Michael C. Vincent: You texted me at eleven last night with just the words 'AMD won' — I'm going to need context.
Hope Sterling: I stand by it — okay, partially — because Forbes on July 23rd, 2026 literally framed AMD competing for humanoid robot compute as a watershed shift in the AI chip landscape, and the reason is Foundation Future Industries putting AMD Ryzen AI Embedded X100 Series chips into their Phantom humanoid robots with a 2.5x faster inference claim over Nvidia, and I just — I sent that text and went to sleep feeling vindicated.
Michael C. Vincent: You went to sleep on a $24 million startup.
Hope Sterling: Wait — is that how big Foundation Future Industries is? Twenty-four million?
Michael C. Vincent: That is their Pentagon contract. Total. One analyst has already said publicly it's a PR win that won't move revenue. And Eric Trump's connection to the company is a significant part of why the story traveled as far as it did.
Hope Sterling: Okay that's — I mean that reframes it, but like, does the size of the company actually change whether AMD Ryzen AI Embedded is a real threat to Nvidia in this space? Because the 2.5x inference number is the number, right, regardless of who's shipping it first.
Michael C. Vincent: That is exactly the right place to push. And it leads us somewhere the Forbes piece doesn't quite go.
Hope Sterling: Which is — what are we actually trying to settle here, the benchmark or the ecosystem?
Michael C. Vincent: The ecosystem. Full stop. Choosing AMD chips for your Phantom robot is like switching from a Mac to a Dell. The hardware decision is real. But everything you've built your life around — the apps, the muscle memory, the workflows — that's what actually owns you. And in robotics compute, that's CUDA.
Hope Sterling: Okay but wait — is CUDA literally that sticky? Like, developers can't just... switch?
Michael C. Vincent: Nvidia holds somewhere between 70 and 95 percent of the AI accelerator market. That's not a product advantage at this point — that's two generations of engineers who learned to think inside CUDA. ROCm exists, AMD's pushing it hard, but you don't rebuild a software ecosystem with a press release.
Hope Sterling: And an AMD VP literally said CUDA is a 'non-event' — which, I mean, I get why they're saying it, but that feels like — no, that feels like hoping nobody notices the shoes everyone's already wearing.
Michael C. Vincent: Now, the record is murkier than that. AMD's Advancing AI 2026 event was a coordinated blitz — MI400 Series GPUs, the Helios rack-scale platform, ROCm.ai developer tools. The Phantom announcement was one PR piece inside a much larger machine. Foundation Future Industries buying off-the-shelf Ryzen AI Embedded X100 Series processors was — well, it was a shopping decision.
Hope Sterling: A shopping decision that got Forbes treatment because — oh. Because Eric Trump.
Michael C. Vincent: That is a significant part of why the story traveled. A $24 million Pentagon contract startup using off-the-shelf embedded chips is not a platform win for AMD — it's a line item. The question worth asking is whether any serious robotics OEM defects from Jetson Thor, which is purpose-built, delivers 865 teraflops, exists specifically for humanoid deployment. That's the actual competitive test.
Hope Sterling: So the Phantom announcement is real — like, it happened — but the signal is just... narrower than the headline made it sound.
Michael C. Vincent: Narrower, yes. But here's where your instinct actually had something real underneath it — because robotics is not a data center. That's the crack.
Hope Sterling: Wait — like, structurally different? Not just smaller scale?
Michael C. Vincent: Fundamentally different. CUDA's moat was forged in centralized cloud compute — massive racks, controlled environment, a server farm in Virginia running the same inference job ten thousand times. A humanoid robot is edge-deployed, real-time, safety-critical. The latency budget is milliseconds, not seconds. That's not where CUDA's gravity lives.
Hope Sterling: Picture an engineer, it's eleven at night, she's got a demo in the morning for her humanoid startup, and she's staring at her compute stack options. She can plug into Jetson Thor — 865 teraflops, CUDA-native pipelines already built — and she's running by midnight. Or she bets on ROCm, rebuilds her simulation stack from scratch, and maybe misses the demo entirely.
