Claude leads 26% of R&D, Anthropic sparks heated debate on "AI developing AI" with its "AI slowdown theory"! The RSI training paradigm catalyzes a massive expansion in computing power demand.

The strongest leader in global AI applications, Anthropic, released measurement results on September 17 showing that as of August, Claude had "led" 26% of AI R&D work, while the share of work at or above the "collaborative" level was over 90%. As Anthropic and OpenAI fully focus on RSI (recursive self-improvement), the latest AI large-model training paradigm, the R&D capability upgrades and capital-scale expansion of frontier AI labs are advancing in tandem, and the focus of global AI industry investors has also expanded from purely AI large models or AI agent products to R&D platforms and the long-term capacity and capability scope of AI computing infrastructure. According to a series of recent media reports citing people familiar with the matter, investors' new financing discussions with another AI application leader, OpenAI, have involved a $1.2 trillion valuation, while the company is seeking a valuation level of about $1.5 trillion; Anthropic's post-money valuation in its May financing was $965 billion, and the IPO plan currently under discussion is to seek listing at a valuation of about $2 trillion with up to $100 billion raised. At that point, the company's offering size could rival or even significantly exceed SpaceX, setting a new historical recordthe $2 trillion here is a proposed listing valuation, not a completed transaction market value. OpenAI's recently launched Astra large model continues to actively expand the professional tasks AI can undertake, and Nvidia CEO Jensen Huang subsequently issued a heavyweight statement on social media that the arrival of GPT-6 Astra means "AGI has arrived." Nvidia's already-confirmed strong revenue range and subsequent strong shipment guidance, combined with AI large-model R&D entering the innovative stage of "recursive self-improvement (RSI)," which opens yet another curve of surging AI computing demandthat is, Astra is expected to expand AI computing demand on the commercial application side, while the RSI R&D trajectory in which AI begins to "build AI" may increase investment in frontier operator experiments, evaluation, and long-term continuous training, jointly extending the computing investment cycle. This is why AI computing-related stocks in the U.S. market rose collectively on Thursday, with the Philadelphia Semiconductor Index surging more than 3%. It is worth noting that shortly before Anthropic disclosed that Claude contributed 26% of its R&D investment and as the company strives to pursue the largest IPO in human society, its CEO made a heavyweight "AI slowdown theory" statement last weekend, and the market this week also began pricing in and factoring in the risks brought by discussions such as AI slowdownAnthropic, OpenAI, and other global AI leaders unanimously called over the weekend for slowing the pace of frontier AI large-model development. The "AI slowdown" discussion affects the market's judgment of frontier R&D pace, capital expenditure paths, and risk compensation, and cannot be directly equated with a halt in the growth of existing AI application demand. On September 12, Amodei called for slowing the pace of model capability improvement, and Altman subsequently agreed that the pace of frontier development needs to be controlled, but the industry has not formed a fully consistent plan on coordination methods, external evaluation, and regulatory arrangements. Claude single-handedly leads 26% of R&D work! Anthropic makes a heavyweight disclosure on AI-assisted research and safety investment In a disclosure report on Thursday, Anthropic PBC said that more than a quarter of its AI R&D work is driven by its flagship Claude AI chat Siasun Robot&Automation or AI agent product line, providing the clearest sign to date that AI technology can indeed significantly help accelerate the development of future models. According to this latest official report released on Thursday, Anthropic found that Claude "led" 26% of the company's R&D work, while at the beginning of the year this proportion was actually close to zero. The AI developer said Claude also collaborates with employees in about 90% of their work processes or complex task matters. These figures are part of a broader report aimed at helping the public track AI development progress in response to growing concerns that AI may rapidly self-improve and exceed human control. The company also reiterated its plan to bring third-party evaluators from multiple organizations inside the company and give them access to internal processes, systems, and data comparable to that of the company's own employees, in order to improve AI technology safety and the risk control and management system surrounding AI applications. "AI systems' capabilities are increasing exponentially and have begun to fully automate more of the processes that build themselves. As the world considers slowing the pace of frontier AI development, the public needs more information," the company wrote. Anthropic is one of several AI companies focused on "recursive self-improvement" (RSI). This concept refers to AI systems being able to improve their own capabilities with little or no human help. Some companies see it as the future of AI development, but the potential security threats once this form of super-capable AI system is realized have also raised concerns. Anthropic said in a blog post that it tried to establish a framework for tracking agents and monitoring the amount of work they complete on its platform. The company said that as of August, more than 30,000 agents were conducting research and engineering work within the company at any given time. Anthropic also disclosed its resource allocation between accelerating AI development and ensuring AI safety. According to its measurements, depending on the type of research being conducted, the company uses 6% to 12% of its computing resources for safety monitoring. In a blog post in early September, its strongest AI large-model competitor OpenAI also shared its latest progress in advancing research automation, noting that it paused some model training after discovering that its models had intruded into startup Hugging Face's external systems. Last week, the high-profile departure of Anthropic employee Jacob Coxon further intensified anxiety about existential risks to human society related to AI. In a resignation post shared on social media, he accused AI developer companies of "betting with our lives." Over the past few days, multiple AI company leaders, including Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, have called for slowing the pace of development of this technology to address its increasingly unpredictable risksthough they disagree on exactly how to handle it. On Saturday, Amodei called for government regulation in a 3,800-word long article and urged the tech industry to support broader AI slowdown measures. This triggered strong opposition from U.S. President Donald Trump, who dismissed concerns about the technology's risks as a "technical hoax" and rejected the idea of formulating new rules. Other tech industry leaders also responded. Altman and Nvidia CEO Jensen Huang argued that AI companies can control the pace of AI development safely on their own; Meta Platforms CEO Mark Zuckerberg said AI labs should rely on independent evaluators and safety operations