OpenAI Kills Sora: $15M Per Day Burn Rate, 6 Months, Dead

On March 24, 2026, OpenAI announced it was discontinuing Sora — the AI video-generation app the company spent two years and an unknown nine-figure sum building. The web and app experiences shut down on April 26. The API follows on September 24. Six months after public launch, the most-hyped consumer AI product since ChatGPT is being put down.

This is OpenAI we''re talking about — the company at $852 billion valuation, the company that is the AI industry. They cancelled their second-biggest product six months in. That has not happened before. It''s worth understanding why, because the answer changes how to think about every AI startup currently raising money.

The Wall Street Journal''s exclusive on the shutdown laid out the numbers, and they''re ugly.

1 million+ downloads in the first week. The launch worked. People wanted to try it. The demos that circulated on X in late 2024 (the dog-walking-on-Tokyo-streets clip, the wooly-mammoth shot) genuinely set a new ceiling for what AI video could look like, and consumers responded.

Active users collapsed to under 500,000 within a few months. Half the people who downloaded the app stopped using it. The remaining users mostly generated novelty videos — meme content, AI-slop edits — not the high-value commercial workflows OpenAI had projected.

Burn rate: ~$15 million per day in compute. Multiple reports peg the operating cost at roughly $15M/day. Video generation is not text generation. Each five-second video requires the model to render millions of pixels across dozens of frames at a quality bar that demands the largest available GPUs for minutes at a time. Compute scales with output size — and video output is enormous.

Total lifetime revenue: ~$2.1 million. Against ~$2.7 billion in projected six-month operating costs at $15M/day. The math is not even close. Sora wasn''t losing money on every transaction — it was a near-pure cost center, with monetization that never approached the daily compute bill.

The $1 billion Disney partnership: cancelled. Disney had signed a multi-year licensing deal in late 2025 that would have let Sora users generate videos with licensed characters. Disney withdrew when OpenAI pulled the plug. That''s a billion-dollar deal evaporated alongside the product.

The ratio is the part that should make every AI investor uncomfortable. Revenue-to-cost of roughly 0.0008. For every $1,000 OpenAI spent running Sora, it took in 80 cents.

OpenAI didn''t shut Sora down to save money in the abstract — it shut Sora down to redirect compute. The freed GPU capacity is going three places.

SPUD and the next-generation language models. OpenAI''s SPUD model was announced in early 2026 with explicit talk of automating economically valuable work — the kind of work that pays. Compute that was rendering Tokyo street scenes is now training and serving a model whose output is enterprise-grade analysis, code, and reasoning.

Enterprise productivity tools. The Wall Street Journal cites internal sources confirming OpenAI is doubling down on the products that make money: ChatGPT Enterprise, the API''s code-generation surface, and the agentic workflows that competitors like Anthropic''s Claude Code have proven companies will pay for.

Robotics research. This was the surprise: OpenAI told reporters it''s using video-generation tech "behind the scenes" for robotics — letting robots simulate physics-based environments before acting in the real world. Sora as a consumer product is dead; the underlying tech survives in a context where the customer is OpenAI itself, not a freemium app user.

The strategic picture is OpenAI getting unsentimental about unit economics ahead of its expected late-2026 IPO. With revenue at $25B+ annualized, the company can afford to be picky. Sora was a drag on the financials any IPO prospectus would have to disclose. Killing it makes the S-1 prettier.

Video generation isn''t just hard. It''s a fundamentally different cost class from text.

A standard ChatGPT-4 response uses tens of thousands of tokens of compute per query. A 10-second Sora video uses the equivalent of a few million. That''s two to three orders of magnitude more compute per output. And consumers expect both to feel free, or near-free, at the point of use.

OpenAI tried to bridge the gap with paid tiers — Sora was bundled with ChatGPT Plus and ChatGPT Pro at $20 and $200 per month respectively. The Pro tier theoretically broke even on heavy users, but the math fell apart in three places:

Free-tier abuse. Users generated their daily quota and stopped, never converting. The funnel was broken.

Quality variance. The cherry-picked launch demos were not representative. Real-user prompts produced output that was 50-80% of demo quality 80-90% of the time — good enough to be impressive, not consistent enough to be trusted in a professional workflow. So commercial users (the people who would actually pay $200/month) didn''t come.

Content moderation costs. Deepfakes of real people, copyright violations from prompts that named films and franchises, AI-generated misinformation. OpenAI had to staff a moderation operation an order of magnitude larger than ChatGPT''s, and even with safety filters baked into the model, every high-profile failure was a brand cost.

