Future of data

1. Record data centre demand

   Small companies are coming up with new AI applications, while the big ones are providing computing power.

   Securing good data infrastructure is a big battle and cloud providers like Google and Amazon and specialised chipmakers like NVIDIA enjoy a disproportionate advantage here.

   As Sifted has reported, not only are startups largely dependent on the hardware and services provided by foreign players, they also have to play by their rules, which can become more constraining as access to GPUs — aka the “compute” used to train AI models — becomes increasingly competitive and waiting lists for key components grow.

   The growth in data-heavy AI applications sees Europe bulking up on server farms: 2023 will be “a record year” for data centre growth in the region, according to analysis from real estate company JLL. Supply is predicted to go up 17%, with Frankfurt, London, Amsterdam, Paris and Dublin to see the most new hubs created.

2. New types of licensing deals

   The growing data needs of AI startups will create new licensing opportunities for content providers, Pablo Ducru, the founder of a new large language model (LLM) called Raive, told an audience in Paris last week.

   He envisages new royalty streams for content providers loaning their data to AI companies for training purposes, rather than redistribution. “The early AI pirates grabbed data without asking permission — it won’t continue that way,” he says.

   Media companies including the New York Times are rewriting their terms and conditions to ban LLMs from scraping their data, forcing AI to go through the front door and sign exclusive licensing deals (see also OpenAI’s partnership with Associated Press).

3. Good data still needs (super)human efforts

   Powerful AI systems will still have to reckon with human foibles. Take the experience of Paris-based Owkin, for example. The company wants to use AI to personalise treatment for every patient. But it needs very good data to do this, which means it needs people like Agathe Arlotti, the company’s senior vice president of partnerships.

   Arlotti’s job is to forge relationships with the people who run Europe’s hospitals and convince them to trade everything from molecular data to slide decks created by the pathologist. “Hospitals are not ready to do this, so we have to send our team to build with them,” she said last week at a Paris conference.

   There’s also data privacy legislation, created by humans, for startups to understand and get around. “GDPR’s slowing down a lot of research products,” Arlotti adds.

4. Beware “poisoned” AI

   Mo data, mo problems. With so much fake data in our future systems, what are the risks that some of it will be used for fraud? Security specialists warn that manipulating the data used to train machines offers a powerful method to get around AI-powered defences.

   This “poisoned AI” is malicious code labelled as “good” by a hacker hoping to trick a neutral network into thinking a piece of software is harmless. Or it might not even be a hacker: we’re seeing the emergence of new tools in the US that let artists subtly tweak their art before they upload it online so that if it’s scraped by an AI company without their say so, it can cause the resulting model to break.

   There’s going to be added focus in the future on ensuring data is clean and trustworthy — look out for a new generation of companies offering to help.

Who early stage startups are up against