Video-to-Text as a Service

I. The Opportunity

The thesis: short-form video (TikTok, IG Reels, YouTube Shorts) is the dominant content format, but AI agents and workflows can’t natively process video. They need text. Converting video→text requires extracting frames, running vision models, transcribing audio — and at scale, this is brutally expensive.¹

The proposed business: batch-process millions of videos using economies of scale (bulk API pricing, GPU batching, cached embeddings), then expose the pre-processed structured text as an API or MCP tool that other AI agents and workflows consume. They pay a fraction of what it would cost to process videos themselves.

II. Market Sizing

The global AI video analytics market is real (US$6.2B in 2026, growing to US$17B by 2031⁴). But this includes surveillance, CCTV, manufacturing QA — most of which is NOT your market.

The actual addressable category is “video understanding APIs for developers and AI agents” — a pre-category market dominated by Twelve Labs (US$107M raised⁶) and a constellation of smaller players. Social media listening (US$10.5B⁵) is the downstream buyer, but they build internally or buy from Brandwatch/Sprinklr, not from an indie API.

III. The Math That Kills It (Unit Economics)

This is the section that matters most. The entire thesis rests on economies of scale — can you process video cheaply enough in bulk to resell at a margin? Let’s find out.

Cost Per Video: Raw Input

Average TikTok video: 43 seconds.² Gemini extracts at 1 frame/second = 43 frames × 258 tokens/frame = 11,094 tokens video input, plus ~500 tokens audio = ~11,600 tokens total input.³

The Full COGS Stack

Processing video isn’t just vision tokens. Here’s the full cost per video:

The Margin Problem

To sell this at a margin, you need to charge more than $0.0034/video. But your competition is the customer doing it themselves:

IV. Competitive Landscape

A. The Funded Players

B. The Scrape-and-Transcribe Layer

A cheaper layer already exists that does 80% of what the thesis proposes:

C. The Big Tech Ceiling

Google, OpenAI, and Anthropic all offer multimodal APIs that process video or image frames. They have three structural advantages you cannot match:

• Zero marginal cost on inference — they own the GPUs and the models
• Native batch pricing — Gemini gives everyone 50% off batch, you can’t get a better deal
• Context caching — 90% discount on repeated content⁷, destroying the cache-and-resell thesis

V. Stress-Testing the “Economies of Scale” Thesis

The core argument: “If lots of people want to process videos, we bulk-process for them with economies of scale, then sell the pre-processed text.” Let’s test every assumption.

The “Pre-Indexed Library” Counter-Argument

The strongest version of the bull case is: “We process ALL popular TikToks/Reels proactively, build a searchable library, and customers query it instantly instead of waiting for batch processing.”

The Real Question: Why Not Just Use Gemini Directly?

This is the question every potential customer will ask. Your answer needs to be one of:

• “We’re cheaper” — You’re not. Gemini 2.5 Flash batch at $0.15/M tokens is accessible to anyone. FAILS
• “We’re faster” — Only if you’ve pre-processed the exact video they need. Otherwise you’re also calling Gemini. FRAGILE
• “We handle the scraping” — Real value. Scraping TikTok/IG is fragile and annoying. But Apify already does this as MCP. COMMODITIZED
• “We structure the output for agents” — This is the only defensible angle — but it’s a thin wrapper, not a moat. THIN

VI. Pattern Matching: Who Succeeded, Who Failed

Succeeded (But Different Business)

The Playbook That Works: Vertical Intelligence, Not Horizontal API

Every success in this space sells insights to a specific buyer, not raw processing to developers:

• Plot sells to brand marketing teams (Warby Parker, Visa)
• TikAlyzer sells to individual creators (hook analysis, viral prediction)²³
• HookScan sells to content creators (first 3-5 second optimization)²⁴
• Videolyze sells to agencies (sentiment reports, client-ready analysis)¹⁵

Nobody succeeded selling raw video-to-text as a horizontal API. The ones that survived either built proprietary models (Twelve Labs) or sold vertical intelligence (Plot, Brandwatch).

VII. Founder-Contextualized GTM

Eric’s Assets

• Dog-food — PCRM genuinely needs TikTok/IG video research
• Existing research skills infrastructure — 22 skills already built (Twitter, Reddit, Threads, App Store, etc.)
• MCP/agent expertise — builds agent tools daily (Donna, Sourcy, claw.degree)
• Agentic backend thesis — knows the while-loop+tool_use pattern

Eric’s Constraints

• 3 active projects — Donna (#1), Wenhao (#2), Dunbar (#3). Sourcy at 10-20% cap. Adding another project contradicts the project discipline commitment.
• $1K/week token burn — unsustainable. Adding a business whose margin compresses with every Gemini price drop makes this worse.
• No go-to-market for developer API — Eric’s edge is network + relationships (Dunbar), not dev tool marketing.

VIII. Red Team: Steel-Manning the Business Case

The strongest version of this business is NOT a generic video-to-text API. It’s one of these:

Shape A: “TikTok Research MCP” (Agent Tool)

An MCP server that agents can call to search TikTok/IG by topic and get structured insights. Not raw text — analyzed, trend-identified, sentiment-tagged content. Apify’s Video-to-Text MCP¹⁶ does the raw version; this does the intelligence version. Price: $0.05–0.10/query (bundle of 50–100 videos analyzed). Margin: 60–70% if you use Gemini batch + Whisper.

Problem: It’s an MCP tool, not a business. MCP tools are meant to be free/cheap. The Twelve Labs MCP server¹² is already available. You’d be competing with free.

Shape B: “Pre-Indexed Social Video Library” (SaaS)

Continuously process the top 100K TikToks/Reels per day. Build a searchable database. Sell to brand teams at $199–499/mo (like Videolyze¹⁵). This is the Plot playbook at 1/100th the scale.

Problem: You’d need to process 100K videos/day = US$167/day COGS (Gemini 2.5 Flash batch) + scraping costs. ~US$5,200/mo infrastructure. Need 11–26 paying customers at $199–499/mo to break even. Harder than it looks — and it’s a SaaS sales grind, not a build-and-sell-to-agents play.

Shape C: “Donna’s Eyes” (Internal Feature)

Build TikTok/IG video understanding as a Donna feature. Use it for Eric’s own research runs, Vincent’s opportunity screens, client reports. The “product” is the research output (charged at $100 HKD/run), not the video processing layer.

This is the only version that makes sense for Eric today. It directly improves the #1 priority (Donna), doesn’t require marketing a developer API, and the “customer” already exists (Eric himself + Vincent + future research clients).