Frenar la carrera de la IA puede dar ventaja a los actores de segunda fila
Una mayor supervisión puede desplazar recursos desde el entrenamiento de modelos hacia la inferencia
The world should be wary when the three most influential figures in AI - Dario Amodei, Sam Altman, and Elon Musk - agree that it is time to slow down. Investors may have fewer reasons to fear the future of humanity, but markets certainly have less to dread at the moment. Amodei's comprehensive text outlines a three-step framework to enhance safety, including external evaluators in companies, greater coordination among frontier model developers, and international cooperation to address new risks.
The specifics of what safeguards to implement and how they will be applied are yet to be determined. However, overall, increased supervision would accelerate the sector's shift from developing and training advanced models to the large-scale use of existing ones, known as inference. Currently, most of the combined computing capacity of Anthropic and OpenAI is dedicated to training, while only 37% is devoted to responding to queries or executing tasks, according to the research firm SemiAnalysis.
That figure has surpassed 29% since the beginning of 2024. The good news is that demand for agents to program, manage emails and calendars, and perform other tasks is skyrocketing, driving demand for chips and infrastructure. By 2030, inference is expected to account for 43% of data center demand, a share far surpassing the 43% currently allocated to AI model training, according to McKinsey.
Consequently, a slowdown in model advancements is likely to redirect a larger portion of global AI spending, estimated at $900 billion this year, towards making this technology more accessible. This could protect the so-called hyper-scalers such as Amazon, Alphabet, and Meta, who dominate the bulk of AI capital expenditure. However, some current leaders in AI, like Nvidia, whose graphics processing units (GPUs) are crucial for training models, and SK Hynix, which supplies wideband memory for those chips, may struggle.
Intel, specializing in central processing units (CPUs), could be one of the biggest beneficiaries. Data centers used for training advanced models typically require one CPU for every eight GPUs; in iterative tasks performed by agents, this ratio can reach one to one. Nvidia's competitors focused on inference, such as Cerebras, could also benefit, as well as Samsung, whose conventional memory products are more suitable for CPUs.
With less emphasis on bringing AI capabilities to the edge, model developers may compete more effectively on efficiency and costs. This will transform the competition and business models. So investors don't need to worry about a sudden market downturn. Whether that's enough to save the world from an imminent apocalypse caused by machines is another matter.
The authors are Reuters Breakingviews columnists. Opinions are their own. Translation by Carlos Gómez Abajo is the responsibility of CincoDías.
Written by urgent.news from El Pais Economia's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
