AI Moves Into Hardware: Amazon, Microsoft, and Google's New Device Wars

The tech industry is collectively rethinking what a computer even is. In the latest episode of The Vergecast, David Pierce and Nilay Patel break down a wave of recent hardware announcements from Amazon, Microsoft, Google, and Apple — and arrive at a shared conclusion: nearly every new device is built around the same core logic — giving AI agents a "surface" to run on.

Amazon's Alexa Tablet: Acknowledging the Past, Betting on the Future

Amazon officially retired the Fire tablet brand this week, launching a new Alexa tablet lineup. It's a complete strategic pivot. Amazon never tried to make a great tablet — it only tried to make an extremely cheap one ($29 kids tablets with a subscription), because competing head-to-head with the iPad was never on the table.

Panos Panay, who oversees Amazon's devices and Alexa business, made two key decisions. First, he overhauled the hardware. The headline feature is a new nano matte display — a frosted texture fused directly into the glass rather than applied as a film. The 8-inch model starts at $229, with the top-end 12-inch at $549, making it highly competitive against iPads that often run into the thousands.

The second decision — and the more fascinating one — is that Amazon dropped its own app store entirely and plugged into the Google Play Store, becoming part of the Google ecosystem. Nilay's read on this is sharp: "A lot of people think apps are the past, but the truth is the app problem has been solved." In other words, apps are still very much the present — but the war over app stores is over, and there's no point fighting it anymore. So Amazon chose to concede the present to Google and focus its firepower on the next battle: Alexa as an all-encompassing AI agent.

What's the cost? Play Store means in-app purchase revenue goes to Google — Amazon is essentially giving up its current monetization model. More concerning is how the subsidy works. Amazon's only playbook for subsidizing hardware is intrusive advertising. When asked how they hit the $549 price point, Panos simply said, "I don't want to talk about how we got to that price" — which speaks volumes.

Nano matte display refers to a manufacturing process that fuses an anti-glare matte coating directly into the glass substrate, as opposed to applying a film after the fact. Traditional matte films can trap micro-bubbles, reduce color saturation, and peel over time. Sintering nano-scale particles into the glass preserves touch consistency while cutting glare, with less impact on color accuracy and brightness. This process has historically been reserved for flagship professional monitors; Samsung and LG use it in some high-end TV lines. Amazon bringing it to a tablet starting at $229 is a rare offering at this price point — arguably the most substantive upgrade in this hardware push.

Microsoft Surface Laptop Ultra: Betting on Local AI

Microsoft also held an event this week, unveiling the Surface Laptop Ultra — its most powerful Surface ever — built on NVIDIA's RTX Spark platform, targeting local AI development, priced around $2,600.

This taps into a resurging debate: where should computation actually happen? Nilay shared his experience building a search tool with Claude while writing a book — he forgot the tool was running locally on his machine, and only realized "where things actually were" when he opened his laptop on a trip. That confusion perfectly captures a cloud-trained mindset being challenged.

What makes this interesting is who's doing it. This is a company that moved Office and Windows to the cloud, whose CEO Nadella once considered renaming Windows to "Azure Edge." Now Microsoft is trying to "bake an entire AI stack into your PC so it works offline."

Both hosts are skeptical about the use case for local AI laptops. If you want local compute and don't want your data in the cloud, the more sensible setup is a headless Mac Studio at home acting as a server you can access remotely from anywhere — not loading models onto a laptop to run locally at a coffee shop. That said, they both acknowledge that even if the pitch is just "Microsoft finally made a laptop that can go toe-to-toe with a MacBook Pro," that alone is exciting.

One notable detail: this is the first device with a built-in magnetic USB-C charging connector — essentially combining MagSafe convenience with USB-C universality. Both hosts were enthusiastic about it.

