AI in 15 — August 25, 2026

Nvidia just paid six billion dollars for a piece of software and a hundred and nine people — and the plan is to give away what they build for free.

Welcome to AI in 15 for Tuesday, August twenty-fifth, 2026. I'm Kate, your host.

And I'm Marcus, your co-host.

Today: what Nvidia actually bought from Poolside, and why it points at Nemotron.

Hugging Face — the neutral ground of open AI — is reportedly exploring a sale at thirteen billion.

Claude designed working proteins for fourteen of fifteen targets, and a lab confirmed it.

Plus OpenAI cuts prices for the third time in a month, Alibaba ships thirty-second video, and a security story about attacking the machine that's running the model.

Marcus, we touched the Poolside deal on Saturday when it was structure. Today it's about purpose. What's new?

Where the technology goes. Reporting now says more than a hundred Poolside people move to Nvidia and work on Nemotron — Nvidia's open-weight model family. The stated ambition is one of the most capable open-weight models in the world, positioned explicitly against DeepSeek, Kimi K3 and Qwen.

So Nvidia is building a model that competes with the models its customers sell.

That's the tension, and it's deliberate. If the model layer becomes free and commoditised, all the value pools in the silicon. Nvidia wins whoever wins. But it does mean the company writing your GPU invoice is also shipping a free alternative to your product.

Why a coding pipeline specifically? Nvidia could have licensed anything.

Because coding is the highest-volume inference workload there is, and it's exactly where open weights currently lose hardest to closed models. If you want open weights to be taken seriously by enterprises, coding is the beachhead. And the backstory explains the price — Poolside was trying to raise around two billion to pay for a forty-thousand-unit GB300 cluster due in January. That raise stalled. So you had a good training stack with no compute to run it on, sitting next to the company that has nothing but compute.

A very expensive coincidence.

Six billion for the software, a billion in equity at twelve billion pre-money, and the shell stays standing. Same pattern Nvidia ran with Groq in December. And I'd note the sourcing here is reporting, not a company statement — that matters for the valuation numbers especially.

Related, and it lands the same week. The Information says Nvidia's in talks to back Perplexity above thirty billion.

More than double where it sat a year ago. The number that justifies it is revenue — annualised past seven hundred and fifty million, up from under two hundred and fifty million at the start of this year. And a meaningful chunk of that is Perplexity Computer, the desktop agent product. Which is a completely different business from answering search queries.

And Perplexity is already in Nvidia's orbit.

Joined the Nemotron Coalition in March. So these aren't two stories. One buys the factory, the other buys distribution for what comes out of it.

Here's my question. Nvidia is now investing in suppliers, model labs, clouds, and applications — all of whom buy chips from Nvidia.

That's the question to hold, and I don't think anyone can answer it yet. It's either the most complete strategic position in the industry or it's a chip vendor financing its own demand. Both descriptions fit the same facts. What would settle it is seeing whether that capital produces revenue Nvidia wouldn't otherwise have had.

Hugging Face. Business Insider reports it's exploring a sale at thirteen billion or more.

Talking to banks to evaluate bids. No buyer named, no deal agreed. Thirteen billion would be roughly triple its four-and-a-half billion valuation from the 2023 Series D — which, note, was backed by Google, Amazon, Nvidia, Intel, Qualcomm, IBM and Salesforce.

Do they even need to sell? That's what confuses me.

Apparently not. Clément Delangue said recently they're close to profitability and have only recently started to touch money raised three years ago. And earlier this year they turned down a five hundred million dollar Nvidia investment at a seven billion valuation — citing exactly the risk that one dominant backer shapes your direction.

So why now?

Unknown, and I'd be careful pretending otherwise. What I can say is what's at stake. Hugging Face is where Meta, Alibaba, Mistral, DeepSeek and thousands of independent researchers all publish to the same registry. Neutrality is the product. Every plausible buyer at thirteen billion is a company with a model of its own to promote.

Which is a strange thing to buy. You'd be paying for the thing you'd immediately break.

That's the sharpest read circulating, and I think it's mostly right. The counter is that infrastructure this central is safer inside a company that can fund it forever than one that can't. But they just told us they're close to profitable, so that argument is weaker than usual here.

OpenAI cut prices again. Third time on this model family in a month.

GPT-5.6 Sol goes to four dollars per million input tokens, down from five, and twenty per million output, down from thirty. Twenty percent off input, thirty-three off output, guaranteed through at least November twenty-first. Applies to Codex credits and ChatGPT Work too.

Guaranteed through November is oddly specific.

It's the whole story, Kate. Time-boxed promotional pricing isn't a margin announcement, it's a retention move. You cut for three months when you're worried about who leaves this quarter. Chinese open-weight models are undercutting on price, Anthropic's taking coding share, and price is the only lever that works immediately.

