Qualcomm came to Maui for their annual Snapdragon Summit with the clearest version of its AI strategy I have heard from the company, and with the same structural limit it has always had. Apple can promise AI that works across devices because it designs the silicon, writes the operating system, and ships the hardware. Qualcomm controls the first of those three. Everything else it delivers through Google, Microsoft, and the phone, PC, and wearable makers that build on Snapdragon. The most interesting thing about this year's Summit was how deliberately Qualcomm is organizing itself around that dependency.

The phone stays at the center

Qualcomm President and CEO Cristiano Amon used his opening keynote to address the question that follows every new AI wearable. "Your phone is not going anywhere," he told the audience, arguing that the phone is unique to you, always with you, and has been personalized to you, which makes it the device that knows you best and therefore essential for agentic experiences. Other devices become endpoints: the agent sits at the center, with your phone, PC, glasses, wearables and car all serving as endpoints for it.

It is the same point John Ternus landed at Apple's iPhone event earlier this month, when he described the iPhone as the intelligent personal hub. As I wrote then, that is an expansion argument. Settling the succession question frees a company to add devices around the phone. Sergio Buniac, speaking on stage in his new Qualcomm role, described smartphones becoming the AI hub for our agents, connected to a platform of devices that touch our lives every day.

Android's Best Shot at an Apple-Like Experience

Apple's cross-device experience works because one silicon team designs the chips in the iPhone, Mac, iPad, Watch and AirPods, and every product is built to work with the others. Android brands have always assembled their experiences on top of a mix of chipmakers and platforms, which makes that kind of continuity harder to deliver. Qualcomm is the closest Android gets to a shared foundation, and in that sense, Qualcomm is the merchant equivalent of Apple silicon.

The announcements this week filled in the pieces. The new flagship phone platforms include a Sensing Hub designed to continuously process voice activity and personal context at very low power, the kind of always-on context a hub needs. Snapdragon Sound Elite Gen 2 supports earbuds, over-ear headphones, open-ear audio, audio glasses, hearing devices, and camera-equipped earbuds. Googlebook, a new laptop category built around Gemini Intelligence, launches with Snapdragon X Elite, with Dell and HP among the first manufacturers. Features like Cast My Apps, which opens apps from an Android phone on the laptop without installing them separately, are the phone-as-hub idea in practice. On the Windows side, Microsoft took the Summit main stage to introduce the next Surface Pro 12-inch and Surface Laptop 13-inch on Snapdragon X2 Plus, putting Snapdragon in the newest laptop launches on both Google's and Microsoft's platforms.

Googlebooks from Acer, Asus, Dell, HP and Lenovo are available for pre-order on both Intel Panther Lake and Snapdragon X Elite, so Qualcomm has competition on the laptop. What Intel lacks is the phone, the earbuds, and the glasses. Snapdragon now spans smartphones, PCs, XR, wearables and automotive. When the same platform sits in the phone someone carries, the laptop they work on, the earbuds they wear, and the car they drive, Android brands have a common foundation for the continuity Apple users take for granted. Apple builds that experience with its own chips across its own products. Qualcomm gives Android the opportunity to build it across brands.

Bringing a technology to market

When I talk about Qualcomm, I always point out that its role in making a technology mainstream has extended well past the chip into everything needed to bring it to market. This week that meant healthcare. Qualcomm and Global Health Connector launched the Personal AI Health Alliance, with founding members including Optum Health, Scripps Health, Merck KGaA, and the Digital Medicine Society. The goal is to speed up commercialization of continuous, privacy-supporting health solutions that use on-device AI to provide wellness insights between clinical visits.

Wearables have been collecting health data for years, and turning that data into clinical value has been slow. An alliance is a starting point. Its value will show in products that clinicians trust and health systems adopt.

A more integrated Qualcomm

Now more than ever, I see a company internally integrated enough to thrive on a better-together story. Buniac's appointment is part of that. He was named group general manager of mobile computing and XR, bringing device experience from his years as Motorola's CEO, which brings phones and XR under one group leader and matches the hub thesis.

The partnerships showed up again. Google's Rick Osterloh, on stage with Amon, said the move to AI-driven computing requires deep collaboration across every level of technology, with Gemini understanding your context whether you are in the car, on your phone, or at your desktop. Mastercard described building a trust layer for agentic commerce. Xiaomi and Motorola announced phones on the new platforms, and Honor showed its AI robot phone. With the Modular acquisition, Qualcomm plans to open source Mojo and MAX so developers can optimize AI across hardware, including other chipmakers' silicon.

Fandom as a trust strategy

The marketing and branding story has become central, and Qualcomm opened the Summit with it. Before any chip was mentioned, Qualcomm's Don McGuire talked about community. The Snapdragon Insiders community has grown to 23 million members, Snapdragon creators reached a combined 405 million people over the past year, and Qualcomm cited research that 92% of tech buyers trust peer recommendations over traditional marketing.