Then competitors arrived. ByteDance''s Seedance 2.0 launched in late 2025 with quality approaching Sora at lower API cost. Runway Gen-4 took the prosumer creative segment. Pika and Luma Ray sliced the indie-creator audience. None of them solved the unit economics either, but they didn''t need to — they''re subsidized by parent companies (ByteDance) or VC money still willing to burn cash on growth. OpenAI, with an IPO on the horizon, no longer had that luxury.

The lesson the WSJ source distilled best: the gap between "demo wow factor" and "daily utility" is the graveyard of AI products in 2026. Sora was wow at launch and useful to almost nobody six months later.

Three things, hard.

Unit economics matter more than benchmarks. It does not matter if your AI feature scores 95th percentile on whatever benchmark if the cost-per-call exceeds what users will pay for the call. AI startups in 2024-2025 raised on benchmarks; AI startups in 2026 will need to defend on per-user margin. If you cannot articulate the cost per active user and the revenue per active user with a positive gap, you do not have a product — you have a research project.

The "demo to product" gap is real, and it''s widening. A demo shows what an AI can do at its best, with cherry-picked inputs, a long inference budget, and human curation of the output. A product is what an average user gets on an average day — at the price they''ll actually pay. The gap between the two has always existed in software, but AI widens it because the most impressive outputs require the most expensive compute.

Standalone AI products are losing to "AI as a feature." Look at what survived the AI hype cycle: ChatGPT (chat is the wrapper), Claude (developer-facing API + a chat front-end), GitHub Copilot (embedded in VS Code), Cursor (a VS Code fork), Claude Code (terminal-embedded). What died: Humane AI Pin, Rabbit R1, and now Sora. The pattern is unmistakable — AI thrives when it lives inside an existing workflow and is dragged in by the rest of the product. AI struggles when it''s a standalone destination users have to choose to visit.

This connects to the Meta Muse Spark pivot (open-source to proprietary for revenue), the Claude Advisor Strategy (cost optimization is the discipline that wins), and the Claude Code 67% dumber incident (Anthropic literally cut quality to cut cost). The whole industry is converging on the same realization: the AI products that survive are the ones whose unit economics work.

Sora joins an increasingly crowded list of high-profile AI product failures, and the pattern is getting hard to miss:

• Humane AI Pin — $700 wearable AI assistant. Returns exceeded sales. Company sold to HP for parts.
• Rabbit R1 — $200 standalone AI device. Reviews were brutal; usage cratered.
• Google Bard — Rebranded into Gemini after consumer-facing flop.
• Alexa+ — Amazon''s LLM-powered Alexa upgrade. Slow rollout, mixed adoption.
• Sora — Now joining this list.

The common thread: a massive launch, a viral moment, an unsustainable cost structure, user drop-off, and a quiet shutdown. The products that did not fail (ChatGPT, Claude, Copilot) all share a structural feature these failures lack: they are embedded in a workflow the user was already going to do. ChatGPT replaced Google for some queries. Copilot replaced Stack Overflow for some queries. Claude Code lives inside the terminal where developers already work. Sora asked users to add a new app to their day — and they didn''t.

While billion-dollar AI products burn cash and die, DevPik''s tools cost essentially nothing to run. They''re 100% client-side JavaScript — no model inference, no GPU bills, no compute scaling crisis. A JSON Formatter, a Random String Generator, a chmod calculator — these have stable, knowable costs. They''ll work tomorrow. They''ll work in 2030. They''re free forever because there''s nothing to charge for.

That isn''t a knock on AI — it''s a different category of product. AI tools that solve real problems will survive. But the era of "throw GPU at it and figure out the business model later" is ending. Sora is the highest-profile casualty, and it will not be the last.

While billion-dollar AI products burn cash and die, DevPik''s tools cost nothing to run — they''re 100% client-side, free forever, and they''ll still be here tomorrow. Try our 49+ free developer tools.

• WSJ — OpenAI Scraps Sora Video Platform Months After Launch
• WSJ — The Sudden Fall of OpenAI''s Most Hyped Product Since ChatGPT
• The New York Times — OpenAI Is Shutting Down Sora
• BBC News — OpenAI closes AI video app and cancels $1bn Disney deal
• TechCrunch — Why OpenAI really shut down Sora
• OpenAI Help Center — What to know about the Sora discontinuation
• Wikipedia — Sora (text-to-video model))