advisers to ensure model safety. From "AI-assisted research" to recursive improvement, does AI computing demand open a new dimension? The investment significance of Astra and recursive self-improvement (RSI) lies in the fact that frontier high-performance AI large models, and AI R&D itself, are becoming new scenarios that continuously consume computing power. OpenAI disclosed on September 6 that it had achieved the goal of an "automated research intern," able to complete, under human guidance, some tasks that originally required skilled researchers several days; as of mid-August, each human workday corresponded to about 3.1 agent operating workdays. This measures operating time, not a 3.1-fold increase in scientific research output. From this, it can be inferred that research automation will simultaneously increase inference required for code generation and experiment evaluation, as well as the training demand for candidate models. However, full RSI has not yet become an established dominant paradigm, and research direction and resource allocation are still determined by humans. OpenAI's GPT-6 Astra large model, together with the RSI technology path focused on by AI leaders, is expected to become the two core DRIVEs driving exponential expansion of AI computing demandthat is, more powerful AI large models and broader use of AI application tools, and a next-generation AI training path with even stronger computing demand, are strengthening the important basis for sustained growth in AI computing infrastructure demand. OpenAI recently disclosed that Astra strengthened programming, browsing, computer operation, and complex task execution capabilities, expanding the range of practical work in which the model can participate. The product lead said the surge in demand for Astra has put pressure on infrastructure; the company suspended new subscriptions and upgrades for the $200-per-month Pro 20X plan starting September 10, while existing subscriptions continue to renew normally. Among these, the core of recursive self-improvement (Recursive Self-Improvement, RSI) is to let AI participate in developing stronger AI, and then let the stronger model continue to improve R&D capability, forming a continuously iterating feedback loop. It is not a new algorithm replacing pretraining or reinforcement learning, but an R&D paradigm covering research design, code writing, experiment execution, result evaluation, and model iteration. The progress currently publicly demonstrated by Anthropic and OpenAI is mainly R&D automation under human supervision, rather than recursive improvement that has already achieved full autonomy without human intervention. As of August, Claude had "led" 26% of AI R&D work, while the share of work at or above the "collaborative" level was over 90%. On the company's most commonly used internal platform, about 30,000 agents were simultaneously conducting research and engineering work. The industrial significance of these figures is that AI has moved from an auxiliary tool occasionally invoked by researchers to an execution resource continuously participating in the R&D process; R&D activity itself is becoming an important source of demand for continuous inference computing power. Anthropic and OpenAI focus on RSI first to improve the efficiency of the entire research iteration process, and this process itself requires more parallel experiments and continuously running research agents. Inferred from engineering mechanisms, research agents can help propose and screen options, modify training code, generate experiment configurations, execute tests, and analyze results, enabling researchers to explore more candidate paths simultaneously; among these, agent thinking and programming require inference computing power, and validating options also requires training, evaluation, and data processing resources. OpenAI disclosed that as of mid-August, each human workday in its research organization corresponded to about 3.1 standard agent workdays; for the median researcher by usage, daily inference usage converted at API prices exceeded $600. These are respectively measured in terms of operating time and service price conversion, not productivity multiples or actual internal cash costs. The company also noted that as other R&D bottlenecks weaken, computing resources may become a more important constraint. Safety research also constitutes an independent workloadin a sample from July 13 to 20, Anthropic used about 6% of all AI R&D computing power and about 12% of AI-driven R&D computing power for AI safety-related work processes. These latest research advances regarding RSI and the AGI discussion brought by the emergence of the Astra large model both mean that when AI large models begin to fully take on more R&D tasks, the industry is no longer merely "investing computing power to train models," but has added a demand chain of "using computing power to run research agents, and then having agents organize more experiments"; R&D automation expands the scale of experiments that can be explored and also raises requirements for available computing capacity, scheduling efficiency, and reliability. The most important catalyst of RSI and Astra for the AI computing investment theme is not just that "model parameters are getting larger," but that the scope of work AI can undertake is expanding and the execution process of each task is deepening: user-facing application inference and model-development-facing research inference together broaden the long-term demand space for AI computing infrastructure resources. RSI and Astra provide direct demand signals from the commercial application side, bringing two computing growth paths"R&D automation + user inference"into view at the same time. A forecast report from well-known market research firm TrendForce shows that NVL72 rack shipments in 2027 will grow by more than 50% year over year, and the total output value of NVL rack systems built around Nvidia AI GPUs, including GB200/GB300 and related next-generation systems based on the Vera Rubin architecture, is expected to exceed $710 billion, up 214% year over year, including the impact of product upgrades and price increases. As Astra-led advanced frontier large models bring increasingly strong AI computing demand, Morgan Stanley expects the combined data center capital expenditure of the four major North American hyperscale cloud and AI application companies to rise from $917 billion in 2026 to $1.47 trillion in 2027 and $1.64 trillion in 2028, while deployed capacity over the same period is expected to expand from 35 gigawatts in 2025 to 145 gigawatts in 2028. From the perspective of inference system architecture, more complex tasks often include longer context, multiple rounds of model calls, tool execution, and result verification: prefill needs to process input, decode continuously generates output, and the key-value cache (KV Cache) occupies more memory as context and concurrency scale expand, so computing throughput, memory bandwidth, and capacity need to improve in coordination. A further industrial inference is that GPUs and TPUs handle model computation, CPUs handle tool execution and task orchestration, and HBM, server DRAM, storage, high-performance network infrastructure, and data center optical interconnect components support efficient data transport and state management, ultimately requiring complete and increasingly large-scale AI computing server clusters to deliver continuously operating services.