AI news5
OpenAI and Broadcom unveil Jalapeño, OpenAI's first custom AI inference chip
OpenAI and Broadcom announced Jalapeño on June 24, OpenAI's first Intelligence Processor: a custom accelerator designed specifically for LLM inference and built in a reported nine-month development cycle that used OpenAI models to accelerate parts of the chip design itself. Early testing shows substantially better performance per watt than current state-of-the-art alternatives. Jalapeño is the first step in a multi-generation compute platform planned for initial deployment by end of 2026, with Celestica handling board and rack integration and Broadcom's Tomahawk networking silicon linking chips at gigawatt data centre scale.
openai.com ↗Why it matters
owning its own inference silicon gives OpenAI the same vertical integration advantage that Alphabet and Amazon already have through TPUs and Trainium. It also signals that the AI compute layer is rapidly consolidating into a small number of companies with end-to-end hardware-software stacks, reducing dependency on Nvidia and reshaping the cost structures that nonprofits and researchers access AI through.
Google makes computer use a built-in tool inside Gemini 3.5 Flash with two new enterprise safeguards
Google has integrated computer use directly into Gemini 3.5 Flash as a native tool alongside code execution, search, and function calling, replacing the previous standalone computer use model. Developers can now build agents that see screens, click, type, and scroll across browsers, mobile devices, and desktops without calling a separate model. Google added two opt-in enterprise safeguards: a confirmation gate that pauses the agent before any irreversible action such as form submission or data deletion, and an automatic halt that triggers if indirect prompt injection is detected. Google's own documentation acknowledges no single safeguard is sufficient on its own, recommending layered defences.
thenextweb.com ↗Why it matters
computer use moving from a standalone model to a tool inside a mainstream fast model lowers the barrier for organisations to build agents that interact with existing software systems without API access. For nonprofits running operations on legacy platforms, this opens a practical automation path that does not require building custom integrations.
Google's AI researcher exodus deepens: two more senior engineers leaving for Anthropic
Bloomberg reported on June 25 that Jonas Adler and Alexander Pritzel, two senior researchers central to Google's Gemini program, are preparing to leave for Anthropic. Adler led Google's AI coding work and Pritzel specialised in training large language models. Their departures follow John Jumper (AlphaFold Nobel laureate, joining Anthropic) and Noam Shazeer (Transformer co-author, joining OpenAI) in the same week, bringing Google's senior AI defections to four within seven days. Google DeepMind CEO Demis Hassabis said talent movement between labs is common in the current market; both Anthropic and OpenAI are approaching IPOs that offer late-stage equity incentives not available inside Big Tech.
telecom.economictimes.indiatimes.com ↗Why it matters
four departures of this seniority in a single week is not normal churn. Anthropic is specifically pulling Gemini pre-training specialists (Adler, Pritzel, Jumper) while OpenAI secured Shazeer for architecture research, suggesting both companies are concentrating on different technical bets. For organisations evaluating long-term model selection, the concentration of pre-training talent at Anthropic increases the probability of Claude pulling ahead on capability within 12 to 18 months.
MIT and Microsoft publish Murakkab, a system that cuts AI agent cloud costs by up to 75%
Researchers from MIT CSAIL and Microsoft Azure Research published Murakkab on June 25, a resource-efficient serving system for multi-step AI agent workloads. Tested on video question-answering and code generation tasks, Murakkab used 35% of the GPU compute, 27% of the energy, and under 25% of the cost of standard cloud deployment methods while meeting user-specified accuracy and latency targets. In one test configuration it reduced energy consumption by more than tenfold with only a 2% accuracy drop. The system works by introducing a declarative abstraction that separates workflow design from execution, enabling cross-layer optimisation that current cloud schedulers cannot perform.
perplexityaimagazine.com ↗Why it matters
agentic AI workflows are materially more expensive to run than single-turn queries, and cost is the most commonly cited reason nonprofits cannot scale AI agent deployments beyond pilots. A 4x cost reduction without accuracy tradeoffs changes the calculus for any organisation currently running agents on cloud infrastructure.
NVIDIA launches BioNeMo Agent Toolkit, giving AI agents PhD-level life sciences tools
NVIDIA announced the BioNeMo Agent Toolkit on June 23, a suite of domain-specific tools enabling AI agents to conduct life sciences research spanning biology, chemistry, genomics, and drug discovery. The toolkit includes NVIDIA Nemotron, NemoClaw, OpenShell, and BioNeMo NIM microservices, and is being integrated by Anthropic, OpenAI, Dassault Systèmes, Databricks, Lilly, Schrödinger, and the University of Washington Institute for Protein Design. Jensen Huang described the toolkit as giving agents the equivalent of a PhD research assistant with access to a supercomputer. It is available now through NVIDIA's developer resources page and GitHub.
nvidianews.nvidia.com ↗Why it matters
the combination of NVIDIA's scientific compute infrastructure with frontier model integrations from Anthropic and OpenAI accelerates the timeline on AI-driven drug discovery and genomics research. For global health nonprofits and research institutions without their own GPU clusters, a cloud-accessible scientific agent toolkit substantially reduces the infrastructure barrier to AI-assisted research.
