Nvidia eyes 70% growth, Clearview's new cop tool, and AI agents flooding public services

Something quietly shifted this week - and it has nothing to do with extinction warnings or math prizes. While the AI safety debate dominated headlines, three stories dropped that say more about where AI is actually landing in the real world: in police departments, in government benefit offices, and in the chip supply chains that make all of it possible. Let's get into it.

Nvidia's H100 GPU became the most fought-over piece of hardware in AI history. But roughly how much does a single H100 GPU cost on the open market?

The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇

Jensen Huang appeared this week to lay out why Nvidia expects to grow 70% next year - a number that sounds almost impossible for a company already valued in the trillions. His argument is straightforward: every major AI lab, cloud provider, and government is racing to build infrastructure simultaneously, and Nvidia has its chips in essentially every one of those deals.

Huang was also careful to address a concern that's been circulating: whether Nvidia's deals are circular, meaning it sells chips to companies that then use those chips to make money that flows back to Nvidia. He pushed back firmly, arguing the demand is genuine and driven by real commercial and research applications.

If Huang is right, the AI infrastructure buildout has years left to run - and anyone betting on a slowdown in AI infrastructure spending is going to be waiting a long time.

Read the full story →

Wired has uncovered a previously unreported prototype from Clearview AI called InquiryIQ - and it goes well beyond facial recognition. The tool runs a photo through Clearview's existing face-matching database, then uses a model from xAI - the maker of Grok - to automatically surface a subject's social media accounts, known associates, and other personal details scraped from across the web.

This is meaningfully different from what Clearview has offered before. Traditional facial recognition matches a face to a name. InquiryIQ attempts to build an instant dossier - connections, activity, and identity - from a single image. The tool is still in testing and has not been widely deployed, but its existence raises immediate questions about consent and the scope of police surveillance.

The pairing of Clearview's scraping infrastructure with a large language model changes what's possible in ways that data privacy advocates have long warned about. One search used to return a name. Now it could return a map of your life.

This is the story to watch - not because it's deployed at scale today, but because the prototype exists, it uses commercially available AI models, and the regulatory frameworks to govern it are still years behind.

Read the full story →

Here's a story that flew under the radar this week: AI agents are filing benefit claims and making service requests with government agencies at a volume those systems were never built to handle. TechCrunch reports that public services are seeing a surge of automated requests - and the twist is that researchers say the vast majority of cases involve people entitled to claim something, claiming for that thing. The agents are just doing it faster and at much greater scale than humans would.

That's an interesting wrinkle. The instinct is to assume AI-driven floods of requests are fraudulent, but the evidence so far suggests these agents are mostly acting on behalf of real people with legitimate claims - they're just removing the friction that previously stopped people from following through. Bureaucratic friction, it turns out, was doing a lot of quiet rationing work.

Government agencies are now having to reckon with what happens when AI removes the paperwork barrier entirely. Systems designed for a certain volume of human-paced applications are suddenly receiving automated bursts they can't process. The infrastructure problem is real, even if the intent behind the claims is not malicious.

AI agents removing bureaucratic friction is genuinely good for people who are entitled to support but historically too overwhelmed to claim it - the problem is that governments now have to upgrade their backend infrastructure to match the new pace of demand.

Read the full story →

Universal Music Group announced a multiyear licensing deal with ElevenLabs to launch a new AI music platform. Users will be able to draw from UMG's catalog to create remixes, mashups, and new interpretations of licensed tracks. Artists can opt in to allow their music to be used and will receive compensation when it is.

This is the licensing-forward model the music industry has been slowly moving toward - participating in AI tools on their own terms rather than simply litigating after the fact. The platform is being built directly by both companies together, which is a meaningful contrast to the approach of training on unlicensed data and cleaning it up legally later.

Speaking of building things fast - if you're working on an AI project and need a site up quickly, 60sec.site is an AI website builder that gets you from idea to live site in under a minute. Worth bookmarking.

The music industry spent years suing its way through the streaming era and mostly lost - the fact that labels are now co-building AI platforms with companies like ElevenLabs suggests they've learned something from that experience.

Read the full story →

Here's a genuinely strange situation: OpenAI is reportedly looking into whether antitrust law would prevent AI companies from coordinating a slowdown in development - even if they all agreed it was the safer path. The concern is that competitors agreeing to limit capability development could be read as anticompetitive coordination, which is exactly the kind of thing regulators normally prosecute.

This sits in odd tension with the AI safety debate dominating the past week. Insiders at multiple labs have publicly said they believe their own systems could pose existential risks. But the legal framework governing competition between companies may make it structurally difficult for them to slow down together - even voluntarily. The AI safety conversation keeps bumping into the AI business conversation, and they don't resolve cleanly.

If you're following AI regulation, this is the most underreported tension in the space right now - the safety argument says slow down, the antitrust argument may say you legally can't.

Read the full story →

The answer is around $30,000! At peak demand in 2024 and into 2025, H100 GPUs were selling on the secondary market for roughly $25,000-$35,000 each - sometimes higher. That scarcity is a big part of why Jensen Huang can credibly project 70% growth: there is still no easy substitute, and the waitlists at major cloud providers have been months long.

The story about AI agents flooding government benefit systems stuck with me this week - the fact that most of the claims are legitimate changes the whole framing. So here's my question for you: have you used an AI agent to handle something administrative that you'd previously been putting off - a claim, a complaint, a form? Hit reply and tell me what happened. I read every response, and I'm genuinely curious how many of you are already doing this without thinking twice about it.

That's it for today - catch more daily AI coverage at Daily Inference. See you tomorrow!