noticias radar ia nvidia openai anthropic

Yesterday's radar ended by asking if Broadcom would make a move, and they have: they are negotiating up to $100 billion in debt to finance AI chips. It is a day of financial backstages: Nvidia is paying $7 billion for the technology and a large part of the Poolside team without buying the company, and Anthropic is preparing the largest IPO in history with $65 billion in annualized revenue. The practical news comes from OpenAI: zero data retention on its API for its flagship models, with automated abuse monitoring and no human eyes.

Nvidia pays 7 billion for Poolside without buying it: your AI provider can hollow out without changing owners

Nvidia has signed a $6 billion non-exclusive license with Poolside, the AI coding model startup, for its Model Factory—the platform it uses to build custom models for software development—plus a $1 billion investment at a $12 billion pre-money valuation, four times higher than a year ago. In the same move, 109 Poolside employees received offers to join Nvidia; the founders are staying and insist it "is not an acquisition or an acquihire," but the pattern is the same as seen with Groq: the giant takes the technology and talent, the company continues to exist, and the operation bypasses antitrust controls because formally nobody bought anything. For your company: the takeaway isn't about the stock market; it's about dependency. If your automation relies on a startup—be it for coding agents or anything else—this type of operation means your provider could become half-empty overnight without any regulator warning or the logo changing. Before signing or renewing, three questions in writing: what happens to the product and support if the key team leaves (continuity committed in the contract, not on the corporate blog), what format do you export your data and flows in if you have to leave, and if there is a code escrow or deployment alternative if the service degrades. And the underlying signal: programming AI is consolidating toward infrastructure giants, so build your flows behind a layer that allows you to switch models without rewriting. Source

What you send to OpenAI's API may no longer be stored: zero retention with surveillance without human eyes

OpenAI announced yesterday Zero Data Retention (ZDR) for eligible API customers even for its frontier models: requests and responses are not kept after processing, OpenAI staff cannot see them, and they are not used for training unless explicitly opted-in. The piece that makes this possible is called Private Safety Processing: an automated system that looks for abuse patterns among related interactions and, if triggered, delivers only the alert category without the underlying conversations—with one exception, child sexual abuse material in images, which is still retained for review and reporting. The rollout and a technical white paper arrive in September, and the specialized press reads it for what it is: a direct attack on Anthropic's enterprise clientele, which made privacy its core argument. For your company: this could unlock the cases that your data officer had stalled—payroll, occupational health, legal, financial—because the objection "OpenAI keeps my prompts" is no longer true by default. But with two conditions before moving a single byte of regulated data: "eligible" means it must be requested and reflected in your contract and data processing agreement, it is not active by default; and the surveillance system rolls out in September, so ask in writing exactly what it covers (intermediate logs, subprocessors, GDPR scope) and wait for the white paper before migrating sensitive data. Note: this applies to the API; the browser-based ChatGPT continues with its own rules. Source

The largest IPO in history will be an AI one: Anthropic is already billing 65 billion annualized

Bloomberg reported yesterday that Anthropic expects its IPO to match or exceed SpaceX's record—$75 billion raised, $86.2 billion with the extension—and that the public prospectus could be filed by the end of this month (the confidential one was filed in June). The engine is in the accounts: annualized revenue of $65 billion at the end of July, compared to approximately $10 billion for all of 2025, driven by enterprise spending on Claude and agents. For your company: this isn't investment advice; the utility is different. When the prospectus becomes public, you will have, for the first time, audited accounts from an AI lab: real margins, compute costs, customer concentration. Read it like you read the accounts of any critical provider, because it tells you two things currently only rumored: if the token prices you pay are sustainable or subsidized by venture capital, and how much quarterly pressure your provider will have to monetize enterprise plans from now on. A listed company with analysts breathing down its neck raises prices and bundles features with a different discipline than a startup burning cash; budget your AI consumption with that hypothesis and not with a curve of perpetual price drops. Source

Broadcom makes a move: up to 100 billion in debt so your AI chips are paid for like power plants

A day after losing part of the Google order to Marvell, Broadcom is negotiating with major investment banks for an AI chip financing package that could reach $100 billion: a senior secured tranche of between $60 and $70 billion plus about $30 billion in junior debt, issued by a special purpose vehicle with a partial guarantee from Broadcom, and with Blackstone and Apollo considering participating. The goal is for clients like Anthropic to secure Broadcom chips without paying upfront; the precedent exists: the AI XPV vehicle already raised $35 billion to buy chips and lease them to Anthropic. For your company: the compute capacity your AI subscription runs on is being financed with debt and off-balance-sheet vehicles, just like power plants are financed. Two operational consequences. First: today's cheap token rests on leverage; if credit becomes more expensive, the price and availability of compute will shift, so do not sign multi-year AI commitments without a two-way review clause. Second: provider risk is no longer just technical; it is financial and spans the entire chain—model, cloud, chip, debt—and the defense at your scale is knowing how many tokens each automated process consumes so you can move it to another provider when the time comes, instead of discovering the consumption when the price hike arrives. Source

What to watch for tomorrow?

If Anthropic files the public prospectus before the end of August: these would be the first audited accounts of an AI lab and the benchmark for judging everyone else's prices. And what Nvidia does with Poolside's model factory—if it appears bundled in its enterprise offering, the consolidation of programming AI will have moved from rumor to catalog.

Watch the 1-minute video