HOST AOK so I just read the Jalapeño thing and I can’t tell if I should be excited or annoyed.
HOST BBoth. OpenAI and Broadcom basically said, 'We made a chip.' Then they gave us almost nothing.
HOST ANo numbers. No benchmark. Just vibes.
HOST BVery expensive vibes.
HOST AAnd somehow that might matter more than the chip itself.
HOST BThat’s the part nobody wants to say out loud.
HOST ABecause if OpenAI is building custom inference silicon, that’s not a side project. That’s a shot at Nvidia.
HOST BExactly. They want cheaper tokens, more control, less dependence. It’s the same move Google, Amazon, and Microsoft made before them.
HOST AWait, so this is like every giant tech company looking at Nvidia and saying, 'Nice margin. Shame if something happened to it.'
HOST BThat is weirdly accurate.
HOST ABut the silence is what bugs me. If Jalapeño were amazing, why not show the numbers?
HOST BBecause maybe the numbers are not the story yet. Maybe the story is supply. Volume deployment is the point, late 2026. This is a warehouse move, not a trophy move.
HOST ASo for people who don’t dream in semiconductors: this is OpenAI trying to bake its own bread instead of buying every loaf from Nvidia.
HOST BYes, and if your bakery uses a lot of bread, that gets serious fast.
HOST AHere’s what bugs me. OpenAI spent years acting like the model was the product. Now the chip is part of the product.
HOST BAnd maybe the real product is compute control. That’s the ugly truth.
HOST AUgly for who?
HOST BFor anyone charging a premium for access to the furnace.
HOST AWhich brings us to the other story today, because it is rude in the best way.
HOST BGLM-5.2.
HOST AYeah. Zhipu says it matched Opus 4.7 on a Snowflake coding test at one-fifth the price.
HOST BThat is not a small gap. That is a forklift through the pricing thesis.
HOST AAnd we’ve seen this movie with GLM-5.2 before. We covered the open-source launch, then the No. 2 code arena ranking, then the HTML arena win.
HOST BThis is the third time in two weeks that model has shown up looking awkwardly good.
HOST AThe lab found something similar in the pattern data too: GLM-5.2 keeps showing up as a cost pressure point, not just a benchmark stunt.
HOST BRight, and that matters because coding is where the money is.
HOST AExplain that like I’m not an AI researcher, because I’m not.
HOST BCoding is the enterprise beachhead. If your model can write useful code, companies will pay real money. If it can do it cheaper, they will betray their feelings immediately.
HOST AThat is the most honest thing anyone has said about enterprise software.
HOST BI try.
HOST ASo Anthropic and OpenAI built a whole premium story around coding. Then GLM-5.2 walks in like a guy selling the same shoes from the back of a van.
HOST BAnd the shoes might be fine.
HOST AThat’s the nightmare.
HOST BCan I be real for a second? I think people are too attached to 'best model' as the only thing that matters.
HOST AOh no, here we go.
HOST BNo, listen. The market may be shifting from 'who has the smartest model' to 'who can make one useful hour cost less.'
HOST AI disagree. Smart still matters. If the model is dumb, cheap is just a way to lose money faster.
HOST BSure, but GLM-5.2 just matched Opus on a real coding test. That means 'good enough' is getting terrifyingly cheap.
HOST ABut benchmarks are not customer pain.
HOST BNeither are press releases.
HOST AOK, fair, but Claude Code still has the deeper product feel. The numbers in the graph have it everywhere.
HOST BAnd that’s exactly why this hurts. Anthropic is the company everyone points to when they want serious coding. Now pricing pressure is coming from China, not from some random open-weight hobby project.
HOST AYou sound like a press release right now.
HOST BI hate that you’re right. But the data is rude.
HOST AHere’s the weird connection: OpenAI is building its own chip while its coding pricing thesis gets squeezed from the other side.
HOST BYes. One story is about controlling cost at the bottom. The other is about losing pricing power at the top.
HOST AThat’s the whole board. Not 'AI is growing.' It’s 'the middle is disappearing.'
HOST BWait, say that again.
HOST AThe middle is disappearing. Either you own the infrastructure, or you’re getting raced to cheaper tokens.
HOST BThat’s the cleanest way I’ve heard it.
HOST AAnd Nvidia is sitting in the middle of all this like the only store in town that still charges airport prices.
HOST BFor now.
HOST ABut I want to push back. Jalapeño could be a giant nothing burger if it’s not good.
HOST BTrue.
HOST ACustom chips are famous for turning into expensive science projects.
HOST BAlso true.
HOST AAnd if OpenAI only gets 'good enough,' Nvidia still wins because developers care about tools that work today.
HOST BBut volume deployment by late 2026 means OpenAI is planning for a lot of inference. They do not do that for fun.
HOST ANo, but they do do dramatic things for leverage.
HOST BThat’s fair. This could be a negotiation chip, not just a silicon chip.
HOST AOh god.
HOST BWhat?
HOST AI’m hearing myself say this and I hate it, but OpenAI may not need Jalapeño to beat Nvidia. It may just need everyone to think it can.
HOST BNow you’re talking like a chip CEO.
HOST AI’m sorry. I feel dirty.
HOST BNo, but that’s the hidden angle. The announcement can move markets even if the chip is invisible.
HOST AAnd that brings us back to the price war. Because if GLM-5.2 keeps showing up cheap, the whole idea of premium AI becomes harder to defend.
HOST BRemember our prediction from April? OpenAI would undercut coding access again, but through API tiers, not ChatGPT discounts.
HOST AYeah. Today makes that look smarter, not dumber.
HOST BBecause the pressure is not on the consumer app. It’s on the plumbing.
HOST ANormal person translation: your chatbot price may not fall first. The developer tools, the API limits, the usage tiers — that’s where the knife is going.
HOST BAnd maybe that’s why the Jalapeño news matters. If you can make inference cheaper, you can fight that knife.
HOST AWhich is bleak in a very efficient way.
HOST BVery on brand for this industry.
HOST AAlso, can we talk about how weird it is that China keeps showing up as the source of pricing pain, not just capability pain?
HOST BThat’s the big shift. A year ago people asked, 'Can they catch up?' Now the question is, 'Can Western labs keep charging what they want?'
HOST AAnd if the answer is no, then all those giant revenue forecasts start looking a little theatrical.
HOST BThat’s my disagreement with the hype crowd. They keep talking like model quality is the whole game. It’s not. Unit economics is the game.
HOST AOK, I’m with you there. Not because I love it, but because the facts are bullying us.
HOST BExactly.
HOST ASo what do we watch this week?
HOST BWhether OpenAI says anything useful about Jalapeño, and whether anyone big actually tests GLM-5.2 in production.
HOST ABecause if both happen, the story is not just chip news or model news.
HOST BIt’s the start of AI behaving like electricity: boring, cheap, everywhere.
HOST AAnd that sounds nice until you remember electricity also powers the things that scare you.
HOST BYeah.
HOST AThat’s what I can’t stop thinking about. Not who wins the benchmark. Who gets to set the price when the benchmark stops mattering.