AI inference

Equinix Inference Exchange arrives Q1 2027 with 200+ open models, two tenancy modes and no disclosed price. Your token-API commitments already clear that date; your GPU and colo deals do not. Cap the term or attach a re-price trigger.

Nvidia's 3,400 tokens/sec on Groq 3 LPX and Cerebras' 4,400 on CS-4 are both single-stream figures. On a 256-LPU rack at 100K context, the memory math caps concurrency near a batch of 12 — so any capacity plan sized off a headline token rate is sized for one user.

Groq certified as an NVIDIA Cloud Partner on August 12 and closed a $350M down round on August 17 with Nvidia planning to participate. If your capacity plan lists Groq as the non-Nvidia leg, that row no longer exists — and the API was built so you would never notice.

AMD signed a definitive agreement on August 6 to buy Taalas, whose chips etch model weights into mask ROM so one chip serves exactly one model. The business model assumes a one-year hardware life, which makes the cheapest inference tier available only to workloads whose model you can name and freeze.

d-Matrix acquired Wallaroo.ai on 3 August 2026, taking ownership of a platform sold as 'any model, any hardware, anywhere'. Neither company committed to keeping third-party silicon supported — which turns a portability guarantee into a roadmap you do not control.

Qualcomm's $3.92 billion acquisition of Modular — maker of the Mojo language and MAX inference engine — is not a chip deal. It's a direct attack on CUDA, the software platform that has locked 4 million developers and their enterprises into Nvidia's ecosystem for nearly two decades. Combined with a reported $8-10 billion Tenstorrent acquisition, Qualcomm is assembling a $14 billion full-stack alternative for the $255 billion AI inference market. Here's how to assess your own Nvidia lock-in and plan a multi-vendor inference strategy.