The end came quietly, as these things often do in Silicon Valley. On April 26, OpenAI shuttered Sora—the web and mobile apps that had, for a brief moment, seemed poised to hand anyone with a text prompt the keys to cinematic video generation. For developers who'd woven Sora into their workflows, the closure felt abrupt. The API? It will linger until September 24, 2026, a kind of zombie infrastructure on life support, but everyone knows what that means.
What's replacing the cinematic gold rush is less glamorous, maybe. But it might actually matter more: infrastructure for structured, pedagogical content. Videos that teach, not just dazzle.
Consider Lamina Labs. It's a two-person team out of Y Combinator's Spring 2026 batch, and their pitch couldn't be further from the Hollywood dream. No sweeping drone shots or noir lighting. Instead, think whiteboard explainers—rendered in seconds, following a deterministic script rather than a diffusion model's aesthetic gamble. Their first product, Simi, ingests PowerPoint decks, PDFs, Word docs, even plain text prompts, and spits out teaching videos with sequential strokes designed for learning.
Sudip Rokaya, an MIT CS and math student on leave, has been posting LinkedIn demos since early June. In one, he claims Simi generated a video in roughly 11 seconds. Those are vendor claims, not third-party benchmarks, but the framing matters. This is infrastructure aimed at EdTech platforms, L&D teams, course creators—anyone who needs explainers at scale, not one-off art pieces.
Third-party trackers have started describing it as a "deterministic whiteboard video API for EdTech agents." That's a mouthful, but it captures the bet: programmatic, schema-driven rendering instead of the unpredictable outputs you get from traditional text-to-video models.
The Cinematic Bifurcation
Sora's exit didn't happen in a vacuum. The AI video market is splitting in two, and the divide is widening fast.
On one side, you've got the cinematic players. Runway raised $315 million at a $5.3 billion valuation back in February and partnered with Lionsgate in June, chasing what they call "world models" and Hollywood-grade workflows. Google embedded Veo 3.1 into YouTube Create, Shorts, and Workspace—AI Ultra subscribers can generate up to 1,000 Veo-powered videos monthly as of early April. In China, Kuaishou's Kling platform reportedly pulled in over RMB 650 million in Q1 2026 alone, a 300% year-over-year jump that suggests mainstream adoption is real.
These are aesthetic engines, optimized for viral clips, marketing assets, creative experimentation. They're excellent at what they do. But ask them to walk through a multi-step problem or illustrate a sequential concept with precise timing? That's where things get awkward.
Educational institutions and corporate training teams need automation, but not the kind that requires you to re-roll a prompt seventeen times hoping the diffusion model gets the diagram right. That's the gap.
NotebookLM added "Cinematic Video Overviews" in early March, turning uploaded documents into documentary-style AI summaries. It's a step toward educational video, sure, but it still leans aesthetic. Golpo AI, a YC Summer 2025 company that raised $4.1 million in February, also claims to build explainers from docs and prompts. Public data is thin, though. The category remains wide open, contested territory without a clear winner.
The Case for Determinism
Here's what makes Lamina's approach different, or at least what they're betting on: embracing determinism as a feature, not a bug.
Traditional text-to-video models are stochastic. You input a prompt, cross your fingers, wait. Maybe the output aligns with your intent. Maybe it doesn't. Iteration is slow, expensive, aesthetically unpredictable. For a teacher or instructional designer who needs a specific sequence illustrated with precise timing, that's not just inconvenient—it's a dealbreaker.
Deterministic rendering treats video as the output of a structured schema. Think of it less like Runway's freeform creativity and more like programmatic animation. 3Blue1Brown's Manim engine—which powers those viral math explainer videos via code—offers a template. Excalidraw's vector diagram JSON schemas are another analog: schema-constrained whiteboard rendering that agents can edit and orchestrate.