AI has an image problem. Consumers are wary of what it does with their data, their jobs, and their attention. Leaning into communities and fandom through Manchester United, Mercedes F1, and creators helps, provided the message stays genuine and transparent. That condition applies to the products as much as the messaging. In the mobile keynote, Qualcomm described an image signal processor that can add demographic information to the pixels it analyzes while identifying and tracking individuals as they move. Capabilities like that need careful explaining to the fans Qualcomm is courting.

Qualcomm has built the audience. The products arriving over the next year have to earn the trust it is asking that audience to give.

Disclosure: Qualcomm covered my travel costs to attend Snapdragon Summit

My configuration has 256GB of unified memory, but memory alone only gets a model onto the machine. The model needs to be fast enough to use, and the rest of the computer needs to keep up with it. M3 Ultra could run capable local models, but I could feel the desktop hitch when I pushed inference and tried to work at the same time. On M5 Ultra I have not had that problem in my use. The CPU is great, the GPU is great, and honestly it is hard to really push this thing. It's funny, a few days before the M5 Ultra was announced I was actually talking to a few friends and was telling them after nearly 18 months I was finally able to really hit the limits of M3 Ultra. Well, I have yet to hit that for M5 Ultra.

It's funny, if you've been following Creative Strategies for a bit, you may notice our website got a slight redesign recently. This was actually done using GPT 5.6 Sol and GPT 6 Astra recently, but both of these were running on the M3 Ultra Mac Studio I've been using for the past year or so. The reason I'm mentioning it is this was actually a full revamp of all the infrastructure hosting it, and part of that was creating a ton of simultaneous containers in OrbStack to test, convert data formats to retain articles when moving from WordPress to Ghost, testing and optimizing our self-hosting optimizations for Railway Docker containers, and help with design for it. Given each container needed its own database, worktree for real time updates for different UI changes, and needed to be available in real time for the agents, it was very memory and CPU intensive. I'm not sure any other computer I have would have been able to handle this as well! It did really push the 512GB memory M3 Ultra to its limits, believe it or not. This isn't super relevant to this report on the M5 Ultra, but it's an anecdote that I feel sorta explains where these ultra-level SoCs and workstations are incredibly useful regardless of where the inference happens. 

You own the CPU too

There’s a lot of debate right now about owning a GPU for inference. Should you buy the hardware and run your own models? Should you rent the GPUs? Should you just use an API? Those are fair questions, and yes, owning the GPU can be great. What gets missed is that with a Mac Studio you also own the CPU that runs everything around the model. The agents, compilers, Xcode, the apps they are controlling, and the rest of your work have a place to run. For me, this is the best CPU and GPU combination I’ve had in one machine. 

What makes the power of this machine so useful is that it is not locked into one job. I can push local inference, then move straight to development, or have the CPU and GPU working at the same time. As a consumer, that makes it my main computer rather than a dedicated AI appliance. For an enterprise, it could be a CPU hub, a GPU hub, or both. If models or inference software move in a different direction, the Mac is still an extremely capable computer. That is a lower-risk bet than committing the whole machine to one workload.

The GPU is obviously powerful too. Apple lists roughly 1.2TB/s of memory bandwidth for M5 Ultra, and its GPU includes Neural Accelerators for matrix operations. That memory bandwidth and compute are why a large local model can feel interactive instead of merely fitting in memory. Of course, the model and the software matter just as much. Apple’s specifications and its explanation of Neural Accelerators cover the hardware side.

Model architectures and inference frameworks change constantly. The machine stays the same. I can tune the software for a model I care about today, then use the CPU for development tomorrow, and I don’t need to predict exactly which model or framework will win three years from now to find the computer useful.

Ok, the local AI part

There are two things that decide whether I actually want to use a local model: can it do the work, and can it do that work quickly enough? Qwen 3.8 Flash Next has been very good for me in OpenCode 2. In my earlier agent testing I saw around 40 tokens per second on M3 Ultra, with waits of roughly 5–20 seconds before an agent turn got moving. On M5 Ultra it was around 75 tokens per second, with waits closer to 1–5 seconds. Over one turn that may not sound dramatic. Over an agent run with tool call after tool call, you feel every one of those waits. In practice this actually ends up being 5-8x faster on M5 Ultra.  

Those are numbers from my agent workflow, not some universal speed claim for every Qwen setup. Context length, cache reuse, quantization, and the inference software all change the result. The 149-tokens-per-second decode figure I get into below is from later kernel tuning, so I wouldn’t put it on the same chart as the earlier 75-tokens-per-second agent observation without labeling the workloads.

I’ve also been able to keep multiple subagents doing real work at once. I haven’t done a controlled concurrency test yet, so I’m not going to give you a neat throughput number. What I can say is that the machine no longer feels like I have to choose between letting an agent work and getting on with my own work. 

M3 Ultra could run these models. That wasn’t the problem. The problem was that I was spending too much time waiting for the next useful thing to happen. M5 Ultra changes that enough that leaving the model loaded in the background now makes sense for how I work.

