Finn Brooks: Okay, weird question before we get into anything — do you think Apple employees had to read the Broadcom SEC filing on July 6th to find out what their own company signed?
Juniper Vale: That is a very weird question. Why are you asking that?
Finn Brooks: Because that's how I found out. Broadcom discloses — through the SEC, not Apple, Broadcom — that they've got new multi-year agreements with Apple running through 2031 for custom ASIC silicon. Apple says nothing. Broadcom's stock jumps 4% premarket. And I'm reading a regulatory filing to understand what one of the world's most famous companies is building.
Juniper Vale: I mean, that's actually pretty normal for supply chain disclosures — the supplier often has more obligation to report than the customer does. But you're right that it's a funny inversion given Apple's whole brand.
Finn Brooks: The brand that says we control everything. Apple Silicon, right — the whole point was we don't need anyone. And then Broadcom gets to announce that actually Apple needs them through 2031, and Apple is 20% of their entire annual revenue — like, that's not a vendor, that's a lifeline going both directions.
Juniper Vale: Both directions is the key part. This is mutual lock-in, not just Apple de-risking supply chain.
Finn Brooks: Which makes me want to ask — was vertical integration ever actually the reality, or was it always just the story Apple told really, really well?
Juniper Vale: That's what I want to dig into, because I think the answer is more interesting than either a yes or a no.
Finn Brooks: But that's the part I keep tripping on — Apple Silicon exists, right, they design their own chips, so why does Broadcom still have them locked in through 2031?
Juniper Vale: Because Apple Silicon and what Broadcom makes are solving completely different problems. Think of it like this — a master chef can build a perfect kitchen, but they're still not going to repair their own HVAC system. Same building, totally separate expertise.
Finn Brooks: Okay wait, say more.
Juniper Vale: An ASIC — Application-Specific Integrated Circuit — is a chip designed for exactly one job. Not computing in general, one job. So a 5G radio chip is optimized purely for processing radio signals. It's wildly efficient at that, and completely useless for anything else. Apple Silicon runs your whole operating system. Those are not the same engineering problem, not even close. The May 2023 deal was specifically Broadcom developing 5G RF components manufactured in the US — that's the foundation the 2031 extension is building on.
Finn Brooks: So Apple can't just... reassign their chip team to figure out radio frequency stuff?
Juniper Vale: I mean, technically they could try — they have the money, they have engineers. But RF design, Wi-Fi, Bluetooth silicon, that's decades of specialized IP. Broadcom's entire identity is built around that. And then in 2024, Apple and Broadcom actually started collaborating on processor designs targeting AI inference specifically for servers, which is a different layer again. So the deal covers RF chips, Wi-Fi, Bluetooth, networking components, and now AI server silicon. That's not one gap Apple has — that's several.
Finn Brooks: Okay that reframes it completely — I was reading this as Apple admitting weakness but it's actually just... Apple being a software and SoC company that doesn't pretend to be a radio company.
Juniper Vale: Right — but don't let that flip too clean either. The question worth sitting with is whether outsourcing this specific layer is a smart division of labor or whether it creates real exposure. Apple is 20% of Broadcom's revenue. That's mutual dependency, not just smart strategy.
Finn Brooks: Mutual dependency, right — but then Craig Federighi stands up at WWDC 2026 and discloses that AFM Cloud Pro, Apple's actual cloud AI model, runs on Nvidia GPUs inside Google's servers. And I just... that broke my brain a little. Because now it's not two companies locked together, it's four — Apple, Broadcom, Nvidia, Google — and Apple is supposedly running the privacy-first AI narrative across all of it.
Juniper Vale: That's the part that doesn't flatten out cleanly. And Federighi said — on stage — that AFM Cloud Pro is comparable in capability to Google's Gemini frontier models. Which is wild, because that means Apple is benchmarking its AI against Google... and running it on Google's hardware.
Finn Brooks: On Google's hardware!
