AI news5
Etched doubles its valuation to $21B in a month on a Jane Street-led round
Etched raised $700 million at a $21 billion valuation in a round led by Jane Street, after the trading firm tested and then bought the startup's AI inference hardware. The valuation more than doubled from $10.3B in July, itself up from the $5B stealth-exit valuation covered previously, and Etched says it has signed over $1 billion in customer contracts. The company builds custom low-voltage prefill chips and cluster-scale memory purpose-built for the decode stage of AI inference.
techcrunch.com ↗Why it matters
a pure-play inference-hardware startup quadrupling its valuation since its stealth exit signals investors now see post-training inference cost, not just model training, as the next big infrastructure bottleneck.
Anthropic posts its first profitable quarter, $11.5B in Q2 revenue, ahead of October IPO roadshow
Anthropic reported $11.5 billion in Q2 2026 booked revenue, a 14-fold jump from $787 million a year earlier, with positive adjusted operating income for the first time at a frontier AI lab. The figures are preliminary and use adjusted, not GAAP, accounting; audited statements arrive with the IPO prospectus at least 15 days before an October roadshow. Anthropic cautioned profitability may not hold for the full year as data-center spending ramps in the second half.
forbes.com ↗Why it matters
a frontier lab moving from "burn rate" headlines to a profitable quarter right before an IPO roadshow changes the investor narrative for the whole sector, not just Anthropic.
OpenAI launches ChatGPT for Teens after mounting lawsuits over chatbot safety
OpenAI released a dedicated ChatGPT experience for ages 13-17 with age-appropriate protections on by default, a new Study Mode that gives guided questions instead of direct answers, and parental controls including Quiet Hours and safety notifications. The launch follows multiple lawsuits tied to teen mental-health harms and comes years after teens began using standard ChatGPT, which has 900 million weekly users.
techcrunch.com ↗Why it matters
litigation risk, not just regulation, is now driving product design at the frontier labs.
Reddit tests turning text posts into AI-narrated short videos
Reddit began rolling out a "Play" mode on 17 August (web) and 18 August (iOS/Android) that converts selected text posts and top comments into short AI-narrated videos with synced on-screen captions, labeled "Real conversation voiced by AI." The test currently covers a limited set of English-language communities while Reddit assesses whether the format helps users engage with long threads.
techrepublic.com ↗Why it matters
a text-native platform betting on AI-generated video is a bellwether for how much of the open web is about to get automatically re-packaged into video.
Unitree's "Superman" robot debuts two days before a record Shanghai IPO listing
China's Unitree unveiled a humanoid robot that can jump 2 meters from standstill and run at 45.6 km/h, developed in just over three months, at the World Robot Conference in Beijing on 17 August. Two days later the company lists on Shanghai's STAR Market after an IPO that raised 6.1 billion yuan (about $905 million) and was oversubscribed more than 8,000 times, a record for the exchange.
aroged.com ↗Why it matters
the world's biggest humanoid-robot maker going public with a record-breaking IPO signals embodied AI, not just chatbots, now has real public-market appetite.
AI in the nonprofit sector1
Stanford HAI and UN agencies launch SimulacraBench to test AI on hard-to-reach-population survey data
Stanford HAI, with the UN Behavioral Science Group, UNHCR, UNICEF, and the World Bank, launched SimulacraBench on 18 August, a NeurIPS competition testing whether AI can accurately simulate survey responses from hard-to-reach populations without releasing the underlying microdata. UNHCR contributed a 4-country, 10,300-respondent Syrian refugee survey; UNICEF a 10-country, 19,800-respondent immunisation survey; the World Bank a 3-country, 6,900-respondent jobseeker survey. Evaluation is closed-loop: code travels to the data, only aggregate scores travel back.
simulacrabench.org ↗Why it matters
a concrete, measurable test of whether AI can extend scarce survey budgets in humanitarian data collection without compromising respondent privacy, backed by real UN operational data rather than a funding announcement alone.