NVIDIA RTX Spark is a discrete GPU platform for ultra-thin laptops, based on the Blackwell architecture, focused on local AI inference. Its key metric is TOPS (Tera Operations Per Second), optimized for INT4/FP8 quantized inference for large language models — a different focus from traditional gaming and graphics GPUs. "Local AI inference" means model weights are stored on the user's device and can run without connecting to a remote server, keeping data on-device with lower latency. The tradeoff is that local VRAM limits the model size — current mainstream locally-runnable open-source models are roughly 7B to 70B parameters, still meaningfully behind cloud-based models at the GPT-4 level.

Google Books: An Android Tablet Pretending to Be a Laptop

Google also launched Google Books this week (five devices). David was blunt in his criticism: this is a platform "where nothing feels finished" — essentially an Android tablet pretending to be a laptop.

The core issue, again, is apps. Take Netflix: the app can only run full-screen or centered at 16:9. You can't resize the window or interact with other windows alongside it — because it's fundamentally a tablet app. David's verdict was pointed: "Android apps are bad everywhere except on a phone." And that flaw is more glaring on Google Books than on any previous device.

But Google Books does get some things right. It introduces a phone-style home screen where you can place widgets and icons, and the control center layout is more intuitive than hunting through Mac settings menus. As for why Google shipped something this unfinished, Nilay's answer is simple: the holiday shopping season is coming, it has to be on shelves before Black Friday, and developers need devices early to drive software updates.

Different Paths, Same Destination: All Hardware Is Paving the Way for AI Agents

The sharpest observation in the episode: whether it's the Alexa tablet, a Siri AI device, a Copilot laptop, or a Gemini-powered Google Books, the reason for all of this hardware is strikingly unified — they're all surfaces for letting your AI agent do things on your behalf.

So what does the ideal "surface" actually look like? Nilay's answer: a tiny cloud Linux box with 8GB of RAM running a browser. He was effusive about Meta's Muse strategy — "it just gives you a little cloud computer, and your brain immediately understands it." This brings up the central question of who owns the compute. For ordinary consumers, the most sensible setup is a cloud-hosted sandbox managed by professional IT — not a Mac mini in the basement running open-claw (one commenter got a virus doing exactly that).

Both hosts also inject a healthy dose of skepticism. All agents are far from perfect right now, and they're brittle in various ways. Amazon is already blocking Muse. Cloudflare and major websites will likely start rate-limiting or charging all agents soon. In e-commerce especially, walls will go up until commercial agreements are sorted — and when that happens, the fun of "letting an agent freely click around a browser" may evaporate.

More fundamentally, these models "hallucinate all day long." That's the root of Nilay's skepticism about all agent promises — the hallucination problem hasn't gone away; the industry has just stopped talking about it. David admitted that after a week with Google Books, he still couldn't build the muscle memory to "invoke Gemini": "Most of the things we do aren't that hard or time-consuming. Learning a whole new system just to outsource them to AI isn't a good trade."

The one genuinely useful case he found: using Muse to help his three-year-old son find a blue-yellow-red balance bike he'd seen at preschool. He cropped a photo, sent it to Muse with "what is this, where can I buy it," and Muse found it on a Dutch school supply website — after his own Google searches came up empty. Hit rate: about 50%. That's roughly the real capability ceiling of AI right now.

This is the one window the tech giants have to "beat the phone." As Nilay put it: "We've never been able to beat the phone. Is this our chance?" The answer is that they have to try — but there's no evidence it will succeed on the timeline anyone is claiming.

AI Agent refers to an AI system capable of autonomously planning and executing multi-step tasks, as distinct from a traditional chatbot that only handles single-turn Q&A. The core mechanism is "tool use": the model can invoke external tools like a browser, code execution environment, file system, or APIs — decomposing complex tasks into sub-steps and iterating until the goal is reached. Muse (Meta) and Claude both have these capabilities. Typical agent use cases include automatically comparing prices across multiple sites and placing an order, reading an inbox and drafting replies, or filling out forms in a browser. However, the "hallucination" problem (the model confidently outputting incorrect information) is amplified in multi-step chain execution — a single error in one step can cascade through every subsequent step, building on a false premise. This is the technical root of the hosts' skepticism about agent reliability.