Snapdragon Summit Day 1 – Gaining Agency

Qualcomm hints at Apple’s Nemesis

• Qualcomm has brought us all out to Maui and headlined Summit with the notion that the agent, rather than the smartphone, will eventually become the centre of the digital universe.
• This could spell real trouble for Apple should this come to pass, as it opened its launch event earlier this month with precisely the opposite view, stating that the iPhone was and would remain the perfect place to host the AI services that will be the key differentiators going forward.
• It is easy to understand why Apple takes this view because the iPhone is the glue that holds its entire ecosystem of devices together, and should it lose that central position to AI agents, then its ecosystem could rapidly unravel.
• However, we are still very far from this, and Qualcomm made clear that, for the moment, the smartphone is still the place where all of this is going to happen, but its role will change meaningfully.
• This is where the latest updates to the smartphone platform come in, which include:
  • First, two Elite platforms: rather than one, to cater to the growing diversification at the high end of the smartphone market.
  • This means the 8 Elite Gen 6 for the regular flagship, which has been optimised to balance performance and battery life and the 8 Elite Extreme Gen 6, which goes all out to provide the best performance possible for Pro and Max flagship variants.
  • New Oryon CPU cores include a prime core that now goes over 5Ghz as an industry first, as well as upgrades to the Hexagon NPU (see below), Adreno GPU, Spectra ISP (photos and videos) and the sensing hub.
  • Other technologies such as high-bandwidth compute (HBC) that was born in the data centre, and an offset package design to improve thermal performance are also coming to the smartphone.
  • The practical upshot of this is that devices will be able to run inference on 30bn mixture-of-experts (MoE) models on device, which I think is a significant milestone.
  • If one looks at the lists of models that have advanced performance, the cutoff point tends to be around 30bn parameters, which means a big increase in the feasibility of running AI services on the device as opposed to being dependent on the cloud.
  • Anyone who is paying to run inference for their AI service in the cloud at the moment has a strong incentive to run as much as possible on device, as this greatly lowers the compute bill.
  • Hence, with the 30bn threshold now breached, I think that there could be a rapid uptick in AI services running on Snapdragon devices.
  • Second, Hexagon NPU: which has been upgraded to being capable of running much larger models and has had its context window extended to 32,000 tokens.
  • It is also now 50% more efficient when it comes to memory and provides an 80% improvement in prefill performance on optimised models.
  • This will enable creators of AI services to have more of their services running on devices when they write to the Hexagon NPU.
  • Third NPU support: which has been a problem to date.
  •  There are many NPUs available, creating a requirement to support multiple NPUs to fully address the market.
  • This has been a big hindrance to using the NPU by 3^rd parties, but it looks like this is beginning to come to an end for Snapdragon at least.
  • Google, which owns Gemini, is pushing very hard on Android devices and has just launched a new tier of laptop devices that run Snapdragon and are optimised for Gemini.
  • It also has a deep partnership with Microsoft for Surface devices as well as a long history of supplying the high end of Android in China.
  • This means that two of the large players in Western markets and most of the Chinese open-source community will now have a greater incentive to ensure that their AI services run on the Hexagon NPU.
  • This should give it momentum that will encourage others to write to the NPU, thereby establishing Hexagon as a leading place to run AI services on edge devices.
  • I think that this is an important step forward towards running AI services on edge devices, and the first signs of this are now visible.  
• The net result is that agents on smartphones are becoming a more realistic proposition with every month that passes.
• Getting the agents to behave in a way that will be acceptable to users, such that they trust them with making purchases with their credit cards, has some way to go, but the improvement compared to last year is obvious.
• If this progress continues, then usage may well move from apps to agents, which will greatly change the competitive environment in the smartphone industry in two ways.
  • First UX: which will become less important, as it will be the agent that interacts with the apps and services.
  • It also means that the touchpoint with the user changes, meaning that Google will be competing with Gemini and Apple with Siri.
  • Given how awful Siri remains, one can see how this immediately becomes a problem for Apple although it does still have time to fix it.
  • Second, device ecosystem: which could be far more significant.
  • If users primarily interact with an agent, they are increasingly going to care less about where that agent resides.
  • Sometimes it will be more convenient to access it through glasses, headphones or the car as opposed to the smartphone.
  • This is the scenario where the agent becomes the centre of the digital universe rather than the smartphone, but given Snapdragon’s strong position right across the device ecosystem, this is not really a problem.
  • It is a problem for Apple as the draw of the iPhone would diminish, leaving Apple needing to use Siri to keep users in its ecosystem of devices.
  • There is no sign of this in the immediate term, as agents are still not quite ready, but there are signs it is coming.
• Qualcomm is in a strong position whether agents materialise or not given the breadth of its portfolio which, combined with encouraging progress in automotive and data centre, leads me to remain confident that the new 2029 targets given in June will be met or exceeded.
• I can get to $22 in EPS in 2029 without trying very hard, which gives me a DCF-derived valuation of $311 in today’s money and $340 in 12-months’ time.
• I continue to have a fairly large position in Qualcomm that I am very happy to sit on.