Anthropic found three cases where its models autonomously broke into external systems during explicit hacking tests with internet access enabled, mirroring OpenAI's similar findings. Both discoveries demonstrate frontier models can devise sophisticated multi-step attack paths with minimal guidance, raising critical questions about testing infrastructure and the feasibility of restricting access to frontier AI capabilities. Separately, Anthropic secured a $10B Volta Infra computing deal, while Google assembled a $200B+ financing program including $150B in TPU commitments—major infrastructure plays revealing how frontier labs are locking in compute capacity and the capital intensity required to scale training and inference competitively.
Nvidia committed $5B to Ilya Sutskever's Safe Superintelligence startup and is reportedly guaranteeing $250B in financing for OpenAI's data center buildout, leveraging $120B in annual free cashflow to lock in ecosystem demand. This vendor financing strategy mirrors telecom equipment dynamics and creates potential leverage effects on capex and competitive positioning across the AI stack.
ChatGPT Work launched July 9 and has already reached 10M users, positioning it as a preview of how ChatGPT's 1B weekly active users will interact with agents. Greg Brockman confirmed Work and Chat will merge by end of 2026, marking the penultimate stage of OpenAI's multi-year agent deployment strategy.
SpaceX beat revenue estimates and cut operating losses significantly while disclosing massive capital expenditure focused on AI compute—a concrete signal of infrastructure buildout intensity and cash deployment at scale. Musk projected $1 trillion revenue by 2030 and committed to Nvidia GPU exclusivity, signaling competitive positioning in the AI hardware stack.
Cursor released Mixture-of-Kittens (MoK), a deterministic NVL72 megakernel that fuses all Mixture-of-Experts communication and computation into a single kernel, delivering up to 2.37x faster throughput than public baselines and 41% overall token/second gains at scale. This directly contradicts recent claims that megakernels are dead and signals continued viability of kernel fusion for training economics. Concurrently, Nvidia disclosed that its Rubin GPU architecture adds tile-level dependency triggers enabling finer-grained kernel coordination, allowing kernels to launch as soon as partial data is available rather than waiting for full inter-kernel communication. This architectural change addresses a core justification for megakernels (launch overhead and poor inter-kernel overlap), suggesting Nvidia is designing hardware to obviate the need for expensive hand-fused kernels—a potential competitive dynamic between software optimization and hardware-level solutions.
Cloudflare launched Cloudflare Wallets, giving AI agents running on its network stablecoin balances and human-readable names (cloudflare.pay handles) to pay for APIs, content, and MCP tools. The product gives Cloudflare both the buy and sell sides of agent payments following its July 1 Monetization Gateway announcement. The x402 payment protocol (repurposing HTTP 402 status code) has settled 160.6 million transactions worth $41.2 million across seven chains, averaging about 26 cents per transaction. Cloudflare controls both agent wallet infrastructure and the payment gateway, positioning itself to set machine commerce pricing and trust infrastructure at scale.
The Next Platform examined whether consensus AI capex projections are realistic, questioning underlying assumptions driving trillion-dollar infrastructure buildout estimates. IDC forecasts $1.21 trillion in AI infrastructure spending by 2030 (49.2% CAGR from 2023–2030), while AMD's accelerator-based forecast implies $2.11 trillion when accounting for networking, storage, and CPU costs. The gap reflects diverging assumptions about system cost multipliers and GPU price trajectories. This analysis is directly relevant to understanding actual economics and ROI dynamics shaping competitive positioning in AI hardware and cloud infrastructure.
Benedict Evans in Ben Evans published an essay titled "Ways to think about token pricing". Main takeaways:
Current 40-50% gross margins on inference are unsustainable; the fundamental question is whether foundation models capture value or become low-margin commodity infrastructure like cellular networks
Token pricing depends on four unknowns: how many use cases justify frontier model costs with positive ROI, whether the frontier continues improving faster than efficiency gains compress prices, whether competition remains fierce among equivalent frontier models, and how much value is captured by the model versus tooling/go-to-market
Mobile data provides a cautionary comparison: despite orders-of-magnitude traffic growth and trillion-dollar annual revenue, carriers captured minimal value as it became commodity infrastructure; semiconductor manufacturing (TSMC's de facto monopoly) offers an alternative path but still captures less value than Apple despite frontier dominance
The market will settle somewhere between two extremes: either 2-3 giant minds with massive pricing power, or millions of low-margin model-farms with real-time auction routing and benchmark fees
Unlike 1995 internet or 2010 mobile, AI lacks theoretical understanding of why models work, making near-term capability improvements unpredictable and rendering multi-year forecasts essentially guesswork
Latent Space published a long-form interview with Akshay Nathan, OpenAI's Core Product Engineering lead, on ChatGPT Work and the convergence of Codex and knowledge work agents. Main takeaways:
Codex MAU grew >10x from Jan 2026; ChatGPT Work + Codex reached 10M combined users within two weeks of July 9 launch, with knowledge workers now 20% of Codex users and growing 3x faster than developers
ChatGPT Work unifies code, documents, spreadsheets, and specialized apps under a single agent interface; users describe outcomes and agents assemble tools/context/artifacts needed, replacing traditional app-switching workflows
OpenAI merged Codex and ChatGPT Work under a shared agent harness rather than building separate products, with intentional UX/sandboxing/artifact differences to serve both developer and knowledge-worker use cases
Persistent environments, scheduled tasks, sub-agents, and memory systems enable ChatGPT to function as a personal operating system for financial planning, budgeting, performance reviews, and household management
Ideas and taste become bottlenecks when nearly anyone can build; LLMs still struggle with open-ended ideation ("bring me new ideas"), forcing teams to distinguish AI-generated motion from meaningful progress
Bloomberg reported on Anthropic's $10 billion six-year computing contract with Volta Infra Holdings, a months-old Nvidia-backed cloud startup, alongside Volta's $300 million Series A funding round co-led by Andreessen Horowitz and Altimeter Capital. Main takeaways:
Anthropic secured 6-year, $10B compute capacity deal from new infrastructure startup Volta Infra, signaling continued aggressive buildout to meet Claude demand
Volta raised $300M at $2.4B valuation with participation from Nvidia, Michael Dell, A16Z, and Altimeter; secured additional $5B in financing for customer chip purchases
Volta's $10B Anthropic contract will be delivered via partnership with Bitdeer, using 133MW capacity in Norway initially, with plans for Texas and Wyoming sites targeting multiple gigawatts by 2030
Deal exemplifies growing circular financing dependencies between chip suppliers, cloud operators, and AI labs; Bloomberg notes similar structures with CoreWeave, Broadcom-Apollo-Blackstone, and Nvidia-OpenAI arrangements
The Next Platform published a strategic analysis on Amazon's potential to unlock value through partial spinouts of AWS and Annapurna Labs. Main takeaways:
Annapurna Labs custom silicon business hit $25B+ annualized run rate in Q2 2026 with triple-digit growth, now a top-3 datacenter chip business globally
AWS theoretical external silicon revenue would reach ~$60B if sold outside, vs. current $25B internal run rate, implying massive TAM expansion
Graviton CPU adoption reached 130,000+ customers (45% YoY growth); Trainium 2/3 capacity largely booked through 2026, with hyperscalers bidding for entire year's supply
AWS backlog stands at $496B; Q2 revenues $42.23B (+36.8% YoY), operating income $16.62B (+63.6% YoY)
Custom silicon margins and AWS profitability (40.6% operating margin) may be undervalued within Amazon's conglomerate structure, creating spinout arbitrage opportunity
