Broadcom lends Anthropic $42B to finance compute infrastructure bet

Broadcom lends Anthropic $42B to finance $125.2B five-year TPU commitment, positioning itself as largest compute supplier by 2027

Broadcom has agreed to lend Anthropic up to $42 billion via convertible notes to finance roughly one-third of a $125.2 billion five-year commitment for Google TPU computing capacity, according to Anthropic's IPO filing. The arrangement locks Anthropic into massive compute demand while making it Broadcom's largest customer in its chip design business by 2027—a relationship that spans compute supply, equipment leasing, and financing. Anthropic disclosed potential conflicts of interest, warning that Broadcom's pricing and hardware decisions could affect its ability to procure sufficient computing infrastructure, and that certain payment defaults could make lease obligations immediately due while limiting access to the financing facility. The deal mirrors Nvidia's strategy of using its balance sheet to boost chip sales and signals the reciprocal spending dynamics that have animated AI skepticism on Wall Street, even as Anthropic targets an IPO valuation near $2 trillion.

37signals stops hand-coding as of March 2026, pivots entirely to AI agents for professional development

Ruby on Rails creator David Heinemeier Hansson announced at Rails World 2026 that 37signals has stopped writing code by hand for professional work, treating manual coding as an exception requiring investigation. The 27-year-old profitable software company is now building native mobile apps (reversing its historical web-first stance), moving backend services to Rust, and relying on AI agents for primary development. This represents a significant signal from a company historically known for code quality and disciplined engineering, suggesting that AI-driven development has crossed a threshold where established software firms are restructuring their entire technical strategy around agent-based workflows rather than treating AI as an augmentation tool.

Decision models emerge as fastest-adopted AI model type, with Jev capturing 13% of Vercel's paid teams in 24 hours

OpenAI, Databricks, and Perplexity all launched decision model offerings this week, positioning this category as a potential next wave after frontier and open-weight models. Decision models are 100x+ cheaper than frontier models while potentially handling 10-15% of current token volume through specialized classification and routing tasks, making them economically significant for the AI stack.

GPT-6.1 Sol cuts costs 30% to $0.72 per task through cache optimization, not token efficiency gains

OpenAI's GPT-6.1 Sol achieves a 30% cost reduction versus GPT-6 Sol ($0.72 vs. $1.04 per task) primarily through improved cache reads and fewer conversation turns rather than raw token efficiency improvements, according to Artificial Analysis benchmarking. At maximum effort, GPT-6.1 Sol costs less than a quarter of GPT-6 Astra ($0.72 vs. $3.26) and 64% less than GPT-5.6 Sol ($1.99), pushing out the cost-efficiency Pareto frontier across all effort levels. The shift signals OpenAI's focus on inference efficiency and cost-per-task metrics as competitive vectors beyond raw model capability.

Recursive Language Models emerge as next frontier beyond reasoning models, with RLM-based harness first to solve ARC-AGI-3

MIT researcher Alex Zhang discusses Recursive Language Models (RLMs) as the next architectural frontier, enabling models to treat their own prompts as objects in external memory for self-improvement and context offloading. An RLM-based harness was the first to solve ARC-AGI-3 before OpenAI's Astra, and the approach is spawning follow-on research like Context Language Models that give models unrestricted control over context management. This signals a shift in how frontier labs are architecting agent systems and harnesses rather than focusing solely on scaling model parameters.

Gemini 4 Argon, FLUX 3, and Solar Mini 4 intensify frontier model competition across text, image, and reasoning

Google released Gemini 4 Argon with revised pretraining and long-horizon post-training; Black Forest Labs launched FLUX 3 with native 4K generation and multi-image reference support; and Upstage released Solar Mini 4 (3B active parameters, 1M context window) at $0.10/$0.40 per-token pricing. Solar Mini 4 scores 24 on Artificial Analysis's Intelligence Index and achieves a new Pareto optimal point for models under 3B active parameters, with particular strength in long-context reasoning (83% on AA-LCR v1.1) but relative weakness in agentic coding (1% on Terminal-Bench 4.0). The rapid release cadence and pricing pressure across modalities underscore intensifying competitive dynamics in frontier AI.

