MINIMAX just announced revenue of $117 million. Up 283.1%. The headline is clean. The market will absorb it as proof that AI video generation has crossed into commercial territory. But I did not build my career reading headlines. I built it by pulling transaction records apart until the truth underneath became too obvious to ignore. So I start with the number the press release did not put in bold: gross margin is 17.8%.
That is the real story. And it is not a nice one.
Hype is a mask; the ledger is the face beneath it. The ledger says that MINIMAX generated $117 million in revenue, but its direct costs ate more than $96 million of it. That leaves $20.8 million in gross profit. Meanwhile, the company still lost $358 million in the period, a loss that narrowed only 11% from the prior year. A narrowed loss sounds like progress. In absolute terms, the company is still burning more than three dollars for every dollar of revenue it brings in.
This is a forensic problem. And it demands a forensic approach.
Context: The AI Video Hype Cycle
MINIMAX is not a crypto project. It does not hide behind pseudonymous founders or token emissions. It is a Chinese AI startup with a real flagship product line, the Hailuo series, built around multimodal generation, especially video. In 2026, AI video is the most crowded, most capital-hungry corridor in artificial intelligence. Competitors include global names like OpenAI’s Sora, Google’s Veo, and domestic Chinese platforms with massive distribution pipelines.
MINIMAX’s revenue trajectory suggests genuine product-market fit. A 283.1% year-over-year increase is not something a company can fabricate entirely. It implies subscribers, API calls, or enterprise contracts are actually growing. But the same set of financials also reveals a business model that is brutally capital-intensive, with gross margins that would be considered unacceptable in almost any mature software sector.</p><p>Compare that to established SaaS companies, which routinely run above 70% gross margin. Even cloud infrastructure providers, which carry enormous physical and energy costs, usually deliver margins in the 30% to 50% range. MINIMAX is sitting near 17.8%. For a company that sells access to software and model inference, that is not only low; it is a Darwinian warning sign.
Core: Disassembling the Growth
Let me walk through the arithmetic. I have done this long enough to know that the fastest way to evaluate a company is to strip out the narrative and recalculate the unit economics from raw disclosed figures. The math here is simple, but the implications are not.
Revenue: $117 million. Gross profit: $20.8 million. Gross margin: 17.8%. Net loss: $358 million.
From these four numbers, we can reconstruct the shape of the business. The loss is 17.2 times the gross profit. That is not an operating company in the traditional sense; it is a research lab with a revenue engine attached. The company is spending 17 dollars on sales, research, administration, and infrastructure for every dollar of gross profit it preserves.
Even worse, the loss implies a forward burn rate of roughly $700 million per year if the second half mirrors the first. At that rate, a company with less than $1.5 billion in cash would face another dilutive round within the next two years. The HKEX listing gives MINIMAX access to public markets, but public markets are unforgiving of companies that cannot demonstrate a path to unit-economic health.
The one genuinely positive signal is the margin trajectory. Gross profit grew 464.8% to $20.8 million. By back-calculation, the prior-year gross margin was roughly 12%. So the company has improved its gross margin from about 12% to 17.8%. That is real, measurable progress. It suggests inference costs per video are falling, or pricing is getting less aggressive, or the revenue mix is shifting toward higher-margin enterprise contracts.
But a 5.7-percentage-point improvement is not a breakthrough. For a video generation company, the direct cost stack includes GPU capacity, cloud compute, electricity, data storage, and content moderation. Even a small cost overrun in GPU procurement can erase an entire quarter of margin improvement.
I have seen this movie before. In DeFi, we called it liquidity mining without a sustainable vault. Projects inflate usage by bribing users with token rewards. The usage metrics spike. The revenue or TVL graphs look amazing. But the foundation is a subsidy machine. MINIMAX is not farming tokens; it is farming market share with below-market pricing, heavy marketing, and an enormous infrastructure bill. That is perfectly legitimate, but investors must not confuse it with a profitable model.
