Morning, folks! AI is moving deeper into the stack today, OpenAI is making its own chips, Hugging Face could be worth $13B, Anthropic is asking outsiders to study AI's impact on wellbeing, OpenAI is automating its own IT desk, and Nvidia is making sure its new AI PCs can actually game. Let's get into it.
Today's Top 5
OpenAI Just Built Its Own Chip. It's Not Ditching Nvidia.: Jalapeño delivers up to 1.9x more AI work per watt and 3.6x lower latency. OpenAI isn't replacing its suppliers; it's taking control of another piece of the stack.
Hugging Face Turned Down $500M From Nvidia. Now It Could Be Worth $13B.: It's reportedly fielding offers at nearly 3x its 2023 valuation. The interesting part isn't the number; it's how valuable the infrastructure around AI models is becoming.
Anthropic Is Paying Outsiders to Study What AI Does to People.: A $5M program will fund independent research into AI and user wellbeing. Anthropic is essentially admitting some of the hardest questions won't be answered inside the lab.
OpenAI Is Quietly Turning Its IT Desk Into an AI Agent.: Its new admin plugin can investigate issues and make authorized changes inside ChatGPT. A similar system already resolves roughly 45% of OpenAI's IT tickets.
Nvidia Wants RTX Spark to Be a Gaming Platform Too.: EA, Ubisoft, and Embark are preparing games for the fall launch, including anti-cheat support. Nvidia knows good hardware isn't enough; the ecosystem has to be ready too.
OpenAI revealed real performance numbers for Jalapeño, its first custom inference chip, and the results are strong, 1.5 to 1.9x more AI work per watt, 1.7 to 3.6x lower latency, tested across OpenAI's own models plus DeepSeek R1 and Kimi K2.5. It went from design to tapeout in just nine months, partly because OpenAI used its own AI to help design it, and some AI-generated chip components actually ran faster than the human-written versions.
This isn't OpenAI trying to cut Nvidia loose, though; their compute stack still runs on Microsoft, Nvidia, AWS, AMD, and half a dozen others. It's about owning one more lever, deployment starts by the end of the year, and Gen 2 is already in the works.
Hugging Face is reportedly fielding offers near $13 billion, according to Business Insider, nearly triple what it was worth in 2023. No buyer's been named, talks are still early. Same pattern as Stripe buying OpenRouter, really. Money's chasing the layer that routes and distributes AI models now, not the models themselves.
Here's the twist though, Hugging Face turned down $500M from Nvidia earlier this year at a $7B valuation. Founder Clément Delangue keeps talking about the company's "long-term responsibility" to its open-source community, which makes you wonder, is this a real sale in motion, or is Hugging Face quietly finding out what it's worth.
Anthropic just launched a $5 million grant program for independent researchers, clinicians, psychologists, or whoever to build open evaluations measuring how AI actually affects user wellbeing. Grantees work on their own and publish everything open source, not locked to Anthropic. Here's the hard part they're naming directly. Advice that's totally fine in one context (diet tips) can be genuinely harmful in another (someone with a history of disordered eating), and that context often doesn't show up until deep into a conversation, not the first message.
Applications close September 21, with funding decisions by October 5. Worth noticing, this comes right after Anthropic admitted there's still no real playbook for AI companionship or mental-health conversations. This grant is basically them handing that problem to people outside the building.
OpenAI launched an admin plugin for ChatGPT Work and Codex, letting admins ask a question, dig into the details, and make an authorized change, all in one conversation instead of switching between five different tools. It's already running inside OpenAI's own IT team, where a similar setup in Slack resolves roughly 45% of support tickets on its own.
Nothing here bypasses existing permissions, though; the plugin just maps a request to an action the admin's already allowed to take and shows exactly what changed afterward. Support volume at OpenAI roughly doubled recently, and instead of the backlog growing with it, it actually disappeared.
NVIDIA is lining up major publishers for RTX Spark ahead of its fall launch, with EA, Ubisoft, and Embark bringing games to the platform. It’s also making sure things like anti-cheat and newer RTX features work properly from day one.
The interesting part is that NVIDIA isn't treating gaming as a side feature. It's trying to make Spark feel like a real PC platform before it ships. That means getting developers and the surrounding software ecosystem on board, not just building a fast piece of hardware.
Other AI Signals:
India's Ringg raised $10M more from Peak XV, now $15.5M total, betting on voice AI moving past simple call center work into things like healthcare appointment booking and KYC checks. It already runs across 1,200 clinics for Practo. Instead of chasing US clients directly, it's partnering with Indian offshore hubs multinationals already rely on.
GTA 6 gameplay clips reportedly leaked online, showing Jason at a nightclub, on the beach, and inside a used-game store called “Game Horde.” Rockstar hasn't confirmed the footage, so for now it's still just another leak making the rounds.
China's bigger robotics story isn't the humanoids making headlines. The country already has more than 2 million industrial robots operating in factories, while schools are training the next generation to work alongside them as the workforce ages and shrinks.
Google is arguing for AI training on publicly available content, with a clear opt-out for creators. Google also says its latest models are 300x more efficient than two years ago, while an Oxford Economics report estimates AI could add up to $6T to global GDP over the next decade. The copyright fight clearly isn't going away anytime soon.
Accelerated Understanding is building AI for physics, not language, using neural operators to predict how things behave across space and time. The company says one query handled 5 trillion data points, with applications ranging from chip design to robotics and weather forecasting.
Today’s AI Tools to Try:
Wan 3.0: Alibaba’s AI video platform for generating and editing high-quality videos with AI.
CanIrun.ai: Checks your hardware and tells you which AI models your machine can realistically run.
Pixel Panda: Turns basic product photos into polished product shots, UGC-style videos, and ad-ready creatives.
Tellie App: A smart teleprompter that follows your speech in real time, even when you pause, skip lines, or improvise.
