Developer Builds Endless AI TV With MiniMax H3 Max

Developer Rehan Sheikh connected MiniMax H3 Max to a livestream and created something that looked impossible a month earlier: an AI television channel that could generate each clip before viewers finished watching the current one. He called it infinite “interdimensional cable,” an unmistakable nod to Rick and Morty, and let viewers steer upcoming scenes through chat.

The viral framing was “TV will never run out of content.” The more useful reading is narrower and stranger. Faster-than-playback video has crossed the minimum technical threshold for continuous generative broadcasting, but the demo also revealed why infinite AI TV is not yet a business: at the quoted 480p rate, uninterrupted inference alone costs about $4,320 every day.

TL;DR: what actually happened?

What Rehan Sheikh built

On August 29, 2026, Sheikh posted that MiniMax H3 Max “generates video faster than you can watch it,” so he connected it to a Twitch livestream. The video showed a rolling sequence of surreal, AI-generated worlds rather than a finite prerecorded loop.

The crucial feature was not simply continuous playback. Viewers could enter commands such as !prompt, changing the direction of future generations. That turns the audience from passive viewers into a distributed writers' room—although one with unpredictable taste, adversarial inputs, and no production standards.

The stream's “interdimensional cable” label draws from the Rick and Morty episodes in which characters channel-surf through improvised television from alternate universes. Some generations reportedly used the show's characters and visual references. There is no indication that the stream was licensed or connected to the show's creators, so “inspired by” is the accurate description—not an official AI adaptation.

The original stream encountered platform friction. The X trend summary says it moved from Twitch toward Kick and Rumble after moderation flags. Sheikh's later replies linked a Rumble page titled “Interdimensional Cable.” Platform migration is part of the story because an unmoderated prompt box connected directly to public video is almost designed to trigger safety systems.

How can video generation run continuously?

The conceptual architecture is surprisingly small:

Livestream chat
      ↓
command parser → moderation → scene prompt + continuity state
                                      ↓
                                H3 Max queue
                                      ↓
                             generated clip buffer
                                      ↓
                         FFmpeg/OBS streaming encoder
                                      ↓
                           Twitch, Kick, or Rumble

The controller must always stay ahead of playback. If a five-second clip takes three seconds to generate, the system has a theoretical two-second margin to download, validate, transcode, and queue the result. In practice, one request at a time is too fragile: network jitter or a single moderation rejection creates dead air.

A resilient version keeps several clips buffered and generates in parallel. It also stores a fallback slate or approved clip that can play when a request fails. The stream is therefore not literally creating the frame currently on screen; it is continuously replenishing a short look-ahead queue.

This is the same pipeline logic used by agentic video systems, except the render destination never stops. A normal video agent can retry a bad shot and deliver late. A broadcast agent has a wall clock and must fill it.

The minimum control loop

A simplified controller might behave like this:

while (streamIsLive) {
  const rawPrompt = await getNextViewerPrompt();
  const safePrompt = await moderateAndRewrite(rawPrompt);
  const scenePrompt = addContinuity(safePrompt, worldState);

  const clip = await generateVideo(scenePrompt);
  const verdict = await inspectFramesAndAudio(clip);

  if (verdict.approved) {
    await playbackQueue.push(clip);
    worldState = updateState(worldState, scenePrompt, verdict);
  } else {
    await playbackQueue.push(getFallbackClip());
  }
}

The hard work hides inside moderateAndRewrite, inspectFramesAndAudio, and updateState. Calling a video endpoint is easy. Preventing a public audience from steering it into copyright infringement, sexual content, hateful imagery, impersonation, or a broken narrative is the product.

The $4,320-per-day cost is real arithmetic

Sheikh published the economic reality after the demo. At a quoted H3 Max 480p rate of $0.05 per generated video-second, continuous generation costs:

86,400 seconds/day × $0.05 = $4,320/day
$4,320 × 30 days = $129,600/month
$4,320 × 365 days = $1,576,800/year

That is inference only. It excludes failed or rejected clips, parallel candidates, an LLM scene planner, moderation models, storage, CDN delivery, streaming infrastructure, logging, engineering, and platform revenue share.

The arithmetic also assumes one generated second for every broadcast second. A quality-oriented system might generate three candidates and select one, tripling the largest cost line before human review enters the picture.

Sheikh estimated that Twitch advertising alone would require a top 10–20 channel—around the 99.9997th percentile of broadcasters—to break even. Sponsorships, paid prompt influence, subscriptions, or branded production might improve revenue, but each also raises the moderation and disclosure stakes.

This is why the demo matters more as an interface experiment than a standalone media business. Compute has become fast enough before it has become cheap enough.

