Today's News
Alphabet raises $85B in equity including $10B from Berkshire as tech giants spend ~50% of revenue on AI capex
Benedict Evans
Alphabet's $85B equity raise, including a $10B investment from Berkshire Hathaway, signals that even massive legacy cashflows are insufficient for AI infrastructure buildout. Meta and Microsoft are following similar patterns, turning to capital markets to fund AI compute competition. This reflects the structural economics of AI infrastructure spending and suggests sustained pressure on capex intensity across the sector.
SpaceX's xAI compute rental business annualizes $26B+ revenue as Anthropic rents $1.25B/month capacity
Benedict Evans
xAI's compute rental business has grown to annualize over $26B in revenue following Google's $920M/month capacity rental and Anthropic's $1.25B/month commitment. Over half of xAI's compute is now rented to third parties rather than supporting internal AI development, revealing xAI's current positioning as a compute reseller rather than a frontier model lab. This shift reflects the founding team's departure and Musk's acknowledgment that rebuilding is necessary.
Microsoft signals frontier model competition with renegotiated OpenAI breakup deal and edge compute strategy
Benedict Evans
Microsoft is positioning itself as a comprehensive platform layer—models, tooling, and edge compute via new $3K+ Surface laptops with Nvidia SoCs—while addressing token cost concerns by pushing smaller "good enough" models to edge devices. This represents a structural shift from exclusive OpenAI dependence to direct frontier model competition.
Anthropic dominates AI lab talent competition with 35% interview prep share and 80% two-year retention
The Pragmatic Engineer
Data from interviewing.io shows Anthropic has captured 35% of coaching requests for interview preparation at any single company, far exceeding OpenAI's 16%, reflecting momentum following its $965B valuation and recruitment of Andrej Karpathy. Workforce.ai data reveals Anthropic achieves 80% two-year retention versus OpenAI's 67%, with Google DeepMind at 78%—outperforming FAANG-level companies. AI engineering headcount at Big Tech grew 60% year-over-year versus 7% for software engineering, signaling a structural shift in where top engineering talent concentrates.
Techmeme + AINews
Anthropic released Fable 5 (general availability) and Mythos 5 (restricted) as its first Mythos-class models, achieving 80.3% on SWE-Bench Pro versus GPT-5.5's 58.6%, and 29.3% on FrontierCode Diamond. Pricing is $10/$50 per million input/output tokens—roughly 2x Opus—with 1M context retained. The model implements safety classifiers that trigger fallbacks in <5% of sessions using prompt modification and steering vectors to limit effectiveness for frontier LLM development, representing a major competitive positioning move delivering frontier-class performance at aggressive pricing with built-in safety constraints.
The Next Platform
Broadcom and Marvell are benefiting disproportionately from AI infrastructure buildout through control of critical interconnect and networking components that enable GPU clusters. Broadcom projects $56B in fiscal 2026 revenue and over $100B in AI revenues in fiscal 2027, with Q2 F2026 semiconductor solutions revenue up 78.5% year-over-year driven by Tomahawk, Trident, and Jericho switch ASICs. Marvell, tracking to $11.5B in fiscal 2027 and $16.5B in fiscal 2028, relies more heavily on AI networking than compute. Both companies benefit from hyperscalers' economic substitution strategy: homegrown AI accelerators cost 40% of Nvidia/AMD GPU retail prices and can be rented at 60-70% of GPU rental rates, allowing operators to retain 20+ percentage points for reinvestment.