Datacenter Backlash Goes Bipartisan Ahead of 2026 Midterms

SemiAnalysis launches GDPx: frontier agentic benchmark across 11 high-ROI domains; Opus 5.5 leads overall, Astra dominates RSI, Gemini 4 Carbon emerges competitive

SemiAnalysis released GDPx, a successor to OpenAI's year-old GDPval benchmark, covering 11 domains where frontier intelligence will capture the first trillion dollars of AI revenue: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering. Built with 14 specialized RL environment partners (Extrapolate, Arithmetic, Latch Bio, Living Context, Chakra, Third Moment, and others), GDPx emphasizes realistic, interactive environments—tasks like navigating 2000-file deal rooms and debugging large-scale RL runs—rather than GDPval's static, prompt-based tasks. Even the best models score only 12%, making it a true frontier-level benchmark. Opus 5.5 leads overall and wins quant trading, hardware design, and chip design. GPT-6 Astra wins convincingly in RSI and materials science. Gemini 4 Carbon emerges as Google's first top-tier showing, leading in biology and edging out wins in software engineering. Safety classifiers meaningfully suppress Opus and Astra scores in cyber, biology, and alignment domains; Muse Wonder Full actually ranks as the best cyber model when safety constraints are removed. The cost-performance Pareto frontier shows Luna and DeepSeek competing at the low end, Opus dominating the high end, and OpenAI claiming the middle. Finance and legal tasks proved so challenging that essentially every model scored zero, indicating the benchmark's difficulty calibration.

Anthropic reaches $11.6B revenue run rate with operating profitability; 3x compute efficiency gains year-over-year

Anthropic's IPO filing reveals explosive growth from $0.4B to $11.6B revenue while achieving operating profitability, with compute efficiency improving 3x year-over-year. SemiAnalysis described this as "the most efficient scaling of a business in the history of capitalism." The company's $1.5-2T valuation requires significant future growth to justify returns comparable to Google or Facebook IPOs, but the disclosed metrics demonstrate rapid value capture and operational leverage.

ICE and trade[XYZ] plan US-regulated perpetual futures on Hyperliquid covering crypto, equities, commodities, and FX

ICE will operate the regulated market and clearing while trade[XYZ] handles market design, pending regulatory approval. This represents major institutional validation of Hyperliquid as infrastructure for regulated derivatives trading, with Kraken's parent company also planning CFTC-regulated markets on the platform.

Frontier LLMs fail at low-latency tasks: GPT-6 Astra achieves only 36% pass rate under wall-clock time constraints

Realtime-Bench reveals a fundamental gap in frontier model capabilities: GPT-6 Astra achieved the highest performance on real-time interactive tasks but this drops dramatically from 70-80% success when time constraints are removed. Only 36% pass rate on wall-clock conditions exposes a critical weakness in temporal reasoning and adaptive decision-making for AI agents, directly constraining near-term deployment of autonomous systems in time-sensitive domains like trading, gaming, and real-time web interactions.

FBI attributes $6.3B Binance hack to North Korea's TraderTraitor unit, same group behind $1.5B Bybit theft in 2025

Binance resumed withdrawals in phases and saw only 2.1% net outflows ($3.64B) in the first 9.5 hours, suggesting retained user confidence despite the breach's scale.

NYT reports Nvidia CEO Jensen Huang targeted in alleged Chinese intelligence operation

A Singapore-based data center executive with close ties to Huang was identified by US officials as a Chinese intelligence asset potentially steering advanced Nvidia AI chips to Beijing, representing a significant counterintelligence concern around chip supply chain security and executive-level espionage.

The Datacenter Backlash Goes Bipartisan: Next Week's Midterms and the 2028 Buildout

SemiAnalysis published an original essay on the political economy of US datacenter permitting ahead of the 2026 midterms. Main takeaways:

• 2027 datacenter buildout (33.8 GW across top 20 states) is largely locked in with ~75% already under construction; 2028 pipeline (62.2 GW) remains vulnerable to permitting changes as 10 of 13 largest-capacity states elect governors on Nov 3
• Texas TCEQ permit freeze (Sept 21) plus interconnection audit poses largest single risk to US buildout; Texas alone represents ~30% of 2028 GW additions, and prolonged freeze could slip projects 5-9 months if tied to 2027 legislative session
• Datacenter opposition has gone bipartisan: 9 of 20 gubernatorial nominees want full pause; 11 want conditions only; both parties converging on three demands (developers pay grid costs, end tax breaks, limit/report water use)
• If pro-moratorium candidates won every race where running, ~30 GW of 2028 capacity would face exposure; conditions-only winners would expose only 5.7 GW; Texas, Oklahoma, Pennsylvania, Iowa favor conditions-only per polling
• Hyperscalers can absorb higher infrastructure costs but cannot easily shift timelines or relocate projects mid-permitting; moratorium impact depends entirely on project-level permit stage and whether approval is still needed

Anthropic's moat is Claude

Noah Smith in Noah Smith's Substack (noahpinion.blog) published an essay titled "Anthropic's moat is Claude". Main takeaways:

• Anthropic's revenue has grown faster than any company of similar size in history, and it now achieves 3x more revenue per dollar of compute spend than one year ago, described as "the most efficient scaling of a business in the history of capitalism."
• Fortune 500 customers of Anthropic spend less than 0.1% of their expenses on the company, indicating massive headroom for adoption growth despite already representing over half of the Fortune 500 as customers.
• Smith argues intelligence itself—the ability to outsmart competitors—may constitute Anthropic's primary moat, distinct from traditional software moats like network effects or brand loyalty, and that this could justify multi-trillion-dollar valuations if sustained.

Realtime-Bench: Can Frontier Models Solve Real-time Interactive Problems?

arXiv published a paper introducing Realtime-Bench, a benchmark suite for evaluating frontier models on real-time interactive tasks with strict temporal constraints. Main takeaways:

• Best frontier models achieve only 44% success rate and 36% Pass@3 on real-time tasks under actual inference latency (Wall Clock), indicating real-time interaction remains a significant capability gap
• Temporal costs and inference latency directly limit success in evolving environments; higher reasoning effort shows mixed effects under time pressure, revealing a quality-vs-speed tradeoff
• Paper distinguishes three evaluation modes (Wall Clock, Paused Clock, Token Clock) to isolate serving latency from models' temporal awareness, providing framework for diagnosing real-time performance bottlenecks