A Chinese AI startup backed by over $1 billion in funding declares it will not build video generation models. In a market where every major lab—OpenAI, Google, Meta—is racing to produce Hollywood-quality clips from text prompts, this feels like heresy. But Kimi (Moonshot AI) isn’t retreating from the frontier. It’s doubling down on a conviction that most of its peers have quietly buried: that scaling visual pixels doesn’t scale intelligence.
Let me state this clearly: Kimi’s decision is not a sign of weakness. It is a signal of deep structural understanding. The team has publicly stated that video generation “does little to improve model intelligence” and that their focus is on “usefulness” and “intelligence ceiling” rather than sensory breadth. This is not a compromise born of resource constraints—though those exist—but a deliberate architectural bet on the next phase of the Scaling Law.
Context
Kimi, founded by prominent AI researcher Yang Zhilin, made headlines in early 2025 with the release of its K3 model, which emphasized software engineering, knowledge work, deep reasoning, and image understanding. The model deliberately skipped any video generation capability. At the same time, rivals like ByteDance, Alibaba, and Tencent poured billions into video-generation models (e.g., Jimeng, Emu Video, and others), chasing the viral allure of Sora-like demos.
But Kimi’s CTO, Zhou Xinyu, was blunt in a recent interview: “Video generation has very limited help in improving model intelligence.” He distinguished between “practical utility” and “intelligence enhancement.” The implication is that the current paradigm of video generation—learning pixel distributions and motion trajectories—does not teach a model causality, physics, or logical deduction. It merely teaches it to interpolate frames convincingly.
Core: The Reasoning-First Thesis
Let’s dissect the technical logic. The AI community broadly agrees that the next major leap in foundation models will come from improved reasoning—the ability to break down complex problems, form multi-step plans, and verify outputs. This is the domain of chain-of-thought, self-critique, and code execution. Video generation, on the other hand, remains a deterministic compression and interpolation task. It’s undeniably impressive, but it doesn’t force the model to grapple with cause and effect.
Based on my own audit experience during the 2017 ICO boom, I learned to spot when projects avoided the hard problems by hiding behind flashy demos. A token sale with a slick website but no smart contract audit was a red flag. Similarly, an AI lab that focuses on video generation while neglecting reasoning benchmarks is prioritizing market attention over technical substance. Kimi is doing the opposite: it’s ignoring the flash to build the brain.
Consider the compute economics. Training a state-of-the-art video generation model requires tens of thousands of H100-equivalent GPUs for months. That’s compute that could otherwise be used to train a model that achieves 95% on MATH-500 or solves 50% of frontier-level code problems. Kimi is effectively arguing that the marginal intelligence gain per GPU-hour is higher for reasoning than for video generation. This is a testable hypothesis. If they are right, their K3 model (and its successors) will consistently outperform larger, more multimodal models on tasks like code synthesis, theorem proving, and deep research.
There’s also the question of data quality. Video data is massively redundant: a one-minute clip of a person walking contains billions of pixels, most of which encode the same background. The signal-to-noise ratio for learning reasoning from video is abysmal. Text and code, by contrast, are highly compressed representations of thought. Kimi is betting that language and code are the most efficient training signals for general intelligence. History doesn’t always repeat, but the trajectory of deep learning has consistently favored models trained on clean, high-signal data over those that try to learn from high-entropy raw signals. Yet many in this industry haven’t seen this pattern clearly enough.
Contrarian Angle: The Blind Spot They May Regret
Here’s the contrarian view that I, as a narrative hunter, must surface. Kimi’s rigid stance could be a blind spot if the next frontier of intelligence comes not from reasoning alone, but from grounded multimodal understanding. There is a growing body of research—especially from the robotics and world-model communities—suggesting that to truly reason about physical interactions, an AI must have a rich internal model of how the world looks, sounds, and moves. Video generation, when done well, forces the model to learn physical priors: occlusion, gravity, fluid dynamics, causal chains of motion. The famous “pizza challenge” in Sora showed that even the best video models fail at basic physics, but that failure reveals the gap. Every failure is a training signal for a better world model.
If an AGI breakthrough arrives from a model that can simulate entire environments internally—think of it as a neural physics engine—then Kimi’s text-and-code-only approach might leave it with a crippled capability to reason about the physical world. That’s the risk they are taking.
Moreover, the decision ignores the economic pull of video generation in the crypto-AI convergence space. Decentralized compute networks like io.net, Akash, and Golem are already seeing demand for GPU rental shift toward video inference workloads. Kimi’s reasoning models, while intellectually superior, may be less attractive to retail token holders who want to generate anime videos and NFTs. The narrative utility of video generation is currently far higher than that of reasoning outputs. But that may change as developers build reasoning-as-a-service on-chain.
Takeaway: Implications for the AI-CryptoIntersection
So what does this mean for blockchain-based AI markets? First, if Kimi is right, the most valuable compute on decentralized networks will not be for video inference but for real-time reasoning—think arbitrage bots, smart contract audits, risk simulations. Projects that allocate GPU resources to support chain-of-thought inference could capture higher-margin demand. Second, Kimi’s focus makes it a prime candidate for on-chain model markets. Its models are intrinsically verifiable: you can easily check a code solution or a math proof, whereas verifying a generated video is subjective. This aligns perfectly with the ZK-proof and verification economy that blockchain enables.
Second, Kimi’s decision highlights a fragmentation in AI narratives. One camp (video-first) chases consumer adoption and cultural impact. The other (reasoning-first) chases enterprise utility and intelligence ceiling. As a crypto analyst, I see the latter as more aligned with the values of trust-minimized systems. Reasoning models can be audited, verified, and composed into smart contracts. Video models are harder to trust. The market will eventually price this difference.
Third, the contrarian danger for Kimi is that video generation models might evolve into reasoning engines themselves—by learning to simulate physical worlds internally, they may achieve a form of common sense that text models lack. If that happens, Kimi will be forced to catch up. But for now, their laser focus on depth over breadth is a calculated bet that the next leap in intelligence will come from thinking, not seeing.
I’ve spent years auditing DeFi protocols that promised the moon but couldn’t handle a simple reentrancy attack. Kimi’s refusal to chase the video hype reminds me of those rare teams that said “no” to easy token listings and instead focused on code quality. That team usually wins in the long run. The market hasn’t priced in Kimi’s bet yet. But it will.
The next 12 months will be telling. If Kimi’s K3 model or its successor tops the LMSYS leaderboard in reasoning categories while its video-focused peers struggle to prove usefulness beyond entertainment, the narrative will flip. And when it does, the same critics will say they saw it coming. They didn’t see it yet.