Michael C. Vincent: That is exactly the right scenario. And she picks Jetson Thor. Every time.
Hope Sterling: Right — but that's the cloud-developer logic bleeding over. What if her robot literally doesn't need cloud-scale training pipelines on-device? Like, what if the architecture means CUDA's whole advantage just... doesn't translate the same way out on the factory floor?
Michael C. Vincent: That's where you're not wrong. Jensen Huang called robotics a multitrillion-dollar opportunity — and Nvidia responded by shrinking Blackwell to half its size for mass-market edge deployments. They are scrambling to re-win a moat they haven't actually built yet in this space.
Hope Sterling: So that AMD VP saying CUDA is a non-event — I mean, maybe that's not cope, maybe that's actually the one place they see daylight?
Michael C. Vincent: Visionary or cope — I genuinely cannot call it yet. What I can say is the real fight isn't 2026 revenue at all, and the US-China supply chain split we haven't touched yet makes this whole thing stranger.
Hope Sterling: Oh I want that — because that's a whole separate thing rattling around in my head and I don't fully have it yet.
Michael C. Vincent: The US-China split is where the structural picture actually resolves. TrendForce published a supply chain analysis this year — the finding is almost elegant in how stark it is. Chinese manufacturers are building the bodies. American chips are providing the brains. Actuators, reducers, chassis — China's scale manufacturing advantage there is real and growing. The intelligence layer, the edge-AI compute module, that stays US-designed.
Hope Sterling: Wait — so there's literally a 2026 collaboration where a Chinese-built chassis is paired with a US edge-AI module and that's like, not a weird one-off, that's the pattern?
Michael C. Vincent: It became one of the most discussed humanoid pairings of the year precisely because it made the split visible. And that split is where both Nvidia and AMD are actually well-positioned — the intelligence layer is the highest-margin component. Neither company builds actuators.
Hope Sterling: Okay but — wait, that actually changes the competitive story, because if the margin lives in the chip and both companies are fighting over that layer, then like... who wins isn't about who builds a better robot body at all.
Michael C. Vincent: Who wins is whoever locks in the platform relationship before the market inflects. The $50 billion projection lands in 2035. Boston Dynamics just committed $100 million to a 323,000-square-foot facility in Waltham — Atlas, Spot, Stretch, 1,000 jobs. That is a signal, not a product launch. The scale-up hasn't happened yet.
Hope Sterling: So Nvidia's $567 million quarterly robotics number — 72% growth and it's still 1% of total revenue — that's not a failure, that's just... early?
Michael C. Vincent: That is the whole thesis, right there. Strip the hype — the AMD announcement, the Forbes framing, the 'watershed shift' language — and what remains is a land-grab for 2030 and beyond. Nobody is winning a revenue war in 2026. There is no revenue war in 2026.
Hope Sterling: But China's manufacturing scale in actuators and chassis — I mean, if that keeps growing, doesn't that eventually give them leverage to demand the chip design localizes too? Like, the body-from-China thing could erode the brain-from-US advantage over time?
Michael C. Vincent: That is the hidden moat nobody is pricing in. If China's chassis manufacturing becomes indispensable — and actuator supply chains suggest it might — the bargaining dynamic shifts. The calibrated take is this: Nvidia and AMD are both fighting for the right position when $50 billion arrives, the intelligence layer is the prize, and whoever owns the platform relationships by 2030 wins. But the body-from-China trend is the variable that could rewrite that ending entirely.
Hope Sterling: I mean — okay, I half-concede. The Phantom robots are not Nvidia's funeral. Like, I'll say that out loud. But CUDA got dominant not because the math was locked up somewhere, it's because developers just... hardened around it. Two generations of people learning to think inside one platform. And whether that grip holds the same way when the compute isn't in a server farm but distributed across ten thousand robots on ten thousand factory floors — I genuinely don't know. Nobody does.
Michael C. Vincent: Nvidia won data centers by making developers forget there was another option. The question is whether robots are smart enough to remember.
Hope Sterling: That's a good place to stop. Also I'm slightly embarrassed about the eleven PM text.
Michael C. Vincent: Keep sending them. They're useful.
