Latest on the Edge of AI, Tuesday, May 5: Apple turns iOS into a model marketplace, Google upgrades Home, and a local AI experiment that makes cloud pricing look silly. Let's get into it.
Apple, the company behind the iPhone, is turning iOS 27 into a choose-your-own-adventure for AI models [14]. Users will reportedly pick which third-party models handle everything from writing to image generation, a move that breaks Apple's walled-garden approach to on-device intelligence and hands the keys to outside labs for the first time. Instead of one house model, you get a menu. The company has spent years building its own on-device AI stack. Now it's opening the door. Developers will integrate their models directly into the operating system. Apple takes a cut. It's a platform play, not a model play. The signal: platform-level AI is shifting from a single-lab monopoly to an open marketplace. My read: Apple's betting that distribution beats model quality. If you control the phone, you control the AI. And the labs will fight for placement. This is the app store moment for artificial intelligence. The question is whether Apple's cut will be thirty percent or something lower. Either way, the model wars just moved to your pocket. OpenAI, Anthropic, Google. They're all competing for the same screen. And Apple holds the gate.
Different beat. Google, the search giant, is upgrading its Home assistant to Gemini 3.1 [1]. The update lets users bundle multiple tasks into a single command, which means you can finally ask your speaker to handle a chain of actions without repeating yourself. Recurring events get better handling. All-day scheduling improves. Last month's natural language fixes are still rolling out. The assistant is finally catching up to what people actually ask it to do. Google has been patching bugs in the Home AI stack for months. This is the first real feature bump. Which matters because: smart home AI has been stuck in single-command limbo for years. You ask it to turn off the lights. It turns off the lights. Groundbreaking. Now it can actually chain actions together. The bar was low.
Now, on the chip side. Cerebras, the AI chipmaker, is requiring limit orders from IPO buyers as demand outpaces supply [6]. Institutional investors must specify share counts and maximum prices, a procedural requirement that signals unusually heavy interest ahead of the public listing. The angle: chip IPOs are still drawing real money. Even after the volatility.
Meanwhile. A Reddit user ran a ten-day test comparing local models to cloud frontier models, and the [unverified] results are uncomfortable for the big labs [2]. The test logged 150 coding tasks across cloud and local setups, tracking token usage, task type, and whether a local model on a single GPU could match the cloud output. File reads, project scanning, code explanation: local matched cloud 97 percent of the time. That covered 35 percent of the workload. DeepSeek V4's pricing at 17 times cheaper triggered the whole experiment. The user ran everything on a single 3090 GPU. No cluster. No enterprise stack. Just a desktop card and an open model. What this changes: the economic case for cloud AI on routine tasks just collapsed. The methodology is transparent enough to take seriously. If even a fraction of that holds up, the cloud pricing model is in trouble.
Three platform shifts in one afternoon. The open one is the one to watch.
That is the edge for today.