MINIMAX's 18-Cent Problem: A Macro View on the Cost Curve of AI Video Generation

MINIMAX's interim report landed on August 26 with all the hallmarks of a carefully staged narrative: revenue up 283.1%, gross profit up 464.8%, losses narrowing 11%. The market will read this as a growth story. Read the numbers again and you will find something else entirely. For every dollar of revenue, MINIMAX spends 82 cents on direct costs. That leaves an 18-cent gross margin. In what universe is that a sustainable business model? This is not a criticism of execution. It is a structural observation about where we are in the AI cycle. We have seen this playbook before. In 2020, during DeFi Summer, I watched protocols celebrate total value locked while their treasury curves inverted. The same pattern is repeating now in AI application layers. Everyone is measuring top-line velocity. Almost nobody is asking about the unit economics of a single video generation inference. Let me be precise about the math. MINIMAX reported $117 million in revenue and $20.8 million in gross profit for the first half of 2026. That implies roughly $96 million was spent on direct costs, overwhelmingly compute. When I audited AI infrastructure costs for a CBDC research project last year, we modeled text inference at fractions of a cent per query. Video generation is a different beast entirely. Generating a single high-resolution clip can cost as much as running thousands of text completions. At this scale, an 82% cost ratio is not an anomaly. It is a feature of the underlying technology stack. The contrarian angle here is not that MINIMAX is doomed. It is that the market is misreading the signal. A 464.8% gross profit growth rate sounds impressive until you realize it is growing off a base that was nearly zero. The real story is that the company has managed to improve its cost structure from roughly 90% cost ratio to 82% in a year. That is meaningful progress, but it is progress from catastrophic to merely terrible. What does this tell us about the broader AI application market? For one, we are still in the phase where AI video generation companies are buying market share with investor capital. The compute-intensive nature of this vertical means that companies like MINIMAX, ByteDance's Jimeng, Kuaishou's Kling, and OpenAI's Sora are all competing in a space where marginal costs are fundamentally higher than text-based AI. This is the opposite of software economics. Traditional SaaS companies hit 70-80% gross margins because their marginal cost of serving another customer approaches zero. AI video generation has no such luxury. My framework for evaluating these companies borrows from my time analyzing liquidity cascades in DeFi. When I mapped the $150 million liquidity crunch during the Compound governance crisis, I learned that leverage ratios matter more than price action. The same logic applies here. MINIMAX's gross margin is its leverage ratio. At 17.8%, the company has almost no room for error. A single pricing war in the AI video API market could push it into negative gross margin territory. The more interesting question is whether this is a structural condition or a temporary one. Based on my experience optimizing inference pipelines, there are genuine levers to pull. Model quantization can reduce memory bandwidth requirements. Distillation can compress model size. Better scheduling can improve GPU utilization. MINIMAX has clearly pulled some of these levers already, given the improvement from roughly 10% to 17.8% gross margin. But the trajectory suggests it will take at least another 12-18 months to reach 30% gross margins, assuming no major breakthroughs in inference efficiency. There is also the regulatory dimension that most analysts will ignore. MINIMAX filed this report through the Hong Kong Stock Exchange, which signals a potential IPO preparation. For a Chinese AI company, this matters. The regulatory environment for AI-generated content in China requires watermarking and content moderation. These compliance costs are not trivial. They add to the operating expense burden that the company is already carrying with its $358 million half-year loss. What I find most revealing is what the report does not say. There is no breakdown of revenue by product line. No disclosure of compute costs. No mention of GPU procurement strategy. For a company that is burning through cash at a rate of approximately $2 million per day, these omissions are not accidental. They are strategic. The 2017 dream of tokenizing everything is today's reality of regulating everything. Similarly, the 2024 dream of AI video generation democratizing content creation is becoming today's reality of enormous compute bills. MINIMAX is running a $7 billion annualized cash burn against a $2.3 billion annualized revenue run rate. This is not a technology company. It is a capital allocation machine. The question is whether the underlying technology can evolve fast enough to justify the capital being poured into it. I have spent the last nine years analyzing the intersection of cryptography and monetary policy. The pattern is always the same. Early adopters overestimate the short-term impact of new technology and underestimate the long-term cost curves. The market is currently in the euphoria phase for AI applications, extrapolating MINIMAX's 283% revenue growth into perpetuity without accounting for the compute requirements that scale linearly with user adoption. From a macro perspective, this matters. The AI video generation sector is becoming a significant driver of GPU demand. As these companies scale, they will put pressure on the entire supply chain from NVIDIA to cloud providers to data center operators. The question is whether the end-market revenue can sustain these infrastructure investments. Based on current gross margins, the answer is not yet. Here is what I am watching. Over the next 12 months, if MINIMAX can push gross margins above 30% while maintaining revenue growth above 100%, the business model becomes viable. If margins stagnate below 20%, the company faces a fundamental structural problem. The market's focus on revenue growth is misplaced. The metric that matters is the slope of the gross margin curve. That is the true signal of technological maturity in compute-intensive AI applications. The deeper irony is that the AI video generation market might be a victim of its own success. As more players enter the space and compute costs remain high, the competitive dynamics could force a consolidation. In that scenario, the winners will not be those with the best models. They will be those with the most efficient cost structures and the deepest access to capital. Based on the current data, MINIMAX has demonstrated the ability to grow revenue but has not yet proven it can build a defensible economic moat. I have seen this movie before. In 2022, I watched the Terra ecosystem evaporate $60 billion in a matter of days because the market believed in an unsustainable yield model. The AI application layer is not facing an imminent collapse, but the underlying economics are similarly stretched. The market is currently pricing these companies on narrative rather than unit economics. That always ends the same way. The winners in this cycle will be the companies that figure out how to decouple revenue growth from compute costs. That means investing in inference optimization, exploring alternative hardware architectures, and building proprietary datasets that reduce the need for expensive retraining runs. MINIMAX has shown some progress on this front, but the current 17.8% gross margin suggests there is still a long way to go. As I look at the broader landscape, I am reminded of a lesson from my time analyzing DeFi liquidity protocols. High growth rates can mask structural fragility. The market is currently rewarding growth at any cost. History suggests that eventually, the market will demand profitability. When that shift happens, companies with healthy gross margins will survive, and those without them will face existential challenges. MINIMAX's 18-cent problem is the single most important metric in its financial report. Everything else is narrative.