Apple Intelligence at WWDC 2026 The Moment It Got Good

0:00Lucas: So today is June 8, 2026. Apple just wrapped its WWDC keynote a few hours ago. And I think we need to talk about the moment Apple Intelligence stopped being a concept and started being something you can actually use. Luna: Yeah — I was watching the livestream. There was that moment when they showed Siri finishing a sentence inside a third-party messaging app, and I actually leaned forward. That's new. Lucas: Right. And I think that specific demo is the episode's angle. Because for years, 'AI on iPhone' has been a slideshow promise. Today they shipped an on-device model that can complete your sentences, edit your photos, and string together multi-step workflows — all without sending your data to a server. Luna: Let's talk about how that actually works. Because I've been reading about 'on-device inference' forever, but this is the first time I've seen it feel fast. Lucas: Before we dig into the architecture — a quick note. We keep this show ad-free, and that's entirely because of listeners who chip in a couple of dollars a month on Buy Me a Coffee. If today's tech conversation gives you something usable, buy me a coffee dot com slash fexingo. It genuinely makes a difference for a show like this. Luna: Yeah, it's a small thing that adds up. Keeps us independent, no sponsors to answer to. Lucas: Exactly. So back to Apple's AI — the big technical shift they announced is a new foundation model that runs entirely on the Neural Engine in the A19 and M6 chips. It's not a distilled version of a cloud model. It's a model trained specifically for the edge, with about three billion parameters. Luna: Three billion parameters feels small compared to what OpenAI or Google are doing. But that's kind of the point, isn't it? Lucas: That is exactly the point. A three-billion-parameter model can fit into the memory bandwidth of a phone and run inference in under 300 milliseconds for a short text prompt. Apple isn't trying to win the benchmark race. They're trying to make the thing you do twenty times a day — typing a message, setting a reminder, cropping a photo — feel intelligent. Luna: And they showed it working across apps. Not just Apple's apps — they had a demo where Siri completed a sentence inside Slack. That requires the model to have a shared context layer. How do they do that without sending data to the cloud? Lucas: They call it 'on-device semantic index.' It's a local vector database that indexes your messages, photos, calendar events, and app data, all encrypted on the device. The model queries that index for context. So when you start typing 'Are we still on for...' in any text field, the model knows you're referring to a calendar event and can auto-complete with the time and place. Luna: That's impressive, but also a little creepy. Apple leaned hard on the privacy angle — they said repeatedly that no data leaves the device. Lucas: They did, and I think that's their competitive moat. Microsoft's Copilot and Google's Gemini are cloud-first. Even with privacy promises, your query goes to a server at some point. Apple is saying: the model lives in your phone, full stop. For enterprise users, that's a big deal. Luna: And then they took it a step further with Shortcuts. They showed a demo where you could say 'Every morning, summarize my emails and read me the top three news headlines' — and Shortcuts built a workflow automatically, chaining an email summary action to a news fetch action to a text to speech action. Lucas: That's the part that might have the longest tail. Because Apple just turned every iPhone user into a potential automation builder. You don't write code — you describe what you want in natural language, and the AI writes the Shortcut for you. That's going to unlock a lot of use cases that developers never bothered to build. Luna: It reminds me of what we saw with the early days of the iPhone App Store. The platform itself becomes the innovation. But let's talk about the market reaction. Because today, while Apple's keynote was happening, a lot of AI hardware stocks were getting crushed. Lucas: Yeah, Broadcom down almost 18 percent in five days. Nvidia down 6.5. AMD down 6.3. The SOXX semiconductor index down 5.7 percent. That's not a reaction to Apple specifically — that's a broader rotation. But Apple's on-device AI announcement adds a narrative: you don't need a data center GPU to run useful AI. You need a good Neural Engine. Luna: Which is bad for the companies selling $30,000 GPUs, but good for Apple and Qualcomm. Though Qualcomm is down 9.6 percent in five days, so the market isn't buying that argument yet. Lucas: No, the selloff in semis this week feels more macro — rate fears, maybe some profit-taking after a huge run. But the Apple news does create a new question: if the most personal device you own can run a three-billion-parameter model locally, do you need as much cloud inference? For some tasks, definitely not. And that's a threat to the cloud GPU buildout thesis. Luna: Microsoft and Google are both betting that inference will shift to the edge eventually. But Apple just made it happen on a device with two billion active users. That's a forcing function. Lucas: Let's talk about what this means for developers. Because Apple also announced that the new AI model is accessible via a native API. So any developer can call the on-device model for text completion, image analysis, or summarization — with no cloud charge. Luna: No per-token cost. That's huge. Right now, if you build an AI feature into your app using OpenAI, you pay per API call. Apple just made it free — well, free with the cost of the device. Lucas: That flips the economics of AI features on its head. Suddenly, a small developer can add smart auto-complete or photo tagging without paying a monthly AI bill. The barrier to entry drops. I expect we'll see a flood of apps with local AI features within six months. Luna: But there's a catch: the model is limited to three billion parameters. It won't write a novel or generate a photorealistic image from scratch. Apple is betting that most users don't need that. They need their phone to finish a sentence, suggest a reply, or find a photo of their dog at the beach. Lucas: Right. And Apple has a separate mechanism for heavy lifting: they call it 'Private Cloud Compute.' If the on-device model can't handle a request — like generating a detailed image — it can send the query to Apple's own servers, which run larger models. But they promised those servers log no data and run on Apple silicon. Luna: That's the part that reassures privacy advocates. It's not all-or-nothing. The device tries first, and only if it fails does it ask the cloud. And even then, the cloud is Apple's own infrastructure, not third-party. Lucas: So the big picture: Apple just entered the AI race with a strategy that's almost opposite to everyone else's. Microsoft and Google are racing to build the largest models and put them in the cloud. Apple is building a small, efficient model and putting it in your pocket. And for most of what people actually do on a phone, the small model is enough. Luna: I want to zoom in on one more thing: the timing. This WWDC is happening during a week when AI stocks are getting crushed. The narrative has been 'AI is a bubble, spending is out of control.' Apple's approach offers a counter-narrative: AI doesn't have to be expensive. It can be a free feature on a device you already own. Lucas: That's a powerful reframe. And if Apple proves that on-device AI drives user engagement and upgrades — if people start buying new iPhones because Siri is finally good — then the whole industry shifts. Qualcomm, MediaTek, and others will race to match Apple's Neural Engine performance. Luna: Intel actually went up 2.5 percent this week, which is notable. Maybe the market is betting that on-device AI benefits the PC, too. Intel's Lunar Lake chips have a decent NPU. Lucas: Yeah, Intel is one of the few green spots in semis. And Apple's announcement actually helps Intel's argument: AI on the device is real, so you need a capable client processor. That's good for Intel and AMD, though AMD is down this week anyway. Luna: Alright, so the concrete takeaway: Apple made AI useful on a phone today, not in a slideshow. What's the one thing you'd tell a friend about this? Lucas: I'd say: Apple just taught your iPhone to finish your sentences, and it does it without sending anything to the cloud. That's the moment local AI became real for two billion people. And if you're a developer, you should start thinking about what you can build with a free, on-device model that runs in milliseconds. Luna: Good note to end on. Next episode we'll look at what the Pentagon's new restrictions on Chinese AI companies mean for the supply chain. Lucas: Yeah, that headliner from today — Alibaba, Baidu, BYD and Unitree being named as supporting China's military — is going to ripple through chip stocks. We'll dig into it.