels, while high-beta names like Palantir and SoundHound behaved more like meme stocks; the A-share computing power chain failed to replicate the broad rally seen at the beginning of the year.

Monstrous Stocks Cumulative Table (as of the end of 2024-12, month-end adjusted closing price compound; Nvidia as the leading comparison)

• OpenAI officially released Sora. Video generation moved from research preview to a subscription product, with watermarking and provenance identification measures enabled by default. OpenAI: Sora is here
• Google released Gemini 2.0, emphasizing the Agentic Era. Multimodal output, tool use, and real-time interaction became the next-generation product direction. Google: Gemini 2.0

This month’s takeaway: The main storyline of 2024 shifts from “what a model can generate” to “whether a model can complete a sequence of actions in real-world environments.”

2025: Reasoning Costs Drop, Coding Becomes the Agent Proving Ground

January: R1 Recalculation Costs, Stargate Increases Infrastructure Investment (688158.SS +155.2%)

Market Record: On January 27, the DeepSeek shock caused Nvidia to drop nearly 17% in a single day, resulting in a rare massive market cap pullback; the U.S. stock computing power chain was sold off first, while A-share concept stocks first saw abnormal movements and then spread after the holiday, with Zhejiang Eastern, Meig Smart, Every Daily Interactive, UCloud, and others becoming representatives of high volatility. This month also saw a split行情 of “selling computing power overseas” and “buying applications domestically.” Reuters: DeepSeek Triggers AI Stock Sell-off

Monster Stock Cumulative Table (as of 2025-01, month-end adjusted closing composite; NVIDIA as the leading benchmark)

• DeepSeek released DeepSeek-R1 and open-sourced its weights and paper. The open reasoning model brought frontier capabilities, training methodology, and unit economics all into the global market spotlight at once. DeepSeek: R1 Release
• The U.S. announced the Stargate project. Partners including OpenAI, SoftBank, and Oracle unveiled a large-scale AI infrastructure plan, with the capex narrative continuing to concentrate around data centers. OpenAI: Stargate Project

This month’s verdict: DeepSeek forces the market to recompute the cost per unit of capability, while Stargate demonstrates that total demand, total investment, and the cost per inference can all rise simultaneously.

February: Research, reasoning, and coding extending task chains (300476.SZ +58.7%)

Market Notes: The DeepSeek theme continues to spread in the A-share market. After Zhejiang Eastern continues to hit the daily limit up/down, Magic Smart, UCloud, QingCloud Technology, and Talkweb Information take turns. The U.S. stock market is recovering from panic, with NVIDIA and cloud vendors returning to the logic of capital expenditure. A-share “demon stocks” and U.S. stock leaders are running at two different paces. Risk warning regarding Zhejiang Eastern and Magic Smart

Monster Stocks Cumulative Table (As of 2025-02, month-end adjusted closing composite; NVIDIA included as the bellwether reference)

• OpenAI released deep research. The model began independently performing multi-step web searches, reading, and synthesis, making research tasks a representative scenario for Agents. OpenAI: deep research
• Anthropic released Claude 3.7 Sonnet and Claude Code. A hybrid reasoning model and a terminal coding tool emerged simultaneously, making software engineering a testing ground for long-horizon Agents. Anthropic: Claude 3.7 Sonnet and Claude Code
• OpenAI released GPT-4.5 research preview. Large-scale pre-training and reasoning models formed two complementary paths, and attention began to shift toward training and serving costs beyond the model name itself. OpenAI: GPT-4.5

Judgment for this month: Reasoning models, web-connected research, and coding agents have collectively lengthened task chains and amplified the costs of context, waiting, and acceptance.

March: Web Search Becomes an API Capability (NVDA -13.2%)

Market Highlights: The market has shifted from the DeepSeek application theme to inference, computing power, and robotics. NVIDIA has regained attention around GTC; A-shares Tuowei Information, Changshan Beiming, Inspur Information, and the optical module chain have experienced high volatility, while Hong Kong tech stocks remain relatively active.

Monster Stocks Cumulative Table (as of 2025-03, month-end adjusted close compound; NVIDIA as the leader benchmark)

• Google releases Gemini 2.5. “Thinking models” have become the core narrative of the Gemini series, with long-context and multimodal reasoning continuing to advance toward productization. Google: Gemini 2.5
• Anthropic launches Web Search for the Claude API. Search capability has expanded from a standalone chat product into an API tool that developers can compose. Anthropic: Claude can now search the web

Judgment for This Month: Connectivity is no longer just a product toggle, but an orchestratable capability within the Agent architecture.

