Android, AI, and Qualcomm Phones
Does Qualcomm's dependency on Android negatively impact Snapdragon-powered AI phones?
In my previous post, we explored Qualcomm’s current smartphone strategy: make an outstanding SoC for both today's and tomorrow’s use cases. We also traced the history of Nuvia, which plays a central role in Snapdragon 8 Elite’s success.
Recently, I’ve been reflecting on Apple's and Qualcomm's differences in AI phone strategies, especially any trade-offs Qualcomm might need to accept as a result of supporting Android.
Let’s explore.
Generative AI and Smartphones
The initial generative AI experience on a smartphone was ChatGPT via the browser and, eventually, the app. The user experience was simultaneously simple and hard. Users open the app, type something, and see magic happen; the challenge for most users is figuring out what to type. A blank canvas is a minimalist UI—peak design simplicity—but places a significant cognitive burden on users.
Benedict Evans explained this problem back in 2013 in Twitter as a Blank Canvas:
Pretty much everyone has the same common experience when they first use Twitter: bafflement. You go, you sign up, you load the account and... what?
The blank screen was really a blank canvas. It can be lots of different things, but it's nothing until you start building. You make your own Twitter.
Hence the stories about abandonment rates: people hear about it, create an account, write a few baffled tweets and get no further. What am I supposed to do with this?
Sound familiar? As noted earlier, by February 2024, less than 25% of Americans had used ChatGPT, with fewer than 20% of those users engaging with it weekly. Thus, only about 5% of Americans interacted with ChatGPT weekly. The numbers have surely gone up, but probably not by much.
Why not? The blank canvas.
When left to figure out how to use ChatGPT, most people start with “How can I apply this to my existing workflows?” instead of “What new work can I do that I couldn’t before?” This often has disappointing results.
Benedict Evans again, this time in Unbundling AI:
Whenever we get a new tool, we start by forcing it to fit our existing ways of working, and then over time we change the work to fit the new tool. We try to treat ChatGPT as though it was Google or a database instead of asking what it is useful for.
ChatGPT and its competitors are helpful, no doubt. But their usefulness is limited by our imaginations.
Even in cases where LLMs can improve existing workflows, switching to a standalone app to access the LLM is burdensome. Instead of sending users to the LLM, we need to bring the LLM to the user. I don’t want to copy a sentence into ChatGPT for an edit. I just want to edit it inline as I write.
Customers want the generative AI features in the right place at the right time without opening the ChatGPT app. On smartphones, that can look like a feature to collapse relevant push notifications into a single summary, like this example of summarizing garage door notifications into the most recent status:
Or suggested improvements for texts:
The former, Summarize Previews, is an Apple innovation. The latter, Magic Compose, is an Android feature. As you’ve guessed, I chose these examples deliberately, as they highlight the issue: Apple has complete control over generative AI features that need to exist at the operating system level. Qualcomm and it’s OEMs do not.
OS-Level Gen AI
If the manufacturers who build phones using Snapdragon 8 Elite want helpful OS-level Gen AI features for their customers, they’re at the mercy of Google; most OEMs use the official Google Android instead of managing their own custom Android implementation, which means they have to wait for Google to prioritize the feature for Android.
Given that Android is building for an entire ecosystem of OEMs (Samsung, Motorola, OnePlus, etc.), a single manufacturer probably has little influence on the priorities of Android’s backlog.
The silver lining is that Google is also building Android for its Google Pixel phone, so Google is especially incentivized to keep up with OS-level feature innovations.
In a Stratechery interview, Google Senior Vice President of Devices & Services Rick Osterloh explains how Pixel, Android, and OEMs work together:
RO: In fact, what I think it allows us to do better is coordinate our overall AI focus so that the Pixel folks aren’t doing their own thing, they can contribute more to the direction of travel with Android overall and I mean, the Pixel team’s going to do the very best job they can at trying to build beautiful devices that people look at and think are awesome and want to buy.
Osterloh emphasizes the value of the Pixel phone as a development platform to benefit the entire Android ecosystem:
BT: What’s the goal of Pixel? Is Pixel supposed to make a lot of money or is it supposed to be a route to monetization for AI?
RO: First off, it’s to try to lead an AI innovation, that’s number one.
We’re trying to gain Android share overall and that’s my principal goal is to try to improve Android competitiveness, improve our innovation, improve the problems we’re able to solve for users at all tiers.
Osterloh’s focus on AI innovation and enhancing the entire Android ecosystem suggests that, in practice, the gap between Apple’s innovations and Android’s fast-followers might not be significant.
Of course, Android might even introduce new features before Apple, especially given that the Pixel is a premium phone trying to take share from the iPhone.
BT: At the end of day, aren’t Pixel users coming from Samsung?
RO: Actually very few of them.
BT: Interesting. So where do they come from?
RO: They’re coming from a large number of people, some of whom who’ve left the market and then also from Apple
Consequently, while Qualcomm and smartphone manufacturers may have limited influence over compelling generative AI features at the Android OS level, the dependence on Google might not be detrimental. In fact, it highlights why choosing the official Android is preferable to developing and maintaining a proprietary version; let Google invest in developing the features to make Pixel competitive and inherit those features on Snapdragon Elite 8 phones for free.
Finally, most smartphone activity today occurs within standalone applications, suggesting that “bring the LLM to the user” will happen more often in apps than in the OS.
When app developers leverage the NPU for inference, they’ll expect strong performance across Android and iOS. If Snapdragon’s AI performance and energy efficiency can keep pace with Apple, Qualcomm should be just fine.