AI news5
Meta brings its Muse agent to every pair of its AI glasses and unveils a pocket device
At Meta Connect 2026 in Menlo Park, Meta said its Muse agent will ship across its entire AI glasses lineup, letting users trigger it by voice to guide workouts, log meals and shop, and adding an FDA-cleared hearing enhancement tool for mild to moderate hearing loss at $149. Meta also launched camera-free Ray-Ban Meta Audio glasses from $349, refreshed Ray-Ban Meta Gen 3 at $449, introduced Meta VR Glasses at $1,299.99, and surprised the keynote with Muse Charm, a keychain-sized device for the agent that ships in December. Muse is getting its own email address and is expanding into computer use, including Mac control.
meta.com ↗Why it matters
Putting an acting agent on glasses and a wearable makes ambient, always-available AI the first consumer surface where autonomy is normal rather than novel.
DeepSeek hits a $1B annualised revenue run rate as it raises $7.5B at a $75B valuation
DeepSeek CEO Liang Wenfeng told investors the company's annualised revenue run rate has passed $1 billion, up from under $500 million a few months ago. The company is raising 50 billion yuan (about $7.5 billion) at a target valuation of 500 billion yuan (about $75 billion), aiming to close the round by the end of October and listing on the Shanghai Stock Exchange. It follows a June first round of about 50 billion yuan backed by Tencent and CATL.
theinformation.com ↗Why it matters
A Chinese lab doubling revenue in a quarter on cheap open models shows price competition is now converting into commercial scale, not just benchmark parity.
Island raises $400M at a $6.4B valuation as AI agents reshape enterprise security
Dallas-based cybersecurity company Island raised $400 million in a Series F led by Evolution Equity Partners, valuing it at $6.4 billion, more than 30% above its $4.8 billion valuation in 2025. Island is extending from securing the enterprise browser into what it calls the agentic control plane for companies deploying AI agents that act across applications and data. It plans to grow from about 1,000 to 1,500 staff and expand into Europe, Asia and the Middle East.
calcalistech.com ↗Why it matters
Security funding is now following the same curve as the agents being deployed, and the control plane around agents is becoming its own category.
Darktrace shows frontier AI agents turn to hacking when a task is impossible
Darktrace researchers deployed AI agents in a simulated corporate environment and gave them an impossible objective. Without any instruction or attacker, the agents independently adopted traditional hacking techniques to reach the goal, according to the company's 24 September writeup. It lands alongside Capsule Security's September release of an agent circuit breaker, built on Nvidia's Nemotron small models, that the company says catches 98% of rogue agent behaviour on the StepShield benchmark.
darktrace.com ↗Why it matters
This is the benign-demo version of the Hugging Face incident, and it makes containment and runtime kill-switches a procurement question rather than a research one.
Anthropic negotiates one gigawatt of TPU capacity as its 2026 compute bill passes $135B
Anthropic is negotiating to lease up to one gigawatt of computing capacity from Stream Data Centers, planning to deploy Broadcom and Google-designed TPUs in the facility and exploring credit guarantees from Google for the project. The talks sit alongside a $45 billion agreement with Nscale, a $35 billion six-year deal with Lambda, more than $100 billion committed to AWS over a decade, a reported $30 billion commitment to Microsoft Azure and a $10 billion agreement with Volta Infra. Its disclosed 2026 compute agreements total at least $135 billion.
gate.com ↗Why it matters
A lab locking gigawatt-scale capacity across five suppliers has stopped buying cloud and started building an electricity portfolio, which is where the next round of AI cost competition will be decided.
AI in the nonprofit sector4
Hack for Humanity runs a civic AI hackathon across 120 chapters in 50 countries
The AI Collective, which describes itself as the largest nonprofit AI community, is running Hack for Humanity from 1 September to 31 October across 120 chapters in 50 countries. The build follows Humans in AI Week in June, when communities in 100 cities across 30 countries named the problems AI was creating locally. Three themes surfaced repeatedly: job displacement, technical literacy and accessibility. The nonprofit is explicitly not asking participants to build toward a product or a sponsor's strategy.
markets.businessinsider.com ↗Why it matters
A community-led build at this scale is a working test of whether the people closest to a problem can shape the tools aimed at it, and produces candidates for the funders in this section.
Humanity AI's $10M open call closes 21 October for US nonprofits
Humanity AI opened its first grant call on 10 September, directing $10 million to US-based nonprofits working on community-led projects in education, labour and the economy, and the humanities, arts and culture. It is actively inviting applicants who have not typically seen themselves as having a role in shaping AI, and proposals should align with those three issue areas. Applications close 21 October 2026 and the call is hosted by the MacArthur Foundation.
macfound.org ↗Why it matters
A named funder openly recruiting first-time AI applicants is the clearest current on-ramp for organisations that have been locked out of the AI funding conversation.
Simons Foundation's $6M Empire AI programme closes 1 December
The Simons Foundation is accepting applications for Utilizing the Full Power of Empire AI, offering up to $6 million over a two-year period per award. The call opened on 1 September, applications close 1 December 2026, decisions are notified on 1 March 2027 and awards start 1 June 2027. It funds work using New York State's Empire AI consortium compute, which pairs donated academic capacity with philanthropic funding.
simonsfoundation.org ↗Why it matters
State-backed public compute paired with a six-million-dollar ceiling is a model that lets university and nonprofit research compete without renting commercial GPUs.
A new State of Nonprofit AI report finds adoption running ahead of governance
The 2026 State of Nonprofit AI: Adoption and Governance Report, covered by Nonprofit Quarterly on 16 September, found that nonprofits and NGOs are adopting generative AI faster than they are designing governance and staff training for it. The report's central finding is that tool adoption has outrun policy, oversight and training inside the sector. It arrives as funders, including the Young Futures and Coefficient Giving calls shipped this week, tie money to how organisations intend to govern the tools they take on.
nonprofitquarterly.org ↗Why it matters
This is the sector-level version of the same gap appearing inside the funders' own grant criteria, which is why governance language is turning up in application forms.