MINIMAX reported $117M in H1 2026 revenue, up 283% year-over-year. Gross profit surged 464% to $20.8M. The headline screams growth. But the gross margin is 17.8%. For every dollar of revenue, 82 cents goes to direct costs—mostly compute. This is not a healthy business. It is a cash-burning machine subsidized by venture capital. The AI industry loves to talk about scalability. The numbers tell a different story: scalability without efficiency is just a larger fire.
I have spent the last decade auditing protocols. I've seen the same pattern in DeFi: high growth, low margins, eventual collapse. The difference is that DeFi protocols had transparent on-chain data. MINIMAX is a black box. But the financial statements are enough to reconstruct the underlying economics. This is not a critique of AI. It is a reality check for anyone betting on centralized AI inference as a sustainable business model—and a signal for decentralized compute networks that promise to cut costs.
MINIMAX focuses on AI video generation. Video inference is compute-intensive. A single 10-second clip can require 100x more GPU cycles than a text response. Gross margin of 17.8% implies that their direct cost of revenue is roughly $96M for the half-year. Assuming NVIDIA H100 cloud rental at $3/hour, that's about 32 million GPU-hours. For a company with $1.17B valuation expectations, that is a staggering burn rate. The gross profit improvement from 10% to 17.8% suggests they are optimizing—quantization, pruning, batch scheduling—but the absolute margin is still dangerously low.
Compare this to a typical SaaS company at 70% gross margin. Or OpenAI at 50-60%. MINIMAX is 2-3x worse. The reason is pure compute cost. AI video generation is a commodity service with low pricing power. Developers will switch to the cheapest API. So MINIMAX cannot raise prices without losing customers. They are trapped in a race to the bottom, where the only escape is reducing compute cost faster than the competition.
Here is where the blockchain angle emerges. Decentralized compute networks—Akash, Render, io.net, Golem—promise to cut GPU costs by 50-70% by tapping into idle hardware. But the MINIMAX data reveals a problem: even if they halved their compute cost, their gross margin would rise to only 59%. That is better, but still below SaaS benchmarks. And decentralized networks have their own overhead: latency, reliability, and coordination costs. The narrative that "decentralized compute will save AI" ignores the fact that AI inference is latency-sensitive and requires high-quality hardware. A 50% discount is not enough if the model takes twice as long to generate a video.
I have verified zk-rollup circuits for a year. I know that complexity is the enemy of security. Decentralized compute adds layers of complexity—consensus, slashing, dispute resolution—that the centralized cloud provider does not have. The total cost of ownership for a decentralized solution may not be lower when you factor in the engineering overhead. Check the math, not the roadmap.
Now, let's look at the loss side. MINIMAX lost $358M in six months, only 11% less than the previous year. That is a net loss of 306% of revenue. They are spending $3 for every $1 earned. At this rate, they need to raise capital or go public. The IPO narrative is clear: the Hong Kong exchange is open to unprofitable tech companies. But the valuation will be based on a multiple of revenue, not earnings. With an annualized revenue run rate of $2.3B (if growth continues), a 30-50x P/S multiple would imply a $70-115B valuation. That is absurd for a company with a 17.8% gross margin. Audits are snapshots, not guarantees. The financial audit says revenue is real. It does not say the business model is sustainable.
Let me contrast this with the blockchain world. When I audited the Bancor V2 contracts, I found that the constant product formula had edge cases that allowed arbitrage losses. The fix was a patch. The core problem was that the protocol assumed liquidity providers would behave rationally. They did not. Similarly, the AI industry assumes that compute costs will keep falling at the same rate as Moore's Law. But GPU prices are not dropping exponentially. NVIDIA's next-generation chips are more expensive per unit. The cost curve is flattening. MINIMAX's gross margin improvement from 10% to 17.8% is impressive, but it is a one-time optimization. The next 10% improvement will require fundamental breakthroughs in model architecture, not just engineering.
Core Insight: The unit economics of AI inference are not improving fast enough to justify the current valuations. The 464% gross profit growth is a mirage—it is driven by revenue scaling, not by margin expansion. The gross profit dollar amount is only $20.8M on $117M revenue. That is a tiny cushion. If revenue growth slows, the loss will widen. The market is pricing AI companies as if they are software companies with infinite margins. They are not. They are compute resellers with a thin margin on top.
Contrarian Angle: Decentralized AI networks are not the solution—they are the same problem with different overhead. The narrative that blockchain will solve AI's cost problem is appealing. But decentralized compute is still unproven at scale. The largest decentralized GPU network, Akash, has less than $10M in monthly revenue. That is a rounding error compared to MINIMAX's compute spend. The real cost savings are not in the hardware but in the software stack. Model compression, speculative decoding, and caching will reduce GPU usage more than any decentralized network. The blockchain industry should focus on verifiable inference, not on commodity compute. If AI models can prove their outputs are correct without re-running the entire computation, that is a real value-add. But that is a cryptographic problem, not a marketplace problem.
Takeaway: MINIMAX's financials are a warning to the entire AI and blockchain ecosystem. The hype says AI is the next internet. The data says AI is a high-cost, low-margin business that burns cash. Investors should ask: how long can this company survive without new funding? What is the plan to improve gross margin to 40%? If the answer is "more optimization," then the risk is high. Complexity is the enemy of security. In this case, the complexity is the business model itself. The simplest explanation is that AI inference is a commodity, and commodities have thin margins. Decentralized compute networks will need to offer more than just a price discount—they need to offer a fundamental improvement in the cost structure of AI. Otherwise, they will face the same 17.8% gross margin problem.
Based on my experience auditing the data availability sampling of Celestia, I know that infrastructure claims must be tested under stress. The MINIMAX numbers are a stress test for the AI investment thesis. It fails. The market is still pricing AI as a winner-take-all opportunity. But the financials show a winner-take-some, with margins that leave no room for error. The next six months will reveal whether MINIMAX can improve its gross margin to 30% or whether it will need to dilute its equity to survive. Either way, the math is clear: this is not a profitable business yet.
Code does not care about your vision. The vision of AI transforming every industry is real. But the unit economics of AI inference are not yet at the point where a standalone company can thrive. The path forward is either vertical integration (owning the hardware) or commoditization (becoming a utility). MINIMAX is currently in the middle, and the financial pain is evident. Decentralized compute networks should learn from this: charge less, but also deliver more than just raw GPU cycles. The future of AI infrastructure is not about cheaper compute. It is about smarter compute—proven by cryptographic guarantees, not just lower prices. That is the real opportunity for blockchain.
Check the math, not the roadmap. MINIMAX's roadmap likely promises AGI. The math says 17.8% gross margin. Trust the math.