Future AI winners will command hardware, training data power

 The current versions of AI are helpful at the “partial automation” of certain tasks, but for them to complete the last mile of training to reach the “full automation” tip of the spectrum, companies and investors should brace for a dramatic escalation of costs.  

If investors want to put their capital to work most effectively, there are some ways to spot an AI winner, according to Neil Thompson, who is an Innovation Scholar at MIT’s Computer Science and Artificial Intelligence Lab and the Initiative on the Digital Economy.  

“About 30 per cent of all the generative AI applications that people put together are going to fail at the proof-of-concept phase [by the end of 2025] because the economics just don’t work out,” he told the Top1000funds.com Fiduciary Investors Symposium at Harvard University. 

“Right now, we’re in an era where everyone is catching up on the things that are already economical to automate [with AI]. We just haven’t done it yet. 

“Then we’re going to have this second phase, where we get to… [AI applications] on the margin.” 

The exploration of AI applications will be a “gradual expansion”, and there are two areas when the competition between companies will be particularly intense. The first one, perhaps unsurprisingly, is competition over hardware such as computer chips.  

“We are scaling up the amount the use of AI way faster than we are scaling up the production of these systems, even though we are doing that as fast as we can,” Thompson said.  

Technology companies have huge reserves of chips. According to AI intelligence firm Epoch AI, Google holds around 1.2 million Nvidia GPUs and its own TPUs (tensor processing unit, a type of deep learning processor), Microsoft has around 600,000 Nvidia GPUs, and Meta almost 500,000.  

“These are chips that are $10,000 to 40,000 each, so these are huge investments that are going into these areas,” Thompson said. 

He highlighted that the scaling law in AI means there is a determined relationship between an increase in computing power and the performance of AI. 

“If we see one firm has an advantage on hardware over another, we could actually try and estimate how much of a benefit that’s actually going to give it.” 

The second area of competition will be around data, specifically how much is available to a company for training purposes. Thompson recalled a conversation with Nvidia about its autonomous driving capabilities, where the company expressed concerns that competitors like Tesla may have a data lead due to the number of cars they already have on the road.  

“So Nvidia made this interesting proposal to them [other car manufacturers]. They said, what we’ll do is standardise the sensors that you put on there. You send us your data, we’ll process it, and we’ll give you back a model that allows you to run it – that meant lots and lots of different car manufacturers were pooling their data at the same time,” Thompson said. 

“Even though they started behind Tesla, collectively, they had more cars on the road.” 

Thompson tipped that this form of data pooling will become more of a common occurrence to close the data gap with, for example, companies like Google which can collect billions of pieces of feedback via their search engine every day. The consolidation of data could lead to a first-mover advantage.  

“We should expect much faster progress in those areas where it’s all the data can be automated, much slower progress in the hands where the data is expensive or rare to get,” he added.  

With the development of AI and the pursuit of more computing power, a next step question is the broader adoption of quantum computers, but Thompson is of the view that it will not supersede classical computers, at least for the time being.

“We’re going to have cases where it [quantum computers] is not useful at all before some date… then there’s going to be some particular problem size and some particular year where it’s going to come in,” he said. 

“For example, if you say I want to crack modern cryptography, there is a moment where you suddenly start to be able to crack any cryptography faster than the classical computer [with quantum computers]. 

“We still need to pay lots of attention to classical computers and AI, but it does mean that in a small number of areas, we should expect big differences… so we can start to think about what that’s going to mean for investing in, say, molecular simulations.”