April: Model Comparison Unit Becomes a Complete Task (PLTR +40.3%)

Market Notes: Tariff and growth concerns triggered a rapid pullback in global tech stocks, with NVIDIA, AMD, Super Micro Computer, and AI infrastructure stocks falling before recovering; A-share high-level computing power stocks diverged, as the market began prioritizing “whether orders can be landed” over thematic hype.

Monster Stock Cumulative Table (as of 2025-04, month-end adjusted closing compound; NVIDIA as the leading benchmark)

• Meta releases Llama 4 Scout and Maverick. The open-weight models add native multimodality and MoE architecture, with the community roadmap continuing to compete for deployment rights. Meta: Llama 4
• OpenAI releases the GPT-4.1 family. Long context, instruction following, and coding become the main selling points of the API release. OpenAI: GPT-4.1
• OpenAI releases o3 and o4-mini. For the first time, the reasoning models combine tools such as search, file analysis, Python, and image generation within ChatGPT, making the Agent form more complete. OpenAI: o3 and o4-mini

Judgment for this month: The unit of comparison for model releases has shifted from a single response to an entire task, as tool use, code modification, and result verification are being incorporated into the same product loop.

May: Claude 4 Competes for Long-Duration Coding Tasks (300308.SZ +56.5%)

Market recap: U.S. AI leaders rebounded on earnings and capex expectations, with NVIDIA, Broadcom, Dell, and cloud providers outperforming most application stocks; in A-shares, optical modules, PCB, and server supply chains reactivated, with speculative stock characteristics concentrated in细分 segments of the industrial chain.

Monster Stocks Cumulative Table (as of 2025-05, month-end adjusted closing composite; NVIDIA as the leading benchmark)

• Anthropic released Claude 4. Opus 4, Sonnet 4, and Agent tool capabilities are further integrated, making long-duration coding tasks an important benchmark in frontier model competition. Anthropic: Claude 4
• Anthropic released Web Search and connectivity capabilities for API. Model vendors are starting to make retrieval, file access, and external service connections into reusable developer infrastructure. Anthropic Newsroom

This Month’s Verdict: The competition in coding is no longer about completing a few lines of code, but about reading repositories, splitting tasks, modifying files, running tests, and handling failures.

June: Terminal Becomes an Agent Working Environment (601138.SS +65.1%)

Market Record: US stocks continue to trade AI capital expenditure, with NVIDIA, Broadcom, and data center suppliers maintaining strength; in A-shares, Zhongji Xudong, Xin-Yi Sheng, Shenghong Technology, and Foxconn Industrial Internet are seeing high-level expansion in the computing power chain. Optical modules and PCB are the closest sectors to “demon stocks” at this stage, but industry prosperity and individual stock prices should not be equated.

Monster Stock Cumulative Table (As of 2025-06, month-end adjusted close compounded; NVIDIA as the bellwether benchmark)

• Google brings Gemini CLI to developers as an open-source AI Agent tool. The terminal is becoming a direct entry point for model invocation, search, and code modification. Google: 2025 AI News Recap
• OpenAI releases o3-pro. Reasoning models with higher compute budgets continue to place reliability, speed, and price on the same usage ledger. OpenAI Model Release Notes

Verdict of the month: Terminals and IDEs are no longer just the display interfaces for models, but the working environments where Agents obtain context, permissions, and acceptance feedback.

July: Agent Moves from Model to Delivery (688256.SS +110.4%)

Market Records: The AI rally has extended from chips to coding, software, and data center infrastructure. Leading U.S. stocks are rotating at high levels, while A-shares in optical modules, servers, and PCBs continue to diverge; high-beta targets such as Zhongji Innolight, Eoptolink, and Victory Giant Technology have become the “meme stock”-type representatives drawing market attention.

Monster Stocks Cumulative Table (as of the end of July 2025, month-end adjusted close composite; NVIDIA serves as the bellwether benchmark)

• xAI released Grok 4. Reasoning, tool use, and real-time search are packaged into a single model product, extending frontier model competition further toward Agents and real-time information. xAI Newsroom
• OpenAI released ChatGPT agent. Research, browser operation, terminal, and connectors are merged into a single Agent capable of executing complex tasks, with models beginning to directly handle requests from “start to delivery.” OpenAI: ChatGPT agent

Judgment for this month: The evaluation of Agents is beginning to approach that of software engineering, where consistently completing tasks matters more than producing a beautiful answer in one shot.

