Does Sora AI use Google?
No. Sora was built entirely by OpenAI, not Google, and Google's competing video model is Veo. OpenAI did sign a Google Cloud deal for general computing capacity, but told Reuters it is not deploying Google's TPU chips at scale. Sora itself is discontinued — the app closed April 26, 2026, and the API ends September 24, 2026.
Why — the first-principles explanation
This question usually packs three different questions into one, and they have three different answers.
"Is Sora a Google product?" No. Sora was OpenAI's, start to finish — OpenAI trained the model, named it, ran it, and killed it. Google's rival text-to-video model is Veo, an entirely separate system from a separate company. The two get conflated because they launched into the same news cycle doing the same trick.
"Does Sora run on Google's computers?" Partly, and this is where the nuance lives. Modern AI companies rent computing capacity from whoever has it, and OpenAI has publicly signed a Google Cloud agreement to supplement its infrastructure. But OpenAI told Reuters it is not deploying Google's TPUs — Google's in-house AI chips — and does not plan to shift significant compute to them any time soon. So "OpenAI uses Google Cloud" is true in the boring sense of renting data-center capacity; "Sora ran on Google's AI chips" is not supported. The distinction matters: renting a landlord's building is not the same as using the landlord's machinery.
"Does Sora search Google to make videos?" No — and this is the most common misconception, worth killing outright. Sora never looked anything up. A diffusion model has no internet connection at generation time. Everything it could produce was already baked into its weights during training, months earlier. When you asked for a cat on a skateboard, it did not fetch a cat photo; it started from random static and denoised it into something statistically shaped like your words. That's also why it couldn't render current events, and why text on signs came out as gibberish — there was nothing to look up, only patterns to imitate.
As of July 2026 the practical upshot: Sora is gone, Veo is Google's live product, and the two were always separate.
An example that makes it click
Think of two bakeries on the same street that both start selling croissants the same month. People start saying "the croissant place," and soon nobody's sure which is which. That's Sora and Veo — same trick, same moment, two different owners.
Now, the second mix-up. OpenAI's bakery rents kitchen space in a building Google owns. That's real — but renting the room isn't using Google's recipes, and OpenAI specifically said it isn't cooking with Google's ovens, the TPUs. And the third: Sora never phoned anyone for ingredients mid-bake. Everything it knew was memorized during training, like a baker who studied a million croissants years ago and now works from memory alone, doors closed, no phone. Ask for today's newspaper on the counter, and they'll draw squiggles that look like text — because they aren't reading anything. They're remembering what reading looks like.
Key facts
- Sora was developed by OpenAI, not Google; Google's competing text-to-video model is Veo.
- OpenAI has a Google Cloud agreement for computing capacity but told Reuters it is not deploying Google's TPU chips and will not shift significant compute to them near-term.
- Sora generated video by denoising random noise in latent space — it performed no web search or lookup during generation.
- Sora is discontinued: the app and web experiences ended April 26, 2026, and the API is scheduled to end September 24, 2026.
- OpenAI announced Sora's discontinuation on March 24, 2026, redirecting compute toward enterprise products.
- Google's Veo remains an active product as of July 2026, while Sora does not.
▶ The 60-second explainer (script)
Does Sora use Google? No. And that question is really three questions, so let's take them in order. First: is Sora a Google product? No — Sora was OpenAI's, built and run entirely by them. Google's competing video model is called Veo. Different company, different system. People mix them up because they launched into the same news cycle doing the same trick. Second: does Sora run on Google's computers? Partly. OpenAI signed a Google Cloud deal for capacity — but OpenAI told Reuters it is not deploying Google's TPU chips, and won't move significant compute onto them any time soon. Renting a room in someone's building isn't using their machinery. Third, and this is the big one: did Sora search Google to make your video? Never. A diffusion model has no internet connection while it generates. Everything it could make was baked into its weights during training, months earlier. Ask for a cat on a skateboard and it doesn't fetch a cat — it starts from random static and un-scrambles it into your words. That's exactly why it couldn't do current events, and why text on signs came out as gibberish. Nothing to look up. Only patterns to imitate. One last thing: Sora shut down April 26th, 2026. Veo is still running.
What authoritative sources say
People also ask
What is Google's version of Sora?
Veo, Google's text-to-video model. It remains available as of July 2026, unlike Sora.
Did you need a Google account to use Sora?
No. Sora used an OpenAI account, the same login as ChatGPT. You could sign up with a Google email address, but that's just an email — not a Google product connection.
Was Sora trained on YouTube videos?
OpenAI never published a full training-data list, and this was a live point of contention. Treat any confident claim about exactly what Sora trained on as unverified.
Does OpenAI use Google Cloud at all?
Yes, for general computing capacity under a publicly reported agreement. But OpenAI told Reuters it is not deploying Google's TPU chips, so it rents infrastructure without adopting Google's AI silicon.
Why did Sora write gibberish on signs if it couldn't search?
Exactly because it couldn't search. It learned what text looks like — shapes, spacing, contrast — without learning to read or write. So it produced text-shaped marks rather than actual words.