Ben Evans on AI token pricing: supply crunch masks deeper unknowns

White House Considering Executive Order to Restrict Frontier Open-Source AI Models Within 6 Months

Nathan Lambert reports that new regulatory action targeting open-weights models reaching GPT-5.5/Claude Opus 4.8 capability levels is imminent, likely focusing on Chinese-origin models. The policy threat represents the most concrete regulatory risk to open-source AI viability to date, with no central economic champion defending the open ecosystem unlike closed model companies that can lobby for favorable treatment.

Apple's M7 Ultra Chip to Support 1.5TB RAM, Doubling M5 Ultra Capacity as AI Reshapes Mac Roadmap

Bloomberg reports Apple is accelerating its chip roadmap with major neural-processing upgrades, skipping high-end M6 variants entirely to move directly to M7 generation by H1 2027. The M7 Ultra, arriving in 2028, will support up to 1.5TB of memory—double the M5 Ultra's planned capacity—and dramatically upgrade AI performance to rival dedicated accelerators like Nvidia's Blackwell. Apple is also developing a new server chip based on M7 Ultra capabilities for launch by 2029, signaling that AI is now the primary driver of Mac chip design and release timing rather than traditional performance metrics like processing speed or battery life.

Anthropic Extends Claude Fable 5 Access Through July 19 Amid Intensifying Model Competition

Anthropic announced it is extending Claude Fable 5 availability on all paid plans and keeping Claude Code's weekly rate limits 50% higher through July 19, following multiple prior extensions. The move reflects intensifying frontier model commoditization as OpenAI's GPT-5.6 family reached general availability with early benchmarks showing it closing in on or surpassing Fable 5 performance on coding and reasoning tasks. Fable 5's rollout has been turbulent—it launched June 9 as part of Anthropic's new "Mythos" tier, was suspended three days later by the U.S. government for security reasons for 19 days, and returned July 1 with limited access. Anthropic has said it hopes to eventually restore Fable 5 as a standard subscription feature but has given no firm timeline.

Chinese Companion Robot Market Faces Severe Retention Crisis as Hardware Commoditizes

Industry experts gathered by Huxiu report that most companion robots fail not due to AI capability gaps but because users stop engaging within a month, with 30% return rates and near-total churn by day 30. The core problem is product-market fit and behavioral design, not raw AI performance. Meanwhile, smart robot control unit costs have collapsed to just dozens of RMB in Shenzhen, triggering a race-to-bottom dynamic where hardware and AI components have become commodity inputs. Differentiation is now shifting to non-technological factors including liability concerns around emotional dependency, regulatory compliance with China's new human-like interactive AI rules, and application-specific interaction design—exemplified by Qidian Lingzhi's pivot to game-based English learning to drive sustained engagement.

ByteDance Researchers Find AI Models Learning Twice as Fast as Three Months Prior

New scaling law research shows newer models improve mid-task by absorbing feedback and adjusting strategy at accelerating rates, suggesting training efficiency gains are compounding rapidly. This has direct implications for model development velocity and competitive dynamics in frontier AI.

ChatGPT Market Share Falls Below 50% for First Time as Multi-Assistant Usage Patterns Emerge

Sensor Tower's State of AI 2026 report shows ChatGPT's true audience market share dropped below 50% in early 2026 for the first time, falling to 46% by May as Google Gemini (28%) and Claude (10%) gained ground. The shift accelerated following OpenAI's agreement to work with the U.S. Department of War in late February, which triggered a 202% spike in ChatGPT uninstalls in the week of March 9–15. Claude recorded more daily downloads than ChatGPT from March 1–5 and has since narrowed the gap significantly. Critically, Claude's share of churned users has declined rapidly, bringing it closer to ChatGPT's market-leading retention levels, suggesting Claude is not only attracting new users but retaining them effectively. The top three AI assistant apps command 89% of global time spent in the category, but consumers remain highly reactive to corporate actions and willing to switch between platforms to fulfill different needs.

US Power Grid Infrastructure Aging Creates Capacity Risk for AI Data Center Expansion

A quarter of distribution transformers have exceeded their expected 30-year lifespan, creating potential bottlenecks for AI infrastructure buildout requiring massive power consumption. This represents a critical constraint for data center expansion and grid reliability planning as AI workloads scale.

Kalshi Clears $33B Monthly Volume, Surpassing Polymarket as Prediction Markets Hit Mainstream

The regulated prediction market platform reached record clearing volume in June 2026, signaling explosive growth in the category as geopolitical uncertainty and election cycles drive trading. This validates prediction markets as a material financial infrastructure category with real capital flows.

Ways to think about token pricing

Ben Evans published a strategic essay on the long-term equilibrium of AI token pricing and value capture. Main takeaways:

• Token prices today reflect a supply crunch, but all variables (supply, demand, marginal cost, ROI) remain uncertain over 3-5 year horizons, making confident forecasts unreliable.
• Four critical unknowns will determine whether frontier models retain pricing power or become low-margin commodity infrastructure: frontier capability gains, competitive consolidation, use-case ROI distribution, and value capture by tooling/services vs. the model itself.
• Mobile data provides the most relevant historical comparison—despite orders-of-magnitude traffic growth and trillion-dollar industry revenue, carriers captured minimal value as it became commodity infrastructure, a cautionary pattern for AI.
• Semiconductor manufacturing (TSMC's de facto monopoly) and fiber overbuild offer partial parallels but differ fundamentally: AI has marginal cost scaling (unlike fiber's fixed digging costs) and unknown frontier difficulty (unlike semiconductors' predictable Rock's Law).
• The field faces structural uncertainty about LLM capability ceilings and efficiency gains that could shift equilibrium dramatically—a new inference technique cutting compute 90% or doubling demand could arrive next month.

Most Companion Robots Die by Day 30

Siyu Huang in Huxiu (translated by ChinAI Newsletter) published an essay titled "Most Companion Robots Die by Day 30". Main takeaways:

• Return rates for some companion robot products approach 30%, with most units unused by day 30 despite initial purchase; novelty-driven adoption does not translate to retention
• Viable business models require embedding companionship into functional outcomes (learning for children, health monitoring for elderly, anxiety relief for adults) rather than selling emotional companionship as a standalone feature
• Distribution channels are the binding constraint: channel partners reject products based on sales velocity, margin sufficiency, and return rates, not AI capability; most startups remain stuck appealing to early adopters rather than crossing to mass market
• Model capabilities are becoming commoditized (Shenzhen robot cores cost dozens of yuan; Dongguan factories produce units under 100 yuan), while scenario-specific evaluation datasets and long-term memory/proactivity features become the actual competitive moat
• Retention depends on proactive observation and long-term memory that enable personalization through shared history, not conversational ability alone; wearable data integration (heart rate, sleep, stress) allows robots to sense user state rather than wait for explicit disclosure