August: GPT-5 and Open Weights Advance Simultaneously (SOUN +26.0%)

Market Records: The U.S. stock market repriced model capabilities and platform ecosystems, with large-cap companies such as NVIDIA, Microsoft, and Meta outperforming most small-cap stocks; in the A-share market, AI capital spread from computing power to applications and end devices, without forming any single, sustainable “demon stock.”

Monster Stock Accumulation Table (as of August 2025, month-end adjusted closing compound; NVIDIA as benchmark leader)

• OpenAI releases GPT-5. A unified system brings fast answers, deep reasoning, vision, code, and Agent capabilities into a single generation of product, with coding explicitly positioned as a core use case. OpenAI: GPT-5
• OpenAI releases gpt-oss-120b and gpt-oss-20b. OpenAI re-enters the open-weight model market, with models that can be downloaded, customized, and run locally or on one’s own infrastructure. OpenAI: gpt-oss

Judgment of the month: Closed-source services and open-weight models are not an either-or choice; model companies may charge via API while simultaneously expanding their ecosystem through open weights.

September: Agent Needs Checkpointing and Rollback (MRVL +33.7%)

Market Recap: Interest rate cuts and software earnings expectations have caused AI trading to shift from hardware to platforms, software, and data, with amplified volatility in Palantir, software infrastructure, and certain robotics stocks; A-share high-level thematic stocks have entered a phase of rapid rises and sharp declines.

Monster Stock Cumulative Table (as of 2025-09, month-end adjusted close composite; Nvidia as the benchmark leader)

• Anthropic releases Claude Sonnet 4.5. Coding, computer use, Claude Code checkpoints, and the Agent SDK were rolled out in the same batch — rollback capability for long-running tasks is starting to become a product feature. Anthropic: Claude Sonnet 4.5

This month’s judgment: For Agents entering production, checkpoints, memory, permissions, and rollback are just as important as model capabilities.

October: Low-latency model offloading (ARM +20.0%)

Market Notes: The commercialization prospects of low-latency models support software and cloud service stocks, while chip stocks are consolidating at high levels; the AI sector in A-shares continues to rotate, with speculative stocks shifting from single-model concepts to crossover themes such as low-altitude economy, robotics, and edge devices.

Monster Stocks Cumulative Table (as of 2025-10, month-end adjusted closing price composite; NVIDIA as benchmark leader)

• Anthropic releases Claude Haiku 4.5. A faster, cheaper model that brings the previous generation’s frontier capabilities down to high-frequency tasks, with model routing and cost control continuing to segment. Anthropic: Claude Haiku 4.5

This month’s take: The industry won’t be left with just one strongest model — low-latency and low-cost models determine whether a large volume of real-world calls can hold up.

November: Gemini 3 Connected Development Platform (300502.SZ +23.9%)

Market Notes: Leading U.S. AI stocks continue to be caught in a tug-of-war between capital expenditure and valuation, with NVIDIA, Broadcom, and cloud providers as the main themes, while software stocks diverge; A-shares see rotation among computing power, applications, and robotics themes, with no single stock able to represent the entire AI market.

Monster Stock Cumulative Table (as of 2025-11, month-end adjusted close compound; NVIDIA as the bellwether reference)

• Google releases Gemini 3. Reasoning, multimodal, and coding are reintegrated, and the agentic development platform begins to become part of product distribution. Google: Gemini 3

Judgment for This Month: The battleground for model vendors has expanded from model APIs to development platforms, search, browsers, and office software.

December: Inference Capabilities Continue to Drive Down Costs (688158.SS +33.6%)

Market Notes: At year-end, the market places greater emphasis on inference costs and real-world usage. AI leaders are shifting from a pure hardware-driven narrative to a layered structure of cloud, software, and applications. High-volatility speculative stocks still exist, but their duration is shortening, and they should not be used as proxies for industry prosperity.

Monster Stocks Cumulative Table (As of 2025-12, Month-end Adjusted Close Composite; NVIDIA as Leading Benchmark)

• Google releases Gemini 3 Flash. A faster and more cost-effective model carries forward the capabilities of Gemini 3, as the price gap between frontier models and large-scale deployments continues to narrow. Google: December 2025 AI Update

This Month’s Takeaway: Reasoning capabilities will only transition from demos to everyday workflows once the per-unit cost is driven down enough.

