Apple M5, the pin to burst the bubble? Last week, Bloomberg published a simple but informative diagram (https://lnkd.in/ehXGP4WA) showing the interdependence between the AI companies: NVIDIA, Microsoft, OpenAI, AMD, Oracle, Intel, XAI, and others. Notably absent was Apple. This is not because it doesn’t have an AI story, but because it is firmly outside of the AI infrastructure bubble. You don’t have to look far to read about this bubble; the Financial Times published an article earlier today titled “The AI bubble is a bigger global economic threat than Trump’s tariffs” (https://on.ft.com/477IKAu). Why is Apple different, though? Apple is making and selling chips with incredible performance (better per watt than NVIDIA) to run locally, in your pocket, on your lap, or at your desktop. The new M5 claims a 3.5x AI performance boost, and the M4 was already fast. Apple is not crearing its own proprietary AI models (LLMs), you can run almost any open-source model on every device, American, French, Chinese, or any other, from wherever you like. They all run beautifully fast on your personal Apple-Silicon device or company servers. The future of AI is small LLMs linked together in agentic networks, not the large proprietary oligarchs’ GPTs running on billions of dollars of cloud infrastructure, which require billions of dollars in power and billions of litres/gallons of water to cool inefficient GPUs. The incumbents have no other path, I suggest, it’s an AI Ponzi scheme. Too much has been invested, and the only way forward is to keep pretending it’s the only way forward. Almost everything we use AI for will run on your laptop, privately and for free. If/when the AI bubble bursts, as with the glitch earlier this year, it will be a realisation that AI doesn’t need (loaned) money spent in this way. Spend 10% of the AI infrastructure budget on R&D or education, and you’ll have a much better return in the long run. Back to Apple and the new M5. When a laptop can do what supposedly requires millions in infrastructure, the emperor has no clothes. Apple is by no means unique in this space, but it is clearly leading it. NVIDIA has obviously seen this and released the $4,000 DGX Spark Desktop yesterday (15th Oct). I was going to order one until I saw the benchmarks. It seems no more than a souped-up Raspberry Pi, not even close to my one-year-old MacBook Pro. A shame as I like new gadgets and I already have loads of Raspberry Pis. I am patiently waiting for the M5 Max or Ultra. Wean yourselves off ChatGPT. With open-source models running locally, your data never leaves your device, your costs are fixed, your latency is near zero, and your privacy is absolute, this scales to enterprise with private cloud too.
Small models for internal orchestration and then use larger API models to query for deeper added context - this is only scalable model for the future. You don’t need more than a 1b parameter model for basic discussion and tool calling.
Small AI models (capable of running locallty) are getting as good as big models; Nvidia and Apple are hardware companies and may survive the bubble burst - OpenAI, Anthropic, xAI will get hit.
Great post. My friends and I were just talking about this last night over a beer. Seems likely the bubble is about to burst all right. Interesting to think about Apple silicon as the way forward.
Aren’t we mixing up concepts? What you’re actually comparing are smaller models versus larger ones. That’s why you’re seeing efficiency gains in decentralized setups. But I don’t see how a centralized, shared infrastructure (say, M5 instances running many small models) could be less efficient than a swarm of decentralized nodes, each idle 90% of the time and running redundant copies of the same models. Computation doesn’t change just because you move it around. If you compare apples to apples, decentralized setups are inherently less energy-efficient than centralized ones. If we’re talking about privacy and oligarchy, then I completely agree, decentralized is better, and I am a big supporter. But that’s the real reason: data ownership.
John Davies Can you connect the dots for me between lower power consumption of M5 with the ability to perform Petaflops of computations to train models? If you can share concrete numbers then that will help. It is clear that eventually the world needs power efficient computation for training and inference, and eventually a smaller form factor of devices that can perform inference and training. This post seems to imply that we are already there or very close to it (with M5?). Is that what you meant?
A good analysis John. I think the future of AI will likely be hybrid, so local inference for privacy-sensitive, latency-critical, or routine tasks, with cloud infrastructure for training, complex reasoning, enterprise collaboration, and scenarios requiring massive scale. I do agree though that the global small language model market will grow as the smaller model capabilities increases, Hardware/GPU's becomes more competitive and companies start to recognize the need for efficient models in edge scenarios.
If you ever want to see if Apple is ahead or behind the curve ball just use Siri. How can a company of such magnitude release something so poor and then neglect as much as they have. This to me is a reflection ot their company, ethos and strategy. So for everyone who says they’re ahead, please…
Open Source LLMs + locally hosted (private) = win
CEO/CTO Technically Focussed Data Leader.
2wThat is an excellent summary and something that most people seem to be missing with the AI hype. There is a huge focus on the cloud models, however what happens when they run out of funding and have to start charging a realistic price to cover their infrastructure cost.