Three drops this week point in the same direction, and it is away from the cloud. NVIDIA unveiled a chip built to run frontier-scale AI models on your own machine. Anthropic shipped a tool that puts its entire platform inside your terminal. And Perplexity launched a desktop computer that runs an AI agent directly on your PC. For two years the default has been that AI lives in someone else's data center and you rent it by the token. This week, three different companies, a chipmaker, a frontier lab, and a search company, all moved the action back onto hardware and software you control. The reasons are cost, speed, privacy, and independence. Here are the three drops and what they mean for you.

At Computex in Taipei on June 1, Jensen Huang unveiled the RTX Spark Superchip, NVIDIA's first serious move into the consumer PC market. A one-petaflop "superchip" with 20 Arm CPU cores, a Blackwell GPU, and 128GB of unified memory, built to run AI models with up to 120 billion parameters entirely on your own machine, with context lengths reaching a million tokens.

This is NVIDIA going after the $200 billion CPU market long dominated by Intel, AMD, Apple, and Qualcomm. The announcement alone sent AMD, Intel, and Qualcomm shares down. Huang said NVIDIA and Microsoft are going to "reinvent the PC," describing a future where billions of AI agents use PCs as tools, which means the world needs far more CPUs than the GPU-centric data center buildout has produced. The chips ship this fall from ASUS, Dell, HP, Lenovo, Microsoft Surface, and MSI, with secure sandboxes co-developed with Microsoft to run agents like OpenClaw and Hermes locally.

The significance is the direction. Running a 120-billion-parameter model on a laptop, with no cloud bill and no rate limit, used to be impossible. NVIDIA just made it a product you can buy this fall. That changes the economics of building with AI for anyone tired of metered cloud pricing.

On June 2, Anthropic shipped ant, the official command-line client for the Claude Developer Platform. Every API endpoint, the Messages API, Claude Managed Agents, sessions, files, and skills, is now callable from a single line in your terminal. It is written in Go, shipped under the MIT license, and the GitHub repo already has hundreds of stars within a day.

The design principle is simple: every API resource becomes a subcommand. Call the Messages API, spin up a Claude Managed Agent in the cloud, and pipe the result straight into a shell pipeline, all without opening an editor, writing a Python script, or hand-pasting JSON into curl. And the detail that matters most: Claude Code already knows how to use ant on its own, through the built-in claude-api skill. The coding agent can now drive the entire platform for you.

For developers, this collapses a whole category of glue code. If you were building agents by calling the API in a loop and managing tool calls by hand, most of that work goes away. It is Anthropic making the Claude platform feel native to the place developers actually live, the command line, rather than something you bolt on through an SDK.

Also this week, Perplexity launched Perplexity Personal Computer for Windows, available now to Max and Enterprise Max users with a waitlist for everyone else. It moves Perplexity from being a cloud search-and-answer tool into being an agent that lives on your machine and operates it for you.

The significance is the instinct, not the feature list. Perplexity built its name as a cloud answer engine. Shipping a desktop agent is the same move as NVIDIA's chip and the same energy pushing developers toward local tools: get the AI off the distant server and onto the device, where it is faster, more private, and operates directly on your files and apps. Every serious AI company is now racing to plant a flag on your actual computer, because the next phase of the fight is happening on the device, not in the browser tab.

The move to local AI isn't neutral. It creates clear winners and losers, and naming them is the useful part.

Winners. Developers and power users, who get speed, privacy, and an escape from per-token metering. NVIDIA, which just found a second $200 billion market on top of its data center empire. And the open-source agents like Hermes and OpenClaw, which were built for local-first from day one and now have the hardware to run on.

Losers. Intel, AMD, and Qualcomm, whose shares dropped the moment NVIDIA entered their market. The pure cloud-rental model, which looks worse every time someone gets a surprise token bill. And arguably the "AI is a website you log into" mental model that defined the last two years, it's being replaced by "AI is a thing that runs on my machine."

The one to watch. The cloud giants, Amazon, Google, Microsoft, are both winners and losers here. They sell the cloud AI that local compute threatens, but they also build the chips and clouds that local-first still depends on. How they play the next year decides whether local AI is a niche for power users or the new default.

Here's the bet we'd put a stake in. Local-first AI is not a fad, but it is not replacing the cloud either. What's actually happening is the split that already happened in every other computing era: the heavy, occasional, frontier-scale work stays in the cloud, and the constant, everyday, privacy-sensitive work moves to the device in your hands.

By the end of 2026, the question stops being "cloud or local" and becomes "which workload goes where." The developers who figure out that routing first, frontier model in the cloud for the hard 20%, local model on an RTX Spark for the constant 80%, will run circles on cost and speed over the ones still renting everything by the token. NVIDIA, Anthropic, and Perplexity all placed the same bet this week. They think the device matters again. We think they're right, and the people who adjust their stack early are the ones who win the next two years.

That's the call. Check back at year-end and hold us to it.

Three drops, one direction: AI is moving onto hardware and software you own. NVIDIA unveiled the RTX Spark Superchip at Computex, a one-petaflop consumer PC chip that runs 120-billion-parameter models locally, taking direct aim at Intel, AMD, Apple, and Qualcomm in the $200 billion CPU market, shipping this fall from every major PC maker. Anthropic shipped ant, an official CLI that puts the entire Claude platform, Messages API, Managed Agents, sessions, and skills, in your terminal, with Claude Code already able to drive it. And Perplexity launched Personal Computer for Windows, a desktop agent that operates your machine directly. For two years AI lived in the cloud and you rented it by the token. This week a chipmaker, a frontier lab, and a search company all moved the action onto devices you control. The cloud era is not over, but the assumption that AI has to live there just cracked. Owning your stack is the new strategy.

Stay building. 🤖