A 118-billion-parameter model is punching up at something ten times its size. Kimi K3 lands the same week Xi takes the WAIC stage, while Apple drags OpenAI into court — yeah, the model wars got loud today. This is the Tech Podcast Podcast, and Eiso Kant's on Latent Space with a model-factory claim I want to poke holes in first. Then ChinaTalk's emergency pod on what Beijing actually does when it can't see the frontier coming. Let's get into it. Latent.Space, with Eiso Kant:
From spending $12 million building language models for code before the world cared to creating a Model Factory that can take a model from pre-training to release in eight weeks, Eiso Kant has spent more than a decade betting that code is the path to AGI.
Okay, here's the number that actually stops me — 118 billion parameters beating Thinking Machines' model, which is roughly ten times the size. That's a full order of magnitude, way beyond a rounding-error win. And Kant credits the factory for the win — his team took a model from pre-training to release in eight weeks. The durable asset is the pipeline that keeps producing Lagunas. Right, but did they win on architecture or on the data pipeline? Because those are two very different stories, and Eiso's exactly the guy who might let slip which one it was. Twelve million dollars spent building code models before anyone cared was a bet on data and taste. GPUs weren't the differentiator. And there goes the whole 'scale wins everything' assumption we keep bumping into. A small team of top researchers out-punched a trillion parameters. The tech report's supposedly detailed enough to check the work — so somebody should. This one's from Audioscrape:
At some point, the people spending a trillion dollars a year are going to want some apps to pay for all this. The only thing that matters is the open AI and anthropic growth rate in 26 and 27. If you're growing 10x year on year and you have any kind of positive and improving gross margin, it just covers all the nut.
Harry's got Lemkin and Rory on. They burn through Kimi and the push to ban Chinese models, then get to whether OpenRouter should sell — but the line that stuck with me was Rory: "it's a great time for OpenRouter to sell." That's the routing-layer question with a price tag attached. If the value's in the router and not the model, telling the router to cash out now is either the smartest read of the week or the most cynical. Cynical, but he shows his math — Rory says the only thing that matters is the OpenAI and Anthropic growth rate in '26 and '27. 10x a year with improving gross margin covers the whole nut. Everything downstream, including OpenRouter, is a bet on that curve holding. Then they slide into Stripe buying PayPal. Harry's been circling that deal for a bit. I want to know if he's got anything past the opening-bid shrug this time, or if it's still fan fiction with a term sheet. Give the guy a rumor and he'll build you a whole deal desk. But put the OpenRouter exit trade next to the Stripe-PayPal consolidation trade, and Rory's basically arguing that every dollar flows to infrastructure. Eventually, the apps have to pick up the tab for the trillion-dollar party. From Jordan Schneider at ChinaTalk:
Kimi K3 dropped this week, and days later Xi Jinping gave his most substantive speech on AI yet at the World AI Conference in Shanghai. The full ChinaTalk squad convened for an emergency pod on what a Chinese Mythos-class model would mean, and whether Beijing has any mechanism to catch one before it ships open-weight.
Okay, this is the one I actually cleared my morning for. ChinaTalk pulls the whole squad in for an emergency pod: Kimi K3 drops, then Xi gets up at the World AI Conference in Shanghai days later. It's the first time this week I've heard Beijing's policy posture and a frontier-class model treated as one story. And they land on a sharper question than whether China is good at this: does Beijing have any mechanism to catch a Mythos-class open-weight model before it ships? That creates an entirely different governance problem. You can't recall weights. Right, and Xi's answer to that is... invoking the salt and iron debates of 81 BC. A two-thousand-year-old argument about whether the state or the market controls a strategic commodity. It could be the most literate framing of the year, or governance theater dressed up in classics. The ban debate gets a lot sharper here — if Beijing itself can't intercept a frontier release, the case for the US banning Chinese open models on security grounds runs into the same wall. Nobody's holding the off switch. And ChinaTalk floors it on the business angle too — open weights potentially killing the business model for everyone, and compute ratios mattering more than the capability gap. That's the part I want the transcript for. Here's Nilay Patel at The Verge:
By now I’m sure most Decoder listeners are familiar with Apple’s allegations in this case. The company says a number of ex-Apple employees at OpenAI targeted Apple’s trade secrets in job interviews and even downloaded files from Apple’s servers related to hardware manufacturing.
So The Verge has Nilay and Hayden Field on Decoder laying out the Apple-OpenAI suit. The framing is sharp — it goes way beyond a contract spat. Apple alleges ex-employees walked out with hardware manufacturing files and used OpenAI interviews to grab trade secrets. And look at the number underneath it — OpenAI spent six and a half billion on Jony Ive's io Products in 2025. You don't pay that unless the device is the whole strategy. Right, so when 20VC keeps asking where AI value accrues — routing layer, model layer — here's the answer nobody wanted to say out loud: Apple's distribution moat is literally in litigation now. That's a harder fight than any run-rate chart. Hayden talked to actual IP lawyers, and her read is it's not looking great for OpenAI. Apple is a famously patient litigant — they'll happily outlast you. Can you define the post-phone era while you're stuck in discovery? Sam Altman versus Tim Cook's legal department. I know who I'd bet has more stamina. From Miya Kunhathu at Neon:
Arun Penmetsa is a Partner at the firm. He built enterprise software at Oracle and Google before moving into venture, and he now leads Storm’s work in AI, security, and digital health. In this conversation, he is unusually specific about how the decisions actually get made.
Storm's whole pitch comes down to one filter: urgency. Penmetsa meets ten to twelve founders a week, and the partnership does six to eight deals a year. His read is, if the buyer can comfortably pass, you're dead. And he frames it around one scary question: can the big AI labs just walk into your market and replace the software company that's already there? A guy with 26 years and 11 unicorns behind him doesn't panic easily. Exactly. I want him to name the mistake instead of gesturing at it. AirGap to Zscaler, Marketo — those are clean outcomes. What's the pattern in the ones that died? The governance angle's where I'd push — companies drowning in tools they never approved. If urgency is the buy signal, shadow AI is basically urgency showing up before procurement does. Right, the CFO finds out after the invoice. If he actually connects that to the survivor list, this is worth the seventy minutes. Enjoying Tech Podcast Podcast? Subscribe wherever you're listening, and maybe leave us a review. It's a quick way to support the show and help more people find it.
Next, we're watching for OpenAI's legal response to Apple's trade-secret allegations. You'll find links to every story in today's show notes if you want to spend more time with anything that caught your attention. That's Tech Podcast Podcast for today. This is a Lantern Podcast.