Hope Sterling: Okay, before anything else — did you see this Jalapeño thing drop last week and just physically react?
Juniper Vale: I did, yeah. I had the same face I make when something sounds too good and I can't figure out why yet.
Hope Sterling: Which is — okay, that's exactly what we're unpacking today. OpenAI and Broadcom unveiled this chip, right, on June 24th, and they're calling it an 'Intelligence Processor' — like, that's its official category. First custom inference chip OpenAI has ever made. And the announcement was — I mean, Hock Tan, Broadcom's CEO, showed up at OpenAI's headquarters to literally hand Sam Altman the samples. In person.
Juniper Vale: It's theater, right.
Hope Sterling: It's TOTAL theater — and then he turns around and says Jalapeño delivers fifty percent lower inference costs than Nvidia GPU deployments. Fifty! And I'm like — okay but you literally just delivered the box, you're not a neutral party here.
Juniper Vale: That's the question I want to sit with — because who else has actually verified that number? ChatGPT runs on Nvidia right now, inference is the expensive part of all of this, and a fifty percent cost drop would genuinely change the economics. But we're taking that from the manufacturer.
Juniper Vale: Think of it like this — training a model is baking the cake. You do it once, it's expensive, it's done. Inference is cutting slices. Every single time someone types into ChatGPT, that's a slice. And OpenAI is cutting millions of slices a day. That's where the money actually bleeds out.
Hope Sterling: Okay that — yeah, that actually landed. So the fifty percent isn't about building cheaper, it's about serving cheaper.
Juniper Vale: Exactly. Which is why the number matters so much — and also why I'd want to stress-test it. Because the only person who said fifty percent out loud is Hock Tan. He also compared Jalapeño to Nvidia Blackwell chips and Google TPUs in the same breath. That framing came entirely from Broadcom, who manufactures the chip.
Hope Sterling: Wait — so who actually checked it?
Juniper Vale: Nobody yet. I mean — the engineering samples physically arrived June 24th. That's the delivery day. The chip hasn't run at scale anywhere. Broad deployment isn't even expected until end of 2026, and some analysts are saying 2027, maybe 2028 for real rollout.
Hope Sterling: So Hock Tan is basically — wait, he's quoting a number for a chip that he just handed over in a box that morning? Like, that same morning?
Juniper Vale: That same morning. And look, the nine-month tape-out timeline using AI-assisted design tools — that part is genuinely impressive. But tape-out means the design went to manufacturing. It is not the same thing as 'this works at scale.' That's still the audition.
Hope Sterling: Okay but wait — and this is the thing that's been eating at me — Jalapeño is inference only. Like, that's all it does. And Greg Brockman went on CNBC that same morning, June 24th, talking about OpenAI's 'full-stack' AI infrastructure vision, and I'm sitting there like... that's not full-stack? Training is still entirely Nvidia?
Juniper Vale: Completely untouched. The Nvidia dependency for training — that's the harder half, arguably.
Hope Sterling: Which — okay, so picture this. There's a startup founder in Austin right now, building an app on OpenAI's API, paying per token, and her margins are just getting squeezed. She sees this Jalapeño announcement and thinks, maybe by 2027 my costs drop. But like... the software stack that would actually deliver that? It doesn't exist yet. Celestica is only just now starting systems integration for datacenter deployment.
Juniper Vale: And that gap is the real story. Google built their TPU software ecosystem over years — we're talking a long, grinding process of making the software actually talk to the hardware reliably at scale. OpenAI is 18 months into a partnership with Broadcom that just produced physical samples.
Hope Sterling: EIGHTEEN MONTHS to get samples. And Meta, Amazon — they've all been doing this for years already.
Juniper Vale: Right. Google, Meta, Amazon — they all have custom silicon and the software maturity that comes with time. That Austin founder isn't just waiting on a chip. She's waiting on an entire stack to prove itself.
Hope Sterling: So 'full-stack vision' is — I mean, it's a vision. It's not a description of what exists on June 24th.
Juniper Vale: And that's the issue — the thing OpenAI is escaping from Nvidia is Nvidia's pricing power, Nvidia's supply constraints. But if Jalapeño works, if Microsoft actually routes workloads through it by 2027, 2028... now the lock-in is OpenAI's own proprietary stack. You're not free. You just changed landlords.
Hope Sterling: Wait — that's it. That's the thing that just clicked for me. It's not an escape from dependence. It's like, a bet that their dependence is safer than Nvidia's. Which — I mean, maybe? But that Austin founder is still dependent on someone.
Juniper Vale: And we won't actually know if the bet paid off until Microsoft has to make a real decision. Route workloads through Jalapeño or default to what it knows. That answer doesn't come on June 24th — it comes in 2027, maybe 2028. Remember what you said at the top? You had physically reacted to this thing dropping, and you couldn't figure out why yet.
Hope Sterling: Yeah. I think I know why now. It's the feeling of a big announcement that is genuinely also just... an audition. Same feeling.
