Welcome to the Halftime Show.

I've been watching MiniMax for a while from here in Shenzhen.

Most people in the US have never heard of MiniMax. Yet their flagship product, Talkie, became the most downloaded AI entertainment app in the United States by mid ‘24.

They are also the fastest AI company in the world to complete an IPO from founding…. (4 years). Backed by Alibaba, Tencent, and miHoYo. Average employee age: 29.

Fascinating company. Even more fascinating to finally see their books.

Let's lock in.

MiniMax went public in Hong Kong in January 2026, raising $620 million with shares doubling on debut day.

On March 2nd, it filed its first full year results.

OpenAI, Anthropic, DeepSeek are all private. For anyone trying to understand the actual economics of building a frontier AI lab, MiniMax just published the most important document in the industry this quarter.

A real P&L, from a real company, in the most competitive AI market on earth.

And the numbers reward careful reading.

→ MiniMax didn't start as a frontier AI lab. First product: Glow, an AI companion app shut down by Chinese regulators in 2023. Rebuilt internationally as Talkie. By mid 2024, Talkie was the most downloaded AI companion app in the United States, with more than half its users American. By its January 2026 IPO, the US was MiniMax's second largest market globally. (MiniMax IPO Prospectus)

→ Over 70% of 2025 revenue came from outside China. From consumers and developers in 200+ countries paying on their own merits. The US alone accounted for 20.4% of international revenue. (MiniMax press release)

→ Disney, Universal, and Warner Bros. sued Hailuo AI in September 2025. MiniMax's video generation model allegedly infringed on their characters and works. Being targeted by the three largest studios on earth signals the level of reach that product had achieved. (Wikipedia)

→ Three model generations in three months. M2 → M2.1 → M2.5. By February 2026, daily token consumption of the M2 series grew to 6x December 2025 levels. M2 was the first Chinese model on OpenRouter to exceed 50 billion tokens in daily consumption. (MiniMax press release)

→ Enterprise API segment runs at a 69% gross margin. 214,000 enterprise customers across 100+ countries. The consumer apps generate the attention. B2B is where the unit economics are already healthy. (Recode China)

MiniMax's 2025 results are one of the first real public windows into frontier AI economics.

It matters less because it is the biggest player and more because it is the first to show what the business model looks like in practice.

The answer: not like a normal software company. More like an AI application company with a model layer underneath. 

Gross margin is the line to watch

Driven by falling inference costs: down approximately 45% YoY. The model gets cheaper to run as hardware improves and architectural efficiency compounds. If that trajectory holds, gross margins reach 40–50% within two to three years.

The enterprise side is even more interesting. Per MiniMax's IPO prospectus, the Open Platform and enterprise services segment posted a 69.4% gross margin in the first nine months of 2025 - far above what most would expect from a company at this stage. That figure suggests MiniMax has achieved real control over inference costs in its B2B stack.

The loss figure everyone cited is wrong

The widely reported IFRS net loss of $1.87B is an accounting artifact- almost entirely driven by fair value changes on preferred shares ahead of the IPO conversion. Every analysis using that number is reading the wrong line.

The operational figure: adjusted net loss of $250.9M. Painful at this revenue scale. Standard for this stage.

This is not a pure research lab

MiniMax behaves more like an AI application company with strong model capabilities than a pure R&D lab. The consumer products are doing most of the heavy lifting.

Consumer products = drive scale

Enterprise services = drive efficiency

That combination is strategically attractive- the consumer side creates reach while the enterprise side improves economics.

The catch: AI companion apps face growing scrutiny globally around content, mental health implications, and age protection. If the consumer side runs into heavier regulatory pressure, the enterprise segment needs to grow materially faster than it currently is.

R&D is not a cost. It is the asset.

MiniMax spent $252.8M on R&D in 2025 against $79M in revenue.

In a foundation model company, R&D builds the model - and the model is simultaneously the product, the distribution mechanism, the moat, and the margin driver. Every dollar spent improving the model reduces inference cost, increases capability, and expands addressable market.

R&D grew 34% while revenue grew 159%. And the ratio is compressing fast. 

Our broader read

MiniMax is one company. Its P&L is a window into an entire category.

Every frontier AI lab - OpenAI, Anthropic, DeepSeek, Moonshot, Mistral - is probably running a version of this same income statement. 

High R&D. Falling inference costs. Expanding gross margins. International revenue growing faster than domestic.

The labs that win this decade will compound gross margin improvement faster than competitors, while keeping R&D efficient enough that the ratio to revenue keeps compressing.

MiniMax's 2025 report is the first data point public markets have on what that looks like in practice. The gross margin line in its 2026 report will tell you whether the thesis is holding.