Tensordyne's R K Anand: beat Nvidia by targeting inference, as Juniper did Cisco

Anand writes from experience: he was a founding engineer at Juniper Networks, the startup that shipped its first core router in 1998 and grew from a fairly anonymous upstart to 17% of the core router market at the end of 1999, then to 38% in early 2001, while Cisco had been the highest market capitalization company in the world in March 2000. He maps that episode onto today's AI chip market, where Nvidia fired the starting gun by hand-delivering its DGX-1 to OpenAI in 2016 and now works with near-exclusive supply and investment deals.

The bridge between the two eras, in his telling, is a shift in what the hardware must do. The initial hyperfixation on training AI models is fading, and the industry is collectively moving on to inference—using trained models—where he says AI's promised productivity gains actually materialize. As evidence of that strain, he cites Google reporting it now processes more than 3.2 quadrillion tokens a month, up from 480 trillion a year earlier, as AI moves from chat to agents that plan and act.

His advice to challengers is less about price than about timing and trust. Build for the market that will come rather than the one the incumbent already serves well, demonstrate and validate each technological edge to compound credibility, and remove the technical friction of switching, since inertia is one of an incumbent's biggest strengths. He also argues small teams can now do what once took an army—Juniper numbered a few hundred people at significant market traction versus thousands at Cisco, and Tensordyne is around 120 people.