Every transaction leaves a scar on the chain. Here, the scar is the 17.8% gross margin.
The missing disclosures are as loud as the disclosed numbers. No R&D expense breakdown. No customer acquisition cost. No retention rate. No split between consumer subscriptions and enterprise API revenue. No cash balance. No capex guidance. In a public filing context, this kind of selective disclosure is a tell. It usually means the story only works when the audience is not allowed to see the wiring.
The Cost Structure Problem
Video generation is not text. A text token costs fractions of a cent. A single high-definition video inference can require thousands of GPU operations. The financial statement is consistent with a company whose majority of direct expenses is compute. That creates two vulnerabilities.
The first is the supplier lock-in. MINIMAX, as a Chinese AI company, faces export controls on the most advanced Western GPUs. It may be leaning on domestic chips or on cloud providers with limited supply. This is not a purely engineering problem; it is a geopolitical risk. Any disruption in the supply chain will hit gross margin directly.
Second, pricing power is fragile. If a competitor releases a model that produces equal or better quality at lower cost, MINIMAX will be forced to match prices, accept lower margins, or lose users. The industry is still in a state where quality differences are visible. But those differences are narrowing, and the history of technology markets suggests that price eventually wins.
Numbers have no emotions, only consequences. The consequence of a 17.8% gross margin in a competitive, price-sensitive market is that the company must continuously improve its model efficiency at a pace faster than its competitors. That is a high-risk operational bet.
Contrarian: What the Bulls Saw First
I am not here to say the entire story is false. That would be lazy analysis. The contrarian position in this case is that the bulls may actually be looking at the right waveform, even if they are ignoring the frequency.
A 283% revenue growth rate in a deep-tech, capital-intensive field is not normal. It indicates that the user pain point is real. AI video generation in 2026 is not a toy; it is increasingly integrated into advertising, game development, film pre-production, and short-form social media. If MINIMAX is genuinely a top-tier video model developer, then its current loss is simply the entry fee for a larger long-term prize.
Gross margin going from 12% to 17.8% is evidence of engineering efficiency gains. This could be the beginning of a path to the 40-50% margins that enterprise AI infrastructure companies eventually reach. The company is also sitting inside the Chinese market, which has both enormous content demand and a domestic cloud ecosystem that may offer cheaper alternative compute than the US market.
I have audited AI-generated code for a lending protocol where the syntax was perfect but the logic was fatally flawed. The difference between syntax and logic is the difference between revenue and profitability. The revenue number is syntactically correct. The logic of the business model is still being tested.
From my work reconstructing the FTX ledger, I also learned that selection is the tool of the narrative designer. The FTX balance sheet had assets that looked real until you followed the counterparty chains. Here, the counterparty is the entire infrastructure supply chain. If MINIMAX can lock in fixed-price compute contracts, or if it can build its own ASIC-level optimization layers, the margin story changes completely. The bulls see optionality. I see a timeline that is shorter than the narrative implies.
Takeaway: The Only Metric That Matters
The next 12 to 18 months will decide whether MINIMAX is a pioneer or a patient in an AI hospital. The revenue number will grow. The loss will eventually narrow. But the only figure that matters is gross margin, measured quarter over quarter, after controlling for price changes and one-time cloud credits.
If gross margin crosses 30% within the next four quarters, this company has a real path to becoming an infrastructure layer. If it sticks at 17.8%, or worse, slips backward, then the $358 million loss is not a temporary investment; it is the permanent cost of living in a losing cost structure.
I do not care about the next funding round. I do not care about the next model demo. I care about the cost per video frame, the dollar per million tokens, and the line item that separates traction from leveraged destruction.
The blockchain was never silent. This balance sheet is not silent either. It is saying that MINIMAX has demand, but not yet discipline. The market should listen to the ledger before it reads the next product announcement.
I will be watching the gross margin line, because that is where the truth scars the page.