Let's start with the simplest, lowest risk scenario for both companies: Apple designs its own ChatGPT app and builds it into iOS 18. This would be like the old days, when the iPhone came with apps Apple designed for YouTube and Maps (with Google on the back end).
In the "ChatGPT by Apple" scenario, Apple could give its own ChatGPT app access to a few private APIs... ...ship iOS 18 with pre-build GPT workflows in the Shortcuts App... But only send data to OpenAI's servers when users explicitly choose to use ChatGPT.
"ChatGPT by Apple" would allow Apple to evaluate how much its customers value and use ChatGPT. ...and keep its privacy-centric brand story intact. For OpenAI, it would be a new and potentially significant revenue stream. Also:
Apple would be paying OpenAI for all the inference from the ChatGPT app on iOS. ...and maybe handling some/all of the ongoing development and maintenance. Instead of OpenAI paying 100% of the costs of operating its iOS app itself.
The "ChatGPT app by Apple" scenario would be the safest, simplest, and lowest-risk partnership for both companies. It would also be the least imaginative. Let's explore the wilder, more interesting scenario: fusion
The fusion scenario is 3 different scenarios: - "SiriGPT" - GPT in Apple data centers, running on Apple Silicon - Why not both? In all three, Apple partners with OpenAI to iterate Apple Silicon towards performance parity with NVIDIA...but with much greater performance-per-watt.
SiriGPT: Apple and OpenAI partner to make GPT models that run efficiently on Apple Silicon ...launch on-device versions of GPT on iOS, MacOS, WatchOS, and VisionOS... And power a version of Siri that can complete complex tasks without touching the cloud.
Apple is very strategic about building third party commercial software into iOS and Mac. It almost never does. Making Siri a thinly-veiled front end for ChatGPT? Not super strategic. Collaborating with OpenAI to get GPT run efficiently on Apple Silicon? Strategic af.
Sam Altman, Mira Murati, and Greg Brockton are no strategic slouches. They would definitely be aware of the risk of Apple using OpenAI's tech to bootstrap its own generative AI models and AI-optimized silicon. (This seems to be what Microsoft is trying to do).
OpenAI would need to see a much bigger opportunity than "powering Siri" on Apple devices 'cause that would be great until Apple replaced GPT with an in-house alternative. But partnering with Apple to get GPT running at scale on Mac servers in data centers? Strategic af
NVIDIA has been making powerful dedicated GPUs for 25 years. It has a huge lead in running them through the most compute-intensive applications. But maximizing performance AND energy efficiency in parallel has never been NVIDIA's focus nor its strength. Apple, tho:
Rumors has it: Apple's experimenting with training AI models on M2 Ultras running in their data centers. I'd guess they're racking servers with their most powerful Silicon and using them to train AI models and produce "spatial media" for the Vision. And since we're speculating:
Why might Apple Silicon servers matter to OpenAI? A single NVIDIA H100 GPU costs around $40K. Some estimates suggest that in one year, every single H100 consumes as much electricity as an entire US household. That's also a lot of heat to consider in your data center design.
Before the M, a maxed out Mac Pro used Intel's 28-core Xeon W CPU. It was $50K. Intel originally developed the Xeon line for servers and data centers. It expanded into workstations because Xeon's features applied well in graphics, video production, 3D modeling, etc
Apple claims M2 Ultra Mac Pro delivers up to 1.8x performance over the 28-core Xeon model for the same compute-intensive tasks: 3D rendering, video transcoding, compiling code, etc. ...for $38K less.
3D rendering, video transcoding, compiling code: Mac Pros have been handling those for years. Apple Silicon's team has had plenty of data from media production use cases to optimize their designs around. Running inference on GPT-grade LLMs? Not so much.
Running GPTs on-device Apple Silicon for a couple years would give Apple the data they need to up-level its entire line of chips for AI. And OpenAI could learn how to get the better performance from their models while consuming a fraction of the power.
Does any of this mean Apple Silicon has any chance of catching up to NVIDIA's dedicated GPUs? Not just on-device, but in data centers?
NVIDIA has invested years in developing orchestration software that allows multiple GPUs running a datacenter to act as like one GPU. As far as the public knows, Apple has nothing like this. But it has had John Giannandrea (who led search and AI at Google) working on something.
IF Apple Silicon achieves parity with NVIDIA GPUs + delivers much more favorable performance-per-watt, it would catch the whole market by surprise. Apple has a lot of incentives to make it so.
IF OpenAI and Apple partner to get GPT running on Apple Silicon, Apple gets: - the market's most privacy-protecting, user-friendly interface for GPT - an answer to GOOG, MSFT and NVIDIA's on-device AIs - a path to optimizing Apple Silicon for the most intensive AI workloads
OpenAI gets: - Its tech in the world's most desirable smartphones, tablets, PCs, and spatial computers - A counterweight that reduces its reliance on MSFT - A potential path to a much lower cost and power-efficient alternative to NVIDIA GPUs for inference and (maybe) training.
There's a lot more to this story, but that's enough for today. In my next thread, I'll explore why Apple shipped the M4 in the iPad Pros less than two months after shipping the M3 MacBook Airs.