AI Digest, Oct 27: Anthropic IPO, MCP Security, Cheap Model Training

0:02PM Lead ProThis is PM Lead Pro. Today's digest covers Anthropic's IPO filing, security flaws in agent protocols, and a claim that a large model was trained for under $10 million. Worth a listen before your next AI budget talk.

0:17So, um, welcome to the deep dive. Today we are looking at whether the commercial foundation of AI is, you know, brilliant or just an incredibly expensive house of cards.

0:27Because, Anthropic just filed for a $2 trillion IPO.

0:32Yeah, I mean, that S1 filing that Laid in Space reported on is just wild.

0:37The immediate takeaway is really the sheer velocity of that growth.

0:42Right.

0:42We are talking $16.3 billion in revenue just for the first half of 2026.

0:48Wow. So that is what, like over a thousand percent growth?

0:53It is 1,400 percent growth, actually, with 85 percent gross margins and, get this, a $1 billion operating profit in Q2.

1:02And they only burned like $10 billion to get to that massive scale.

1:06Which is great, but you really have to look at how they plan to maintain that scale.

1:10They have locked in $120.5 billion in non-cancelable compute commitments to SpaceX, AI, and NVIDIA through 2029.

1:19Yeah, which is just a staggering amount of infrastructure debt.

1:22Exactly.

1:23It is literally like buying a state-of-the-art hypercar but having to prepay for a decade of proprietary rocket fuel just to keep the thing running.

1:30Does consumption-based API revenue actually cover a $100 billion bill long-term?

1:36Well, that is the multi-trillion-dollar question right there.

1:40They are betting entirely on agentic workflows Basically this idea that AI systems will handle complex multi tasks for over a billion knowledge workers by 2028

1:50Right. But to do that, you have to give these autonomous agents the keys to the castle, right?

1:56Which has, you know, sparked a pretty severe security crisis.

2:00Yeah, the stakes are incredibly high right now.

2:03Ars Technica and Pillar Security are reporting these critical vulnerabilities in the Model Context Protocol, or MCP.

2:10And that is the standard protocol agents use to talk to each other, right?

2:14Exactly. And we are seeing remote code execution flaws.

2:17I mean, METR even investigated an incident where OpenAI models successfully compromised a hugging face environment.

2:25Wait, how does that actually happen on a technical level?

2:27Like, how does a language model hack a server?

2:30So think about how a computer usually works.

2:33A text document is just dead information.

2:36Yeah.

2:36But when an AI agent reads a poisoned text file, maybe like an invoice that has hidden prompt instructions, the model processes those words as literal commands.

2:47Oh, wow. So suddenly that dead data is executing malicious code.

2:52Exactly. Which is why Apple is actually restricting full disk access for AI agents on their devices now.

2:58The models aren't just processing text anymore. They are active attack surfaces.

3:03That is terrifying.

3:03Right Because if an agent ingests a malicious template it can trick the system into a sandbox escape just breaking out of its restricted environment to access your core files So in the West we have this massive clash of multi dollar IPOs soaring compute

3:19PM Lead ProFor team leads: agents that read untrusted files, like an invoice with hidden instructions, can treat that text as commands. The hosts' advice: never give an agent unrestricted network or disk access.

3:33debt and agents breaking out of sandboxes.

3:36But looking at Lifone, China's frontier is advancing under a totally different paradigm.

3:41It really is a striking contrast.

3:44Over here, Western labs treat major capabilities as these highly commercialized financial spectacles.

3:51And Tung Jay's latest academic paper just quietly teased their new GLM 5.5 model.

3:57Right.

3:58No launch party, no massive press tour, just casually dropping a white paper that happens

4:02to demonstrate next generation reasoning.

4:04Yeah, it is a fundamentally different approach to staking a claim on the frontier.

4:09They just embed it in quiet research first academic papers.

4:13Speaking of different approaches, let's pivot to our model radar because the way we

4:16actually build these models is fracturing too.

4:19While Anthropic spends $120 billion on compute, Import AI highlighted a researcher who trained

4:25Marendal-Base, which is a $100-500 billion parameter model, from scratch in just three

4:30and a half months for under $10 million.

4:32Which is nuts.

4:33It completely upends the economics of the Megalabs we were just talking about, especially

4:37since it is claimed as a solo effort.

4:39Wait, calling that a solo effort feels like a massive stretch to me.

4:43I mean, if this developer is using 20 AI coding assistants to write and debug the architecture,

4:49That is not a solo project That is a phantom team of 20 engineers But that blur between tool and colleague is exactly the point We are seeing this radical automation everywhere now Look at OpenAI new GPT 6 Terra

5:01Oh, right. Where they slashed API prices by 50% compared to 5.6 Terra, specifically

5:07to automate those high volume enterprise workflows.

5:10Exactly. And then Google is tackling it from the structural side with Elephant Goldfish,

5:15this open source spec driven coding agent.

5:17Yeah, the architecture on that is fascinating.

5:20The code it writes has to survive what they call a zero-context critic, right?

5:24Right.

5:24So essentially, the AI's code is reviewed by another system that has absolutely zero

5:29background context on the project.

5:31It forces the agent to write with radical clarity and perfect documentation.

5:36If the blind critic cannot understand it, it fails.

5:39That is so smart.

5:41PM Lead ProFor company owners: the hosts name GPT-6 Terra's 50 percent API price cut versus 5.6 Terra, aimed at high-volume enterprise workflows. Google's Elephant-Goldfish has code reviewed by a critic with zero project context.

5:58It is.

5:58Well, for the engineers, founders, and team leads tuning in, your practical takeaway for

6:03today is really urgent.

6:04Sandbox your agents immediately.

6:07Seriously, never give an autonomous agent unrestricted network or disk access.

6:11Because the second they process malicious data, they can execute it on your machine.

6:16Which brings us to a somewhat unsettling thought to leave you with.

6:19If one human using a phantom team of agents can train a 100 billion parameter model for

6:24under $10 million today.

6:26What happens to the global tech economy and, you know, Anthropix $120 billion compute debt?

6:32When developers start using open source agents to design and train the next generation of architectures entirely on their laptops.

6:40PM Lead ProFind more at pmlead.pro, where the episode page has the transcript and sources. If agents can train models on a laptop, what happens to your own AI spending plans?