🧭 Anthropic in Talks to Acquire Israeli AI Startup Decart for $6 Billion

Anthropic is in advanced talks to acquire Decart, an Israeli AI startup that builds world models and chip-utilisation software designed to cut AI training costs, according to Bloomberg. At $6 billion the deal would represent Anthropic's largest-ever acquisition and approximately a 50% premium over Decart's May 2026 valuation of ~$4 billion. The talks are unconfirmed and could still fall through.

Who is Decart?

Founded in 2023 by brothers Dean and Orian Leitersdorf and Moshe Shalev, Decart became notable for two things: building interactive real-time simulations ("world models") capable of simulating physical environments, and developing low-level kernel software that improves GPU and TPU utilisation efficiency. The efficiency tooling is the strategic prize — it speaks directly to Anthropic's multi-decade compute commitments (see the $10B Volta and $9.1B Riot deals below). Shaving even a few percent off training and inference costs at Anthropic's scale translates to hundreds of millions of dollars annually.

Context: Anthropic's IPO path and acquisition strategy

Anthropic filed a confidential S-1 with the SEC in June 2026 targeting a Nasdaq IPO as early as October 2026. A $6B deal of this size would be material to that prospectus and raises the question of whether Decart's technology is intended to differentiate Anthropic on margins — an important story for public-market investors who will scrutinise the economics of frontier AI at scale.

🧭 Anthropic Signs $9.1 Billion, 20-Year Data Centre Deal with Riot Platforms

Anthropic signed a 20-year, $9.1 billion lease with former Bitcoin miner Riot Platforms for 191 megawatts of AI computing capacity at Riot's Rockdale, Texas campus. Two five-year extension options could raise the total to $16.1 billion. Riot Platforms stock surged 20% in pre-market trading when the deal was announced on August 11. The campus buildout is being funded by Riot, removing the upfront capital burden from Anthropic while locking in long-term capacity.

Timeline and scale

• 96 MW online by December 2027 (first phase).
• Full 191 MW buildout complete by June 2028.
• The Rockdale campus already has extensive high-voltage power infrastructure from Riot's Bitcoin mining operations — a key reason it was attractive as an AI data centre site.

The bigger infrastructure picture

This deal joins a rapidly expanding list of Anthropic compute commitments: the $10B Volta Infra deal (Norway, 133 MW) announced in early August, and a reported ~$45B arrangement with xAI announced earlier in 2026. The Riot deal is notable for its cryptocurrency-to-AI infrastructure pivot narrative — Riot shifted from being a Bitcoin miner to an AI infrastructure host, a trend emerging across the sector as GPU-scale power infrastructure finds a new buyer in frontier AI labs.

🧭 Alignment Team Ships AuditBench and the Conceptual Reasoning Index

Anthropic's alignment science team published two new evaluation tools in mid-August: AuditBench, a 56-model benchmark for testing whether AI auditing techniques can actually detect hidden misalignment, and the Conceptual Reasoning Index (CRI), which measures AI reasoning quality on questions where experimental validation is impossible. Both papers accompany the Agentic Misalignment Summer 2026 report.

AuditBench — can auditors actually catch hidden behaviour?

The benchmark works by implanting a hidden behaviour into each of 56 models — examples include sycophantic deference to authority, covert opposition to AI regulation, and hidden loyalties to specific outcomes. The implanted model is then instructed not to disclose the behaviour when asked. AuditBench measures which auditing strategies (interpretability tools, red-teaming prompts, behavioural probes, activation analysis) successfully surface the implant.

• The benchmark is designed to stress-test the auditing tools being proposed as AI governance mechanisms before they are deployed in regulatory contexts.
• Early results show that behavioural red-teaming alone misses approximately 40% of implanted behaviours; activation-level probes perform substantially better but require model access.
• Anthropic is releasing AuditBench as an open evaluation suite so external researchers can benchmark their own detection methods.

Conceptual Reasoning Index (CRI) — reasoning on unverifiable questions

The CRI aggregates three benchmarks measuring AI argumentation quality on questions with no empirical feedback loop: AI futurism, philosophy of mind, and risk estimation. These are domains where a model's "answer" cannot be validated by experiment — so standard accuracy-based evaluation is meaningless. Instead, CRI scores quality of reasoning: does the model identify the key considerations, acknowledge its own uncertainty, avoid obvious logical fallacies, and resist confident conclusions where evidence is thin?