The developer reaction was interesting though — not really about money.

The top comment on Hacker News was someone saying the fact that models can be so easily distilled and replicated is a stroke of luck — that fifteen years ago they'd have assumed there'd be a massive moat. That's the real observation. And practically, one developer noted stacking OpenRouter's promotional discount gets you an effective two dollars in, ten out. Frontier-class inference at what was mid-tier pricing a year ago.

So if you're building on these APIs?

Assume unit economics improve faster than your roadmap does. Most product plans I see are still priced off last year's rates.

Alibaba shipped Wan3.0 — thirty seconds of video in a single pass.

Double the fifteen-second ceiling of Wan2.7, and it takes documents, spreadsheets, PDFs and web content as input, not just text prompts. Single-pass duration is the genuinely hard part — stitching clips together is easy, keeping a scene coherent for thirty seconds is not.

And the timing?

One day after a ten-point-two billion dollar share placement in Hong Kong, earmarked for AI. But the number I'd sit with is different. Alibaba's quarterly net profit fell seventy-five percent, with capex at sixty-seven point seven billion yuan in the quarter. They've committed over three hundred and eighty billion yuan — about fifty-six and a half billion dollars — across three years.

Seventy-five percent. That's not a dip.

It's a deliberate conversion of current earnings into AI capacity, funded partly by selling equity. That combination tells you they think the window closes. Whether public shareholders in a Western market would tolerate that trade from a US company is a fair thing to wonder about.

This next one is my favourite of the day. Anthropic says Claude designed working protein binders.

Claude Opus 4.8 and Mythos Preview, working with Adaptyv Bio and Twist Bioscience. Thirty candidate binders for each of fifteen biological targets — thirteen hundred and twenty designs — physically synthesized and tested in a lab. Three hundred and fifty-four bound to their targets, across fourteen of the fifteen cases.

Bound meaning it actually worked.

Actually worked, in wet-lab testing by third parties. Hit rate was twenty-two to thirty-five percent against a field-typical ten to fifteen. Mythos Preview hit thirty-five point one percent overall when targets were handled independently, and forty percent on one target specifically, beating prior competition results there.

Why does this matter more than the usual AI-for-science headline?

Two reasons. It's not a benchmark — external partners synthesized the molecules and measured what happened, which is a much higher bar than most of these claims clear. And it's a general-purpose language model, not a specialist system like AlphaFold. That's the surprising part.

Your caveat?

Fifteen targets is a small sample. I'd want to know how those targets were chosen before I accept a doubling of the field baseline as a real advance. That's not skepticism about the result, it's just the question the next paper has to answer.

Security. There's an essay arguing a model could attack the machine running it.

And the important thing is what it's not saying. This isn't about escaping an agent sandbox. It's about the model attacking the inference engine itself — vLLM, SGLang, llama dot cpp — by emitting token sequences crafted to exploit parsing bugs in the software reading its output.

Is that theoretical?

There's a real precedent. CVE-2025-9141, a vLLM flaw where the XML parser for Qwen3 Coder passed nearly every tool-call argument to eval. Arbitrary code execution. And a commenter pointed out the fix was force-merged by the lead maintainer over automated warnings.

Which is a very human failure mode.

Under performance pressure, always. And think about what that host is — it holds the weights, serious compute, and a privileged network position, while parsing untrusted input at high speed. That combination has produced incidents in every other category of software. The proposed fix is architectural: split GPU execution from token parsing across machines, have the GPU host emit only logits, treat everything it produces as untrusted. Well understood, almost nobody has done it.

There was a companion piece about backdoors in open models.

Time-release backdoors — dormant behaviour until a trigger fires. The mechanism is real, there's published research including Anthropic's own sleeper-agents work. But the framing got pushback, and correctly. This is a general supply-chain problem, not an open-source one. With a closed model you don't need a back door — you have the front door.

Quick one. Claude went down for about three hours yesterday.

Started 05:06 UTC, hit the API, Claude Code, Cowork and Workspaces, resolved around 08:30. No public cause given. The community reaction was practical rather than angry — run multiple providers with automatic failover, Bedrock stayed up throughout, and older model versions often keep working when the newest ones don't.

Three hours used to be nothing.

Coding agents are production infrastructure now, so it isn't. But notice the mitigation everyone reached for. Multi-provider failover only works because these models are close enough substitutes to swap at runtime. That's a competitive fact dressed up as an ops tip.

One to watch: OpenAI's public S-1. The confidential draft went in back in May, the earliest realistic window for the public filing is basically now, and it would be the first audited look inside a frontier lab — which every valuation in today's episode is quietly priced off.

I'll take the other side on timing. A company projected to lose fourteen billion this year gets to pick its weather, and it will wait for a good day.

That's your AI in 15 for today. See you tomorrow.