{"slug": "how-apple-pushed-the-frontier-of-ai-hardware-again", "title": "How Apple pushed the frontier of AI hardware again", "summary": "Apple's M5 Ultra Mac Studio and M6 Mac mini extend the company's lead in consumer and professional AI hardware, according to a first-look test by Jason Hiner, who reports the M5 Ultra Mac Studio carries 256GB of RAM and 5x the memory bandwidth of Nvidia's DGX Spark and AMD's Ryzen Halo AI boxes. Hiner says both machines can run open models locally — including Google Gemma, Nvidia Nemotron, DeepSeek, Alibaba's Qwen, Kimi and GLM — to cut AI token costs, speed up AI jobs, and keep sensitive data such as PII on-device. Apple's prior-generation Mac mini and Mac Studio have faced long backorders since early 2026, and Hiner says the new machines should remain future-proofed for the next 2-3 years, with full benchmarks still to come.", "body_md": "Apple continues to push the cutting edge in one part of the AI industry that could emerge as the next frontier of the ecosystem over the next 6-12 months.\n\nI've been testing the M5 Ultra Mac Studio and the M6 Mac mini and both devices have extended Apple's lead at the frontier of AI hardware for individual consumers and professionals. The M6 Mac mini is better than ever for running an always-on agent like [Perplexity Computer](https://www.perplexity.ai/products/computer), [Hermes](https://hermes-agent.nousresearch.com/), or a variant of OpenClaw. The M5 Ultra Mac Studio is a workhorse that runs at the speed of a racehorse for the most demanding AI builders and small teams.\n\nBoth machines can save you a ton of money on AI tokens by running the latest open models locally, such as the ones from Google Gemma, Nvidia Nemotron, DeepSeek, Alibaba's Qwen, Kimi, GLM, and others. But beyond the cost savings, the hardware can often run AI jobs a lot faster. And of course, since none of the data leaves your machine it’s a lot more private and secure, which is critical for working with PII and sensitive data.\n\nApple's last-generation Mac mini and Mac Studio boxes were already terrific for AI and have faced long wait times for backorders since early 2026. Apple didn't have to make a new generation of hardware. It could have just increased production of its last-generation products and it would have likely sold every device that it could make.\n\nBut I'm glad that Apple didn't rest on its laurels and chose to push the envelope instead. I haven't started fully benchmarking the machines yet, but I have no doubt that when I do, the numbers are going to be eye-popping. The TLDR is that you can be confident that if you buy one of these machines, they are going to be future-proofed for the next 2-3 years.\n\n*You can also daisy-chain up to 4 Mac Studios. | Photo: Jason Hiner*\n\nI'm also currently testing the [Nvidia DGX Spark](https://www.nvidia.com/en-us/products/workstations/dgx-spark/) and the [AMD Ryzen Halo](https://www.amd.com/en/products/processors/desktops/ryzen/ryzen-ai-halo.html). Both are excellent little AI boxes that sit somewhere in between the Mac mini and the Max Studio and have many of the same benefits. These machines are based on the same GPU hardware that runs much of the world's most popular AI chatbots and agents in data centers. And what's wild is that the highest-end Mac Studio has 5x the memory bandwidth of the Nvidia and AMD boxes to deliver bleeding edge performance. That speaks to Apple's lead in chip design and vertical integration in consumer AI devices.\n\nStill the AMD Ryzen Halo is great if you want a machine running Windows and the Nvidia DGX Spark is perfect if you want a headless desktop AI appliance running Linux. And both Nvidia and AMD are working with hardware vendors to build their own AI computers to compete with Apple in the years ahead. But, make no mistake, they are still largely playing catch-up.\n\nFor now, I'm running the M6 Mac mini as a dedicated agent machine, running [Perplexity's Personal Computer with Hybrid Compute](https://www.thedeepview.com/articles/perplexity-s-hybrid-agent-is-a-win-for-ai-privacy), one of the most user-friendly AI agents and one of the best orchestrators between different models. I also plan to try it with [NanoClaw](https://nanoclaw.dev/), a secure implementation of OpenClaw.\n\nFor the M5 Ultra Mac Studio, I'm throwing a ton at it, including running coding agents Claude Code and Codex while also editing video, running virtual meetings, running multiple monitors, and running six different Mac spaces. I was already doing much of that with a M2 Ultra Mac Studio (with 128GB of RAM) and could barely make it blink so I'm going to double down on running big models locally on this M5 Ultra Mac Studio with 256 GB of RAM to increase the pressure. I'll follow up with another story on how those tests go.\n\n## Our Deeper *View*\n\nOn-device AI is expected to be one of the most important AI trends of the next year, for all three of the reasons I mentioned above: cost, performance, and privacy. While the past year has been about agents and AI getting a lot more useful, the result has been an explosion of token use and skyrocketing costs. One CTO I spoke with recently said that for each engineer her company is spending 1.5x their total compensation in token costs. That could quickly net out to $15,000 to $20,000 per month. A maxed out Mac Studio costs $18,299, so it's easy to see where it could pay for itself pretty quickly. But beyond those extreme use cases for AI builders, on-device AI has excellent potential to become a much bigger part of the future beyond just saving money. As the AI models and harnesses get smarter and more capable, there are more things they could do to be useful every day. For example, I'd love to use them to run a workflow every 10 minutes scanning specific sites and dropping updates into a Slack channel, but that job burns through too many tokens using today's cloud-based AI. I'd also love to use AI to scan my home security cameras and my health data and ping me when there are important updates or anomalies, but that's highly sensitive data that I wouldn't trust to send to any of today's leading AI providers. On-device AI can solve those problems and a lot more like them, and Apple's new desktop Macs remain the friendliest and the most powerful ways to take advantage of it.", "url": "https://wpnews.pro/news/how-apple-pushed-the-frontier-of-ai-hardware-again", "canonical_source": "https://www.thedeepview.com/articles/how-apple-pushed-the-frontier-of-ai-hardware-again", "published_at": "2026-09-21 13:00:00+00:00", "updated_at": "2026-09-21 15:23:38.266693+00:00", "lang": "en", "topics": ["ai-infrastructure", "ai-products", "ai-agents", "ai-chips"], "entities": ["Apple", "M5 Ultra Mac Studio", "M6 Mac mini", "Nvidia DGX Spark", "AMD Ryzen Halo", "Jason Hiner", "Perplexity Computer", "OpenClaw"], "alternates": {"html": "https://wpnews.pro/news/how-apple-pushed-the-frontier-of-ai-hardware-again", "markdown": "https://wpnews.pro/news/how-apple-pushed-the-frontier-of-ai-hardware-again.md", "text": "https://wpnews.pro/news/how-apple-pushed-the-frontier-of-ai-hardware-again.txt", "jsonld": "https://wpnews.pro/news/how-apple-pushed-the-frontier-of-ai-hardware-again.jsonld"}}