The cache still matters. A lot.

Here’s the part that gets lost when people only compare tokens per second: an agent isn’t one continuous stream of text. It reads context, calls a tool, gets something back, and starts another turn. Whether it can reuse the work it already did on that context makes a massive difference.

Prefill is the work of processing the input before the model generates anything. If the server can reuse cached context, there is less to process on the next turn. If that cache breaks, the model has to chew through those input tokens again. It doesn’t matter that decode is fast if you’re sitting there waiting for prefill.

OpenCode 2 has given me better cache reuse in my setup, which is one of the reasons I’ve liked using it with Qwen 3.8 Flash Next. But I’ve still seen a cache break turn into a 30–40 second wait on a long context, even on M5 Ultra. You notice very quickly how much the software matters.

That’s also why I don’t think it makes sense to talk about the hardware as if the performance is fixed on launch day. The silicon gives you a ceiling, but the kernels and the rest of the inference stack decide how close you get to it. I’ve been working on that part directly.

Writing kernels for M5 Ultra

A model is a series of operations on numbers: matrix multiplications, attention, normalization, and so on. A GPU kernel is the code that actually carries out one of those operations, or a group of them, across the GPU. How you divide the work, reuse data, and move it through memory has a lot to do with the speed you end up seeing. MLX supports custom Metal kernels, so you can tune the operations that matter for the model you are running.

Take a matrix multiplication. A good kernel breaks it into tiles, reuses data in fast on-chip storage, and keeps the matrix hardware doing useful work. You can also combine compatible operations instead of writing an intermediate result to memory just to read it back. That’s fusion. Metal’s tensor operations can use the Neural Accelerators inside M5-class GPU cores, but you still have to feed that hardware well. The same chip can give very different results depending on the software.

This is one of my favorite parts of using the Mac Studio. I’ve had Opus 5 and GPT 6 Astra writing and optimizing Metal kernels for Qwen 3.8 Flash Next. Give the agents the Metal 4.1 programming guide, a token budget, and the actual M5 Ultra to test against, and they can write code, compile it, run the model, look at the result, and try again. The GPU runs the model, but the CPU is hosting the whole loop that makes the GPU faster.

In this round of tuning, I took prefill from around 2,400 to 4,000 tokens per second on the same machine. That’s a 66.7% increase without changing the hardware. Decode is around 149 tokens per second in this optimized setup. I’m still working on it, and I think there’s more to get out of the chip.

What does that mean for an actual task? At those prefill rates, the same input takes about 40% less time to process. If your agent spends a lot of time on long prompts or cache misses, that is a meaningful change. If it spends most of its time in tools, the overall improvement will be smaller. This is why an enterprise deploying a particular model has a reason to optimize for its own context lengths and workload, not just run one benchmark and call it done.

What I mean by owning the CPU

A good example of what I mean happened yesterday. I was on a plane, using the Mac Studio remotely while I had agents working on the Qwen model. The CPU was running the agent harness, compiling Metal code, managing the tools, and hosting the rest of the development environment. The GPU was where I was testing the optimizations. I could check in while traveling, but the interesting part was how much the machine itself was doing.

Then the optimization work finished and the machine moved on to my hobby projects in Xcode. It didn’t need a new purpose. It was already a macOS development machine with a great CPU, memory, storage, and all the apps I use. That’s the part I keep coming back to when people frame this as a question of whether to own a GPU. You’re also getting the host that can do real work before, during, and after inference.

On M3 Ultra I could sometimes feel the desktop hitch or lag when a model was running and I was trying to use the computer for something else. I haven’t noticed that on M5 Ultra in my use so far. I’m sure there are workloads that could push it there, but being able to leave a model loaded, call it when I need it, and keep working normally changes how useful the whole setup is.

For a consumer, this can just be your computer. For an enterprise, maybe it is a developer workstation, maybe it is a local inference node, maybe it is the CPU hub coordinating several jobs, or maybe it is all of those things. That flexibility matters because nobody knows exactly which models, inference frameworks, or deployment patterns will make the most sense in a few years. If one part of the stack changes, the other part of the machine doesn’t suddenly stop being useful.

And yes, the CPU is fast

My Geekbench 7.0.0 Pro result was 3,666 single-core and 51,448 multi-core, on macOS 27.0 build 26A428. The screenshot identifies the machine as Mac17,15. That is a strong CPU result, and one reason I don’t want to treat the CPU as a footnote in a discussion about inference.

A benchmark score is nice, but what I actually notice is the computer doing the work around the model. It is building code, running Xcode, hosting agents, and staying responsive while inference is happening. I’d need a comparable M3 Ultra Geekbench 7 run before using these scores to make a precise generation-over-generation claim.

One result that really stood out was building a Rust project from scratch. It took about six minutes on M5 Ultra versus 20 minutes on M3 Ultra. That’s a strange enough difference that I think something else may be going on, but it was striking to see in my own work.