AI in the nonprofit sector4
Ad Council and Microsoft launch AI chatbot to help parents prevent gun injuries in children
The Ad Council, Microsoft, DEPT, and RSM launched a conversational AI chatbot on AgreeToAgree.org on June 17, extending the "Agree to Agree" youth firearm injury prevention initiative now in its second year. Built using Microsoft Copilot Studio and Azure OpenAI, the chatbot provides personalised guidance to parents and caregivers on having gun safety conversations with children, while also triggering crisis response resources if distress is detected. The initiative addresses what is now a fourth consecutive year as the leading cause of death for US children and teens aged 1 to 17. Among Americans aware of the campaign, 66% have had a gun safety conversation in the past year compared to 47% among those unaware of it.
adcouncil.org ↗Why it matters
this is a textbook responsible-AI deployment in a high-stakes social impact context: narrow scope, human escalation built in, transparent about what the chatbot will and will not discuss, and backed by measurable behaviour change data from the broader campaign. For AIH members looking for case studies on deploying conversational AI for sensitive topics, this is a ready-to-reference public example.
Singapore Red Cross uses Dataiku to automate disaster surveillance and forecast disease outbreaks
Singapore Red Cross partnered with Dataiku under its AI-for-Good Program to automate two humanitarian workflows. First, disaster surveillance across Southeast Asia: SRC previously relied on manual collection, cleaning, and consolidation of data on natural and man-made disasters across the region; automating this improved speed, accuracy, and data quality while enabling the team to layer in climate trends and weather patterns. Second, leptospirosis forecasting in Thailand: machine learning models trained on weather and environmental data now predict outbreaks of the waterborne disease endemic in the country, enabling earlier resource planning for vulnerable communities. Dataiku's AI-for-Good Program provides nonprofit organisations with free platform access and pro bono data scientist support.
techedt.com ↗Why it matters
this is a concrete example of AI moving a humanitarian organisation from reactive to predictive. The dual-use case (surveillance automation plus disease forecasting) from a single platform relationship is the kind of operational efficiency model that is practical for mid-size nonprofits operating in multi-country environments without large data teams.
Windle International uses Microsoft Copilot to deliver hybrid learning to 10,000 students in Kenya refugee camps
Windle International Kenya and Somalia deployed Microsoft 365, Teams, SharePoint, and Copilot across the Kakuma and Dadaab refugee camps to serve roughly 10,000 learners with hybrid in-person and digital instruction. Teachers use Copilot for translation, lesson design, and administrative tasks. In the year since deployment, female enrolment in upper-primary centres rose 14%, addressing early-marriage dropout pressure through flexible learning pathways. In 2025 Kenya Certificate of Primary Education trial exams, participating centres improved mean scores by 7 percentage points. The model is now being considered for adaptation in Bangladesh and Jordan's Za'atari camp.
windowsnews.ai ↗Why it matters
the 14% female enrolment increase is the most concrete gender-equity data point from an AI-assisted refugee education programme published this year. It demonstrates that lowering the logistical barriers to attendance through hybrid delivery has direct impact on who stays in school, which matters for any nonprofit operating in fragile or displacement contexts.
GiveDirectly delivers cash to 2,108 Cyclone Fytia families in Madagascar 18 days after landfall using AI-assisted field tools
GiveDirectly published its full account on June 25 of its emergency response to Cyclone Fytia in Madagascar. Within 24 hours of landfall, GiveDirectly activated a response; within 3 days a field team was deployed to a country where the organisation had never operated before. Over 5 days, local volunteers registered affected families at community sites using tablet-based surveys with an AI-assisted ID verification tool (Persona), with 81% of participants automatically passing eligibility without additional review. First payments went out 18 days after the cyclone, compared to 100 to 120 days for traditional flood responses. 2,108 families received approximately $67 each via Orange Money and Airtel Money.
givedirectly.org ↗Why it matters
18 days versus 100 days is the headline, but the mechanism is the story. AI-assisted ID verification cut the manual review workload that slows emergency cash responses, and the 81% automatic pass rate meant field staff could focus human effort on the 19% edge cases. For disaster-response nonprofits evaluating where AI saves the most time, intake and eligibility verification is the answer in this data.