2026: Model companies begin to face scrutiny from the public market and real-world workflows

January: Constitution and Automatic Browsing (NVDA +2.5%)

Market Records: AI companies in the Hong Kong stock market now have publicly tradable samples, and Zhipu (02513) and MiniMax (00100) have shown significant volatility after listing; the U.S. stock market continues to be dominated by Nvidia, Microsoft, and cloud providers. The high turnover of newly listed stocks can be recorded as “zombie stock”-style market action, but it cannot replace revenue and valuation analysis. Hong Kong Stock Exchange: MiniMax Newly Listed Securities Admission Notice

Monster Stocks Cumulative Table (as of 2026-01, month-end adjusted closing composite; Nvidia as the bellwether for comparison)

• Anthropic publishes Claude’s new Constitution. Model values, training goals, and behavioral boundaries are now publicly available as a fully readable document, with safety and governance moving from the appendix into product materials. Anthropic: Claude’s Constitution
• Google summarizes its January 2026 AI updates. Gemini further integrates into Chrome, Search, and automated browsing, with AI products evolving from answering questions to handling multi-step tasks. Google: AI Updates January 2026

Judgment of the Month: The competitive focus of 2026 has shifted to “whether you can advance a task for the user” rather than “whether you can generate another paragraph.”

February: Reasoning Models Move Toward Research Collaboration (2513.HK +20.6%)

Market Records: Hong Kong stock trading of Zhipu and MiniMax has shifted from listing sentiment to circulating shares, unlocking, and valuation games, with a divergence in Hong Kong tech stocks; U.S. AI leaders are still priced around capital expenditure and inference demand, with no new globally unified “monster stock” emerging for now. Hong Kong Exchange: Zhipu and MiniMax Approved for Short Selling

Monster Stocks Cumulative Table (as of 2026-02, month-end adjusted close composite; NVIDIA as the bellwether reference)

• Anthropic releases Claude Opus 4.6. Coding, Agents, computer use, tool calling, and search are pushed into higher capability tiers, as model companies continue to compete for complex workflows. Anthropic Newsroom
• Google updates Gemini 3 Deep Think. The dedicated deep-reasoning mode is extended toward science, research, and engineering tasks. Google: Gemini 3 Deep Think

Judgment for this month: Reasoning models are shifting from exams and leaderboards toward research collaboration, but the real value still depends on whether experts can review and reuse them.

March: Agent Completes Retrieval and Memory (688256.SS +72.9%)

Market Record: The divergence among AI stocks continues to widen. Large-cap platforms and computing power suppliers in the U.S. stock market remain relatively stable, while the intraday volatility of Zhipu and MiniMax in the Hong Kong stock market, as well as the computing power chain in the A-share market, is greater; “demon stocks” have shifted from single model names to targets with smaller circulating shares and rapidly changing expectations.

Meme Stock Cumulative Table (As of 2026-03, month-end adjusted close composite; NVIDIA as the benchmark leader)

• Google released Gemini Embedding 2. Text, images, video, audio, and documents enter a unified multimodal vector space, with retrieval and classification beginning to become fundamental layer capabilities of Agents. Google: Gemini Embedding 2
• Gemini’s Search Live, Canvas, and cross-application capabilities continue to expand. Search begins to simultaneously handle real-time conversation, long-task workspaces, and personal profile connections. Google: AI Updates March 2026

This month’s takeaway: The infrastructure for an Agent is not just a set of large models, but also includes retrieval, memory, file system, search, and cross-application permissions.

April: Long Tasks Encounter Product Life Cycle (2513.HK +83.8%)

Market Notes: The market has begun incorporating the retention, cost, and cash flow of AI products into stock price narratives, and the valuation elasticity of concept stocks has decreased. Leading U.S. stocks, Hong Kong-listed AI IPOs, and A-share computing power stocks are each following their own trends, without forming new cross-market monster stocks.

Monster Stock Cumulative Table (as of 2026-04, month-end adjusted closing composite; NVIDIA as the leading benchmark)

• Anthropic released Claude Opus 4.7. Stronger coding, Agent, vision, and multi-step task capabilities continue to push model competition toward long-horizon tasks. Anthropic Newsroom
• Sora product discontinued. The technical impact of generative video persists, but the user base, cost, and commercial return of a single product are not automatically guaranteed by model demos. OpenAI: Service status note on the Sora release page

Judgment of the Month: AI products must also undergo scrutiny based on lifecycle and unit economics; having strong model capabilities does not guarantee that the product will endure in the long term.

