Spending on artificial intelligence could hit a staggering $1tn, according to analysts concerned about whether there will be a return on such a spree. Mark Zuckerberg’s answer this week to such jitters was to release his latest AI system for free.
Meta’s Llama 3.1 405B is its most powerful yet, it says, and one of the most capable in the world. While the tech company didn’t disclose how much it cost to train, Zuckerberg, its co-founder and chief executive, has previously disclosed a $10.5bn (£8.9bn) investment in just the chips required to power its AI data centres – with the rest of the electronics, the electricity itself, and the physical building an additional cost on top of that.
Yet despite the exorbitant outlay, the parent of Facebook and Instagram will charge you nothing for it. If you can get hold of a computer powerful enough to run it, you don’t have to pay Zuckerberg a cent.
Whether that gamble will pay off matters to more than just Meta, though: a big bet by investors and Meta’s tech peers is hingeing on the same question.
In June, analysts at Goldman Sachs published a note with the sceptical title: “Gen AI: too much spend, too little benefit?” It pointed to a $1tn investment in AI over the next few years by the tech industries, other companies and utilities on infrastructure, including chips and power grids. That prompted the question: “Will this large spend ever pay off?”
The note covers a range of views on whether the spree will bring acceptable returns, including an economics professor at Massachusetts Institute of Technology, Daron Acemoglu, who argues that “truly transformative changes” brought about by AI “won’t happen quickly”. In other words, the beneficial economic return from this boom might take longer than investors expect.
The research also questions whether power supply can keep up with the demand related to training and operating AI systems – and asks the same about the chips needed to power those models.
There are also more optimistic takes from Goldman analysts, who argue that AI will ultimately automate 25% of all work tasks in the US (thus making the economy more productive but also creating new tasks and products) and that the spending boom is not out of whack with previous tech investment sprees.
However, Sequoia Capital, an early investor in ChatGPT developer OpenAI, has made it clear that AI companies need to work hard to pay back all that investment in infrastructure such as chips and data centres. A prior estimate that AI companies will need to earn $200bn to pay back their investment has risen to $600bn, wrote Sequoia partner David Cahn.
Cahn stressed that backing AI so strongly would “almost certainly” be worthwhile but “the road ahead is going to be a long one”.
Benedict Evans, a tech analyst, asked in a note this month whether large language models – the technology underpinning tools such as ChatGPT – “might also be a trap”, because while ChatGPT might look like a full-fledged product, it isn’t.
Evans draws an analogy between LLMs and the first iPhones and the appearance of the internet – the potential is there, but other things need to happen as well.
“When most technologies first appear they aren’t ready yet and it’s not clear why they’re useful, and they need a bunch more work,” he says. “The iPhone didn’t have 3G and it didn’t have apps. The web arrived in the early 90s, but no one had a modem, let alone broadband. A whole bunch of stuff had to happen before the web could take off.”
Evans adds that chatbots have worked well for people who need time-saving fixes for coding and marketing, but it has yet to bring fixes in other areas.
“There are very specific cases where this already saves a huge amount of time, but it hasn’t generalised to everybody yet.” He adds: “The feeling that I think most people have looking at ChatGPT is, OK, this is amazingly cool … but what am I supposed to do with this?”
This week, OpenAI announced it is testing a search engine in the US, in what could become a direct challenge to search behemoth Google, which has launched AI-generated search answers, too. OpenAI has also hired a former Meta and Twitter executive, Kevin Weil, as its chief product officer to help answer the question: “What am I supposed to do with this?”
This week, tech news site the Information reported that OpenAI could lose as much as $5bn this year, based on analysis of internal financial figures and interviews with sources, as it amasses $8.5bn-worth of operating costs including $4bn a year on renting servers from Microsoft and $3bn a year on training models.
Full-year revenue could be anywhere between $3.5bn and $4.5bn per year the Information reported, implying a loss of up to $5bn. San Francisco-based OpenAI makes money from charging companies to build consumer products using its models, or letting companies build in-house chatbots with its technology, as well as generating revenue from charging users $20 a month to access a more powerful version of ChatGPT. Google and Anthropic also sell subscriptions to AI systems for $20 a month, while Cohere sells its models to business customers for tasks such as coding or data analysis.
The Information admitted its OpenAI calculations were “guesstimates” but if they are right, the company needs to raise cash over the next 12 months. OpenAI declined to comment, although with Microsoft as its main financial backer it could call upon a wealthy benefactor – albeit with regulators circling.
But even as tech companies have been fighting for domination at the “frontier” of AI, investing billions in building bigger and better systems, a second front has opened up in a more traditional battle: cost. OpenAI’s most recent release isn’t GPT-4o, the headline-grabbing “multimodal” system that sounded so much like Scarlett Johansson that it sparked a lawsuit.
Instead, it’s GPT-4o mini, a stripped-back version of the same AI that the company offers to third-party developers at less than 5% of the cost of its frontier system – undercutting the previous cheapest models, Google’s Gemini 1.5 Flash and Anthropic’s Claude 3 Haiku.
On Tuesday, Meta’s release of the latest version of its Llama system undercut OpenAI. Again, while the attention was on the frontier version of the model, dubbed 405B, the company was equally eager to push its smallest LLM, 8B. Unlike its competitors, Llama is available for anyone to download and run on their own systems, with cloud providers such as Groq offering it for a third of the price of OpenAI’s competitor systems. Llama 3.1 8B underperforms those OpenAI equivalents, per benchmarks from Artificial Analysis, but for the price, who can complain?
Meta calls Llama “open source”, although critics dispute the claim: it’s possible to download the model and use it with a fairly free licence but the training data remains entirely closed, and the model is not free for anyone to use for any purpose. Some of the restrictions are practical: the copyright status of training data for large language models is, at best, controversial. Meta keeps the specific data it trained Llama 3.1 on a secret, and almost certainly lacks the licence to redistribute it free of charge.
Other restrictions have more commercial weight behind them. By keeping restrictions on Llama 3.1’s use in place, Meta places itself at the middle of any future sector that grows around its AI: even if it doesn’t charge directly for access to the model, it gets to control the direction of development, and can always close future avenues if competitors grow too big using its tech.
On Tuesday, Zuckerberg said: “I believe the Llama 3.1 release will be an inflection point in the industry… I hope you’ll join us on this journey to bring the benefits of AI to everyone in the world.” For investors and other tech companies, those benefits need to produce a meaningful return.