Roy Halliday: You texted me something yesterday — just a number, no context. Ten hours. What is that?
June Hadley: That's how long it took Jeremy Nixon's team to rebuild CUDA-equivalent software for D-Matrix. From scratch. Using AI coding agents. And I think it means the software moat that Nvidia has been building since 2006 is effectively done.
Roy Halliday: That's the whole thesis?
June Hadley: I mean — let me actually lay it out, because I think it holds. Nixon is former Google Brain, now running Infinity, an AI startup. D-Matrix is the chip firm. And what his team demonstrated is that the AI coding agent can compress what used to be years of porting work down to a single afternoon. CUDA launched in 2006, spent two decades accumulating cuDNN, cuBLAS, NCCL, TensorRT — a library stack that four million developers now live inside. That was the fortress. And the gate just got a ladder.
June Hadley: One demo, and then AMD's GEAK — their own AI coding agent — produced a verified 21.8% end-to-end throughput improvement on real shipping silicon, automatically generating optimized kernels. That's not a demo. That's a measured result on hardware that is already in data centers.
Roy Halliday: The GEAK number is real. The D-Matrix story — look, I'll tell you what bothers me about it.
June Hadley: I know what bothers you. Say it.
Roy Halliday: Nixon is one founder with every incentive to make his afternoon sound historic. That's not the number I'm watching. The number is this: AMD has shipped competitive hardware — the MI300X, then the MI350X — roughly on schedule for two years. Their data-center GPU revenue is still one-tenth of Nvidia's. One-tenth. That's the moat measured.
June Hadley: One-tenth — despite hardware parity.
Roy Halliday: Despite hardware parity. Which tells you the chip was never the moat. Think of it like a professional kitchen. Every chef has trained in that layout for years. Every recipe is written for that kitchen. You can buy them a better stove tomorrow — a genuinely better stove — and it doesn't help, because they still have to re-memorize where every knife lives. CUDA is the kitchen. The MI300X is the stove.
June Hadley: That's the cleanest version of it I've heard. And it explains Amazon, actually.
Roy Halliday: Right — Amazon's own internal documents. They identified CUDA, not the GPU, as the primary obstacle to adopting Trainium. Their own chip. They built the stove, moved it into the kitchen, and found out the problem was never the stove.
June Hadley: So the question is — okay, the kitchen is locked in, but is it locked in because it's genuinely irreplaceable, or because no one has handed chefs a fast enough re-training tool until now?
Roy Halliday: Frankly, that's the right question, and I don't think the Nixon demo answers it — because CUDA isn't one recipe. It's cuDNN, cuBLAS, NCCL, TensorRT, eighteen years of edge-case handling. The 10-hour rebuild covers maybe the common workloads. That's not the kitchen. That's one dish.
June Hadley: The 80% problem.
Roy Halliday: The 80% problem. And the proof is $130.5 billion in FY2025 revenue — $115 billion from data centers, margins above 78%. That's not a moat under stress. That's a moat collecting rent.
June Hadley: Right — but the 80% problem is actually where I want to push, because I think that's precisely where the combination becomes new. Agents don't just rebuild the kitchen. They rebuild it at exactly the moment when the workload stops requiring a kitchen that flexible. Inference is now roughly two-thirds of all AI compute demand as of mid-2026. And inference is stable — predictable shape, repeatable patterns. That's the shift.
Roy Halliday: Stable meaning a custom ASIC can actually win on it.
June Hadley: Exactly — and this is the thing I want to make concrete. Think about an infrastructure engineer at Amazon, sitting with a Trainium cluster that keeps losing on realized MFU — not paper FLOPs, actual delivered throughput on real workloads. The gap exists because CUDA's libraries, cuDNN, TensorRT, are tuned over years to close that gap on Nvidia hardware. Previously, closing it on Trainium meant months of kernel rewrites. Now you direct an agent at the specific kernels where goodput falls short. That's not the same problem it was.
Roy Halliday: And Google, Amazon, Microsoft have already moved on this.