May: AI Applications Become Execution Systems (SMCI +68.2%)

Market Recap: AI capital expenditure remains the main theme in U.S. stocks, but capital is starting to seek out software and applications that can convert calls into revenue; in A-shares, computing power, optical modules, and PCBs are rotating, while Hong Kong AI IPOs are highly volatile, and the gap between meme stocks and leading stocks is becoming more pronounced.

Monster Stocks Cumulative Table (as of 2026-05, month-end adjusted close composite; NVIDIA as benchmark leader)

• Google I/O 2026 releases Gemini 3.5 Flash and Gemini Omni. Gemini enters the “agentic era,” with models simultaneously handling input, reasoning, action, and video creation. Google: I/O 2026 Overview
• Google integrates Agent, Search, Gemini App, and developer platform into the same product update cycle. The boundaries between search, shopping, code, and personal assistant continue to blur. Google: I/O 2026

Judgment of the month: “AI applications” are increasingly looking less like a standalone chat box and more like an execution system that overlays search, office work, devices, and development tools.

June: On-device Models Meet the Open Market (MRVL +45.3%)

Market Recap: On-device models have brought Apple, mobile phone, and storage supply chains back into AI pricing, while the public market has begun scrutinizing model companies’ financing and losses. A-shares such as Zhongji Innolight, Eoptolink, and Victory Giant Technology have experienced high-level volatility, and the unlocks and changes in floating shares of Hong Kong-listed Zhipu and MiniMax have become new trading variables.

Monster Stock Cumulative Table (as of 2026-06, month-end adjusted closing composite; NVIDIA serves as the leader benchmark)

• Apple Released the Third Generation of Apple Foundation Models. The on-device model, server model, and the foundation model system developed in collaboration with Google are incorporated into the new generation of Apple Intelligence. Apple Machine Learning Research
• Anthropic Released Claude Sonnet 5. Lower-cost models begin to take on browser, terminal, tool-calling, and long-task execution, with Agent capabilities continuing to penetrate into larger-scale everyday usage. Anthropic: Claude Sonnet 5
• Anthropic Submitted Confidential IPO Filing. The company has not yet determined the offering size, pricing, or timing of the listing; what is confirmed is that model companies are beginning to consider the public market as a viable capital path. AP: Anthropic confidential IPO filing

Judgment for this month: On-device models and the public market represent two different kinds of constraints: on one side, power consumption, privacy, and latency; on the other side, revenue, losses, financing, and cash flow.

July: Coding Agent Becomes a Clear Main Theme (NVDA +5.4%)

Market Notes: As of July 13th, leading US AI stocks continue to rotate at high levels around computing power, cloud, and software platforms. In the A-share market, optical modules and PCBs show greater momentum, while Hong Kong stocks Zhipu and MiniMax are even more volatile. Coding agents represent a new market narrative, but it has yet to be proven that they can drive all AI stocks higher in unison.

Monster Stock Cumulative Table (as of 2026-07, month-end adjusted closing composite; NVIDIA as benchmark leader)

• July 8: xAI releases Grok 4.5. The launch page lists coding, agents, and knowledge work as the primary scenarios, with availability in Cursor, Grok Build, and the API. xAI: Grok 4.5
• Anthropic releases a retrospective on Claude Code. The journey from an internal CLI to a coding agent is documented in public, with software engineering continuing to serve as an important workflow for validating long-horizon agents. Anthropic: The Making of Claude Code
• OpenAI releases ChatGPT Work. Agents begin to carry complex projects forward across applications and files, as the product narrative shifts further from “answering questions” toward “breaking goals down into outcomes.” OpenAI: ChatGPT Work

Judgment for this month: As of July 13, the clearest product direction is still the coding Agent; its advantage lies not in being able to write code, but in the fact that tasks can be split, run, tested, and rolled back.

What to Watch After the Timeline

The Entry Point Has Shifted from Chat to Workflow

ChatGPT first proved that natural language could become the entry point for software. GPT-4, plugins, multimodal capabilities, and search then connected that entry point to files, the web, and external tools. After 2025, Deep Research, Claude Code, Codex, and various terminal Agents have continued to extend the task chain. What users now deliver is no longer just an answer, but a piece of research, a set of modifications, a collection of test results, or a project that can keep running.

