For 20 years, advantage was a question of who you hired

For 20 years, competitive advantage was a question of who you hired. Having the best engineers would produce the best products — or at least that’s the story Google sold, and many of us bought it.

The logic was simple. Talent was scarce, so the company with more of it shipped better software.

§On Nvidia’s own ground

Now frontier labs are taking on Nvidia on its own ground. In October 2025, OpenAI announced with Broadcom a plan to deploy 10 gigawatts of OpenAI-designed accelerators. On 25 August 2026 it published first results for Jalapeño, its first custom inference chip. In its own tests, the chip delivered 1.5 to 1.9 times more throughput per rated kilowatt than Nvidia’s GB200 and GB300 systems. Anthropic has moved part of its compute away from Nvidia too, expanding to up to one million Google TPUs, well over a gigawatt of capacity in 2026. Both are software companies whose main product is a model.

Now frontier labs are taking on Nvidia on its own ground.

But now, instead of the best engineers, they have the blend of best engineers and best models, and that is enough to beat a company nobody thought was beatable.

The blend is literal. When the Broadcom deal was announced, Greg Brockman said: “We’ve been able to apply our own models to designing this chip.” Humans had already optimised the components. The model then found its own improvements, including “massive area reductions” that he said would have taken engineers weeks. Google has a longer record here, as its AlphaChip system has generated layouts for three generations of TPU, producing them “in hours, rather than taking weeks or months of human effort.” The model helps design the chip that will later train and serve the next model.

The Jalapeño figures come from OpenAI’s own tests on an engineering sample, with price and yield still unknown. Nvidia says it remains “a generation ahead of the industry”.

§A hybrid culture

It’s mesmerizing to watch organisational culture turn into something hybrid.

Culture used to mean how people work together. Now it also covers how they work with the model. Which tasks go to the model and which stay with people? Who reviews its output, and how closely? Do engineers share what works, or keep private tricks? Two teams with similar CVs can get very different results from the same model.

The differentiator moves fast from who you hired into how your team is using the model.