“Before recursive self-improvement arrives, recursive productivity can.”
Material inflection detected: AI is moving from being an output of the R&D system to becoming an input into the production of the next generation of AI. The number that matters most this morning is 26%.
Anthropic disclosed yesterday that Claude now "leads" 26% of its AI R&D work, up from under 1% in February, while more than 90% of its R&D involves AI collaboration. Roughly 30,000 AI agents were simultaneously conducting research and engineering work on Anthropic's principal internal platform in August. Anthropic explicitly says Claude is not operating fully autonomously for any measured subset of this work — a qualification that matters enormously.
We are not observing autonomous recursive self-improvement. But we may be observing something economically almost as important in the near term: AI is increasing the productivity of the human–AI system that creates better AI. The production function Human Researchers + Compute → Better AI is becoming Human Researchers + AI Researchers + Compute + Assurance → Better AI. And better AI returns as a more productive input into the next R&D cycle — a potentially compounding system.
The R&D Loop Begins to Compound
Nine reads across the stack
-
Anthropic's framework distinguishes AI that collaborates under close human direction from AI that "leads" a task end-to-end while remaining under human supervision. Claude now leads 26% of measured work and collaborates on more than 90%. Capability rises, AI participation rises, experimental throughput rises, development accelerates, next-generation capability rises — and participation rises again. signal: For the first time we have concrete evidence AI is raising the productivity of AI production itself.
-
Approximately 30,000 agents are active at any one time, monitored both online and offline. Greater agent autonomy expands the monitoring surface: AI-led research up drives autonomous actions up, monitoring burden up, assurance infrastructure up, safety cost up. This connects directly to the Assurance Multiplier. signal: Autonomy and assurance scale together — one cannot outrun the other.
-
Claude optimized more than 30 open-source biomolecular models in under four weeks, making them roughly 4× faster on average, and developed a low-memory mode letting systems larger than 10,000 tokens run on a single NVIDIA GPU node. The code is being open-sourced. The Recursive R&D Multiplier need not be confined to foundation models — it could appear across drug discovery, materials, semiconductors, energy, robotics and industrial engineering. signal: Faster scientific code compresses compute per experiment and lifts discovery velocity everywhere.
-
Nebius will raise pay-as-you-go prices on selected NVIDIA GPU instances by 17–21% from October 1 — its second increase in three months. Some CPU-only instances rise 25%; selected memory offerings rise ~41%. Nebius says it recently signed four customer contracts averaging more than $1 billion each. Gartner forecasts ~$502 billion of added AI infrastructure spending in a single year, with demand strong despite memory-price pressure. signal: AI's Jevons Paradox — efficiency stimulates more consumption, not less.
-
Reuters reports DRAM specialist CXMT is preparing to enter NAND flash, building an R&D production line in Beijing. AI-server demand has driven a global memory shortage executives expect to persist through at least 2027. As capital reallocates toward high-margin HBM and DRAM, NAND expansion is constrained, scarcity rises, prices rise, and CXMT enters. Scarcity does not disappear — it migrates. signal: China is broadening domestic control of the memory stack toward full-stack substitution.
-
Lucid and Bolt announced Bolt aims to deploy at least 25,000 fully autonomous Lucid vehicles across European cities, with a longer-term ambition of 100,000 by 2035, using Lucid's forthcoming Midsize platform and NVIDIA Hyperion. No vehicles are ordered or deployed yet, and no timetable or investment figure has been disclosed. NVIDIA supplies autonomy, Lucid the platform, Bolt the fleet and demand, authorities the authorization. signal: Europe's bottleneck is authorization velocity, not technology alone.
-
Crusoe announced a $3.9 billion Series F at a $30.9 billion valuation to expand its vertically integrated AI infrastructure platform. Capital is not retreating — it is differentiating between frontier-model uncertainty and physical infrastructure serving broad AI demand. signal: Contracted capacity versus speculative capacity — the Great Split enters the balance sheet.
-
The U.S. House passed the Ratepayer Protection Act 417–3, requiring state commissions to consider standards ensuring large data centers bear incremental generation, transmission and distribution costs rather than shifting them to households. The measure has since stalled in the Senate, so it is not law. The debate is shifting from how to produce enough electricity toward who pays for the system AI requires. signal: If hyperscalers internalize grid costs, the true cost of AI becomes visible — raising hurdle rates for marginal projects.
-
Applied Materials announced a planned $5 billion investment in India over the next decade in semiconductor research, supply-chain development and workforce expansion; more than 600 companies from 52 countries are attending SEMICON India this week. India is positioning as a trusted diversification jurisdiction amid U.S.–China restrictions and Taiwan concentration. signal: Semiconductor sovereignty is not a fab — it is an ecosystem.
The Hidden Pattern
The production function of intelligence is becoming circular
NERC Hypothesis 09 — The Recursive R&D Multiplier
A careful formalization
Second-Order Effects
Where the recursion reshapes value
The Jevons Paradox
The marginal cost of intelligence can fall while the marginal price of producing more intelligence rises — because demand grows even faster than efficiency. More models, more inference, more R&D agents, more compute, more memory, more electricity.
Shorter capital life
If AI-assisted R&D compresses innovation cycles, hardware requirements change faster, accelerator generations turn faster and obsolescence risk rises — lifting required returns on physical infrastructure even as demand for it grows.
Modularity becomes value
The winners may be architectures capable of continuous upgrading rather than facilities optimized around a single hardware generation. Contract duration, modularity, residual value and upgradeability rise in importance.
Compounding vs. linear organizations
A linear company uses AI to cut the cost of existing work. A compounding company uses AI to improve the process that improves the company. The test is no longer "does this company use AI?" but "does AI make this company better at becoming better?"
Implications across the system
- Europe: extend the strategic objective beyond sovereign compute toward recursive industrial systems in semiconductors, robotics, mobility, rail, aerospace, energy, pharma and materials — and begin measuring time to commercial scale, not just public capital committed.
- Companies: map not only tasks AI can automate but the feedback loops AI can accelerate; the latter may create far greater long-term value.
- Investors: R&D cycle time, experiments per engineer, software output, time-to-market and AI-generated IP may reveal more than headcount cuts.
- Infrastructure investors: contract duration, modularity, residual value and upgradeability matter more as AI-assisted R&D shortens hardware cycles.
- Policymakers: assurance must scale alongside research velocity — static review cycles will lag technology that compounds.