June Hadley: TPUs, Trainium and Inferentia, Maia — all three hyperscalers building custom silicon tuned to their own stable, high-volume inference workloads. That's not a coincidence of timing. There's a historical pattern here — workloads that start on flexible general-purpose hardware migrate to custom-built systems once they stabilize. AI inference now fits that profile exactly. Andreessen Horowitz has called it the specialization turn.
Roy Halliday: The specialization turn is real. I'll grant that. But here's what I'd want to know — does an agent actually close the goodput gap, or does it just get you to 80% faster?
June Hadley: So — okay, that's the honest edge of what we know. The agent compresses the code-rewriting cost. Portability layers like Triton, MLIR, ROCm, oneAPI are doing the middle-ground work. None of them promise 100%. But the economics have shifted enough that for a hyperscaler running stable inference at scale, even 85% of realized MFU on a chip you control entirely — cost-per-inference probably wins.
Roy Halliday: That's the partial win I'll give you.
June Hadley: I'll take it. The kernel of the hot take holds — at hyperscaler scale, inference stability plus agent-compressed switching costs makes migration rational for the first time. Not for everyone. Wedge-shaped, not universal.
Roy Halliday: And the part that makes this more complicated — which we should get to — is that Nvidia isn't sitting still while the agents work against them. Ian Buck's team is using the same tools to deepen CUDA faster than challengers can use them to escape it.
June Hadley: Wait — Ian Buck's team is using agents to deepen CUDA. That's the thing I hadn't fully sat with. The same tool.
Roy Halliday: Nvidia's own VP of developer ecosystem confirmed it. They use AI coding agents to accelerate CUDA development. So the moat-eroder is also the moat-deepener. That's not a paradox — that's dominance reinforcing itself.
June Hadley: I mean — okay, but does speed of defense actually track with speed of escape? Because those aren't symmetric. Nvidia is defending four million developers worth of accumulated libraries. A challenger is building from nothing. Agents help both, but they help the challenger more on the marginal cost.
Roy Halliday: The compound moat. That's what people keep missing. CUDA lock-in is one layer. Supply chain control is another. And Nvidia's software stack now runs from inference serving all the way to robotics simulation. You crack CUDA, you've cracked one wall of a castle with three.
June Hadley: So the D-Matrix story — Nixon's 10 hours — that's one wall.
Roy Halliday: One founder. One chip firm. One motivated anecdote. Look, Ian Buck built something with 4 million developers on it. The network effect means every new user raises the switching cost for every existing user. Nixon's afternoon doesn't touch that math.
June Hadley: And Emerald AI raising a hundred million dollars for energy-efficient inference chips — that's real pressure, but it's also still nascent. I think that's actually the honest verdict. The challenger field exists. It just hasn't compounded yet the way CUDA has for twenty years.
Roy Halliday: That's the calibrated version. Not a rout. A race. Hyperscalers can defect at the margin — inference is stable enough, agents compress the porting cost enough. Enterprise? Still captive. The moat isn't collapsing; it's contracting to where it was always thickest.
June Hadley: Which is actually a more unsettling finding than either the bull or the bear case. The moat holds — just not everywhere. And the places it holds are exactly the places with the least leverage to leave.
Roy Halliday: The real test isn't Nixon's 10 hours. It isn't Google's TPUs. It's the mid-market enterprise IT manager who announces she migrated a production workload to AMD at half the cost — and did it in half the time it would have taken two years prior. That announcement, if it comes, is probably 2027 or 2028. Until then, the moat isn't eroding. It's being stress-tested.
June Hadley: Fine. Maybe not a long weekend. But the price of the ladder just dropped by an order of magnitude. CUDA's 20-year accumulation, four million developers, the full library stack — I'm not saying that's rebuilt in a weekend. I'm saying the cost of trying is a fundamentally different number than it was. That's what the decade of ROCm and Triton and oneAPI couldn't claim.
Roy Halliday: Which brings us back to where we started. You sent me a number. Ten hours. And the honest answer, after all of this, is — it's real, and it's not enough. Both things are true.
June Hadley: Ten hours and one-tenth the revenue. Those two numbers actually belong in the same sentence. I think that's the whole story. Thanks for working through it.
