14 Sep 2025

AI News Digest

🤖 AI-curated 7 stories

Today's Summary

OpenAI is shaking things up with Microsoft, cutting partner revenue shares down to around 8% and potentially freeing up billions in revenue. This renegotiation could have ripple effects on how they both approach cloud services and AI competition. Meanwhile, a photonic chip from the University of Florida might just revolutionize AI’s energy efficiency, slashing power use for convolution operations by up to 100 times, which could be a game-changer for data centers.

In the realm of AI search, Perplexity’s latest funding boost to a $20 billion valuation shows investors are still keen on backing search innovations in the AI space. And on the international scene, the UK and US are gearing up for a tech pact that could reshape AI infrastructure and investment landscapes, with heavyweights like Nvidia and OpenAI in the mix. Also, Coursera’s new AI-driven learning tools are making skill-building more accessible, while Microsoft’s Windows 11 updates are expanding AI features for webcams and improving user interaction through Copilot.

Stories

Report: OpenAI to Cut Partner Revenue Share to ~8% as Microsoft Deal Is Renegotiated

Reuters reports that OpenAI is projecting it will share roughly 8% of its revenue with commercial partners (notably Microsoft) by the end of the decade, down from the current ~20% — a change that could free up tens of billions of dollars of additional revenue for OpenAI. The article says the companies are also negotiating how much OpenAI will pay Microsoft to rent servers, and notes the firms signed a non‑binding memorandum of understanding to revise their partnership terms. Why it matters: this marks a major commercial re‑balancing between two of the AI industry's biggest players, with implications for OpenAI’s path to raise capital or pursue an IPO, Microsoft’s cloud/AI business economics, and broader competition for cloud compute and model access.

Light-powered silicon photonic chip slashes AI energy for convolution ops (≈10–100×)

Researchers at the University of Florida published a study (reported via SPIE/ScienceDaily) describing a silicon photonic chip that uses etched microscopic Fresnel lenses and laser light to perform convolution operations — a core, energy‑hungry building block of many ML models. The prototype achieves near‑zero energy convolution with classification accuracy comparable to conventional chips (≈98% on handwritten digits) and can process multiple wavelengths in parallel (wavelength multiplexing). Why it matters: this is a concrete, experimentally demonstrated hardware advance toward on‑chip optical computation for ML, offering a potential pathway to drastically reduce the electricity footprint of training/inference pipelines and to enable new high‑throughput, low‑latency AI accelerators. Impact: if the approach scales (fabrication, integration with existing electronics, and generalization to larger, real‑world models), it could shift part of the hardware roadmap for AI accelerators and relieve growing data‑center energy pressure.

Read more → ScienceDaily (reporting SPIE / University of Florida research)

EU Joint Research Centre: 'Nine challenges' taxonomy warns limits of current AI benchmarking

The European Commission’s Joint Research Centre (AI Watch) published a meta‑review (to be presented at AIES 2025) identifying nine major problems in current AI benchmarking practice — from weak construct validity and rapid obsolescence to commercial incentives that encourage ‘SOTA‑chasing’ and gaming. The paper recommends clearer task definitions, transparency, dynamic benchmarks that track model drift, and assessments that measure errors and harms alongside performance. Why it matters: as frontier models rapidly outpace static tests, regulators, researchers and purchasers need more robust, policy‑ready evaluation methods to accurately assess capabilities and risks. Impact: the taxonomy and recommendations are likely to influence how academic benchmarks are designed, how procurement/regulatory bodies evaluate models, and how conferences and workshops approach benchmark reliability and update cycles.

Read more → European Commission — AI Watch / Joint Research Centre

Perplexity reportedly closes $200M at a $20B valuation as AI search funding streak continues

Perplexity — the conversational AI search startup behind the Comet browser — is reported to have secured about $200 million in fresh funding that values the company near $20 billion. The raise (reported by multiple outlets and summarized here from TechCrunch's coverage) underscores continued investor appetite for AI search players and gives Perplexity more firepower to scale product development, expand its API and browser ambitions, and compete with incumbents like Google and OpenAI. For the industry, the round highlights how late‑stage capital is still flowing to AI-native consumer search and browser plays, and it may accelerate consolidation and further aggressive moves (partnerships, acquisitions or product bundling) in the AI search space.

UK and US to sign multibillion‑dollar tech pact — AI and data‑centre investments in focus during state visit

Governments in the UK and US are set to sign a major technology agreement tied to AI, semiconductors and telecoms during President Trump’s state visit, Reuters reports. The package is expected to include multibillion‑dollar commitments to build AI infrastructure (including new data centres) and strategic investment — with reports that firms such as Nvidia and OpenAI figures will be involved and that investors like BlackRock may put in roughly $700 million for UK data‑centre capacity. The deal signals deeper transatlantic coordination on sovereign AI compute and supply‑chain resilience, and it could accelerate infrastructure investments that shape where large model training and cloud AI services are hosted — with implications for regulation, national security reviews and vendor partnerships.

Coursera unveils Role Play, Course Builder for partners and new Skill Tracks to teach GenAI and coding skills

At Coursera Connect (Sept 9, 2025) Coursera rolled out a set of AI-powered learning and authoring tools aimed at making AI and coding training faster and more measurable. New features include Role Play — interactive AI personas for practicing soft and job-specific skills — an expanded Course Builder (GenAI authoring tools) for education partners, AI-graded questions and peer review, and four Skill Tracks (GenAI, Software & Product, IT, Data) that map verified learning paths to employer needs. Why it matters: these product moves make it easier for organizations and instructors to create scalable, assessment‑backed AI/coding curricula and help learners gain verified, job‑aligned skills more quickly.

Windows 11 testing expands Copilot Studio Effects to external webcams and adds handy Copilot UI upgrades

Microsoft has started testing Windows Studio Effects on Copilot+ PCs for external USB and docked webcams (rolling out in preview builds), plus other Copilot improvements like an 'Ask Copilot' button in File Explorer and on‑device 'fluid voice' dictation. Why it matters: the updates make built‑in AI camera enhancements and Copilot file/query features usable in more realistic developer/remote‑work setups, improving productivity and accessibility for creators and developers working with multiple monitors and external peripherals.