AI news6
Google split its TPU into separate training and inference chips, taking a direct shot at Nvidia
Google announced on Wednesday 22 April that its eighth-generation TPU will ship as two distinct processors: one tuned for training, one for inference. The new inference chip TPU 8i carries 384MB of SRAM (triple the seventh-gen Ironwood) and the training chip delivers 2.8x Ironwood performance at the same price. Anthropic has committed to multiple gigawatts of Google TPUs, and all 17 US Department of Energy national labs run AI co-scientist software built on the chips. DA Davidson estimated the TPU and DeepMind business at $900B in September.
cnbc.com ↗Why it matters
With Amazon, Microsoft, and Meta all shipping custom silicon, the cost-per-token of running frontier models is starting to decouple from Nvidia's pricing power.
Google put $750M behind consultancies at Cloud Next 2026, the largest single hyperscaler partner fund
Google announced a $750M fund on Tuesday 21 April to finance partners' agentic AI development. Accenture has built 450+ agents on Google Cloud, KPMG committed $100M of its own capital, PwC committed $400M, Deloitte described its investment as "largest yet" in any single cloud AI platform, and NTT DATA dedicated 5,000 engineers. The structure is co-investment, training subsidies, credits, and go-to-market funding rather than venture capital. Partners capture up to $7.05 in services revenue for every $1 a customer spends on Google Cloud.
thenextweb.com ↗Why it matters
Enterprise AI competition has shifted from selling cloud infrastructure to financing the consultancies that decide which platform Fortune 500 companies adopt for agent deployments.
Anthropic's Mythos cybersecurity model accessed by unauthorised users for two weeks
Bloomberg reported on Tuesday 21 April that a Discord group has had access to Anthropic's Mythos AI model since 7 April, the same day Anthropic announced limited testing access. The group obtained access through a third-party Anthropic vendor environment and used knowledge from a recent Mercor data breach to make an educated guess about the model's online location. Mythos is Anthropic's most powerful cybersecurity tool, capable of identifying and exploiting vulnerabilities in every major operating system and browser. Official access is limited to Nvidia, Google, AWS, Apple, Microsoft, and select governments via Project Glasswing.
theverge.com ↗Why it matters
Vendor-chain access is now the soft spot in frontier-model security; restricting public release does not stop determined groups when third-party contractors hold credentials.
OpenAI launched ChatGPT Images 2.0 / gpt-image-2 with native reasoning and 2K resolution
OpenAI launched ChatGPT Images 2.0 on Tuesday 21 April via a noon PT livestream, available in the ChatGPT app, the API as gpt-image-2, and Codex. The model adds native reasoning (it plans visibly before generating), 2K resolution output, multi-image consistency across a series, multilingual text rendering that is actually legible, web search to inform images, and multi-image-from-one-prompt. Hands-on testers (VentureBeat, ZDNet, TechCrunch) flagged infographics, slides, maps, and manga as the strongest categories.
community.openai.com ↗Why it matters
This is the first image model from a major lab where infographic and slide-grade output is reliable, which collapses the gap between AI image generation and design tooling.
Alibaba released Qwen3.6-27B, flagship-level coding in a 27B dense model
Alibaba released Qwen3.6-27B on Wednesday 22 April, framing it as flagship-level coding performance in a dense 27-billion-parameter model. The release follows Qwen3.6-Max-Preview earlier in the week which topped six major coding benchmarks, and Qwen3.6-35B-A3B which Simon Willison ran locally on his laptop where it beat Opus 4.7 on his pelican-drawing test.
qwen.ai ↗Why it matters
A laptop-runnable dense model at frontier-class coding performance pushes the cost floor for serious agent-coding work down another tier and weakens the case for closed-API spend on routine coding tasks.
OpenAI launched Trusted Access for cyber defense, scaling defender-side model availability
OpenAI announced a Trusted Access program this week to scale model availability to cyber defenders. The program lands in the same window as Anthropic's Mythos breach and broader concerns about offensive AI capability outpacing defender access. Details on which models are covered and what trust verification looks like are still emerging.
openai.com ↗Why it matters
A defender-access program from OpenAI is the counterweight to the offensive-capability narrative dominating AI-security coverage this month, and signals that lab-mediated access tiers are becoming the working compromise between open release and full lockdown.
AI in the nonprofit sector4
Gates Foundation opened a Grand Challenges RFP for AI in charitable giving
The Bill & Melinda Gates Foundation announced a new Grand Challenges Request for Proposals this week titled "AI to Accelerate Charitable Giving." The call seeks bold and practical projects examining how AI can help donors give more and give sooner. This is the first major Gates Foundation AI-for-philanthropy RFP framed around donor behaviour rather than nonprofit operations.
linkedin.com ↗Why it matters
A Grand Challenges call from Gates carries both funding weight and signal weight; whatever framing wins this RFP shapes how the broader funding sector talks about AI for the next two years.
AWS Imagine Grant UK 2026 opened, deadline 5 June
Amazon Web Services opened the 2026 intake of its Imagine Grant for UK nonprofits, with a deadline of 5 June 2026. The grant funds nonprofits integrating AI and cloud computing into mission-driven work. The Australia/New Zealand version of the same program is also live.
www2.fundsforngos.org ↗Why it matters
AWS Imagine is one of a small number of nonprofit AI grants that explicitly funds cloud and AI deployment costs alongside the project itself, which closes the affordability gap that blocks most small-org AI adoption.
Forbes Nonprofit Council: the social-sector workforce that powers AI adoption needs investment first
A Forbes Nonprofit Council piece on Wednesday 22 April argued that the social sector relies on deep individual commitment but has not built systems that sustain the people delivering on that commitment. The piece frames AI tooling as downstream of the workforce question, not a substitute for it.
forbes.com ↗Why it matters
This reframes the standard nonprofit AI adoption argument: tools alone do not increase capacity if the workforce around them is burning out, which means AI training programmes need to be paired with retention and wellbeing investment to actually work.
Anthropic shared a $3M Claude API credit pool with four Australian research institutions as part of the AU government MOU
Following the Anthropic-Australian Government MOU signed last week, Anthropic confirmed $3M in Claude API credits split across four Australian research institutions including the Australian National University. The credits are tied to research partnerships on AI safety capabilities and risks, not commercial deployment.
linkedin.com ↗Why it matters
This is the first material funding flow into AU research-sector AI work tied to a major US lab, and it sets the template for how other labs may engage with sovereign AI partnerships in mid-tier markets.