AI Roundup: After Anthropic's Cutoff, New Shifts in Chinese Models and Startup Paradigms

Anthropic's Service Cutoff Opens Opportunities for Domestic Coding Models

On September 5, Anthropic published a notice on its website announcing that it would stop providing services to companies in regions where its products are not approved, and prohibit companies or organizations whose ownership structures are controlled by jurisdictions in such unapproved regions (such as China) from using its services, including entities more than 50% directly or indirectly owned by companies headquartered in unsupported regions ^[1]. For many developers and enterprises, this meant being shut out overnight from the world's leading large models, and even from the world's best coding models ^[1]. Claude is already the absolute mainstay model in the coding market, and in the B2B market, Anthropic has also surpassed OpenAI to become the model provider most used by enterprise users; Menlo Ventures' annual report shows Anthropic at 32%, ahead of OpenAI's 25% ^[1].

DeepSeek V3.2 Released: Long-Context Breakthrough and Halved Pricing

On September 29, DeepSeek launched the experimental model DeepSeek-V3.2-Exp, introducing DeepSeek Sparse Attention (DSA) technology on the basis of V3.1-Terminus, aiming to greatly improve long-text processing efficiency ^[2]. The official app, web version, and mini program have all been updated to DeepSeek-V3.2-Exp, while API service prices have been reduced by more than 50% ^[2]. DeepSeek says DSA is the first to implement a fine-grained sparse attention mechanism, improving long-text training and inference efficiency without almost affecting model output quality; on public evaluation sets across various domains, DeepSeek-V3.2-Exp performs essentially on par with V3.1-Terminus ^[2]. The model has been open-sourced on Huggingface and ModelScope, and the paper has also been released simultaneously ^[2].

Compliance for AI Going Global: Get the Four-Part Playbook Right First

After a product going global finds PMF, the next step is to address compliance and legal issues ^[3]. In the sharing session, the guest summarized the planning and preparation needed before going global into four stages: legal country-by-country research, overseas transaction structure design, domestic approval procedures, and local compliant operations ^[3]. Among these, domestic approval procedures include filing the project feasibility with the National Development and Reform Commission (NDRC) and registering with the commerce department ^[3].

ARR Is Being Distorted; AI Startups Need a New Narrative

Growing ARR from 0 to seven figures in just a few months is a common narrative in today's AI startup circle ^[4]. But for most AI products/businesses, there is revenue, but it is not recurring—especially in the early stage, where high user churn and new product testing are more like an experiment ^[4]. Real pricing power lies in the hands of giants such as OpenAI and Anthropic; they may adjust costs at any time and may also completely change an AI startup's unit economics ^[4].

Plaud's PMF: Selling Hardware Can Also Generate Over $100 Million a Year

Plaud made more than $100 million last year and can make more than $200 million this year ^[5]. Its roughly $150 recording card has been sold to more than 1 million people worldwide, attracting imitators among big companies; some U.S. teams have even come to China to find suppliers ^[5]. Mo Zihao is CEO of Plaud China and a partner responsible for AI products and R&D; before joining, he saw the real data—fourfold growth in a year—and believed this was PMF ^[5].

Seedream 4.0: New Competition in Image Models

Volcano Engine launched Doubao Image Creation Model Seedream 4.0, featuring combined input of text and images, enabling core capabilities such as multi-image fusion creation, reference-based image generation, combined image generation, and image editing, with subject consistency greatly enhanced ^[6]. It is also the first model to support 4K multimodal image generation ^[6]. Against the backdrop of successive hit products in the AI image generation field, Seedream 4.0 is regarded as a more suitable Nano Banana alternative for young Chinese users ^[6].

Google Cloud Report: Application Companies Don't Need to Build Their Own Models

Google Cloud released a trend report for AI startups, inviting entrepreneurs and investors to discuss AI trends, give advice to founders, and predict the direction of AI development ^[7]. The report's recommendations include: move fast to seize market opportunities; price based on delivered value rather than limiting yourself to per-seat charges; prioritize evaluations and start them immediately; and instead of focusing on general AI, go deep into verticals to solve specific difficult problems ^[7].

A New Startup Paradigm: Find the Prosumer First

At a roundtable at the Google Cloud Startup Summit, Geek Park founder Zhang Peng spoke with guests including Hillhouse Venture Capital partner Wang Bei and Fotor AI founder Duan Jiang ^[8]. They proposed that today's AI startups need to find their own prosumers: people with some understanding of technology, a higher tolerance than ordinary users, and a willingness to spend more patience researching and tinkering with AI; if they feel a product is valuable for work and life, they may even be willing to pay for a subscription ^[8]. Between time and cost, the guests argued that time matters more than cost, and it is not yet time to cut costs and improve efficiency ^[8].