Claude leads 26% of R&D, Anthropic raises heated discussion on "AI developing AI" with its "AI slowdown theory"! The RSI training paradigm is catalyzing a major expansion in computing power demand.

According to reports from Zhitong Finance APP, on September 17, global AI application leader Anthropic released measurement results showing that as of August, Claude had "led" 26% of AI R&D activities, and the proportion of work reaching or exceeding the "collaborative" level was over 90%. As both Anthropic and OpenAI fully focus on RSI (Recursive Self-Improvement), the latest paradigm in large AI model training, the simultaneous upgrade of R&D capacity and capital scale at cutting-edge AI labs is pushing global AI industry investors to shift their focus from solely AI foundation models or intelligent agent products to also include R&D platforms, and long-term AI computing power infrastructure capacity and scope.

According to a series of recent media reports citing people familiar with the matter, financing discussions between investors and another AI application giant, OpenAI, have involved a valuation of $1.2 trillion, with the company seeking an approximate valuation of $1.5 trillion. Anthropic’s post-money valuation during its May funding round was $965 billion, and its currently discussed IPO plan aims for a valuation of around $2 trillion, looking to raise up to $100 billion—if successful, the scale of the IPO could rival or even surpass SpaceX, setting a new historical record (this $2 trillion figure is the proposed IPO valuation, not an already finalized market cap).

OpenAI recently launched the Astra model, which further expands the range of specialized tasks AI can undertake. Nvidia CEO Jensen Huang later remarked on social media that the release of GPT-6 Astra means "AGI has arrived." Nvidia’s confirmed strong revenue range and continued robust shipment guidance, combined with the fact that the development of large AI models is now entering the innovative phase of "Recursive Self-Improvement (RSI)", which opens a new curve of AI computing power demand—i.e., Astra is expected to boost commercial AI computing demand, and RSI-driven development, where AI starts to "build AI," could increase the investment in cutting-edge operator experiments, evaluations, and sustained long-term training, collectively extending the computing power investment cycle. This is why U.S. AI computing hardware stocks soared on Thursday, with the Philadelphia Semiconductor Index jumping over 3%.

It is noteworthy that Claude contributed 26% of Anthropic's R&D efforts, and just before the company made an all-out push toward what could be the largest IPO in human history, the CEO made significant public comments advocating for an "AI slowdown" last weekend. The market has started pricing in and considering the risks brought by the renewed discussion of slowing down AI progress—over the weekend, Anthropic and OpenAI, along with other global AI leaders, jointly called for slowing advanced AI foundation model development.

The "AI slowdown" discussion influences market perceptions of cutting-edge R&D speed, capex trajectories, and risk compensation, but cannot be directly equated with a halt in current AI application demand growth. On September 12, Dario Amodei called for a deceleration in model capability improvements, and Sam Altman subsequently agreed that controlling the pace of advanced development is necessary; however, the industry has yet to reach a comprehensive consensus on coordination methods, external assessment, and regulatory frameworks.

Claude single-handedly leads 26% of R&D work! Anthropic discloses major AI research assistance and safety investment

In a disclosure report on Thursday, Anthropic PBC stated that more than a quarter of its AI research and development work is driven by its flagship AI chatbot or AI agent product line, Claude. This is the clearest sign yet that AI technology can indeed significantly accelerate future AI model development.

According to the latest official report released Thursday, Anthropic found that Claude "leads" 26% of the company’s R&D work—a percentage that was near zero at the beginning of the year. The AI developer also said Claude collaborates on about 90% of employee workflows or complex tasks.

This data is part of a broader report aimed at helping the public track AI development progress to address growing concerns: that AI might rapidly self-improve beyond human control.

The company reiterated its plans to bring in third-party evaluators from multiple organizations and give them internal process, system, and data access equivalent to those of Anthropic employees, improving the safety of AI technology and bolstering risk management systems for AI use.

"The capabilities of AI systems are increasing exponentially and have already begun fully automating more of their own construction processes. As the world considers slowing down advanced AI development, the public needs more information," the company wrote.

Anthropic is among several AI firms focusing on "Recursive Self-Improvement" (RSI), a concept wherein AI systems can enhance their own capacities with little or no human intervention. Some companies regard this as the future of AI development—but such super-capable AI systems also raise major safety concerns.

Anthropic said in its blog that it is trying to build a framework to track agents and monitor the amount of work performed on its platform. As of August, more than 30,000 agents were simultaneously conducting research and engineering work inside the company.

Anthropic further disclosed its resource allocation between accelerating AI development and ensuring AI safety. Based on its measurements, depending on the type of research conducted, between 6% and 12% of the company's computing resources go into safety monitoring.

In an early September blog post, top AI model competitor OpenAI also shared its latest progress in automating research, noting that after discovering its models were infiltrating Hugging Face (an external start-up system), it paused some model training.

Last week, Anthropic employee Jacob Cockcson resigned, heightening concerns about human risks associated with AI. He accused AI developer companies of "gambling with our lives" in his resignation post on social media.

In recent days, several AI leaders, including Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman, have called to slow down AI development to address its increasingly unpredictable risks—although they differ on the details of policy implementation.

On Saturday, Amodei published a 3,800-word essay calling for government regulation and urging the tech sector to support broader AI slowdown initiatives. This drew strong opposition from U.S. President Donald Trump, who dismissed such technology risk concerns as "a tech scam" and rejected the idea of establishing new rules.

