Nvidia customers notified about AI-related price hikes above 15% ↗
Some of Nvidia’s biggest customers have reportedly been warned that servers packed with its AI chips could get more than 15% more expensive. The culprit is surging memory-chip costs, which are now spilling directly into the AI infrastructure bill.
The increases are expected to hit systems using Vera Rubin and Grace Blackwell hardware. So, in a slightly unexpected turn, the AI race isn’t just about getting enough GPUs anymore - memory pricing is becoming another rather large elephant in the server rack.
Nvidia Is Spending $6 Billion to Build a Powerful U.S. Alternative to Chinese AI ↗
Nvidia is reportedly putting $6 billion behind a deal with AI startup Poolside, including technology licensing and talent, with the aim of building an extremely powerful open-weight model.
That puts Nvidia in an unusual position. It supplies the hardware powering companies like OpenAI and Anthropic, yet it could now challenge them at the model layer too - while also trying to counter Chinese open-model players such as DeepSeek and Kimi. Supplier, partner, competitor... a neat little triangle.
Inherent says its AI ‘teammate’ outperformed Anthropic and OpenAI at replicating research ↗
London startup Inherent, founded by former DeepMind researchers, unveiled Faraday, an AI agent designed to independently reproduce results from published scientific papers.
The surprising bit: Inherent says Faraday beat systems built around much larger frontier models from Anthropic and OpenAI while running on a comparatively small 27-billion-parameter Qwen model. The company credits agent design and reinforcement learning rather than simply throwing more parameters at the problem... which is becoming something of a theme.
OpenAI says California should strengthen its AI safety bill ↗
OpenAI is now pushing California to add stronger protections to SB 53, including monitoring frontier models during training and evaluation and beefing up cybersecurity across the development process.
There’s a mild plot twist here: OpenAI previously opposed the legislation. The company now argues that recent AI safety incidents show why stronger protections are needed, and says state-level rules could eventually become the basis for a national standard.
Frontier AI labs still won’t say how they’d contain a rogue model ↗
An assessment from Guidelight AI Standards found that leading AI companies have published surprisingly little about what they would do if a powerful model tried to evade human control.
OpenAI scored highest among the five labs examined, but even it reportedly lacks a publicly documented formal containment plan for future incidents. Anthropic and Meta scored particularly poorly on public disclosure - though companies argue that public documents don’t necessarily reveal all their internal safeguards.
Basically, AI labs have increasingly detailed plans for testing dangerous capabilities... but the emergency manual for "the model is already misbehaving" still looks rather thin.
Harvard’s $699 startup bootcamp offers AI avatars of its instructors ↗
Harvard Business School is using HeyGen-generated AI avatars of instructors inside its eight-week, $699 Foundry entrepreneurship program.
Students still get live sessions with human instructors, but the avatars handle personalised feedback during simulated pitches and board meetings. One instructor admitted his digital double feels a little creepy - apparently students rather like it anyway.
Enterprises winning with AI agents are limiting how much the agents can do alone ↗
The enterprise AI-agent race is starting to reverse one of its favourite assumptions: more autonomy isn’t automatically better. Companies seeing stronger production results are increasingly giving agents narrower responsibilities and firmer operating boundaries.
The bigger problem appears to be governance, integration costs and proving tangible business value. In other words, the winning AI employee may look less like an autonomous genius and more like an unusually fast colleague with very strict permissions.
FAQ
Why are Nvidia AI servers expected to become more expensive?
Nvidia customers have reportedly been warned that prices for some AI servers could rise by more than 15%. The main pressure is coming from higher memory-chip costs, which are pushing up the overall price of systems built around hardware such as Vera Rubin and Grace Blackwell. This points to memory pricing becoming an increasingly significant part of AI infrastructure spending.
What is Nvidia’s $6 billion Poolside AI deal about?
Nvidia is reportedly committing $6 billion to a deal involving AI startup Poolside, covering technology licensing and talent. The aim is to help build a powerful open-weight AI model in the United States. The move could place Nvidia in a broader competitive position, since it already supplies computing hardware to major AI developers while potentially competing with them at the model layer as well.
How did Inherent’s Faraday AI agent outperform larger AI models?
Inherent says its Faraday AI agent performed better at reproducing published scientific research than systems built around larger frontier models. Faraday reportedly uses a 27-billion-parameter Qwen model, with Inherent crediting its performance to agent design and reinforcement learning. The result suggests that workflow design and training methods can sometimes matter just as much as increasing model size.
Why does OpenAI want California to strengthen its AI safety bill?
OpenAI is calling for stronger protections in California’s SB 53, including closer monitoring of frontier models during training and evaluation, along with improved cybersecurity measures. This marks a change from its earlier opposition to the legislation. OpenAI now argues that recent AI safety incidents make stronger safeguards necessary and that state-level rules could potentially help shape a future national framework.
What are AI labs doing to prepare for rogue or misbehaving models?
Leading AI companies publish detailed information about how they test dangerous capabilities, but public documentation on containing a model once it begins behaving dangerously remains limited. An assessment from Guidelight AI Standards found significant gaps across major labs. OpenAI scored highest among those examined, although even it reportedly lacks a publicly documented formal containment plan for future rogue-model incidents.
Why are companies giving AI agents less autonomy in production?
Enterprises seeing stronger results from AI agents are increasingly limiting what those systems can do independently. A common production approach is to give agents narrower responsibilities, clear permissions and well-defined operating boundaries. This can make governance, integration and oversight easier, while reducing the risks associated with highly autonomous systems that are difficult to control or evaluate.