AI news5
Nvidia is in talks to invest in Perplexity at a valuation above $30 billion, as the AI search startup's revenue triples
Nvidia is negotiating an investment that would value Perplexity at more than $30 billion, over 50% above its previous round. Perplexity's annualised revenue has tripled to more than $750 million since the start of 2026, up from under $250 million, growth the company attributes largely to Perplexity Computer, an AI agent product that consumes far more compute per query than traditional search.
the-decoder.com ↗Why it matters
a chipmaker investing directly in one of its own biggest customers blurs the line between supplier and stakeholder, and a 40x revenue multiple shows how much investor appetite still exists for AI-native products with real usage behind them.
Microsoft's AI cloud business is heavily concentrated in a handful of customers, with OpenAI alone driving about 70% of AI revenue
A Bloomberg analysis found OpenAI accounts for roughly 70% of Microsoft's total AI revenue, with ByteDance's TikTok, Adobe, Perplexity, and Sierra rounding out the next tier of top spenders. TikTok's share of Azure OpenAI revenue has fallen from around 25% toward 15% as Microsoft's customer base has diversified, but the overall picture remains a small number of accounts driving most of the growth.
bloomberg.com ↗Why it matters
Microsoft's AI narrative to investors depends heavily on a handful of relationships it doesn't fully control, which is a real concentration risk if any one of them (especially OpenAI) shifts spend elsewhere.
Xiaomi unveils its own 3nm AI mobile chip, aiming directly at Qualcomm and MediaTek
Xiaomi's new Xring O3, built on TSMC's 3nm process, packs a 10-core CPU (two cores up to 4.35GHz), the first mobile support for LPDDR6 memory, and an NPU rated at 200 TOPS. Xiaomi says it beats its own prior chip by 31-61% on Geekbench, and the chip debuts in the Xiaomi 18 Fold and Pad 9 Pro Max in September.
androidheadlines.com ↗Why it matters
every major phone maker building credible in-house silicon narrows the market for Qualcomm and MediaTek and signals that on-device AI performance is now a competitive battleground, not just a cloud one.
Google's A2A agent-communication protocol formally joins the Agentic AI Foundation alongside Anthropic's MCP
The Linux Foundation-directed Agentic AI Foundation announced on 20 August that Google's Agent2Agent (A2A) protocol has moved under its umbrella, sitting alongside Anthropic's Model Context Protocol. The Foundation has grown from 49 to more than 250 members in under a year, with platinum backers including AWS, Anthropic, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI.
axios.com ↗Why it matters
rival labs consolidating their competing agent standards under one neutral body makes it more likely a single interoperable protocol stack wins, rather than developers having to pick a side.
Anthropic says Claude can now design working proteins, with lab tests confirming a 27% hit rate
Anthropic reported that Claude autonomously designed de novo protein binders that were then synthesised and tested in a wet lab, achieving a 27% experimental hit rate across most targets, in some cases beating human expert designers. The result comes out of Claude Science, the AI workbench Anthropic launched in June, and is Anthropic's most concrete evidence yet that a language model can do real, physically-validated scientific work rather than just literature synthesis.
anthropic.com ↗Why it matters
a hit rate that beats trained scientists on a wet-lab task, not a benchmark, is a genuinely different class of claim than most "AI does science" stories, because someone actually checked it against physical reality.
AI in the nonprofit sector1
Gates Foundation opens a call for US colleges to test AI-driven institutional transformation, concepts due 11 September
The Gates Foundation is inviting broad-access US higher-education institutions (community colleges, public regional universities, minority-serving institutions) to submit concepts for AI transformation work, with concept submissions due 11 September 2026 by 5pm PDT. The Foundation expects to fund up to eight institutions at roughly $150,000-$200,000 each over a nine-month learning partnership running Fall 2026 through Summer 2027, explicitly framed around organisational transformation rather than just funding tool adoption.
ailearninggrants.org ↗Why it matters
this is a funder betting that the hard part of AI adoption is organisational change, not access to the tools themselves, and the shared-learning-agenda structure means whatever these eight institutions find out is meant to be reusable by the wider sector.