Morpho and Aave credit markets more than doubled while memecoin activity collapsed 39-52% in Arc's first two weeks

Arc's initial casino phase (83K tokens launched day one) rapidly gave way to real credit demand, with $175M borrowed against bitcoin at near-zero rates on Morpho and institutional yield vaults already operational. This concrete shift from speculation to credit markets demonstrates product-market fit for institutional DeFi primitives.

SpaceX's Starship reaches orbit and deploys Starlinks for first time on 14th flight attempt

On its 14th flight attempt, Starship achieved orbital velocity, released 26 Starlink V3 satellites, and demonstrated engine control to leave orbit before splashing down in the Pacific. This marks a critical inflection point for rapid reusability and cost reduction in launch infrastructure.

Broadcom to Lend Anthropic Up to $42 Billion; Becomes Central Infrastructure Partner

Reuters reported on Anthropic's IPO filing disclosures revealing Broadcom's $42 billion financing commitment and deepening infrastructure partnership. Main takeaways:

• Broadcom will lend Anthropic up to $42B to finance infrastructure spending, covering approximately one-third of Anthropic's $125.2B five-year TPU lease commitment with Google/Broadcom
• Anthropic projected to become Broadcom's largest compute customer in 2027, with Broadcom forecasting $115B AI semiconductor revenue in FY2027 and $230B in FY2028
• Financing structure includes convertible notes that could be converted to Anthropic shares; Broadcom can designate a financing partner for the debt instruments
• Anthropic disclosed "potential conflicts of interest" in relying on Broadcom for both hardware supply and financing, with risks around pricing, procurement capacity, and payment defaults triggering lease acceleration
• Arrangement mirrors Nvidia's strategy of using balance sheet strength to boost chip sales, with analyst noting concentrated bet on two companies generating sufficient revenues to support AI financing ecosystem

Recursive Language Models

Alex L. Zhang, Tim Kraska, Omar Khattab (MIT CSAIL) in arXiv published a paper titled "Recursive Language Models". Main takeaways:

• RLMs treat long prompts as external environment variables, enabling LLMs to recursively call themselves over prompt snippets via symbolic code execution rather than loading full context
• GPT-5 RLM achieves 26% median improvement over vanilla GPT-5 on long-context tasks (S-NIAH, OOLONG) while maintaining comparable cost, with performance stable up to 1M tokens vs. severe degradation in vanilla models
• RLM-Qwen3-8B post-trained on 1,000 samples improves 28.3% median over base model and approaches vanilla GPT-5 quality on three long-context benchmarks, demonstrating rapid improvement with minimal training data
• Outperforms context compaction (26% gain), CodeAct with sub-calls (130% gain), and Claude Code (13% gain) on evaluated benchmarks across four diverse long-context tasks

OpenAI's Software Factory: How Codex Became the Backbone of Engineering at Frontier Labs

The Pragmatic Engineer published an essay on OpenAI's internal engineering transformation driven by AI coding agents. Main takeaways:

• Codex adoption surged from ~0% to 90% across non-engineering teams (finance, legal, recruitment) in four months after desktop releases, driven by /goal feature enabling long-running tasks
• CI/CD systems experiencing 10x load increases in six months as PR volume accelerates; traditional code review and pull request workflows being reimagined with agentic review and deployment agents
• IDE usage declining since January 2026 as Codex becomes primary development interface; native mobile app deployment emerging as critical bottleneck in acceleration pipeline
• OpenAI fully dependent on single shared harness; minor outages trigger alerts across entire organization as work flows exclusively through Codex and ChatGPT Work
• Engineering specializations disappearing; judgment, agency, and ability to manage agentic systems becoming primary value drivers over language polyglot expertise