Essays

The third era is that of Web 2.0 and online collaboration. After the dotcom bubble burst, renewed excitement about digital media clustered around the social web.

Kramer and coauthors collaborated with Facebook on an experimental tweak to the newsfeed algorithm. Some users received a higher dosage of negatively valenced Facebook posts from their friend networks; other users received a higher dosage of positively valenced Facebook posts. They discovered a miniscule but statistically significant effect on users’ posting behavior. If you see sad posts in your newsfeed, you become slightly more likely to perform sadness in your own postings; if you see happy posts in your newsfeed, you become slightly more likely to perform happiness.

We produce mountains of Twitter and website research. We produce molehills of Facebook research. We produce practically no research on email, Reddit, or the algorithmic choices of the major platforms themselves. And this is entirely because Twitter has, for several years, made its data more easily accessible to researchers than Facebook. Websites can be crawled and scraped, while email lists are closely guarded by civic and political organizations. In the era of big data, most of the research community has flocked to the types of big data that are most accessible.

While many nations invoke the language of “AI sovereignty,” only two—the United States and China—enjoy anything close to full-stack control, which ranges from chip design and fabrication to hyperscale cloud infrastructure and leading AI enterprises. Most other nations connect with points along global supply chains that determine where they operate within a digital hierarchy, rendering sovereignty more mirage than reality.

Access to capital, chips, and cloud infrastructure—not patriotic ambition—determined InstaDeep’s trajectory.

The implications are sobering. Only three countries house 60 percent of the world’s top supercomputers: the United States, China, and Germany. While talent can emerge anywhere, computational capacity remains stubbornly concentrated in the wealthy north.

Third, it complicates the notion of “sovereign AI.” A country might develop its own AI models, but if they run on cloud infrastructure hosted abroad or require imported chips, their sovereignty remains compromised. True “sovereign AI” would require control over the entire vertical stack—from semiconductor design and manufacturing through data centers and energy infrastructure.

Fourth, it creates new forms of dependency. When InstaDeep relocated from Tunisia to London, it wasn’t just about market access—it was about proximity to computing resources. Similarly, when researchers in Global South countries send data abroad for processing, they establish relationships of digital dependency that mirror historical patterns of resource extraction and processing.

Far from transcending national boundaries, AI and its development may reinforce existing geopolitical power structures rather than disrupt them, with control of computing infrastructure serving as a new form of strategic advantage.