noticias radar ia chatgpt hugging face publicidad ia

Monday's radar was about small models beating the big ones; Tuesday's is about who gets paid for them. As of yesterday, ChatGPT is showing ads in Spain and thirty other European countries — only on the free and Go plans — and Hugging Face, the repository where much of the open AI you might already be using lives, has hired a bank to sound out buyers for over $13 billion. In between, Claude went down for hours right during the first block of the European morning, and Alibaba raised $10.2 billion for its own AI stack and released generated video at five cents per second.

ChatGPT now shows ads in Spain: if you sell to businesses, your buyer won't see them

Starting yesterday, Monday, OpenAI is displaying advertising inside ChatGPT in 31 European countries, including Spain, in the largest expansion of its ads business to date. The ads appear below the response, marked as sponsored and visually separated, and are only seen by those using the Free plan or the 7.99 euro per month Go plan: Plus, Pro, Business, Enterprise, and Education remain ad-free. Targeting starts off as contextual — approximate location, device type, time, and language — and is personalized only if the user opts in. Advertisers do not gain access to conversations, and OpenAI claims that advertising does not alter responses; for now, ads are purchased through their sales team, agencies, and tech partners, with self-service coming later. For your business: the important thing is who is on the other side. If you sell to consumers, you have a new channel exactly where people compare alternatives before buying, and it’s worth checking if your category already has ads. If you sell B2B, the conclusion is the opposite and saves you budget: your buyer is almost always on a paid plan and won't see a single ad of yours, so your only way to appear in that conversation remains organic — having the model cite you — and that involves working with content that answers specific questions, not bidding. And one obligation that does fall on you no matter who you are: since August 2, European Regulation requires you to label AI-generated content, and that includes ad creatives. Source

Hugging Face explores $13 billion sale: the warehouse of open AI could change hands

Business Insider reported on Sunday that Hugging Face — the repository where open models are downloaded, featuring over three million public models and nearly a million datasets — has hired a bank to sound out buyers in a deal that would value it at $13 billion or more, nearly triple the $4.5 billion from its last round in 2023. No buyer has been named nor an agreement signed, and the shortlist of candidates is awkward: Nvidia, Google, Amazon, IBM, and Salesforce are already shareholders from that previous round. For your business: this is the flip side of what we discussed yesterday, when an agent built on a small open model outperformed the leaders from OpenAI and Anthropic. If your team has started using open models to lower costs, check exactly where they come from: today, it's common for a deployment to download the model from the repository at startup, and that is a third-party dependency that could end up inside a giant and change conditions, licenses, or pricing. The defense is cheap and done once: keep your own copy of the models and datasets you use in production, with fixed versioning and checksums, and ensure the deployment pulls from your copy and not the Internet. Besides protecting you from a change in ownership, it protects you from the other thing that has already happened there — manipulated models uploaded to the repository. Source

Claude went down right at the start of the morning in Spain: the Plan B isn't another provider, it's a procedure

Anthropic acknowledged widespread errors yesterday affecting several Claude models — Opus 5, Opus 4.8, Fable 5, and Mythos 5 — across web, mobile, desktop, and also the API, with the incident resolved by 8:30 UTC (10:30 AM in mainland Spain). In other words, the outage covered the first few hours of the European workday, and not just the chat: any automation calling that API was returning errors during that period. For your business: the useful response isn't to hire two providers just in case — that doubles cost and work without solving anything if no one has decided what happens when one fails. Make a list of your automated processes and split it in two: those that have a customer waiting (support chat, forms, automatic quotes) need an automatic failover — an alternative model already tested with your own cases or an honest degraded response like "I can't help right now, a person will get back to you in X minutes"; those that don't (summaries, email classification, reports) just need a queue with retries that recovers on its own when the service returns. And measure one thing almost no one measures: how many requests were lost without a trace. If you can't tell, your automation doesn't have sufficient logging. Source

Alibaba raises $10.2 billion and sets generated video at five cents per second

Alibaba announced a share placement in Hong Kong for 80 billion Hong Kong dollars — about $10.2 billion — to fund its "full-stack AI capabilities," from chips to infrastructure, and its shares fell Monday following the news. The same day, it deployed Wan3.0, its video model: up to 30 seconds in a single pass, it accepts text, image, audio, video, and for the first time documents — PDF, PPT, DOC, XLS — as input. On its platform, it costs $0.05 per second at 480p, $0.10 at 720p, and $0.20 at 1080p, roughly six dollars for a half-minute high-definition clip. For your business: the data point that changes things isn't the placement, it's the price. Turning a product page or a presentation into a thirty-second video for the price of a coffee moves video from a "budgeted project" to a "process line," and that affects how you plan commercial content and internal training for the rest of the year. Before putting it into your workflow, two checks that are on you: it is a Chinese provider, so look at the contract to see where data is processed and what happens with international transfers before uploading client material, images of people, or anything confidential; and remember that synthetic video you publish must be labeled as such. Start with a case involving no personal data — catalog, product, how-to — and measure if the generated video converts before replacing anything. Source

What to watch tomorrow?

Nvidia reports earnings: the market expects about $92 billion in quarterly revenue, and that figure is the thermometer for whether the money funding the entire chain — and with it, the price of the tokens you pay for — continues to flow at the same rate. And Anthropic's public prospectus is still pending before the end of August: it would be the first audited accounts of an AI lab.

Watch the 1-minute video