AI has awakened: Fastweb’s supercomputer arrives to develop an Italian ChatGPT

AI has awakened: Fastweb’s supercomputer arrives to develop an Italian ChatGPT

In a telecommunications world experiencing a crisis in margins and profits, Fastweb try to launch yourself into the gold rush ofArtificial intelligence. And they do so by purchasing an Nvidia supercomputer, the custom-built machine for training and operating large linguistic systems such as the one underlying ChatGPT, which will be the first of its kind in Italy. The computing power of this supercomputer will be made available to customers, both private and public administration, to develop their AI applications. But Fastweb will also use it to train a younger brother of ChatGPT’s “Italian mother-tongue, in theory better equipped to master the nuances of the national idiom than American passport competitors. The project rides the economic-political trend of the time, which puts “national sovereignty” over key data and infrastructure first. And it is thus a candidate to be one of the poles of the Italian supercomputer network project announced by the Meloni government, which should take shape in the coming months.

Nvidia’s custom-built cars

In detail, Fastweb announced that it has purchased a SuperPod system composed of 31 Nvidia DGX workstations, each of which contains eight processors H100, for a total of 248 chips. These are some of the most advanced microprocessors of the American company, which thanks to the ability to manage a multiplicity of calculations in parallel have established themselves in recent months as the infrastructure underlying Artificial Intelligence models generative developed by the American giants of the sector, from OpenAI, Google, Meta. In the wake of AI-mania, a global race to grab these chips began, which saw large states such as Saudi Arabia and companies such as the French telco move Iliad, which a few months ago purchased a thousand of them by investing 200 million euros to create the most powerful European supercomputer for generative AI. The Fastweb machine, which will be installed in a data center in Lombardy powered entirely by renewable sources and should be operational in the first part of 2024, has a more limited number of processors – and therefore computing power. But considering that each of them costs tens of thousands of dollars, the overall investment, which the company is not disclosing for now, should be several tens of millions.

AI as a service

What will Fastweb do with this supercomputer? First it will sell access to the infrastructure “as a service” to customers who want to use generative AI. The fields of application are the most varied, from research, to financial services, to interaction with customers, and relying on an Italian infrastructure guarantees that the data always remains on the national territory. “Fastweb strengthens its vocation as an enabler of the country’s digital transformation and lays the foundations to allow the development of artificial intelligence services through an approach of national sovereignty and sustainability,” said the company’s CEO. Walter Renna.

Speak as you eat

And from this perspective of a “tricolor” AI also goes the project to develop – thanks to the platform that Nvidia itself makes available to customers – the first large natively trained linguistic model in Italian language. The theme is not secondary, given that several studies have shown how the fact of being designed and developed by American companies, one could say that they “think” in English, makes ChatGPT and its brothers much more competent in that language than other languages. The Fastweb model will hardly be able to compete at the frontier of linguistic understanding with the amazing American models, but it will struggle to compete even with the smaller European challengers (the French Mistral, the German Aleph Alpha), which have teams of engineers on their side highly prepared and hundreds of millions if not billions of dollars in funding. Fastweb, however, is betting that it can still be competitive for those looking for less sophisticated applications but with the ability to “capture nuances of the Italian language and national cultural specificity”.

The Meloni government project

The multinational is thus also a candidate to be one of the private entities involved in the national strategy for AI announced by the government, which should be put on paper – with related funds – in the coming weeks, in time to make it one of the key themes of the G7 under the Italian presidency. One of the chapters already announced, which should replicate what has been done with the national cloud and similar projects in Switzerland and the United Kingdom, will be the creation of a national supercomputer network, with Leonardo’s large calculator in Bologna at the center – which has even more powerful Nvidia processors inside, namely the A100s – and alongside other public bodies and partner companies. Fastweb, which together with Aruba had won the Pnrr tender for the national cloud, only to then see it illegitimately assigned to the alternative consortium composed of Tim, Leonardo and Cdp, will try to enter this new game.

Source link