The vision, as Lamina sees it: LLMs generate lesson plans, which compile into whiteboard sequences with stroke-by-stroke control, rendered near-instantly. Microsoft Research published a paper in March on StreamWise, a real-time multimodal generation service architecture designed for podcast video with sub-second startup times. Academic work earlier this year showed LLM-to-Manim pipelines improving student post-test scores over text-only baselines in controlled settings. The technical pieces are converging, at least on paper.
Rokaya's co-founder, Kartikesh Mishra, rounds out the technical side with MIT EECS degrees (BSc 2024, MEng 2025). The team is young, scrappy, building in San Francisco on YC's standard deal—around $500,000 in total funding as of late June, with no additional rounds disclosed publicly. That's lean even by early-stage standards. It also means Lamina is racing to define a category before better-capitalized players show up.
The EdTech Appetite
The AI-in-education market hit $7.52 billion in 2025 and is projected to reach $10.6 billion in 2026, according to ResearchAndMarkets data published in April. That's a category far broader than video—chatbots, adaptive learning platforms, grading automation—but it signals institutional appetite for instructional content automation.
The AI video generator market itself, estimated at $788.5 million in 2025 by Grand View Research in January 2026, is forecast to hit $946.4 million this year and $3.44 billion by 2033, growing at roughly 20% annually.
Avatar-based training platforms already dominate corporate L&D. Synthesia, which reported $100 million in annual recurring revenue in 2025 and is targeting $200 million this year per Forbes, powers multilingual training videos for companies like Zoom and Electrolux—90% faster than traditional production methods, they claim. HeyGen and Colossyan serve similar workflows, with customers spanning HP, BASF, BMW, Novartis.
But avatars are talking heads. They narrate. They don't draw diagrams or walk through problem-solving step by step. That pedagogical gap is where whiteboard infrastructure fits—or where Lamina hopes it fits. They're positioning Simi for EdTech platforms that want to embed explainer generation directly into lesson authoring tools, LMS ecosystems like Canvas or Moodle, AI tutoring agents that need to "show while explaining."
Whether that market is ready to pay for it remains an open question.
Regulatory Side Benefits (Maybe)
There's another wrinkle, though Lamina isn't advertising it: the regulatory environment around AI-generated content is tightening.
YouTube rolled out stronger disclosure requirements in May, requiring creators to label realistic AI content and deploying internal detection signals. TikTok and Meta tightened synthetic media policies throughout 2024 and into 2026. The EU AI Act's Article 50 mandates machine-readable watermarks and labeling for AI-generated audiovisual content, with enforcement ramping up later this year. China's CAC measures and national standard GB 45438-2025 impose similar requirements.
For startups building educational content, these rules matter less if the output is clearly pedagogical—whiteboard diagrams and step-by-step illustrations don't masquerade as realistic footage. The risk profile is lower. The compliance burden lighter. That's not a selling point Lamina is leading with, but it's a structural advantage worth noting.
The Road Ahead (or Not)
Lamina Labs is still pre-product-market fit by conventional metrics. No public customer list. No disclosed pricing. No non-YC funding as of late June. The company's YC profile lists founding year 2025, and the Spring 2026 batch suggests Demo Day arrived recently or is imminent. Whether the deterministic video thesis resonates with EdTech buyers—and whether Simi's speed claims hold up at scale—remains unproven.
But the timing is notable, perhaps more than the founders expected. Sora's shutdown removes what could have been a platform competitor. Google, Luma, and Runway are focused on cinematic workflows and creator tools, leaving the pedagogical niche underserved. NotebookLM's video overviews normalize AI-generated learning content but stop short of offering editable, schema-driven infrastructure. Microsoft and other research teams are exploring real-time streaming diffusion and causal attention architectures, which could enable agents to "draw while talking" in product support and onboarding flows.
If Lamina can nail the developer experience—clean APIs, reliable latency, integration hooks for LMS platforms—there's a path. The alternative to cinematic AI video isn't no video. It's structured, fast, repeatable explainers that scale with the lesson plan, not the render queue.
That's less sexy than Hollywood. But it might be what the market actually needs.
