Maya Chen: Jonathan, okay, so — I was in a coffee shop yesterday and the person next to me was on an earnings call, headphones in, and at one point they just went 'huh' very quietly to themselves. Just once. And I couldn't hear the call, but I thought about that sound all day, because I think that's basically what I did when I looked at AMD's quarter.
Jonathan Ingles: That sound being — 'this doesn't add up.'
Maya Chen: Exactly that. Eleven-point-five billion in Q2 2026. Earnings beat. AMD runs this whole 'Advancing AI' event — new hardware, customer commitments, Barron's writes that they're setting a high bar for Nvidia. And then the stock... drops.
Jonathan Ingles: The market looked at a genuine beat and said: insufficient.
Maya Chen: Which is — I mean, what does that actually mean? Like, what is the market seeing that the results aren't showing?
Jonathan Ingles: The fact is, that's the episode. Not 'is AMD's hardware good.' We can stipulate that. The question is whether good hardware, real customer wins, a splashy event — whether any of that converts into durable market share against Nvidia. And the stock falling after a beat is the market saying: we're not convinced.
Maya Chen: And we're sort of — we're not going to resolve that cleanly, are we.
Jonathan Ingles: Not cleanly. But we can name what the actual tension is, which nobody seems to want to do.
Maya Chen: But naming the tension doesn't tell me what AMD is actually doing differently. Like — what is the move here, structurally?
Jonathan Ingles: Okay. Plain language. There are two completely different jobs an AI chip does. Training — that's building the model, massively compute-intensive, that's been Nvidia's house. And then inference — that's running the finished model at scale, millions of times a day. Cheaper workload, enormous volume. AMD is betting the industry's center of gravity is shifting from the first kitchen to the second.
Maya Chen: The second kitchen being — faster and cheaper once the recipe's already written.
Jonathan Ingles: Exactly that. And here's where Taalas comes in — acquired August 6th, this year, Reuters and CNBC both had it. The technology doesn't process the AI model on the fly. It hardwires the model directly into the silicon itself. That's not iterating on what Nvidia does. That's a different architectural bet entirely.
Maya Chen: Wait — baked in. Like, the model isn't running on the chip, it's sort of... living inside it?
Jonathan Ingles: That's the bet. And Bank of America drew a specific analogy — AMD's AI accelerator trajectory now rhymes with how AMD took server CPU share from Intel. That's not a casual comparison. Intel looked unassailable too.
Maya Chen: Right — but the part that doesn't fit is Intel actually stumbled. AMD didn't just show up with a better chip; Intel handed them the opening. So the BofA analogy only works if Nvidia makes a comparable mistake, yeah?
Jonathan Ingles: Which is the complication nobody wants to say out loud. AMD can execute perfectly on inference, land Taalas, win hyperscaler pilots — and if Nvidia doesn't stumble, the analogy just... doesn't complete. The market fell seven percent on a beat because investors are pricing that exact gap between the narrative and the actual condition required for it to be true.
Maya Chen: But that gap — Nvidia not stumbling yet — I think it's weirder than just 'they're still ahead.' Because the moat isn't really the hardware anymore, is it? Like, the thing that makes CUDA sticky isn't that it's technically superior at this point. It's that... engineers built their whole careers around it.
Jonathan Ingles: Sunk costs. That's the actual lock-in.
Maya Chen: Right — so picture an ML engineer, three months from a production launch. Their entire pipeline is CUDA. AMD comes in and says 'we'll give you twenty percent off.' And that engineer just... doesn't move. Not because the AMD chip is worse, necessarily. It's — rewriting the stack mid-cycle isn't a procurement decision, it's years of migration.
Jonathan Ingles: And that's exactly what Rubin is designed to exploit. Nvidia's proprietary rack system — it's not just tighter hardware-software integration. It's a deliberate bet that even if AMD's ROCm software layer closes the gap, the hardware entanglement is already so deep that switching costs compound. You're not migrating software. You're migrating the whole physical stack.
Maya Chen: Oh. So Helios — AMD's open-standard answer to Rubin — the 'open' framing is almost beside the point if no one can actually tell us whether hyperscalers are adopting it in volume?
