I've always wondered what would happen if we could create AI chips that could think not twice, but a thousand times faster of a human brain. Today, this goal is becoming a reality before our eyes. Advances in AI chip design are accelerating at such a dizzying pace that yesterday’s boldest predictions already seem obsolete.
Computing power is growing exponentially, and with it the possibility of developing artificial intelligence that far surpasses human cognitive capabilities. But are we ready for this “evolutionary leap”?
The Vertical Growth of AI Chips
Something extraordinary is happening in the world of AI chips, with XAI on the shields. In just six months, between Grok 2 e Grok 3, we have seen a 15-fold increase in available computing power. A leap that shattered all previous predictions, including the one that hypothesized an increase of "only" 5 times a year. And that's not all: XAI has already installed another 100.000 GPUs, effectively doubling its computational capacity.
But the real game-changer will come with the new chips B200, which promise performance 5 times higher than current ones H100. With an installed power of 490 megawatts, XAI will be able to power as many as 400.000 of these computing monsters. And that's just the beginning: By the end of 2025, the goal is to reach 1 million B200s powered by 1,2 gigawatts of power.
Computing power is growing exponentially, and with it the possibility of developing artificial intelligence that far surpasses human cognitive capabilities.
Hardware Evolution in 2026-2027
I am particularly fascinated by the roadmap planned for the two-year period 2026-2027. During this period, we will witness the introduction of chips Ruby e Dojo 3, each capable of delivering 5 times the performance of the B200. With one million of these chips powered by 1,2 gigawatts of power, we will reach a computing power of 100 zettaFLOPS.
And that's not all: thanks to the Tennessee Valley Authority, the available power could triple by 2028 to 3,6 gigawatts, allowing for the installation of 3 million Dojo 3 chips. This would translate into a computing power of 300 zettaFLOPS, a number that makes your head spin just saying it.
The real turning point, however, could come with customized hardware FPGA e ASIC. By removing software overhead, these chips could offer performance gains 100 to 1000 times faster than the current CUDA stack. We are talking about just 3 years from now to 2028. Realize that the scope of this transformation is unimaginable.
The Future of AI Mega-Clusters
The plan for 2029 It's even more ambitious: two sites of 10-12 gigawatts each, one in northern Alberta and one in Texas, will house 10 million AI chips. Dojo 4 e Dojo 5. These processors They will be 10 times more powerful than their predecessors, reaching 1 exaFLOP per chip.
The total computing power of the two sites will reach 20 yottaFLOPS. And with FPGA/ASIC hardware acceleration, we could see a further increase of 100-1000 times, reaching a computing power between 2.000 and 20.000 yottaFLOPS.
In other words, from now to 5 years, by 2030 the available computing power could be 100 million times greater than that of Grok 3.
The impact on AI performance
The implications of this increase in AI chip power? They are astonishing. According to the laws of scaling, a 10-16% reduction in loss compared to Grok 3 would translate into an artificial intelligence with a QI estimated between 900 and 1200, about 11-24 standard deviations above the human average.
To give you an idea, we are talking about an intelligence capable of solving complex scientific problems or simulating reality in real time. An evolutionary leap that would take us well beyond the current capabilities of AI, towards what we could define as a true superintelligence.
AI Chips, the Avalanche of the Future
Of course, we cannot ignore the challenges that lie ahead. Data scalability will have to keep pace with increasing computing power. Innovative methods for synthetic data generation and video data acquisition will need to be developed to feed these massive AI clusters.
The future that awaits us, if I had to describe it in two words, is both exciting and intimidating. As AI chips continue their exponential evolution, we must ask ourselves not only if we are technically ready, but also if we are prepared as a society to deal with the implications of an artificial superintelligence with capabilities so far beyond our own.
