Over the past 72 hours, a quiet signal emerged from the AI video production space: Preview, a platform that raised $12 million in two rounds (a $2M pre-seed from General Partnership followed by a $10M seed from Sequoia six months later), is now onboarding over 100 studios and has a waitlist of 3,000 more. The numbers are impressive, but as a protocol developer who has spent the last seven years dissecting the gap between narrative and infrastructure, I see a more fragile picture beneath the surface. Preview positions itself as a "central control panel for AI video production"—a unified workspace where scripts, storyboards, shot lists, AI generation, review, and feedback coexist. Sequoia’s thesis is that AI video needs a "video version of Cursor." The funding is real, but the architectural assumptions underpinning this model are dangerously similar to the composability mistakes I witnessed during the 2020 DeFi summer.
Let me be precise: Preview is not a blockchain product. It is a centralized SaaS platform. Yet the dynamics unfolding here mirror the systemic fragility I mapped in Ethereum’s composability stack. The core insight is that Preview’s value proposition relies on infinite interoperability between different AI models—Claude, GPT-4, Midjourney, Runway, ElevenLabs, and others—all orchestrated through a single API layer. This is the same pattern that gave us flash loan attacks and liquidity fragmentation. The platform records every frame’s provenance: who generated it, which model, what parameters. On the surface, this is a metadata revolution. But metadata without a trustless settlement layer is just a centralized ledger that can be rewritten by the company’s database administrator.
I have audited enough smart contracts to know that when a system claims to be a "unified control panel," the real question is: who controls the control panel? Preview’s architecture stores all generation provenance in a centralized database. The company owns the infrastructure. The models are accessed via API keys that can be revoked. The character, scene, and prop management is a relational database behind a login wall. This is not a composable ecosystem; it is a walled garden with a beautiful facade. Fragility is the price of infinite composability—but only if that composability is built on a permissionless foundation. Preview’s foundation is permissioned, meaning the composability is an illusion maintained by the company’s uptime and goodwill.
Let me calibrate the technical risk. When Sequoia says "AI video needs a video version of Cursor," I hear a dangerous analogy. Cursor is a code editor that hooks into large language models, but it operates on local files and open-source plugins. Preview, by contrast, operates on cloud-hosted assets and proprietary integrations. The difference is between a local compiler and a remote execution environment. In crypto terms, this is the difference between a self-custodial wallet and a centralized exchange. I have seen what happens when centralized exchanges promise composability: they create a honeypot for attackers and a single point of failure for users. Preview’s waitlist of 3,000 studios is a honeypot waiting to be exploited.
The contrarian angle here is not that Preview is a bad product—it is likely a good product for the current market. The blind spot is the assumption that centralization is a stable state. Based on my experience analyzing the Terra/Luna collapse, I recognize the pattern: a platform that aggregates external resources (AI models) and internal resources (user-generated scripts, storyboards, assets) creates a systemic fragility that is invisible during the growth phase. The moment Preview’s database experiences a write failure, or a model provider changes its API terms, the entire production pipeline breaks. Every studio that depends on Preview becomes a leaf in a brittle tree.
I have a specific worry: the metadata provenance feature. Preview records who generated each frame, which model, and what parameters. This is a fantastic feature for accountability and copyright tracking. But without a decentralized, immutable ledger, this metadata is legally meaningless. If a studio uses Preview to generate assets for a Fortune 500 ad campaign, and later a dispute arises over ownership or model usage, the only evidence is Preview’s database. The company could be subpoenaed, go bankrupt, or simply decide to modify the records. I have seen similar issues in NFT metadata storage during the 2021 BAYC analysis—centralized IPFS fallback URLs that rendered the assets worthless when the server went down. Preview’s metadata is the same vulnerability, just with a different surface.
Sequoia is betting that Preview will become the operating system for AI video production. But operating systems require trust minimization at the infrastructure layer. Preview’s architecture is a web app with a database, not a protocol. The company’s $12 million in funding is 10x less than what is needed to build a decentralized alternative, but it is enough to create a dependency lock-in. Over 100 studios are already using it, and the waitlist is 3,000 deep. That is a lot of creative capital flowing into a centralized funnel. The question is not whether Preview will succeed—it will, for a while. The question is whether the market will realize that hype creates noise; protocols create history before the first major outage or data breach.
Let me ground this in a technical audit perspective. I have spent the last year analyzing the custody solutions for Bitcoin Spot ETFs, specifically the threshold signature schemes used by BlackRock. The lesson I learned is that institutional trust is not a substitute for cryptographic guarantees. Preview’s users are trusting the company’s operational security, not a mathematical proof. The same studios that demand provenance for copyright will eventually demand decentralized provenance. When they do, Preview will need to either become a protocol or face a fork.
The market is currently in a bear phase for tech funding, but AI video is still attracting capital because it promises a step change in productivity. Preview’s $12 million is a seed round, not a series A. The company has six months of runway at most. The pressure to monetize will push them to extract value from the data they collect. Every frame, every script, every model parameter is a data point that can be used to train competing models or sold to third parties. The privacy implications are severe, and the regulators are watching. I have seen this playbook before: start with a free tier, build network effects, then monetize the data. Preview is not a protocol; it is a data aggregator with a creative interface.
The takeaway is not a prediction of failure, but a vulnerability forecast. Preview will likely grow to 10,000 studios within the next 12 months. At that scale, the centralized architecture will become a target. The company will need to hire security engineers, implement zero-knowledge proofs for provenance, and migrate to a decentralized storage layer. The question is whether they will do it before the first exploit. Based on the current funding and the speed of the AI video market, I estimate a 40% chance of a critical infrastructure failure before the series A. The fragility is baked into the design.
For the studios on the waitlist: trust, but verify the source code. Unfortunately, Preview is closed-source. You are trusting a black box. That is not a protocol; it is a promise. And in the world of production, promises are not finality. Fragility is the price of infinite composability—but only if the composability is real. Preview’s composability is real only as long as the company stays solvent and honest. I have seen too many protocols die because they assumed the same. The market will wake up, but the network will already be asleep.