Premium News To Stay Up-to-Date, March's End

TLDR;

Welcome back! This week, we discuss the troublesome reality of AI finance, the secret ingredient to AI startup success, and cool products, including Cline's first-ever agent orchestrator.

Enjoy!

PUBLIC MARKETS

Anthropic, October IPO?

According to Bloomberg, Anthropic, the company behind Claude and super-popular products like Claude Code or Cowork, is exploring an initial public offering that could happen as early as October 2026.

TheWhiteBox’s takeaway:

The move comes after a sharp rise in Anthropic’s valuation and revenue, which sat around $19 billion in annualized revenue as of February (now probably well over $20 billion), making it one of the fastest-growing companies in the world (and ahead of OpenAI in terms of growth).

I’m not a fan of this company’s leadership, we both know it by now, but I need to call balls and strikes, and, unless OpenAI finds a way to effectively monetize its soon-to-be 1 billion customer base, there’s no way on Earth the latter is worth two times what Anthropic is worth.

Maths reveal this. Assuming OpenAI is at $30 billion in annualized revenue (it’s probably a little below that) and Anthorpic is at $20 billion (it’s surely above that), and considering their last valuations of $840 and $380 billion, respectively, that still gives multiples over sales of 28 and 19 (they would surely IPO way above that).

Despite both being pretty rich valuations, I have to say it feels hard to find any meaningful justification for that spread, especially given that both are unprofitable (unless OpenAI finds a way to monetize its free users).

SMARTPHONES

Apple will train Gemini models

According to The Information, Apple reportedly has broad access to Google’s Gemini models and is using that access to build smaller AI systems that can run directly on Apple devices.

The idea is to use Gemini to generate high-quality answers with reasoning traces, then distill that behavior into lighter models that are cheaper to run and better suited for on-device tasks (a process called distillation). In plain English, Apple is licensing Google’s AI technology, and it’s free to use (and improve) it at its will.

Apple can also adjust how those models behave, although it has apparently run into challenges because Gemini has been optimized more for chatbot and coding use cases than for all of Apple’s product needs. This work is tied to Apple’s plan for a more capable Siri in iOS 27, while Apple continues developing its own separate foundation models for longer-term use.

TheWhiteBox’s takeaway:

It’s totally clear at this point that Apple will not participate in the foundation model battles at all and instead hopes to license that tech to others and fine-tune it to their needs.

This may seem outrageous, but it could make sense; Apple is a consumer hardware company at the end of the day, and all their use cases will involve power-constrained models (i.e., small models).

Yes, they could spend billions of dollars in training huge models and then use these as teachers to train smaller models that…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.