Juniper Vale: So picture a developer, January 2027, trying to decide whether to build on Apple's AI server stack. She pulls up the Broadcom announcement from July 2026 — Broadcom ASICs for the server layer, targeting deployment that same year. But then she cross-references the WWDC disclosure and realizes the cloud inference is Nvidia GPUs in Google data centers. And she genuinely cannot tell which layer her app will actually touch.
Finn Brooks: That's — yeah. That's the hot take holding up. Apple Silicon on device, Broadcom ASICs potentially in servers, Nvidia in Google's racks for inference — those aren't redundant layers, they don't even obviously connect.
Juniper Vale: And Private Cloud Compute was specifically Apple's answer to the trust question — the whole premise was that your data processes on Apple-controlled infrastructure. I mean, that promise is doing a lot of work once Nvidia and Google are in the room.
Finn Brooks: No, I don't buy that Apple just... forgot about that. There has to be some architectural separation they're claiming. But I don't know what it is.
Juniper Vale: Neither does anyone outside Apple, which is actually the problem. And the chip story — wait, this part matters — that's only half of what Apple is juggling here. The geopolitical layer of how Apple is sourcing across different continents, that carries its own separate category of risk that we should get into.
Finn Brooks: Oh I want that. Because CXMT is sitting right there and I've been holding it.
Juniper Vale: CXMT is — yeah, that's the other end of the bet. So Apple is testing ChangXin Memory Technologies chips, which is Chinese state-backed DRAM, for devices sold inside China. And simultaneously lobbying Washington to allow broader use of CXMT products. Both things, at once.
Finn Brooks: Wait — lobbying to expand CXMT access while locking in US-manufactured Broadcom components on the other side?
Juniper Vale: That's the actual shape of it. The May 2023 deal with Broadcom was specifically about manufacturing 5G RF components in the United States — that's not incidental, that's a domestic supply chain argument Apple is making to regulators. And then over here, they're running a parallel track with a Chinese state-backed chipmaker. Those aren't two versions of the same strategy, they're opposite bets.
Finn Brooks: Okay that is — I mean, that's not hedging, that's contradiction. How do you square that publicly?
Juniper Vale: You don't, publicly. You call both of them supply chain resilience and hope nobody reads the full parts list. But here's what doesn't survive scrutiny — locking in Broadcom through 2031 concentrates critical AI and connectivity infrastructure in one external supplier right at the moment US-China tech friction is at its peak. That doesn't reduce geopolitical risk, it just trades concentration in one geography for concentration in another.
Finn Brooks: Both ends are exposed simultaneously.
Juniper Vale: Right — regulatory risk on the CXMT side if Washington tightens restrictions, single-supplier risk on the Broadcom side if that relationship breaks. And Apple's vertical integration story, the Apple Silicon narrative, it survives only if you never look past the device layer. The moment you include the server stack, the memory sourcing, the cloud inference — I mean, the actual full picture — it's a company running two opposite geopolitical bets and calling both of them strategy.
Finn Brooks: That's the defensible version. Not that vertical integration is fake — it's that the map Apple shows you stops at the edges of their own silicon.
Juniper Vale: And the test of which story is true is actually pretty simple, when you think about it. If those Broadcom ASICs show up inside an actual Apple server rack — custom silicon, purpose-built for AI inference, not just Wi-Fi and RF — then the 2031 commitment means what Apple wants it to mean. If they don't, then the deal is connectivity continuity dressed up in AI language, and the AI story and the hardware reality are just... drifting further apart.
Finn Brooks: Okay, fine. Maybe it's not a confession — maybe it's very expensive strategic humility. I'll give them that much. But right now, today, the most privacy-forward AI company on the planet is running its frontier model on Nvidia GPUs inside Google's data center. And that sentence should feel weirder than it does.
Juniper Vale: It really should. I think we've just gotten used to Apple's framing moving faster than Apple's infrastructure.
Finn Brooks: March 2027 is going to be a very telling moment. Thanks for actually walking me through all of this — I came in thinking supply chain drama, and I'm leaving with something stranger.
