Nvidia completes the inference loop: partners with Equinix and Together AI to open model inference to enterprises

NVIDIA is extending its AI infrastructure blueprint from the training side to inference.

On Wednesday, NVIDIA reached a partnership with Equinix, the world's largest data center colocation provider, and AI inference platform company Together AI to jointly offer open model inference services to enterprise customers, filling the final gap from model training to inference deployment.

According to the division of responsibilities, Equinix provides data center colocation, Together AI offers the inference platform, and NVIDIA supplies GPUs and the software stack. The three parties bundle hardware, software, and colocation capabilities to directly target the inference phase of enterprise-level AI applications.

This collaboration is a key complement to NVIDIA's inference-side ecosystem. Previously, the company had already established a leading position on the training side; now, by extending open model inference to enterprise clients, it further broadens the coverage of its computing power ecosystem.

For Equinix, the cooperation also means it has found a differentiated entry point—enterprise-level open model inference—in the trillion-dollar AI data center construction boom.

Division of Responsibilities: A Combination of Colocation, Platform, and Computing Power

In this partnership, Equinix is responsible for data center colocation, Together AI provides the inference platform, and NVIDIA delivers GPUs and the software stack.

In the trillion-dollar AI data center construction boom, Equinix, as the world's largest data center colocation provider, focuses its positioning on the enterprise-level open model inference niche through its partnership with NVIDIA and Together AI, becoming another notable beneficiary in this collaboration.

The core objective of the three-way partnership is to help enterprise clients run inference workloads for open models, instead of limiting computing power to only a few leading model vendors.

NVIDIA’s logic for betting on open models is clear. CEO Jensen Huang recently published an open article explaining the importance of open-source AI, which was co-signed by nearly all major AI companies. In his view, AI models and applications are complementary to NVIDIA GPUs, and the prosperity of open-source models means more accurate and broader demand for computing power.

Inference Revenue Has Surpassed Training: NVIDIA Completes the Loop

According to information disclosed at NVIDIA’s investor conference, about 18 months ago, NVIDIA’s training and inference revenue each accounted for roughly half; now inference revenue has surpassed training, and the gap is expected to continue widening.

Meanwhile, AI computing infrastructure revenue contributed by emerging cloud service providers has exceeded 50%, with growth momentum spreading from traditional hyperscale cloud providers to the broader AI computing power ecosystem.

On the training and architecture side, NVIDIA has previously made intensive moves.

Through a technology licensing agreement, NVIDIA has brought in Groq’s core team and LPU technology, and plans to deeply integrate LPU in the next-generation Vera Rubin architecture (Groq 3 LPX); it also intends to acquire open-source AI platform Hugging Face for $12.9 billion.

Now, partnering with Equinix and Together AI is a key step for NVIDIA in extending its layout to the inference side, enabling its computing power ecosystem to cover the full chain from model training to inference deployment.