Anthropic committed to $1.25 billion per month ($15B annually through May 2029) for access to SpaceX's Colossus 1 and expanded Colossus 2 capacity—220,000 NVIDIA GPUs (H100s, H200s, GB200s, GB300s) delivering 300 megawatts of power. The deal, disclosed in SpaceX's S-1 IPO filing, includes a 90-day termination clause making the guaranteed commitment ~$3.75B, but signals major competitive realignment: Anthropic secures training capacity while SpaceX books $45B in revenue ahead of a rumored $2T IPO valuation. Anthropic's CFO stated he spends 70% of his time on compute acquisition, underscoring the constraint-driven economics of frontier AI labs. The partnership also commits Anthropic to exploring AI data centers in space, validating SpaceX's core IPO thesis.
Separately, OpenAI's reasoning model independently solved the 80-year-old Erdős unit distance problem—discovering a better dot-packing configuration than the previously optimal square grid using algebraic number theory, without specific mathematical training. Leading mathematicians verified the result, demonstrating concrete capability gains in mathematical reasoning and proof generation with implications for model reasoning architectures.
AI21 Labs laid off 110 people (61% of staff) with explicit messaging that selling language models is "not a sustainable revenue stream," marking a major competitive signal that pure model licensing is no longer viable as a standalone business. The pivot reflects a broader industry shift where model labs are repositioning around agent products and closed-loop workflows rather than API access. Frontier labs including OpenAI and Anthropic are now signaling they will serve third-party builders at the app layer rather than compete directly, realizing that going down the stack (chips, datacenters, power) and building platforms is more scalable than competing in first-party applications.
DeepSeek made its 75% V4-Pro discount permanent, placing blended inference costs at ~$0.18/M tokens—estimated 3x cheaper than Gemini 3.1 Pro, 12x cheaper than GPT-5.5, and 19x cheaper than Claude Opus 4.7. This materially shifts the inference cost frontier and validates the "intelligence too cheap to meter" thesis, forcing frontier labs to compete on inference efficiency and agent economics rather than raw model capability licensing.
Anthropic is projecting $559M in operating profits for Q2, with quarterly revenue exceeding its entire lifetime revenue to date and profitability arriving two years ahead of schedule. This represents extraordinary growth momentum in the frontier AI market and positions Anthropic as the first major AI lab to achieve profitability while scaling inference and training operations.
Lisa Su forecasts CPU demand will accelerate to 35%+ annual growth through 2031, a dramatic acceleration from historical 3-4% growth, driven by AI inference workloads and agentic AI demand. The projection represents a structural demand shift that could reshape semiconductor economics and competitive positioning across the stack, signaling that CPU-centric inference and agent execution environments are becoming material drivers of semiconductor demand.
Nvidia reported Q1 FY27 revenue of $81.6B (up 85% YoY) with data center revenue of $75.2B (up 92% YoY) and gross margin of 74.9% (up 23.8 percentage points). Operating margin reached 65.6%, up 33.6 percentage points YoY. The results reinforce Nvidia's dominance in AI infrastructure as the primary beneficiary of hyperscaler and enterprise AI capex buildout.
Coinbase announced perpetual-style equity index futures launching June 8, 2026, allowing traders to go long or short on equity sectors using perpetual futures structures (no expiration, funding-rate mechanics) standard on crypto exchanges. This marks the first time a US-licensed exchange offers perpetual-style futures on traditional equity indices, forcing regulators to ratify an offshore-native product model for regulated US markets.
I read the full report. Main takeaways:
- Trump postponed a planned executive order on AI that would have created a voluntary federal pre-release review system for advanced AI models, citing concerns it would slow innovation and harm US competitiveness vs. China
- David Sacks, Trump's former AI czar, directly intervened with Trump to derail the order hours before signing, arguing voluntary review could become mandatory and would hamper the AI race
- Industry pushed back on specific terms: shortening model submission window from 90 to 14 days, shifting review authority from NSA to intelligence community, and preserving voluntary (not mandatory) framework
- OpenAI supported the order's framework; Anthropic and Meta sent lower-level executives instead of CEOs to the planned signing ceremony
- The reversal signals Silicon Valley's effective veto power over AI regulation and Trump administration's prioritization of speed/competitiveness over safety oversight
I read the full essay. Main takeaways:
- Agent Labs (Cursor $29B, Perplexity $20B, Cognition $10B, Sierra $10B) are a distinct business model from Model Labs, prioritizing agents and outcomes over model capability alone
- Model Labs allocate <30% of compute to inference and pay applied engineers 50-70% of research engineer salaries; Agent Labs charge $2000+/month outcome-based pricing with 10x+ margin advantage
- OpenAI and Anthropic are explicitly pivoting to AI Cloud platforms for third-party developers rather than competing via first-party apps, creating structural opportunity for Agent Labs
- Post-training (RL) now rivals pretraining in frontier lab spend; Agent Labs can start from open-weight models and close capability gaps through continued training, as evidenced by Cursor's Ray Summit presentation
- Model diversity (Chinese/US open models + frontier labs) means Agent Labs can "capabilitymaxx" across multiple models, creating demand for specialized model selection and orchestration services