Open Weights and Closed-Source Services Each Bear Different Costs

“Open” does not mean free, and “closed-source” does not mean no ecosystem. Real deployment decisions come down to licensing, GPU memory, context length, inference speed, tool calling, fault handling, and accountability.

AIGC Moves from Content Generation to Credibility Competition

Text requires fact-checking; images require verification of source and copyright; videos require confirmation of timing and source material; and voice/audio requires consideration of identity boundaries. As the cost of generation decreases, the cost of verification becomes even more important. Visual completeness cannot replace the chain of evidence, and the model’s confident tone cannot replace original documents and reproducible processes.

Coding is the earliest Agent scenario that can be accepted

Code completion is only the starting point. A real coding Agent needs to read repositories, understand constraints, modify files, execute commands, run tests, handle failures, retain checkpoints, and allow humans to review and roll back. Software engineering has clear delivery boundaries, so it’s easier to form subscriptions or enterprise budgets than general chat; but this also means every line of code generated by the model must be tested and maintained.

AI Stocks Need to Move from Computing Power Orders to Cash Flow

AI stocks do not all rise in a straight line together. In the early market, investors bought the entry point and the vision; starting in 2023, they bought GPU orders and data centers; and from 2025 onward, they care more about unit inference cost, application-level payments, capital expenditure, and the financing ability of model companies. The upstream can receive money first, while the application layer still has to go through pilots, procurement, renewals, and profit validation.

Looking back at the monthly “Market Records,” three clear re-anchorings can be observed: before May 2023, the market was pricing in “whether generative AI would become the gateway”; after NVIDIA’s earnings report, compute orders and data centers became the primary theme; after the DeepSeek shock, the market began to ask again whether training and inference could become cheaper. By 2026, model companies such as Zhipu and MiniMax entered the Hong Kong stock market, and the question shifted to revenue, losses, free float, and lock-up expiry, rather than simply comparing model launch events.

The following three tables are worth tracking more closely:

1. The capital expenditure, orders, and remaining contract liabilities of cloud vendors determine how long infrastructure demand can keep rolling.
2. The revenue, losses, inference costs, customer concentration, and cash flow of model companies determine whether capability can be turned into budget.
3. The usage depth, rework rate, test pass rate, and renewal rate of coding and Agent products determine whether “being able to complete a task” truly equals “having commercial value”.

Market Data and Record Boundaries

• Nasdaq Composite Historical Data and Nasdaq Market Activity
• Yahoo Finance: NVIDIA Historical Data, Super Micro Computer Historical Data, C3.ai Historical Data, SoundHound AI Historical Data, Palantir Historical Data and Arm Historical Data
• Yahoo Finance: Zhongji Innolight Historical Data, Eoptolink Technology Historical Data, Cambricon Technologies Historical Data, Foxconn Industrial Internet Historical Data and Zhipu Historical Data
• Shanghai Stock Exchange: Stock Data, Shenzhen Stock Exchange: Market Data
• Hong Kong Exchanges: Securities Prices

The monthly commentary and cumulative tables in this article are based on Yahoo Finance’s publicly available monthly data feeds and news from the corresponding months. The tables use adjusted closing prices. For the current month, data is current through July 13, and indices, currencies, and trading regimes from different markets have not been forcibly merged. Since Yahoo Finance does not provide an available monthly series for MiniMax (00100.HK), this article does not fabricate price changes for it. The term “妖股” (meme stock) is used here only to look back on market sentiment and does not constitute investment advice. For matters involving listing dates, short-selling eligibility, share unlock schedules, and corporate disclosures, please refer to the official announcements from the relevant exchange.

References

• OpenAI: Introducing ChatGPT
• OpenAI: GPT-4 research
• OpenAI: GPT-4o
• OpenAI: o3 and o4-mini
• OpenAI: deep research
• OpenAI: GPT-5
• OpenAI: gpt-oss
• Anthropic: Claude 3 family
• Anthropic: Claude 3.5 Sonnet and computer use
• Anthropic: Claude 3.7 Sonnet and Claude Code
• Anthropic: Claude 4
• Anthropic: Claude Sonnet 4.5
• Google: Gemini
• Google: Gemini 2.5
• Google: Gemini 3
• Meta: Llama 2
• Meta: Llama 3
• [Meta: Llama 3.1](https://ai

This article only compiles publicly available materials and personal observations, and does not constitute investment advice. Models, products, prices, stock prices, and company disclosures are subject to continuous change; for content related to July 2026, the cut-off date is July 13th, and you should refer to the company’s latest announcements and original market data thereafter.