New Transformers, Taiwan, & Systems

In partnership with

THEWHITEBOX

TLDR;

Welcome back! This week, we take a look at a lot of market data, including both Taiwan and the recent Venezuelan lore. We also discuss new algorithmic improvements coming out of China in great detail, billion-dollar acquisitions, the ‘eternal’ Tesla vs Waymo debate, and how a small AI startup is kicking the ass of large AI Labs.

Enjoy!

IPOs

The Revealing (and Concerning) Truth about China’s AI Lab Financial Struggles

In December 2025, Zhipu AI and MiniMax filed Hong Kong IPO prospectuses days apart (hearings Dec 16–17; prospectuses Dec 19–21), with filings highlighting fast growth but faster cash burn. In other words, these companies are in quite a bit of trouble and losing tons of money (dollar conversions below use $1 ≈ ¥7.0367).

At first, the numbers look like heaven on Earth.

Zhipu's revenue rose from ¥57 million (~$8.1 million) in 2022 to ¥312 million (~$44.3 million) in 2024, a 447% increase.

MiniMax reported $30.5 million in revenue in 2024 (up 782%) and $53.4 million in the first nine months of 2025.

These guys are printing money, or so you would think. Moreover, gross margins, how much they charge per token relative to the costs of producing that token, were also decent: Zhipu 56% (2024); MiniMax B2B 69.4% (first three quarters of 2025). This means Zhipu charged per token 2.27 times the cost of producing that token, while Minimax, being more enterprise-focused, charged 3.3 times the cost.

Strong pricing power. In principle, this signals a healthy business; one that is offering something at a price well above the costs of producing it.

But then, why are these companies hemorrhaging money?

The problems start when you factor in the cost of maintaining that pricing power. In AI, you can sell tokens as long as your model is on the frontier of what your customers can access. Otherwise, customers churn on the spot.

Yes, companies like OpenAI have that brand value, or brand moat, that makes them a choice by default even if their models aren’t necessarily the best, but that is not the case for the rest of the world.

And the costs of creating these models are eating companies alive on both sides of the Pacific, making the overall losses much larger than revenues.

Zhipu adjusted net loss grew from ¥97 million (~$13.8 million) (2022) to ¥2.47 billion (~$351.0 million) (2024).

MiniMax losses rose from $7.37 million (2022) to $465 million (2024).

As you may have guessed, spend concentrated in R&D/compute. Zhipu R&D hit ¥2.2 billion (~$312.6 million) in 2024 (vs ¥84 million, or ~$11.9 million, in 2022). Of these, ¥1.55 billion (~$220.3 million) went for compute services (about 70% of R&D), with marketing also having a strong presence of ¥387M (~$55.0M) in 2024.

Not only is the product expensive to improve, but it is also effectively commoditized, meaning you also have to spend a lot on marketing to differentiate it.

MiniMax, on the other hand, reported $142 million in training-related cloud compute costs in…

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