The question I’ve been asked most over the past week, by students and clients alike, is “why has OpenAI dropped Sora?”
If you’re not aware, OpenAI is winding down its consumer-facing generative AI video app, the first time we’ve seen a major retreat by the AI company.
The most honest answer would be “I don’t know why” – and few even within OpenAI probably know exactly what happened. After all, just days before they canned Sora, they published a blog post about creating safely with it.
Follow the AI money
But I can take an informed guess, and that would be “cost”. We talk a lot about the energy and water cost of AI, usually in an environmental context. But, of course, those costs come with a financial cost, too. And the price of energy is spiking thanks to Trump’s war on Iran.
A standard ChatGPT query costs 100 times that of a Google search. So, think how much more energy a video clip creation costs… This is a phenomenally expensive business for OpenAI – and one with no clear route to monetisation. Adobe, meanwhile, has bundled it into Creative Cloud — with limits.
One of the elephants in the room of AI – and there are several – is that almost nobody, at a consumer level at least, is actually paying the true cost of the AI inference they’re using. Essentially, the price (business or consumer) we’re paying is being heavily subsidised by the VCs who are betting on a company to win the AI race.
The true cost of generative video
That situation won’t last forever – and it’ll be interesting to see how many people will stay as keen to use AI as they are now if they have to pay the full price. But that’s the future’s problem – we’re not there yet. And efficiencies in AI and changes in our power supply might alleviate that. If nothing else, the current situation with oil has made renewables look ever more attractive…
Today’s problem is that the cost of video generation is, in all probability, vastly out pf proportion to the potential revenue – and so OpenAi, mindful of the likely energy cost spiral, and the need to extend their runway of investment money, are pulling back.
The Wall Street Journal is reporting that Sora was costing OpenAI $1m a day to keep running. There was zero chance they could recoup that any time soon – especially with use of the app dropping.
There are persistent rumours of an OpenAI pivot towards a business customer model – on the basis that businesses are far more likely to be willing to pay the real cost and a profit margin – of AI than consumers. And this might be the first indication that there’s some truth to those rumours. It’ll be worth watching what (if anything) emerges from OpenAI’s hardware collaboration with Sir Jony Ive.
Is this the AI bubble bursting?
Is this the first sign of the AI bubble starting to deflate? Perhaps. There are other signs, too, including a decline in RAM prices, as demand from the AI companies cools.
Make no mistake: we are in a bubble. To whatever degree AI is a transformative technology (and I don’t believe it’ll drop from our radar like the metaverse did), the promises being made of it cannot be supported by the reality of the technology. We’ll adjust, we’ll reset our expectations, and in time, we’ll use the technology for the things it’s actually good at. Maybe video will eventually be part of that – but it’s exponentially more complex and expensive than images, so it could be a while.
Meanwhile, we get to laugh at, for example, Matt Goodwin, once an academic, now a failed Reform candidate, writing his book with AI, and getting busted:
Suicide of a Nation lacks citations, elementary facts are incorrect and quotes appear to have no basis in reality – a classic sign of AI hallucination. What’s more, on those occasions where the failed Reform candidate has bothered to list a citation, he has sometimes left ChatGPT in the URL, making it clear where his research had come from.
This is another good reminder that AI is, essentially, a text comprehension and next word guessing machine (at least, LLMs are). It’s terribly misnamed because there’s no real “intelligence” as we humans would understand it there. It’s all about probabilities.
There are places where it makes perfect sense to use it. (Being able to get Ghost to do what I want by just asking? Yes, please!) But trusting it as a research assistant and co-writer for your book is not one of them. Personally, I’d say exactly the same about letting it write your journalism for you – and I bet we’ll see more journalists and publications embarrassed by doing so in the coming months.
Advice for AI in newsrooms
So, the Sora news doesn’t fundamentally change my basic AI advice, apart from adding a new line. The original:
- Experiment and understand this new technology
- Deploy it where it demonstrably brings speed, efficiency or creativity benefits, once checking is taken into account
- Be very wary of putting AI content in front of audiences — hostility to it is only growing
And the new one:
- Don’t build anything mission-critical that would become unsustainable if AI costs spike.
Because, you know what? Sooner or later, they’re going to…
An AI trilogy
If you want a deeper picture of my thoughts about AI, I’ve recent completed an “AI trilogy” of posts for my long-term client NEXT Conference (a small outpost of the mighty Accenture Song). It’s written for a very different audience than you people reading me here, but you might find them of interest:
Adam Tinworth
Adam Tinworth
Adam Tinworth