I wouldn’t attribute all of that to the silicon without repeating it with the same project revision, Rust toolchain, dependencies, caches, and build settings. It is a workflow observation for now. Still, it’s another reminder that this CPU is doing real work for me, not just sitting there to launch GPU jobs.

It’s a whole computer

Another example: I used a local model through OpenCode to build a benchmark app for Apple’s Foundation Models framework in Xcode. The model ran locally, called tools in Xcode, wrote the app, built it, tested it, and kept going through that loop on the Mac Studio. Think about what is happening there. The model and the application it is building are on the same machine, and the CPU and GPU each have work to do.

I already have an M3 Ultra cluster that I use remotely through a personal inference proxy, so I’m not opposed to separating the model from my working computer. I can access that cluster from my M5 Max MacBook Pro or the Mac Studio. With M5 Ultra, though, I’ve found more value in having a capable model and the rest of my work on the same machine. Even when I’m away from it, I can reach the full computer remotely.

None of this means I think every agent should run on a local model. For many tasks, I still want a frontier cloud model doing the main reasoning, with a local model as a tool when that makes sense. But there are also jobs I want to run entirely on my own machine, and I don’t want to stitch together a bunch of smaller models just to make that possible. Having enough memory and speed to run one capable model when I choose to is a big deal.

The same goes for privacy. Local isn’t automatically the right answer for everyone. Privacy is about choosing where the data goes and who you trust with it. If a team needs an air-gapped system, or wants certain documents and tools to stay on hardware it controls, this Mac can host the model and the applications together. If the workflow can use a cloud model, it can do that too. The point is having the choice.

To be blunt, this thing is insanely good. Qwen is fast enough for real agent work, the CPU can build and host everything around the model, and I can keep using macOS as my main computer without feeling it slow down. I can use it for inference, make it a CPU development hub, or do both at once. There are certainly heavier workloads that could push it, and I want to try clustering for larger models, but in my work so far the M5 Ultra Mac Studio feels frankly overpowered.

There has been a lot of discussion about what comes after the smartphone. AI creates new ways to interact with technology, so naturally people want to know what new hardware those interactions will require. But I think it is worth taking a step back and asking what we actually need from that hardware.

It already has so much of what would make personal intelligence useful. Your messages, calendar, photos, contacts, and the apps you use to get things done are accessible through it. Your Watch and AirPods connect to it. It has cameras to understand what you are looking at, microphones to capture what you hear, and a display to show you something when a spoken answer is not enough. You already carry it, charge it, and know how to use it.

Having data on a phone does not automatically mean an AI system can access all of it. Permissions, app integration, and the ability to retrieve the right information still matter. But the iPhone is already where many of those relationships come together. That is a significant starting advantage.

In my earlier piece on agents, I argued that what will sell is products and the features on those products. We can build increasingly capable agents, but consumers still need a reason to use them. Asking someone what they want to automate in their life can be surprisingly difficult. Helping them remember something they just heard is much easier to explain.

Apple’s Audio Intelligence features are a good example of this. Live Rewind lets you bring up text from the previous 15 seconds by double clicking the Watch’s Digital Crown. Someone mentions a book, and you miss the name. You can catch it and save it. Siri Recap serves a different purpose, producing high-level takeaways from conversations that you can view in the Siri app.

These are small, understandable things. They address moments people already experience, on a device that is already on their wrist. You do not have to figure out an entire new workflow before finding a use for them. Apple also does it in a far more secure way, with a new Secure Exclave that holds a ~15s audio buffer, sends it to the iPhone directly to transcribe it, and the iPhone sends the text back to the watch and the buffer disappears. There is no saving to storage, there are no uploads. From watch to phone to watch, and gone. 

The Watch also makes the broader hub idea more concrete. The iPhone does not need to be in your hand for its ecosystem to be useful. The Watch gives you an interaction on your wrist. AirPods give you a way to talk and listen when your hands are full. The phone provides the screen and access to your apps when you need to go further. Different devices can do what suits them while remaining part of the same experience.

This is why I think new form factors can actually strengthen the phone’s position. A new way to capture information or interact with an assistant can connect to the device that already has your context. It does not have to recreate that entire relationship from scratch. The question for new hardware becomes what useful role it plays in your life and how well it works with everything else.

There is another distinction here that matters: the location of the product experience and the location of the intelligence doing the work are separate decisions.

Calling the iPhone a personal intelligence hub does not require every model to run on it. Apple described both on-device models and more advanced models accessed through Private Cloud Compute. The phone can bring together the user’s context, permissions, and interactions while different workloads run where they make sense.

That is consistent with the broader point I made about local models. There are useful jobs for smaller models running close to your data. Sound recognition is one example Apple described running on the Watch itself. More complex work can have different requirements. We should evaluate those choices based on whether they deliver a reliable feature, at an acceptable cost, with appropriate privacy protections. The user’s experience is what ultimately needs to work.

Audio makes the personal context story especially interesting because so much of our lives never becomes a message or a calendar entry. A recommendation from a friend, something discussed in a meeting, or a detail from a conversation may be useful later. Giving people a simple way to retain some of that information could make their devices considerably more helpful.

