Good morning. Today's stories split cleanly into two camps. On one side, the AI platform layer is consolidating fast — toward models that understand the physical world and agents that run privately on your own machine. On the other, a backlash is hardening: people opting out of AI-saturated search, and companies putting fences and price tags around the data that trains it. Read NVIDIA's Cosmos 3 release for where foundation models go next; the RTX Spark announcement for the on-device agent push; DuckDuckGo's traffic surge and Strava's API update for the backlash; and The Verge's Build preview for the week ahead. Prefer this once a week? Subscribe to the weekly brief.
1. NVIDIA opens Cosmos 3, the first omni-model for physical AI
NVIDIA released Cosmos 3, which it calls the first open omni-model for physical AI, available on Hugging Face today. Where earlier Cosmos releases split the work across separate models for world generation, scene understanding, and policy, Cosmos 3 folds all of it into a single Mixture-of-Transformers model that can generate physically plausible video worlds, reason about motion and causality, and predict future action sequences in one pass. It ships in two sizes — Cosmos 3 Nano (an 8B-parameter model tuned to run on workstation-grade GPUs) and Cosmos 3 Super (a 32B model for large-scale synthetic-data generation) — alongside a Hugging Face Diffusers integration, post-training scripts, and a set of open synthetic-data datasets for robotics, driving, and warehouse scenes.
The substantive read is that the center of gravity in foundation models is shifting from text and pixels toward the physical world — robots, vehicles, and "smart spaces" that need to understand physics, not just language. Unifying generation, reasoning, and action in one model is the architecturally interesting part: it removes the brittle hand-offs between specialized models that have made robotics pipelines hard to build. The open release (weights, datasets, and training scripts) matters as much as the architecture, because it puts a world-model foundation in the hands of teams that could never train one from scratch.
Why it matters. If you work in robotics, autonomous systems, or simulation, Cosmos 3 is worth a hands-on evaluation — the Nano model is sized to run on a single workstation GPU, which lowers the bar to experimenting. If you follow the model landscape, note the strategy: NVIDIA is seeding an open ecosystem for physical AI that runs best on its own hardware, the same playbook that built CUDA's moat. And if you're weighing claims, "first open omni-model" is NVIDIA's framing; the verifiable facts are the unified architecture and the open weights — judge the capability against your own benchmarks, not the launch copy.
2. RTX Spark and OpenShell push secure agents on-device
At GTC Taipei during COMPUTEX, NVIDIA unveiled RTX Spark — a class of Windows PCs purpose-built for personal AI agents — alongside the OpenShell runtime, built on new Microsoft security primitives, for running agents locally under user control. NVIDIA says RTX Spark offers up to 1 petaflop of AI compute and 128GB of unified memory, enough to run capable agents entirely on-device; OpenShell adds policy controls that let users define what an agent can do, route sensitive queries to local models, and mask personal information before anything reaches the cloud. The announcement also bundled a claimed 2x inference speedup for open models in llama.cpp via multi-token prediction, plus expanded support for computer-use agents from H Company.
The substantive read is that the industry's answer to the biggest objection to AI agents — "I'm not giving software that much access to my machine and my data" — is to move the agent off the cloud and onto hardware you own, wrapped in OS-level guardrails. That's a genuinely different trust model from cloud agents, and it's notable that the privacy pitch (local models, masked queries, user-defined policy) is now front and center rather than an afterthought. The caveats are the usual ones for a keynote: performance figures are vendor-reported and tied to specific models, and most of the consumer-facing pieces ship "this fall," so the demos are ahead of the shelves.
Why it matters. If you've held back from agents over privacy, an on-device option with OS-level containment is the development to track — it changes the calculus for granting an agent access to your files and apps. If you build agents, the Windows security-primitive plus OpenShell stack is worth understanding early, because it may become the default container for shipping a desktop agent. For the underlying distinction between an assistant that suggests and an agent that acts, our AI agents vs AI assistants explainer lays out what you're actually authorizing.
3. DuckDuckGo's "no-AI" search traffic triples
DuckDuckGo is making its "no-AI" search experience easier to set as a default, launching browser extensions for Chrome and Firefox as traffic to the feature surges. The no-AI mode strips AI-assisted answers, chat prompts, and most AI-generated images from results. According to the reporting, traffic to the page roughly tripled to a new high after Google leaned harder into AI search at I/O in May, with visits settling well above their old baseline rather than spiking and fading — and U.S. weekly app installs rising alongside. DuckDuckGo says it will also add AI-search controls to its existing Privacy Essentials extensions across Chrome, Firefox, Edge, and Opera.
