Microsoft wants to make AI cheaper. Will investors actually see the savings?
THE SO WHAT
Custom chips may lower Microsoft’s unit costs, but the AI P&L is dominated by model complexity, inference load, and how much of that cost is subsidizing growth. Enterprise buyers should expect pricing experiments and bundling games—negotiate based on workload economics, not just list prices or chip narratives.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAnthropic says Claude traced NIST's ECDSA curve seeds to a pseudorandom number generator, challenging decades-old suspicions of an NSA backdoor in the algorithm
If Claude can reverse‑engineer the provenance of NIST’s ECDSA curve seeds, LLMs are now practical tools for auditing foundational cryptography and standards history, not just app code. Security and infra teams should start treating frontier models as first‑pass reviewers on legacy primitives and protocols you’ve long taken for granted.
Applied AI$137 Million Has Been Spent on AI-Generated Ads Ahead of Midterms. Only 27% Included a Disclaimer
Nine‑figure AI ad spend with only 27% disclaimers means synthetic content is now a normalized—but opaque—part of political persuasion. Any consumer‑facing brand using AI media should assume regulatory and reputational scrutiny is coming and get ahead of it with clear labeling and internal approval rules.
Applied AIFive of the web's biggest data-scraping brands now sit under one $3.6 billion umbrella, and AI agents are the reason why
Quadrupled demand from AI agents turning the web into a machine-readable firehose is consolidating scraping into industrial-scale infrastructure. If your product depends on third-party web data, treat providers like core infra vendors — redundancy, contracts, and rate-limit risk now matter as much as accuracy.
Applied AI'Difficult to Overstate' Volatility of White House AI Messaging, Says Greg Allen
Volatile federal AI messaging — swinging between deregulation and existential risk — makes long-dated compliance bets fragile. Treat US AI policy as a moving target and design governance, logging, and model-choice decisions to be reconfigurable on 6–12 month cycles.