But we should be precise about what Apple has described. This is not yet evidence of an assistant that remembers every conversation and automatically acts on everything it hears. The current actions for recaps are saving and sharing them. Creating a calendar event directly from a recap is not a supported action. There is still work between capturing something useful and reliably doing something with it. I think in the current environment until this technology is better understood by the public, going on the side of caution at the risk of having a less useful proactive product is the right decision. 

Apple also described a deliberate limit on what it retains. Its approach to recaps favors high-level notes without speaker attribution or a retained audio recording. That is a meaningful product decision. People may want help remembering a conversation without wanting a permanent recording of it.

That choice also makes the quality of the AI important. If the takeaway is wrong and there is no recording to check, the convenience starts to fall apart. Privacy, accuracy, and utility have to work together. Simply having the right devices and access to context does not solve that.

This is where Apple still has to execute. The personal intelligence hub is a compelling direction, but people will judge it through the moments when they ask for help. Does it find the right information? Does it understand what they mean? Can it complete the action without creating more work? Those are the tests that turn a good strategy into a product people rely on.

What I find encouraging is that the strategy starts with how people already live. They have their apps, their devices, their conversations, and their habits. Making those things work better together is a very large opportunity on its own.

The iPhone is exceptionally well suited to sit at the center of that. It already connects so much of your life. Apple’s opportunity is to make those connections useful in ways that previously required you to do the work yourself. That is a future for personal AI I can see people actually wanting.

And on a slightly different note, I have been arguing this idea that we don't need a new form factor and fundamentally a smartphone, smartwatch, and earbuds are a perfect starter pack for AI experiences. Voice and conversations with earbuds in, it can use the display on the watch to show and alert you to information, if you need to show the AI something pulling out your phone to use the camera is natural, it handles privacy well by matching actions the consumer normally takes, and isn't trying to sell extra hardware without a propose. I'm glad Apple started this route. I think consumer trust for dedicated AI hardware is non-existent, I think smart glasses are not going to end well and the experience isn't really there quite yet, and just making your phone feel more intelligent in an age where AI is being blamed for near universal price inflation in technology is bridging the gap better than I think many would expect. Apple did a good job and I'm glad they're going on what I think is the right path. 

And the price

The Duo starts at $1,999 for 256GB, running up to 2TB. Speculation had run to $2,500, which would have placed it above every Android foldable and framed it as a halo device. At $1,999 Apple lands it where everybody else is. It also sits $700 above the iPhone 18 Pro Max at $1,299. The price signals that while Apple might be supplied constraint it will sell as many as it can make. It also says that this is a category Apple is seriously pursuing.

The hardware holds up

The hinge is over a hundred precision components with carbon fiber support and a custom variable torque profile, plus integrated magnets that give it a sharp close.The display work is where the crease question gets answered. Apple brought its nano-texture finish to iPhone for the first time, using a custom polymer it claims is up to 40 percent stiffer than what the industry uses, with high-strength glass above and below the OLED panel, viscoelastic adhesives between the layers, and a titanium plate underneath. The result is matte, perceptually flat, and resistant to the dents and fingerprints that have affected plastic-topped foldables.

The software is the differentiator

Android foldables have had capable hardware for several generations. What they have lacked is an operating system that treats the second screen state as a design target in its own right.In iOS 27, the essential controls move to the side to preserve vertical space for scrolling content. The dock moves with them. The status bar has been redrawn as a compact element that tucks into whichever corner it needs. Liquid Glass was reworked for the new aspect ratio.There is no Face ID on the Duo and to some people this seems like a step back. But considering all the different orientations you can hold the phone at Touch ID built into the side button is the least frictionless experience especially as you can enroll different fingers depending on how you hold the device. Considering the changes of Duo bu=yers also having an Apple Watch are high, Apple added Apple Watch unlock for added conveninece.Separately, the FaceTime camera on the inner display sits under the panel and stays hidden until a call comes in.

App support? Check!

Netflix is the clearest example. Its clips feature runs on the closed display for scrolling through content, and opens to the inner display to watch. The app renders its own reflections that respond as you move the hinge, which is the kind of detail developers only build when they expect the hardware to matter. Zoom put participants and shared content on screen together. Slack showed the full workspace alongside messages. Apple also opened a Core AI framework for on-device models, which a content creation app was already using in the labs.Developers showed up for the buyers. A $1,999 device selects for exactly the customers a subscription business wants, and that math works at volumes well below iPhone Pro.

Who buys it, and what Apple did not say

Most iPhone Duo buyers will come from inside Apple's own base, upgrading from Pro or Pro Max. Some Android switchers will cross over for the software and the app support. Samsung can match the panel and the hinge.Apple did show productivity. Split View puts two apps side by side on an iPhone for the first time, and you can pair apps to return to them, run two Safari windows, and drag content between them. The framing around it stayed light. The center of the pitch was the same one Apple used at the original iPad launch: you already love this device, here it is with more room.That framing carries risk. iPhone customers have been pragmatic about industrial design for fifteen years and grew up on the glass rectangle. Android buyers showed more appetite for form-factor experimentation, which is part of why foldables took root there first.

