California enacts strictest US laws on AI and social media for kids
California just put the country's strictest guardrails between AI chatbots and its children — and the industry now has a live template for what compliance looks like. Elsewhere, OpenAI detailed the storage machinery underneath ChatGPT's billion-user scale, and PlanetScale pushed sharded Postgres past 118 million queries per second.
California Governor Gavin Newsom signed a package of child-safety laws Thursday that together form the nation's tightest restrictions on AI chatbots and social media for minors. The rules ban "addictive" social media features for users under 16 — the kind of infinite-scroll and algorithmic feeds at the center of the design debate — and require AI chatbot operators to run risk assessments before deploying their products. Companies found to have harmed children through their platforms face fines of up to $1 million per child, and families can opt out of school-issued laptops. The package also updates child exploitation law to cover AI-generated sexual imagery of minors, closing a gap that was written for a pre-generative era.
The signing lands in a state that is now regulating AI from two directions at once: earlier this week Newsom signed Anthropic-backed legislation creating a framework for independent safety assessments of AI models and a registry of AI auditors. Civil-liberties groups are pushing back — the Electronic Frontier Foundation calls the under-16 social media restrictions a "functional ban" that cuts young people off from communities and speech, and that fight will now move to the courts and to other statehouses copying the template. With roughly one in eight American children living in California, and no federal law in sight, the practical default for every major platform is to build to California's rules and ship the same product everywhere. Tech companies have promised this fight before; what's different now is the volume — chatbot risk assessments, age-gated feeds, per-child liability — and that each piece has a named enforcement mechanism.
What to watch: whether other states clone the under-16 addictive-features ban before year's end, and which chatbot operator publishes the first risk assessment under the new law.
OpenAI published a rare engineering look inside the storage platform that keeps ChatGPT online — a system now handling more than 70 million requests per second for over 1 billion weekly users across nearly 40 regions. The platform, called Habitat, started in mid-2024 as a simple Python library in front of one database; it now serves more than 500 petabytes of data behind products from ChatGPT to Codex. The interesting part of the post is the sequencing: OpenAI knowingly ran a Python service at a scale where that choice looks wrong, accepting the performance tax to unblock product teams, then bet that its own coding models would make the eventual rewrite cheap. That bet paid — in Q2 2026, two engineers with Codex and GPT-5.5 rewrote the whole service in Rust, which now handles 95% of production traffic and, by OpenAI's numbers, runs 6x more CPU-efficient and 15x more memory-efficient than the Python version. The company says Python gets fully deprecated in the coming weeks.
The story doubles as a data point on how AI is changing infrastructure work itself: the rewrite that used to be a year-long team project became a two-person job with agents in the loop. It's also a rare admission of how close to the edge hypergrowth pushes design decisions — OpenAI says it grew Habitat's load more than 10x year-over-year for three straight years, so the team kept making tactical squeezes while fending off capacity crunches.
What to watch: part two of the series, where OpenAI says it will detail the multi-tenancy reliability work and how it scaled the Azure Cosmos DB layer under that 70-million-request load.
PlanetScale says its new sharded Postgres platform, Neki, sustained 118 million queries per second with 1.22 pebibytes of data across 512 shards. The company's blog post is careful about the asterisks — the run was read-only, primary-only with no replicas, and no failover was tested during the measurement window — but even with those caveats, the number is a statement about where database engineering is heading for AI-scale workloads. PlanetScale built its name scaling MySQL with Vitess; Neki is the same sharding philosophy applied to Postgres, which remains the default database for most new AI application stacks. The 118M QPS figure lands a week after PlanetScale's Postgres launch made noise across the database world, and it's aimed squarely at the same buyers who are discovering that AI-agent traffic multiplies read load by orders of magnitude. The engineering write-up of how they got there is promised in a future post.
What's your read — is California's per-child liability model the fix families have been asking for, or a lawsuit magnet that chills products minors actually need? Tell us in the comments.