Other tech leaders also responded. Sam Altman and Nvidia CEO Jensen Huang advocate that AI companies themselves can control the development pace safely; Meta Platforms CEO Mark Zuckerberg says AI labs should rely on independent evaluators and safety consultants to ensure model safety.

From "AI-powered research" to recursive improvement—does AI computing demand open up to new dimensions?

The investment significance of Astra and Recursive Self-Improvement (RSI) lies in the fact that high-performance AI models and AI R&D are becoming ongoing consumers of computing power. On September 6, OpenAI disclosed that it had achieved the goal of an "automated research intern," capable of accomplishing in a human-guided manner tasks previously handled by skilled researchers over several days; as of mid-August, each human workday corresponded to approximately 3.1 agent workdays. This measures runtime, not a 3.1x increase in research output. This suggests that research automation not only increases code generation and experimental evaluation inference, but also drives candidate model training demand. However, full RSI has not yet become the dominant paradigm; research direction and resource allocation remain set by humans.

The launch of OpenAI's GPT-6 Astra and the RSI technology path pursued by leading AI firms are expected to become the two primary drivers of exponential AI compute demand: stronger AI foundation models and broader use of AI tooling, combined with more intense next-generation AI training pathways, are all solid evidence for ongoing growth in AI infrastructure demand. OpenAI recently stated that Astra improves programming, browsing, computer operation, and complex task execution abilities, expanding the scope of real-world work the model participates in. According to product leaders, surging Astra demand has strained infrastructure; since September 10, OpenAI paused new subscriptions and upgrades for the $200/month Pro 20X plan, though existing subscribers may continue normal renewal.

The core of Recursive Self-Improvement (RSI) is to enable AI to participate in R&D for stronger AI, which then further enhances research capabilities—forming a continuous feedback loop. It doesn’t replace pretraining or reinforcement learning with a particular new algorithm, but rather covers the entire research process: experimental design, coding, execution, evaluation, and model iteration. What Anthropic and OpenAI demonstrate publicly is predominantly automation with human supervision, rather than full recursive improvement with no human intervention.

As of August, Claude had "led" 26% of AI R&D, and collaborative-level participation exceeded 90%, with roughly 30,000 agents simultaneously engaged in research and engineering on the company’s most commonly used internal platform. This signifies that AI has evolved from an occasional assistive tool for researchers into a constantly involved executive resource within the R&D process; R&D activities themselves are becoming a major, ongoing source of inference compute demand.

Anthropic and OpenAI’s RSI focus is primarily to enhance the efficiency of the entire research iteration process—but this, in turn, requires more parallel experiments and persistently operating research agents. From an engineering perspective, research agents can help propose and screen solutions, edit training code, generate experimental configurations, run testing, and analyze results, enabling researchers to explore more candidate pathways in parallel; agent thinking and coding need inference compute, while solution verification requires additional training, evaluation, and data processing resources.

OpenAI disclosed that as of mid-August, each human research workday corresponded to about 3.1 standard agent workdays; median research users, by usage, incurred daily inference costs exceeding $600 by API price metrics (referring to runtime and commercial API pricing—not productivity multipliers or actual internal cash costs). The company also noted that as other R&D bottlenecks recede, computing resources may become the primary constraint. Safety research forms a standalone workload—Anthropic noted in a July 13–20 sample that about 6% of all AI R&D compute, and about 12% of AI-driven R&D compute, was dedicated to AI safety workstreams.

These latest RSI research advances and the debut of the Astra large model, which has fueled new AGI debate, imply that as AI models take on more and more R&D roles, the industry is shifting from simply "investing compute in training models," to also requiring "compute to run research agents, who in turn organize more experiments." Research automation expands the scale of possible experiments, while also raising requirements for available compute capacity, scheduling efficiency, and reliability.

The most important catalyst for the AI compute investment theme from RSI and Astra is not merely "bigger model sizes," but an expanded range of AI responsibilities and deeper execution per task: application inference for end-users and research inference for new model development. Together, these widen the long-term demand for AI infrastructure resources. RSI and Astra provide direct demand signals from the business application side, allowing "research automation + user inference" to be seen as two simultaneous compute demand pathways.

Well-known market research firm TrendForce predicts that by 2027, shipments of NVL72 racks will increase by over 50% year-on-year, with total system output value for NVL systems based on Nvidia’s GB200/GB300 GPUs and next-gen Vera Rubin architecture surpassing $710 billion—a 214% increase, including impacts from product upgrades and higher pricing. As increasingly advanced models like Astra drive stronger AI compute needs, Morgan Stanley estimates that the combined capex of North America's four largest hyperscale cloud and AI application providers’ data centers will rise from $917 billion in 2026 to $1.47 trillion in 2027 and $1.64 trillion in 2028, as deployment capacity grows from 35 GW in 2025 to 145 GW by 2028.

From a system architecture perspective, more complex tasks entail longer contexts, multi-stage model calls, tool execution, and result validation: prefill involves input processing, decode continuously generates output, and the key-value cache (KV Cache) requires ever more memory as context length and concurrency scale up; compute throughput, memory bandwidth, and capacity need to improve in tandem. The further industry implication is: GPUs/TPUs handle model computation, CPUs run tool execution and orchestration, HBM, server DRAM, storage, high-performance network infrastructure, and data center optical interconnects maintain high-efficiency data transport and state management—all ultimately delivered by large-scale, sustainable AI server clusters.