Key Shifts

  • Chinese AI models rapidly gaining share among US companies: US companies’ token usage on Chinese models via OpenRouter has held above 30% weekly since February, peaking at 46% — compared to an 11% average over the prior 12 months and just 4.5% in H1 2025. DeepSeek and Z.ai’s latest models are closing the performance gap with Anthropic and OpenAI while remaining up to 20x cheaper per token. CNBC

  • FTC issues draft policy statement on suppressing accuracy in AI systems: The Federal Trade Commission proposed that AI companies which steer or deliberately suppress model accuracy for ideological reasons — including to comply with state laws like Colorado’s revised AI Act — may be engaging in deceptive practices under Section 5 of the FTC Act. Public comments are open through July 31. Federal Register

  • Code maintainability plummets in the AI coding era: GitClear and GitKraken analyzed 623 million real-world code changes from 2023–2026 and found duplication up 81%, code reuse down 70%, and legacy refactoring down 74% compared to pre-AI levels. AI-assisted commits now account for 25% of all commits, and the pattern of “create a new package every time” is rapidly accumulating technical debt. LeadDev

  • Microsoft joins cost-cutting trend, shifts to in-house MAI models: Microsoft has begun using its homegrown MAI models for a portion of AI responses in Excel and Word instead of relying exclusively on OpenAI and Anthropic. The move is part of a broader belt-tightening across Amazon, Uber, Meta, and Accenture as AI costs continue to surprise enterprise buyers. TechCrunch

Startup / Product / Platform Radar

  • Claude Cowork expands to mobile and web — from coding tool to AI office coworker: Anthropic launched Claude Cowork on mobile and web for Max subscribers. Tasks started on desktop can be monitored on mobile, and work continues in the cloud even with the laptop closed. Usage data from 1.2 million anonymized sessions shows coding accounts for only 8.7% of usage; business process operations (33.4%) and content creation (16.4%) dominate. TechCrunch · The Verge

  • Savi Security raises $7M seed to protect consumers from AI voice-cloning scams: Founded by brothers Patrick Coughlin (ex-Cisco/Splunk) and Ryan Coughlin (ex-Apple/Spotify) after their mother was targeted by an AI-cloned kidnapping scam. The iPhone and Android app launched Tuesday. Led by Acrew Capital, with Magnify Ventures, TTCER, and Resolute Ventures. TechCrunch

  • Discord AI moderation bug wrongfully bans 8,000+ users: Harmless images — spreadsheets, chessboards, game textures — with grid patterns were flagged as harmful content over two months. A bug caused the system to bypass human review and permanently ban accounts immediately. All affected accounts are being restored. TechCrunch

AI Future Signals

  • Frontier vs. open source: substitution or lifecycle?: Decagon CEO Jesse Zhang posits a new framework: frontier models own discovery of new use cases, while open source models own production for mature use cases. Vercel dashboard data backs this — DeepSeek leads in token volume but Anthropic still captures over half of total AI spend. For startups, the strategic implication is clear: prove with frontier, scale with open source. TechCrunch

  • AI data center energy demand collides with US manufacturing revival: Rust Belt manufacturers in the PJM Interconnection grid region are seeing electricity costs soar as AI data centers consume growing power capacity. The 141-year-old Belden Brick Company saw monthly electricity bills jump from $1,600 to $12,000; steelmaker Metallus reports a 70% increase since 2024, costing an extra $15M per year. The tension between courting data centers and reviving domestic manufacturing is no longer theoretical. Ars Technica

  • First American autonomous ground vehicles deployed in combat: Forterra has been operating 100+ autonomous Lancer ATVs in Ukraine for nine months for logistics and supply transport. Gas-powered with 750kg cargo capacity, they address a critical gap in drone-dominated battlefields where human drivers face constant surveillance and strike risk. TechCrunch

Realistic Opportunities / Experiments

  • Build a frontier-to-open-source pipeline: The pattern of proving new AI capabilities with frontier models and shifting validated use cases to cheaper open source alternatives (DeepSeek V4 Flash at $0.06/M tokens vs. Opus 4.8 at $1.37/M tokens) is spreading across the industry. Startups should design multi-model strategies from day one rather than betting on a single provider.

  • Design AI moderation with mandatory human-in-the-loop checkpoints: Discord’s incident shows how automated moderation can catastrophically bypass human review due to a single bug. Any product handling user-generated content with AI moderation must ensure human review steps cannot be skipped by software faults.

Uncertainties / Keep Watching

  • FTC enforcement trajectory on AI accuracy: The FTC has opened the door to treating ideologically motivated accuracy suppression as a deceptive practice, but specific enforcement criteria and precedents are absent. The potential collision with state-level AI laws like Colorado’s creates a regulatory risk vector that startups in regulated sectors should monitor closely. Comments close July 31. Federal Register

  • Sustainability of Chinese AI model dependence: US companies are flocking to Chinese models on cost grounds, but geopolitical risk and potential regulatory barriers (export controls, security reviews) could intervene at any moment. The OpenRouter trend is unmistakable, but whether this is a structural shift or a temporary cost arbitrage remains an open question. CNBC