Why this is not yet an endless television show

H3 Max's launch limits were short: five-to-15-second clips, 768p maximum, and text-to-video or image-to-video generation. Those constraints fit a stream of disconnected sketches. They do not create episodes.

There is no durable narrative memory

The model does not remember an indefinite broadcast. Every clip needs external state describing:

• recurring characters and their visual references;
• what just happened and which objects moved;
• unresolved goals and relationships;
• camera position, lighting, location, and time;
• dialogue already spoken;
• facts the story must not contradict.

A scene planner can summarize recent clips, but summaries do not guarantee visual identity. Reference images help, yet a live channel has to decide which context deserves limited input space and how to recover when a character mutates.

The audience can attack the prompt surface

Public chat is adversarial by default. Users will test blocked topics, protected characters, political impersonations, slurs, sexual material, and prompt-injection tricks. Keyword filters catch the least creative attacks.

A viable stream needs layered moderation before generation and visual/audio inspection afterward. It also needs a delayed broadcast buffer so approved output—not raw output—reaches the platform. The experiment's platform movement is evidence that latency cannot be the only optimization target.

Copyright risk is not hypothetical

The Rick and Morty resemblance makes the demo instantly legible and instantly risky. A channel operator is responsible for what it broadcasts even if a viewer supplied the prompt and a model synthesized the pixels.

The commercially defensible version uses original characters, documented style guides, licensed voices and music, and a blocklist for protected properties. “The chat asked for it” is not a rights-management system.

What developers can learn from the experiment

The most useful lesson is not “clone this Twitch channel.” It is that video models are becoming responsive components.

Generation can now sit inside an interaction loop

When output took minutes, a product submitted a job and sent a notification. At three seconds for a five-second clip, an application can react to a vote, game event, sensor reading, lesson answer, or agent decision while the user remains engaged.

That enables interactive story rooms, live product visualizers, adaptive learning scenes, game interludes, personalized sports explainers, and event graphics. The deeper project list and API patterns live in explainx.ai's H3 Max developer coverage; this experiment supplies the proof that the loop can run continuously.

Playback is now a form of backpressure

Once generation outruns viewing, the audience becomes the slow consumer in a producer-consumer system. The software needs queue limits, selection policies, and cancellation. Generating more clips than anyone can watch wastes money unless a critic ranks them or multiple channels consume them.

Continuity becomes infrastructure

Long-form generative media will be won by systems that maintain state, not merely by the fastest clip model. Character bibles, scene graphs, reference retrieval, episode memory, and automated continuity checks become first-class data products.

This resembles loop engineering for agents: the model call is one step; the reliable loop is the application.

Could endless AI TV become viable?

Yes, but probably not first as a general 24/7 entertainment channel.

The economics work better when at least one of these conditions is true:

1. The channel is intermittent. Generate during a scheduled event rather than paying for empty overnight hours.
2. A sponsor funds the compute. Branded interactive experiences can justify a cost that ad revenue cannot.
3. Viewers pay to influence scenes. Prompt priority, character choices, or branch votes create direct revenue—if moderation remains strict.
4. Clips are reused. A generated segment can serve the live stream, social cuts, highlights, and personalized edits.
5. Inference price falls sharply. A 10x reduction turns $4,320/day into $432/day, still material but compatible with a small production budget.
6. Generation is event-driven. Keep authored or cached footage playing until interaction demands a new scene.

The likely near-term product is therefore not infinite television. It is selectively generative television: mostly controlled programming, with AI producing the moments where personalization or audience participation creates enough value to pay for inference.

What remains unverified

The viral X summary is useful but not a technical report. No public repository, full architecture, prompt-moderation policy, uptime log, or independently audited cost statement accompanied the original demo.

The $4,320 figure is transparent arithmetic based on Sheikh's quoted 480p rate, not a published cloud invoice. The 2.5–3-second generation claim aligns with fal's H3 Max launch demonstrations, but end-to-end livestream latency includes queueing, transfer, validation, and encoding. Treat the project as a credible proof of concept, not a production benchmark.

Related on explainx.ai

• ViMax: agentic video generation guide — planning, generating, and reviewing video in loops
• AI video generation in 2026 — model and workflow comparison
• Claude FM and 24/7 developer radio — a different kind of continuous developer stream
• How AI video generation works — prompts, reference images, iteration, and limitations
• Loop engineering with coding agents — bounded autonomous loops and failure handling
• AI prompts for video production — structuring shots and production instructions

Primary sources: Rehan Sheikh's original H3 Max livestream post, the X trending summary and discussion, and the linked Interdimensional Cable stream page.

Project details, platform availability, model pricing, and social engagement are accurate as of September 2, 2026. The stream and provider prices may change; cost figures are estimates based on the developer's quoted rate, not audited expenses.