[mushufasa]

I'm skeptical of this. They have been doing this for a decade already. So far, those same companies have just become bigger and bigger customers of NVIDIA.

NVIDIA is very strategic about building product to avoid commodification -- both by building out network effects where software is tied to their proprietary sdk libraries, and by always focusing on being at the cutting edge of product.

Both these things can be true: a large company should try to build their own hardware to reduce supplier risk, and a large company should be open to suppliers that have better product that delivers business value.

So far, these large companies' internal hardware has been useful internally but never a complete replacement for NVIDIA, which keeps staying at the cutting edge of new capabilities.

NVIDIA already faced existential risk when Intel was commodifying all the dedicated motherboard components in the late 90s, 2000s, (like sound cards etc), so they're hyper-aware of this.

[voidspark]

> So far, these large companies' internal hardware has been useful internally but never a complete replacement for NVIDIA

Google use their own hardware for AI/HPC. Nvidia hardware is offered to external customers who demand it on Google Cloud Platform.

The AI datacenter explosion has only occurred within the past two years. We are talking about plans that will take years to implement. The hyperscalers are trying to cut out Nvidia as soon as possible.

[loufe]

I'm baffled seeing this. I'm absolutely convinced of Google value prop, knowing Gemini 2.5 pro was trained and inference is performed on their TPUs is enormous. It's far and way the best model (Claude still has my heart, but their usage limits and limited context can't compete).

The company that can allow collaboration up the value chain only like Google can in this space right now is going to win. The author went over similar pushes by Meta, AWS, and Microsoft.

Now that their eyes are on the prize, and the prize is so staggeringly big, I'm convinced of the threat to their moat.

[tough]

Google's Gemma new QAT training comes from google using their own TPU's

don't know much about the rest tho

[nsteel]

I agree. In the exact same way that hyperscalers now build some of their own network stuff, they still buy an absolute boat load from traditional (specialist) vendors.