Why now

Apple does not enter markets early. It waited on the larger-screen iPhone until the market had proven the demand, and it has done the same here. Foldables have been shipping on Android since 2019.What has changed is what a bigger screen is for. Apple has spent years building a content business it wants people to actually sit and watch. It now has a version of Siri that, from early beta, looks materially more capable than what it replaced. Those two things put pressure in the same direction, toward more display, and toward a device you interact with conversationally as much as by tapping.The framing across the whole lineup was iPhone as the intelligent personal hub, and John Ternus used it to close out the question everyone keeps asking about what comes after the phone. His answer is that nothing comes after it. The phone is the hub, and more devices arrive around it.That is an expansion argument. Watch unlock on the Duo is a small feature and a large signal: a device you already own authenticating a device you just bought. The audio intelligence work on Series 12 comes at it from the other side, putting ambient AI on the wrist and leaning on the phone for the context that makes it useful. Each device Apple adds pulls from the same personal context and gives something back to it.So the question for next year stops being what replaces the iPhone and becomes what attaches to it.

There is no iPhone 18

This is the first iPhone launch where only the higher end of the lineup arrives in the fall. Pro, Pro Max, and Duo. The iPhone 17 stays on sale as the entry point.One reading is scheduling. Apple brings the base model in the spring, closer to when the mass-market Android devices land, and gets a second moment in the calendar instead of one crowded September.The other reading is about cost. Memory prices have moved against everyone building phones this year, and a base iPhone 18 would have to absorb that. Raising the price of the model that sells to the segment already telling us they do not need a new phone every year is a hard sell. Holding the 17 at a $100 more serves the buyer better than a more expensive 18 would.

October is the number that matters

IPhone Pro preorders open Saturday with availability September 18. Duo preorders open October 16, with availability October 23 in more than 70 countries and October 30 in 28 more. That is a five-week gap, and nobody buys a foldable without trying it for size. Buyers who would otherwise take a Pro now have a reason to wait, and it has nothing to do with price.If Pro preorders come in soft, the temptation will be to read weak iPhone demand. The likelier explanation is deferred decisions, and the number to watch is combined Pro and Duo sell-through in November.Apple has a partial hedge here. iPhone Handoff lets one number live on two iPhones, active on whichever you pick up. At launch that is T-Mobile in the US and Deutsche Telekom in Germany, with more carriers follow.

The camera is still the compromise

The Duo runs a dual camera system: a 48MP Fusion Main camera with an integrated optical-quality 2x telephoto, a fast aperture, a large sensor and sensor-shift OIS, plus the same 48MP Fusion Ultra Wide as the 18 Pro, which also handles macro. It shoots 4K120 in Dolby Vision. Apple built form-factor-specific features around it: Duo FaceTime, Duo Preview so your subject can see the framing on the outer display, and Smart Take, which watches the scene and fires when people are posing.There is no dedicated telephoto. But Apple said they learned from iPhone Air about how many people depend on the ultra wide. That is a fair account of the tradeoff, and it is still a tradeoff.Battery is the better story. Two cells with custom silicon rebalancing them so they present as one, delivering 31 hours of video on the inner display, 44 on the outer, and 24 hours of typical use per charge with the displays used equally. Fifty percent charge in around 20 minutes wired, 30 minutes on MagSafe or Qi2. That puts it level with the iPhone 18 Pro on everyday use, which is not where foldables have historically landed.Both objections are smaller than they have been on any previous folding phone.

What is left is the question Apple cannot engineer past: whether its own customers, told for fifteen years that the glass rectangle was the answer, want a phone that folds.

Google has a new strategy around this, where the news embargo and videos from tech creators come out a few hours before the actual event. The last year it was around 2 hours before, this year it's 8 hours before. These events were also hosted by celebrities and influencers (who are also celebrities).

Normally, events are pretty simple: live stream with a keynote and the news comes out when it starts (Samsung) or as it's going (Apple). Google did this up until last year as well! The reason I consider this important context is this is the standard launch, everyone does it like this. Every tech brand did it like this, even car brands and others like Taco Bell started to copy this format. Taco Bell example may be a bit off because it was more of a parody of it, but you get the point.

These events for tech brands tend to work well because the market already knows about the products. Samsung and Apple are the largest smartphone brands in terms of volume, together they ship around 40-45% of all smartphones, around half a billion phones annually. There is interest in these phones from demographics that don't know or understand technology, you find normal people (think your parents, grandparents, friends, random person on the NYC subway) talking about the new iPhone and to a lesser extent Samsung Galaxy, but they are still aware of it.

