Nvidia's dominance in AI chips resembles Cisco's stranglehold on internet routing in the late 1990s, but history shows challengers can break through by optimizing for future market conditions rather than competing on the incumbent's terms.
R.K. Anand, co-founder and chief product officer of Tensordyne, an AI inference company, draws parallels to his time at Juniper Networks. Juniper shipped its first core router in 1998 and captured 17% of the core router market by end of 1999, then 38% by early 2001, despite Cisco's dominance. Cisco held the highest market capitalization in the world in March 2000.
The key difference between Juniper's approach and a losing strategy was architectural innovation. Juniper didn't build a cheaper router. Instead, it "introducing an entirely new architectural approach to how routers are built, along with an innovative design philosophy focused on performance and reliability," Anand writes.
Today's equivalent opportunity lies in AI inference. Google processes more than 3.2 quadrillion tokens a month, up from 480 trillion a year earlier. Inference workloads impose different constraints on memory, power, and cost than model training, where Nvidia has concentrated its efforts.
Anand outlines three principles for challengers. First, optimize for the market that will emerge, not the one already well-served by incumbents. Second, build credibility through demonstrated technological edges since buyers default to market leaders. Third, minimize switching friction by making migration easy.
Tensordyne operates with around 120 people. Juniper had a few hundred employees when it achieved significant market traction, compared to Cisco's thousands.