OpenAI's Jalapeño chip beats NVIDIA on inference efficiency by up to 104x

Claude Fable 5.1 leads SciUniverse benchmark with 45% pass rate on real lab automation tasks

A new benchmark from C5R measures frontier models' ability to operate physical scientific equipment across chemistry, biology, and materials science. Claude Fable 5.1 achieved 45.3% accuracy on Level 1 tasks at $40.61 per attempt, outpacing GPT-6 Astra (32.5%, $52.37) and Claude Opus 5 (30.5%, $46.31). The benchmark tests end-to-end scientific work including sample preparation, instrument control, protocol adaptation, and real measurement interpretation—not just downstream analysis. Models repeatedly failed on physical reasoning: attempting to pipette frozen samples, vortexing open plates, and reusing pipette tips across contamination-sensitive wells. This signals AI systems are transitioning from theoretical scientific capability to real-world lab automation, though significant gaps remain in physical task execution.

AI infrastructure investment drove 92% of US GDP growth in H1 2025 despite construction workforce constraints

Data center buildout has become the primary engine of US economic growth. Jason Furman's analysis shows information processing equipment and software investment accounted for nearly all GDP growth, with non-AI GDP expanding at only 0.1% annually. Construction employment and real wages both rose since mid-2022, with construction wages outpacing national average growth even after Trump deported 13% of the undocumented construction workforce. Goldman Sachs estimates 500,000 additional construction jobs will be needed by 2030 to sustain the infrastructure buildout. This underscores how dependent current macroeconomic stability is on data center expansion offsetting policy uncertainty.

Nebius acquires Inferize for $100M-$150M to optimize GPU utilization and reduce inference costs

Nebius, the Nasdaq-listed AI cloud company, has acquired Israeli startup Inferize to address a critical inefficiency in AI inference: GPU idle time during model loading and demand fluctuations. Inferize, founded by former Granulate executives and operating in stealth for 10 months with 17 employees, developed technology to reduce cold-start delays and improve GPU utilization when model weights are loaded or updated. The acquisition reflects competitive pressure to improve AI economics as compute costs remain a critical bottleneck. Nebius is simultaneously building out software layers around its expanding infrastructure through acquisitions: Tavily ($275M initial, up to $400M with performance targets) for AI agent search, and Eigen AI and Clarifai's core team for inference optimization. The company is also expanding physical presence in Israel with data center capacity agreements and selection for a national AI supercomputer.

Multi-agent swarms trade token efficiency for speed with 3-5x returns per 10x agent increase

Research on swarm scaling shows multi-agent systems require roughly 2x total tokens versus single agents but complete tasks in half the time through parallelization. Diminishing returns follow an economic "stepping on toes" parameter similar to human team coordination dynamics. This introduces a new scaling dimension beyond compute and data, with implications for RSI-driven intelligence explosions and the economics of distributed AI reasoning.

Harvey's Legal Agent Benchmark reveals gap between component accuracy and real-world legal work quality

Harvey has open-sourced a 1,200+ task legal agent benchmark (LAB) spanning 24 practice areas and evaluated by 75,000+ expert-written rubric criteria. The benchmark mirrors real law firm workflows: agents receive loose instructions, navigate client matter files, and produce reviewable work product graded against partner-level scrutiny. Leading models like Muse Spark 1.2 achieve 94.5% accuracy on individual legal criteria but only produce acceptable finished work 3.8-25.4% of the time, revealing a significant gap between component performance and end-to-end deliverables. Despite this limitation, law firms report substantial efficiency gains: Harvey users report 89% can take on more work, Legora users cite 42% winning new business, and Laurel Employment Law increased weekly demand letters from 48 to 104 after deploying an AI drafting agent. While only 0.6% of Claude queries involve legal tasks (ranking 38th of 718 occupations), the benchmark provides a shared foundation for measuring progress on long-horizon legal agents and identifying where human review remains essential.

DeepMind proposes agentic economies for autonomous science with proof-of-ideation and fractional lab access

DeepMind argues that as AI reasoning capabilities improve, physical resources and empirical validation—not compute—will constrain AI-accelerated science. The paper proposes market mechanisms where AI agents trade access to labs and equipment based on risk-adjusted research proposal valuations, including proof-of-ideation, ex-ante forecasting, and fractional licensing. The framework anticipates infrastructure and governance challenges of vastly expanded scientific output, requiring deterministic APIs, smart contracts, and cryptographically verifiable reputation for multi-agent coordination across institutions. This reframes autonomous scientific discovery as an economic problem of allocating scarce physical resources among competing agents with divergent objectives, rather than a purely cognitive one.

Richard Ho on OpenAI's Jalapeño: Architecture, Design Process, and Hardware-Software Co-Optimization

More Than Moore published a detailed interview with Richard Ho, OpenAI's hardware lead, on the company's custom inference accelerator Jalapeño. Main takeaways:

• Jalapeño achieves 1.5x-1.9x better performance-per-watt and 1.7x-3.6x better latency vs. NVIDIA GB200/GB300, with up to 104x advantage at operating points NVIDIA struggles to reach
• OpenAI designed a single balanced chip for entire inference workload (prefill, decode, speculative decode) rather than splitting across specialized hardware like competitors
• Significant portions of chip design—compute architecture, circuit fitting, kernel optimization—were designed with OpenAI's own models; one attention implementation improved from <1% to ~90% of memory/compute roofline in 40 hours
• Ho's team pulled together talent from D.E. Shaw, Google TPU program, Meta, Amazon; team-first approach prioritized founder mindset and creative problem-solving over predetermined architecture
• Ho emphasizes memory supply and manufacturing capacity as the real industry constraint, not chip design innovation

Data Centers and Taxation: State Tax Policy's Role in AI Infrastructure Location Decisions

Tax Foundation published a report on state and local taxation of data center infrastructure. Main takeaways:

• Sales tax on data center machinery and equipment (M&E) costs a $1B facility ~$58.3M in year one and $9.7M annually; most states exempt this but often contingent on job/investment thresholds
• Tangible personal property taxes on data center equipment exceed 20% of total federal, state, and local tax burden across 12 examined jurisdictions
• Equipment replacement cycles for AI-oriented data centers shortened to 3 years (vs. 5-6 years for traditional cloud), making sales tax exposure on $1B+ annual M&E spending a material location decision factor
• Hyperscaler projects now routinely exceed $5-10B in initial investment; tax structure differences between states can create binary location choices for operators
• States with aggressive data center taxation risk losing major infrastructure investment to competing jurisdictions with more neutral tax treatment

Estimating time spent on work tasks

arXiv published a paper on task-based labor economics and AI exposure measurement. Main takeaways:

• Prior AI exposure studies (Eloundou et al. 2024) weight tasks equally; this paper introduces time-weighted exposure using O*NET frequency data and constraint satisfaction methods to estimate per-task duration
• Time-weighting substantially reshuffles top-25 most-exposed occupations: 11 of 25 shift away from clerical work toward analytical/research roles, widening gaps between least and most exposed
• Massage therapists drop from highly exposed (44% of tasks) to moderately exposed (16% of work time) when accounting for actual time allocation, demonstrating material impact on labor displacement estimates
• Methodology factors task exposure into expected frequency × duration per instance, validated against worker surveys and constraint satisfaction solutions
• Findings suggest prior widely-cited AI labor impact studies may overstate exposure for occupations with high-exposure but low-time-allocation tasks