The substantive read is that "AI in everything" has produced a measurable, durable opt-out market — not a fringe of holdouts, but enough sustained demand for DuckDuckGo to build product around it. That's a useful signal precisely because it cuts against the prevailing narrative: a meaningful slice of users experience AI-saturated search as worse, not better, for the task of simply finding a link. The tripling is real and notable, but worth keeping in proportion — it's growth on a smaller base, not a dent in Google's dominance, and "tripled traffic to a feature page" is a different claim from "tripled market share."
Why it matters. If you publish on the web, this is a reminder that not every reader wants an AI summary standing between them and your page — there is renewed audience for clean, direct results, and for being the source rather than the thing being summarized. If you build search or assistant products, the lesson is that AI features should be opt-in and reversible; the backlash is loudest where AI is hardest to turn off. And if you simply want a less cluttered search, the no-AI mode now takes two clicks to set as default.
4. Strava puts a price on its API to fight AI scraping
Strava is overhauling its API agreement, replacing free, tiered developer access with a flat subscription — reported at $11.99/month, with regional variation and a 90-day transition window — as part of a broader clampdown on AI scraping, as TechCrunch reported. The company is also tightening website security so that certain data is visible only to authenticated users, retiring some API endpoints, and says it plans to add support for the Model Context Protocol so AI assistants access data through a controlled, structured channel rather than scraping. The moves come as Strava prepares for a potential IPO.
The substantive read is that platforms sitting on valuable proprietary data are moving from "open by default" to "metered and gated," with AI scraping as the stated trigger and an IPO as the backdrop. There's real tension here: a paid, locked-down API protects users' fitness data and the company's asset, but it also raises the cost and friction for the small developers who built the third-party ecosystem around Strava in the first place. The MCP support is the forward-looking piece — an acknowledgment that the answer to uncontrolled AI access isn't only walls, but a sanctioned, observable doorway.
Why it matters. If you build on third-party APIs, treat Strava as a template for where data platforms are heading — assume free, unmetered access is ending and price it into your plans. If you run a platform, note the two-part playbook: gate the open scraping surface, then offer a structured MCP channel so AI access becomes a product you control rather than a leak you fight. And if you're a user, the underlying shift is toward your data being shared only through authenticated, auditable paths — generally a good thing, even when it makes some apps you liked stop working.
5. What to watch at Microsoft Build
Microsoft heads into its Build developer conference this week, and The Verge expects new AI models and Windows improvements as the company continues to reshape its business around AI and works to win developers back. This is a preview, not a disclosure, so treat the specifics as informed expectation rather than fact — but the framing is worth flagging: Build is shaping up as a pivotal moment for how Microsoft positions Windows as a platform for AI development and on-device agents, the same battleground NVIDIA staked out at GTC Taipei this morning.
The substantive read is that the Windows-as-AI-platform contest is the throughline connecting today's hardware news to this week's software news. NVIDIA brought the silicon and the on-device runtime; Microsoft owns the operating system and the developer relationship. Whatever Build ships, the question to ask is the same one that runs through this brief: how much of the AI experience moves onto the user's own machine, under the user's own control, versus staying in the cloud.
Why it matters. If you develop on Windows, Build is the keynote to actually watch this week — the model and dev-mode announcements will set the near-term roadmap. If you're tracking the platform war, hold today's RTX Spark news and this week's Build news side by side; together they tell you where the on-device agent stack is consolidating. We'll cover the concrete announcements once they land — expectations are not yet evidence.
What to take from today
Five stories, two currents. The platform layer is consolidating toward the physical world and the local machine: Cosmos 3 opens a world-model foundation, and RTX Spark and OpenShell push agents on-device — a trend Microsoft Build looks set to extend. At the same time, the backlash is getting concrete and measurable: DuckDuckGo's tripling no-AI traffic and Strava's paid, gated API show users and platforms drawing lines around where AI gets to reach. The common thread is control — over your data, your machine, and what AI is allowed to touch. Judge each move on that axis.
Tomorrow's brief lands at 15:30 UTC. If you'd rather read this in your inbox once a week — just the stories that actually matter — subscribe here.