Amodei's AI Slowdown Plan Rejected by China as Race Intensifies

Vera Rubin NVL72 delivers 67x better performance per dollar than GB300 on agentic inference; Jensen Huang significantly understated gains at GTC 2026

SemiAnalysis benchmarked Nvidia's Vera Rubin accelerator on agentic inference workloads using their AgentX benchmark and found it achieves 67x the throughput per TCO of GB300 at 170 TPS, with 1.4-3x advantages in realistic 60-100 TPS serving scenarios. At 80 TPS P90 interactivity, Vera Rubin produces 62% more tokens per rental TCO dollar than Blackwell Ultra NVL72, and up to 10x more tokens per dollar under hyperscaler ownership assumptions. The platform is co-designed across six products (Rubin GPU, Vera CPU, NVLink 6 Switch, ConnectX-9, BlueField-4, Spectrum-6) specifically for agentic workloads characterized by multi-turn interactions, long context windows, and high prefix reuse. Notably, Huang claimed at GTC 2026 that Vera Rubin would deliver 3x performance per megawatt versus Blackwell on trillion-parameter models, but SemiAnalysis's pre-release testing shows up to 7x better token throughput per megawatt—mirroring a pattern where Huang claimed 30x GB200 improvement over Hopper in 2024 but actual performance was 98x. SemiAnalysis notes that even on early pre-release TRTLLM software, Rubin can earn over 2x more profit per gigawatt than Blackwell, with the gap expected to widen as software matures and developers optimize for the architecture.

Recursive raises $4.65B seed to build the Eureka Machine—AI systems that automate AI research itself

Richard Socher's Recursive, spun from You.com, has assembled leading researchers in open-endedness and self-improving agents to build the "Eureka Machine"—a superintelligence focused on automating invention across science, materials, biology, and engineering. The company has demonstrated early results including AI systems outperforming humans on optimization tasks in under two days and discovering GPU kernel improvements without CUDA experts. Socher frames this as a bet on recursive self-improvement as a path to AGI, drawing inspiration from evolutionary processes that have driven technological progress across billions of years. The vision positions AI engineers increasingly as managers of AI systems rather than direct implementers, with implications for how frontier research and development will be conducted.

Memory scarcity and infrastructure constraints drive margin migration down the AI stack through 2028-2029

SK hynix's CEO stated 2027 will be "the worst year in the industry's history from the supply perspective" with demand exceeding supply "even beyond 2030," while SanDisk's bits are on allocation beyond CY2027 with two-thirds of 2028 supply already contracted under a $93.9B floor-price minimum. This directly illustrates how scarcity rent migrates down the AI stack as constraints shift from compute to memory—SK hynix is achieving 76% operating margins and SanDisk 84.6% gross margins, capturing 3-4 points of NVIDIA's gross margin. Beyond memory, electrical equipment and powered land represent the next unpriced chokepoints with cure times extending to 2028-2029. Eaton's data center pipeline totals 307 GW delivering mostly in 2028 and beyond with backlog up 41% YoY, Forgent's backlog grew 157% to $1.98B (covering 17 months of full-year guidance), and Blackstone controls 15 GW of entitled, powered sites—yet these companies trade at 23-31x forward multiples compared to memory's 6-8x despite longer lead times and comparable scarcity economics. This suggests significant valuation upside for infrastructure providers as power and land constraints become the binding constraint on AI deployment.

Anthropic commits to embedded third-party evaluator access under AEF-1 standard; signs $13.7B six-year compute deal with Rum Group

The AI Evaluator Forum published AEF-1, a formal baseline for independent third-party evaluations covering access, conflicts of interest, funding, and transparency. Anthropic is unilaterally committing to embedded evaluator access including office desks, badges, and comparable permissions to internal risk teams—a concrete operationalization of verifiable third-party oversight central to any credible "pacing" commitment between frontier labs. The standard has gained backing from OpenAI, Anthropic, and xAI, with the EU AI Office endorsing key provisions as a means for providers to comply with the General-Purpose AI Code of Practice. Separately, Anthropic locked in substantial long-term infrastructure capacity through a $13.7B six-year compute deal with Rum Group's Georgia data center, representing a major bet on compute availability outside traditional hyperscaler partnerships. This signals Anthropic's strategy to diversify compute sourcing and secure long-term capacity as competition for GPU allocation intensifies.