Google is in a different boat, while they have been growing Pixel for 11 generations now, Pixel is still not a well known brand in the same way the others are. It is more in a way that those people are aware that Google makes phones and those phones are called Pixel, but not nothing about them. When someone will go to buy a phone, this makes choosing a Pixel uncomfortable vs. the reliable devices, Samsung and Apple. It becomes a familiarity thing.

Making these devices more familiar to the masses becomes extremely difficult. It's a leap of faith to upgrade devices, let alone swap the brand you have been using for years, regardless of that experience being good or bad. A known bad experience is better than an unknown for a lot of people.

Now, the reason this preamble is included is pretty simple: because of the aforementioned market conditions, Google needs to build trust, they need to build familiarity, and they need to build recognition. There are many ways to try to do this, one way could be by becoming a fashion brand and start with high style designs and lifestyle, this is what Nothing is trying. Another could be going for the total extremes in their products to make the best of the best and market themselves, this is what a lot of Chinese OEMs try.

Google, on the other hand, is taking an approach I think I like a lot more. They are trying to appeal to consumers with public figures and celebrities that people know and trust, building products with and around their needs, and using them to grow awareness. The Steph Curry Fitbit Air and Pixel Watch 5 are perfect examples. If Steph Curry is helping make it, you can trust that it's good. If Alex Cooper starts to talk about the Pixel 11 and it's camera, you can trust that the camera is good. These events are now a showcase of the products and how the products are used, rather than a keynote trying to explain tech specs.

You can try to combine these into a single event and keynote, and it's extremely difficult. Arguably a single fruit-company is the only one to do it well. Google decided to separate it, in the morning (10am today) Google released all of the specs, reviewers posted their hands-on videos, and media/influencers/press received their devices. This afternoon at 6pm, Trevor Noah hosts the Made by Google event. There is an 8-hour window where media, press, fans, enthusiasts can all be excited and enjoy the devices, like they otherwise would be.

This is an exciting launch, regardless of the initial coverage kinda calling it boring. I haven't used the phones yet (I have no clue when I will get mine), but over the past few days I talked to a ton of people who used it at the early content capture. Every comment was about how the camera processing is really good to the point they consider this a multi-generational leap back to what the Pixel's used to feel like around Pixel 4, the thing felt fast, didn't overheat, and battery life was good. These are basically my complaints from Pixel 10 Pro XL. This initial coverage gives the enthusiasts (hello, that is me) time to be excited! I don't need to see someone on stage for this to be true and exciting.

This afternoon, with actual live event, that's time to get others excited and interested. I know about the Pixel series, do I need to see Steph Curry, Trevor Noah, Alex Cooper, and everyone else whose names I am not familiar with using the features I already know about? Absolutely not, but I will enjoy every second of it.

You know who needs to see that? My mother who is thinking about getting a Pixel 11 Pro to replace her aging Galaxy S24. My girlfriend who didn't know Google made phones until I told her about this event. My neighbor who hasn't looked at anything but an iPhone in years. We as tech enthusiasts and creators and fans of technology take for granted the knowledge and education we have about these devices from pure genuine interest. These are names they recognize, maybe people they follow or care about seeing. Maybe it's just curiosity to see how what Google must be spending a fortune on.

Google needs to build that same genuine interest and trust to make Pixel a trusted choice over the known devices. I think this is a great way to do it. Some find it cheesy, but it's not meant for them. That's completely ok. If you're reading this, it's probably not meant for you either. Also ok. What is important is understanding why they make these choices. I could be totally wrong, but this is a strategy that I think is working exceptionally well for Google and makes the launches just far more interesting than looking at specs.

and just to acknowledge it, yes the specs are minor iterative upgrades but the reality is there really isn't that much more to improve with current technology. Tensor G6 has new IP in the SoC to help power the new imaging pipeline, that imaging pipeline has a massive leap in image quality. The battery life is better thanks to the chip too, even without silicon carbon batteries. The display is brighter, the thing runs cooler, there are more intelligent Gemini features. The Pixel Watch 5 uses a new large sensor model architecture for health tracking that can detect insulin resistance trends and is far more accurate. I don't care if it doesn't look different, that's huge! What else could you want or need? What else would make it better? I can't think of anything that I'd want. The iterative upgrades look small on paper and might not be exciting but they create new experiences and solve pain points that existed before, and no longer exist. That is exciting. Smartphones are mature, the market is mature, how to market them, how to talk about them, how to get people interest in them is different in a mature market. Google did that well.

That sequence explains why The next measure of AI momentum is work transformed, published by Jared Spataro on the Microsoft 365 blog the morning after FY26 Q4 earnings, spends its customer proof points on agent counts. Premera Blue Cross has built more than 900 agents, most of them by employees outside IT. Microsoft’s own supply chain team runs more than 70 across planning, sourcing, fulfillment and logistics. Jared Spataro closes on the growth and governance of agents as the measure that matters next, which is a long way from the hours-saved framing Copilot launched with.