Jonathan Ingles: The research doesn't give us adoption numbers. Frankly, nobody does. So the open-versus-proprietary argument sounds compelling and we can't stress-test it.
Maya Chen: Which is — mm, that's uncomfortable. Because the whole 'open standards appeal to hyperscalers wary of lock-in' story is doing a lot of work in how people frame AMD's shot here, and we're sort of... we're just taking that on faith.
Jonathan Ingles: Look, the behavior problem is this: institutional knowledge doesn't show up on a spec sheet. The Bank of America Intel analogy only closes if there's a crack from inside the ecosystem — Intel stumbled on execution, which gave engineers permission to leave. CUDA hasn't given anyone that permission yet.
Maya Chen: And there's a version of this story where Amazon and Google diversifying isn't AMD winning at all — it might be hyperscalers deciding the entire dependency structure is the problem, which lands somewhere very different. That's the part I want to get to.
Jonathan Ingles: And that's exactly what the Amazon story actually says, if you read it straight. Amazon isn't qualifying AMD because AMD beat Nvidia. They're qualifying AMD because they've decided depending on any single supplier — including AMD — is the structural problem. The NYT had Amazon and AMD as 'emerging contenders' back in December 2024. That was eighteen months ago. Nvidia is still the primary platform.
Maya Chen: Wait — so Amazon diversifying to AMD isn't AMD winning the race. It's Amazon deciding to stop running the race at all.
Jonathan Ingles: They're building their own silicon. Custom chips. Because they're done waiting for anyone to dethrone Nvidia. That's not a vote for AMD. That's a vote of no confidence in the whole duopoly.
Maya Chen: Which means — mm, okay, that reframes the hedging completely. The hyperscalers aren't picking a winner between AMD and Nvidia. They're sort of... opting out of having to pick.
Jonathan Ingles: And Broadcom is sitting right in the middle of that. Their custom TPU work for hyperscalers — that's a third vector nobody in the AMD-versus-Nvidia coverage is accounting for. It's not AMD, it's not Nvidia, it's a bespoke solution built for exactly one customer's workload.
Maya Chen: Broadcom just — yeah, Broadcom barely shows up in the headline narrative, and they're right there doing it already.
Jonathan Ingles: Then layer in Chinese chipmakers gaining ground from Beijing's domestic tech push — that's another vector, almost entirely absent from the coverage. The frame of 'AMD versus Nvidia' is missing at minimum two serious competitive forces.
Maya Chen: So the market dropping seven percent on a beat that looked real — I wonder if that's actually the market pricing all of this at once. Not just 'can AMD sustain it,' but 'does AMD still get to be the challenger story when the challengers multiplied.'
Jonathan Ingles: That's it. The earnings beat was already priced in. What the stock drop is pricing is execution risk at the speed AMD's valuation demands — and a world where hyperscalers building custom silicon means AMD's ceiling just got lower than anyone in the headline story has said out loud.
Maya Chen: AMD's ceiling got lower. That's — I keep turning that over. Because if the ceiling is maybe fifteen, twenty percent of total AI accelerator revenue, and it stays there for years... that's not nothing, but it's also not the story anyone's telling when they say 'inflection point.'
Jonathan Ingles: The one thing I'd actually watch — not the hardware announcements, not the customer wins — the next AMD earnings report. Specifically data center revenue. Does it sustain a growth curve, or does it flatten the moment Helios hits any execution or software maturity friction in Q4 2026, Q1 2027? That's the whole question in one number.
Maya Chen: And the Bank of America analogy either completes or it doesn't, right there. Because if the curve flattens — I mean, AMD beat Intel eventually, but Intel had to actually hand them the opening. That took years. So the answer to 'is this an inflection' might just be... we find out in 2028.
Jonathan Ingles: Which is maybe where we should just sit. Neither of us has a verdict. The market doesn't have one either — that's what a seven-percent drop on a genuine beat actually is. It's not a verdict. It's an open question priced as doubt.
Maya Chen: That person in the coffee shop, headphones in, just went 'huh' — I think that's still the right sound. Not panic, not excitement. Just... the math not quite landing yet. That feels like where AMD is. Thanks for thinking through it with me.