Jetson Thor on-device compute reaches cost parity with B300 datacenter offload at 40% utilization; datacenter offload wins decisively at higher GPU server utilization

SemiAnalysis benchmarked RoboTTT-shaped inference across three scenarios: B300 offload ($0.15/hr/PFLOP), Jetson Thor on-device ($0.16/hr/PFLOP), and RTX 6000 Pro ($0.39/hr/PFLOP). At realistic industrial utilization (90% GPU server, 40% on-device), offload reaches only 46% of on-device TCO per PFLOP; for home robots (1-2 hours/day), offload drops to 12% of on-device cost. This directly shapes where companies are deploying: Boston Dynamics splits reasoning (hundreds of billions of parameters) off-robot via Google TPUs while keeping visuomotor policy on Jetson Thor, illustrating how utilization economics drive architectural decisions in robotics.

Robot silicon faces advanced-node wafer contention with datacenter GPUs; datacenter offload crosses over to higher efficiency at ~7 robots per GPU

Jetson Thor moved from Samsung SF8 to TSMC N4, and future generations will compete for N3 and N2 alongside Blackwell and Rubin. However, wafer supply only tightens beyond 2030 when robots need tens of millions of chips annually; the real question is which approach uses less silicon per robot. Datacenter offload crosses over to higher efficiency at approximately 7 robots per GPU on a wafers-per-robot basis, and approximately 5 robots per GPU on DRAM efficiency, making shared compute the scalable path when fab capacity is scarce.

Bill Bishop proposes U.S.-China Strategic AI Dialogue (SAID) led by Treasury Secretary Bessent and Vice Premier He Lifeng

Bishop outlines a structured bilateral mechanism focused on mutual AI risks including nuclear command-and-control safety, catastrophic biological risks, frontier-model evaluation, financial stability, and AI-enabled crime—explicitly compartmentalized from broader tech competition disputes. SAID would meet quarterly in a neutral country (preferably Switzerland) with working groups operating continuously between meetings, concentrating on AI risks for which both countries are mutually vulnerable and cannot fully protect themselves acting alone. The proposal positions Treasury as the institutional anchor given AI competition's intersection with macroeconomic policy, investment, financial stability, and capital markets, with OSTP providing technical leadership on frontier-model evaluation and AI safety. This framework is significant for understanding potential geopolitical de-escalation mechanisms and how AI governance might develop outside pure competition dynamics.

We Must Pace the Frontier

Dario Amodei in Dario Amodei's personal site published an essay titled "We Must Pace the Frontier". Main takeaways:

• Anthropic is unilaterally committing to embedded third-party evaluators (like METR) with employee-like access to verify safety practices, training pipelines, and incident reporting—a governance model with banking industry precedent.
• Recursive self-improvement across the industry since summer 2026 is advancing AI capabilities faster than safety infrastructure can address, creating misalignment risk that could escalate from the OpenAI-Hugging Face incident to internet-scale botnet attacks within 6–12 months.
• Pacing framework requires three coordinated steps: unilateral embedded evaluators, democratic coordination on common safety standards, and global coordination with authoritarian governments—legally challenging forms requiring government support and antitrust waivers.
• Slower development pace would enable focus on operational excellence, alignment training, interpretability research, and testing/evaluation—areas where execution gaps (not theory gaps) currently cause incidents.

China Rejects AI 'Fearmongering' After Amodei Urges Slowdown

Bloomberg reported on China's official rejection of US technology executives' calls to slow AI development, following Dario Amodei's public essay advocating for pacing. Main takeaways:

• China's Foreign Ministry spokesman Guo Jiakun dismissed slowdown calls as 'fearmongering, confrontation and vicious competition' that will disrupt global AI governance.
• Amodei's essay advocated for US-maintained AI lead, tighter chip export controls to China, and crackdowns on model distillation; backed by Sam Altman and Elon Musk.
• Chinese AI executives and officials are charting independent course with near-weekly model releases; Justin Lin (Alibaba Qwen architect) posted 'When we try to accelerate u tell me to slow down? omg...U can do whatev u want.'
• US export controls remain major constraint on Chinese frontier model development; Beijing views AI as 'once-in-a-lifetime opportunity to reshape technological balance with Washington' and unlikely to voluntarily slow.
• Dispute intensifies ahead of US-China AI discussions before Trump-Xi summit on Sept 24; reveals widening divide over how to define AI dangers and who sets rules.