What the seat actually installs

On the January earnings call Satya Nadella described the data underneath Microsoft 365 as the most important database any company running Microsoft has, pointing to the tacit information it holds about people, relationships, projects and artifacts. Copilot Chat users on the bundled tier do not get Work IQ grounding. The paid seat is the thing that turns that database into something an agent can query with permissions attached.

This makes broad deployment a technical requirement in its own right. A partial rollout produces a partial context graph, and agents grounded in a partial graph produce generic output. Enterprises that bought a pilot allocation of 500 seats to test ROI were, without knowing it, testing the version of the product least likely to work. The economics that looked unjustifiable at $30 per user for time savings look different when the seat is the prerequisite for an agent that closes a workflow.

Agent 365 extends the same logic to identity. Agents carry their own credentials and permissions, Entra governs them, and Microsoft Scout, introduced in June, runs in the background under that model. The governance layer is the asset competitors cannot assemble quickly, because it depends on already sitting inside the enterprise directory.

The trajectory, with two qualifiers

Microsoft had never published a Copilot paid seat count before January 2026. It disclosed 15 million that month, 20 million in April, and more than 30 million this week, with net adds going from 5 million to 10 million quarter over quarter. Customers running more than 50,000 seats grew more than sevenfold year over year. NHS England is deploying to 505,000 clinicians and staff, HSBC has committed to 200,000 seats, and EY is putting the E7 suite in front of 400,000 employees.

Q4 is Microsoft’s fiscal year end, when large deals cluster to capture year-end discounting on exactly the product the vendor wants to push, so the 50 percent sequential jump needs one more quarter before it reads as a run rate. And the attach rate stays small. Against a Microsoft 365 commercial base of roughly 464 million paid seats, derived from the 6 percent annual growth Microsoft reported against the 450 million figure it last published in January, 30 million works out to about 6.5 percent. The seat count doubled in six months and more than 93 percent of the installed base has not bought.

Comprehensive stopped being the differentiator

Microsoft’s portfolio breadth is real. Microsoft 365 Copilot, GitHub Copilot, Security Copilot, Copilot Studio, Agent 365, Dynamics, Power Platform, Azure AI Foundry. No competitor matches that span from developer tooling through security to the knowledge worker.

Breadth stopped sorting the field sometime last year. Google has Gemini Enterprise, Workspace, the Gemini Enterprise Agent Platform that replaced Vertex branding at Cloud Next, Code Assist, and a security line that now includes Gemini 3.5 Flash Cyber. Andy Jassy has been describing AWS share gains in terms of a top to bottom stack since 2025, and Quick Suite put Amazon directly into the knowledge worker layer. Every hyperscaler claims full coverage now. What separates Microsoft is the installed base and the directory, which is a distribution argument.

What this means for a Copilot evaluation

For anyone still running a Copilot pilot, the allocation size is the thing to reconsider. Testing 500 seats against a time-savings baseline measures the configuration least likely to produce a return, because the context graph underneath it is thin and the agents built on it stay generic. An evaluation that produces a usable answer covers a full function, targets one workflow that crosses systems, and measures completion rather than minutes saved.

Fiscal year end matters to the negotiation. The cohort above 50,000 seats grew more than sevenfold this year, which tells you volume pricing exists at thresholds Microsoft has an incentive to reach. Buyers sitting between 5,000 and 50,000 seats have more room than list price suggests, and the leverage is largest in the quarter Microsoft is trying to close.

Microsoft is competing with its own supplier

The most revealing line in the Copilot post sits in footnote 2. Microsoft ran 125 test runs across 12 prompts comparing Copilot Cowork against Claude Cowork with the Microsoft 365 connector, both running Opus 4.8, and reports its own product came in 30 to 40 percent cheaper. A post celebrating 30 million paid seats does not normally carry a competitor cost benchmark in its footnotes.

The threat that footnote names is a model provider reaching Microsoft’s own surfaces through a connector, without Microsoft’s seat and without Microsoft’s margin. Copilot Cowork runs on agentic technology Microsoft brought in from Anthropic, and Microsoft booked a $3.2 billion gain on its Anthropic investment in the same quarter it published the comparison.

Nadella made the architecture explicit on the same call. Answering UBS analyst Karl Keirstead, he said the platform design requires you to “keep your harness separate from the model,” with memory and context held externally so that any model is swappable at any time. He described models as an input to the enterprise, and the goal as a firm that controls its own human capital and token capital.

That is the grounding argument stated as vendor strategy, and it names the layer Microsoft intends to own. It also explains the benchmarking. Footnote 2 puts Copilot Cowork against Claude Cowork on cost. On the same call Nadella claimed MAI-Cyber-1-Flash outperforms a much larger Mythos model at half the cost when paired with Microsoft’s multi-agent security harness. Two comparisons against Anthropic in the same week, from a company holding a stake in it.

Cowork’s multi-model design lets a task pull whichever model fits, which is Microsoft declaring models a commodity input it intends to arbitrage. The position holds while the harness stays proprietary and the context stays inside Microsoft’s directory. Thirty million seats are what keep it there.