$AMZN $NVDA $INTC $PANW $CRWD Amazon's CEO says AWS is running the Graviton pl… · @carrioresearch · TwKit

@carrioresearch

$AMZN $NVDA $INTC $PANW $CRWD Amazon's CEO says AWS is running the Graviton playbook on AI chips. Graviton ended up taking more than half of AWS's new CPU capacity, three years running. 1. Andy Jassy told a16z that when Amazon bought Annapurna in 2015, Intel "had virtually all of the CPU share," and "when you have a partner who has virtually all the share, they're not often thinking about how to drive price performance." So Annapurna built Graviton. He says 99% of AWS's top 1,000 EC2 customers now use it "very pervasively," at about 40% better price-performance than x86. 2. Here's what that did to x86 inside AWS. Last December AWS said that for the third year in a row, more than half of the new CPU capacity it added ran on Graviton. Intel stayed a partner the whole time and x86 is still all over the fleet, but most of the new racks went to Amazon's own chip. 3. Then he said this: "a few years ago, we saw that same movie in the AI space where you really had one dominant provider who's very good." He never said Nvidia, and he said that partnership will last "for as long as I can foresee." Trainium is the answer he's betting on, at about 30% to 40% better price-performance than comparable processors. 4. Why he cares so much about cost: he thinks inference is "still an order of magnitude too expensive, in my opinion, for it to be used as pervasively as possible." Enterprises won't rewrite their CPU apps for accelerators until that gap closes. Cheaper inference grows the whole market, and Amazon wants as much of that growth as possible running on its own silicon. 5. So for $NVDA, the useful precedent is x86 at AWS. The installed base stayed. The new capacity went somewhere else. That's the version of this story that would show up in the numbers, and Amazon doesn't break out what share of its new accelerator capacity is Trainium yet. 6. The case against: a CPU is close to a commodity, and Graviton still needed roughly five years from its 2018 launch to win more than half of new capacity. AI accelerators come with a software stack and a developer habit that x86 never had to the same degree, so the same movie could run a lot slower. 7. Thursday, Oct 29 after the close, is the first earnings print since he said all this. Three numbers matter. AWS grew 36.7% last quarter on a $169B run rate, by his own count, so the first one is whether that speeds up. The second is the ~$220B capex plan. He said they're "leaving a meaningful amount of demand on the table" and "we could spend a lot more, it wouldn't be a bad use of capital," so whether that number moves is worth a look. The third is anything new on Trainium, whether that's capacity, commitments or customers. 8. The other thing he said that ends up in budgets is security. "I predict that companies will have to rewrite almost all their code to be agentically capable." His reasoning: models now find far more vulnerabilities than people did, so fixes have to ship "in minutes," and "if you don't actually have your applications rewritten to be able to be updated by agents, you're going to have no chance in that new world." That's a spending cycle $PANW and $CRWD sell into. AWS is selling into it too. He said it has already built a service that finds vulnerabilities and writes the patches, so who captures that spend is still open. 9. Graviton got to more than half of AWS's new CPU racks while Intel was a partner the whole way. Jassy just put Trainium on the same track, in public, a week before earnings. The share of new AI capacity running on Amazon's own chips is the number to track through 2027.

曝光 359 · 评论 2 · 点赞 4 · 书签 1