Nvidia is extending its AI infrastructure dominance from the training phase into the inference stage.

On Wednesday, NVIDIA announced a collaboration with Equinix, the world's largest data center colocation provider, and Together AI, an AI inference platform company, to jointly deliver open-model inference services for enterprise clients. This partnership completes the final link in the chain from model training to inference deployment.

Under the division of labor, Equinix provides data center colocation, Together AI supplies the inference platform, and NVIDIA contributes its GPUs and software stack. The three parties bundle hardware, software, and colocation capabilities to directly target the inference segment of enterprise AI applications.

This alliance is a crucial strategic move for NVIDIA to solidify its position in the inference ecosystem. Having already established a leading edge in the training sector, the company is now leveraging open-model inference to reach corporate clients, further broadening the scope of its computing ecosystem.

For Equinix, this partnership also provides a differentiated entry point into enterprise-grade open-model inference amid the trillion-dollar AI data center construction boom.

Division of Roles: Colocation, Platform, and Computing Power

In this collaboration, Equinix is responsible for data center colocation, Together AI offers the reasoning platform, and NVIDIA supplies the GPUs and software stack.

Within the trillion-dollar surge in AI data center development, Equinix, as the global leader in data center colocation, aligns itself with NVIDIA and Together AI to focus specifically on the enterprise open-model inference niche, emerging as a key beneficiary of this agreement.

The core objective of the tri-party collaboration is to enable enterprises to run inference workloads for open models, rather than confining computing power to a handful of leading model developers.

NVIDIA's bet on open models is clear. CEO Jensen Huang recently published an open letter advocating for open-source AI, which garnered co-signatures from nearly all major AI companies. In his view, AI models and applications are complementary to NVIDIA's GPUs, and the proliferation of open-source models translates into broader and more precise demand for computing power.

Inference Revenue Surpasses Training as Nvidia Completes the Loop

According to NVIDIA's investor communications, roughly 18 months ago, revenue from training and inference was nearly balanced. Now, inference revenue has overtaken training, and this gap is expected to widen further.

Meanwhile, emerging cloud service providers now account for over 50% of AI computing infrastructure revenue, with growth momentum shifting from traditional hyperscale cloud vendors to a more diverse AI computing ecosystem.

On the training and architecture front, NVIDIA has already made dense moves. Through a technology licensing agreement, it brought in the core team from Groq and its LPU technology, with plans to deeply integrate LPU into the next-generation Vera Rubin architecture (Groq 3 LPX). It also intends to acquire open-source AI platform Hugging Face for $12.9 billion.

The alliance with Equinix and Together AI represents a critical step for NVIDIA to extend its footprint into inference, enabling its computing ecosystem to cover the complete path from model training to inference deployment.