The race for AI is reshuffling the cards in the semiconductor industry
The craze for generative AI is boosting the
semiconductor industry, particularly to meet growing demand from businesses.
Between Intel, Samsung and TSMC, the race is on.
• This market boom, which is expected to double within six years to reach
nearly $1 trillion annually, is propelling companies into a frantic race for
investments.
• They are nevertheless encountering obstacles such as the shortage of talent
and geopolitical issues that are weakening this flourishing industry.
The exposure of the uses of generative AI and the interest
in companies are causing an unprecedented emulation in a semiconductor industry
that is facing many challenges. Because, in terms of AI, the players
manufacturing traditional processors – CPUs (Core Processing Units) – are
outdated: it is the graphics processors, GPUs, that are used to train and run
AI models. " Demand will continue to increase and,
already, the leader Nvidia is delivering its GPUs on a contingent basis ",
explains Carlo Reita, Director of Research and Innovation Alliances at the
Technological Research Department of Nvidia's success is notably due to a
well-oiled hardware/software articulation, the key to the implementation of
new-generation computers. For the researcher, the craze is due to a vision
centered on the cloud, because generative AI is too heavy to be embedded:
" Companies need these tools to improve their production
and their offer by exploiting their databases. And in Industry 4.0, there is
also a significant interest in AI. " The enthusiasm is such
that players like Amazon , Google and
many others have decided to design their own semiconductors dedicated to AI.
When we talk about the AI market, we're talking about building factories that cost between $10 billion and $20 billion.
Pharaonic financial needs
OpenAI
CEO Sam Altman's announcements made headlines when he said it would
take $7 trillion to restructure the semiconductor industry. Estelle Prin,
founder of the Semiconductor Observatory, believes that the American
entrepreneur " is asking the right questions, because
this corresponds to an industrial necessity and he is facing problems that all
GAFAMs and non-civilian players encounter in semiconductors. We are talking
about a production market that was worth $527 billion in 2023 and that will
increase to $1 trillion per year in six years, hence the need for all players
to be able to increase their production. "
Artificial intelligence requires dedicated production lines
and, for the time being, it is the Taiwanese company TSMC that
produces 90% of the chips dedicated to training and operating generative AI .
For its part, Samsung
has decided to challenge TSMC by offering 2nm chip engravings .
" When we talk about the AI market, we are talking about
building the most expensive and sophisticated semiconductor factories in the
world, which cost between 10 and 20 billion dollars ,"
underlines Estelle Prin. Faced with the Taiwanese giant, Intel
has taken the radical decision to become a foundry again, that is to
say to manufacture its own chips, and is aiming for 1.4nm engravings in 2029.
The American multinational wants to become the second largest manufacturer in
the world by 2030. For Carlo Reita, " Intel can get back
in the race ". To achieve this, its first factories
will have to be operational in 2027. Finally, in addition to computing units,
the development of other components dedicated to AI is booming, such as HBM
(high-bandwidth memory) memories which are taking off thanks to the graphics
processor (GPU) market.
Talent shortage
Beyond market needs, the race for talent will define the
success of the players. “ There is a real shortage of
engineers and PhD students when we talk about AI, and there was already a
shortage of talent in traditional semiconductors ,” notes
Estelle Prin. This
penalizes Taiwan , which sees the number of unfilled jobs increase
year after year. “ There is a lack of skills across the entire
chain,” laments Carlo Reita. From clean room operators
to system application designers, it is difficult to recruit and we feel it at
the CEA. The risk is that supply cannot keep up with demand despite the fact
that it is European talents who have the most advanced skills. ”
A race for sovereignty
In Europe, players such as the CEA are investing in embedded
solutions, known as edge, for AI. In the long term, the exploitation of AI will
be shared between the cloud and the edge.
Internationally, this race for AI has geopolitical
implications. Taiwan and South Korea, the main players in the market, are
threatened by China and North Korea respectively. This is one reason why TSMC
is building a factory in Arizona subsidized by the US government. This will not
resolve the issue of access to raw materials in a context where China controls
60% of gallium production and 80% of germanium production, both essential for
chip design. In other words, the fastest growing market in the world is also
one of the most vulnerable.
