New Way Now: Imgix processes 8 billion daily media requests 50% faster with Google Cloud’s G4 VMs

*Featured in this video:* Alfonso Acosta, Head of Engineering at Imgix

*Executive summary:* Imgix, a leading real-time visual media platform, evolved its infrastructure from private data centers to a GPU-based environment on Google Cloud’s AI Hypercomputer. To eliminate infrastructure bottlenecks, Imgix transitioned its real-time processing pipeline to G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs. This upgrade yielded a more than 50% reduction in processing latency and boosted throughput per node by up to 6x — all without making a single change to core application code. Imgix also integrates Google’s multimodal Gemini models to pre-screen media assets to prevent unnecessary, high-cost AI processing for its customers. Behind the scenes, Cloud Spanner delivers single-digit millisecond response times cost-efficiently, while BigQuery consolidates data from multiple pipelines to analyze performance and usage in real time.

*Challenge:* Imgix needed infrastructure capable of processing 8 billion daily requests and thousands of concurrent, real-time transformations. Relying on private data centers forced its lean team to waste valuable time managing procurement and frequent on-call incidents rather than on building new features. Processing large, high-bitrate video magnified these constraints, introducing delays that impacted real-time playback performance.

*Solution:* Imgix upgraded its core processing fleet to Compute Engine G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs, using Terraform on Google Cloud to automate multi-region deployments and deliver performance gains without changing a single line of code. To support its core processing pipelines, Imgix relies on Cloud Spanner to achieve consistent, single-digit millisecond response times cost-efficiently, while BigQuery helps unify and analyze real-time performance and customer usage data. Imgix also uses Gemini’s multimodal capabilities to help protect customers from high compute costs, evaluating media assets and then routing them to the best models for the task.

*Results:* Building on Google Cloud’s AI Hypercomputer, Imgix has delivered significant performance gains: a more than 50% reduction in processing latency — cutting median render times from 100 milliseconds to around 40 milliseconds — and an up to 6x boost in throughput per node. Now, Imgix has the foundational power to weave advanced AI capabilities into real-time workflows, enabling new ways for customers to tell their stories at global scale.

*Key takeaways and highlights from our interview with Alfonso Acosta, Head of Engineering at Imgix:*

→ “My team was spending a lot of time managing the hardware and the procurement pipeline. On-call was getting paged multiple times a week, if not a day. We couldn't keep doing this forever. So, moving into the cloud, all those problems now they’re Google’s problems.

→ “What we've done on the stack is use GPUs provided by Google and focus really on keeping the majority of the operations on GPUs so it can be really fast. We moved up to the G4 architecture. We saw our median render times drop from 100 milliseconds to about 40 milliseconds, and we virtually have to change no code.”

→ “For our customers, if we're using the most expensive model every single time, that's going to be a big bill for them. So we use a cheaper model, like Gemini to figure out, ‘Hey, does this image or video even need this AI feature?’”

*Google Cloud products used:* AI Hypercomputer, Compute Engine, Cloud Spanner, BigQuery, Gemini models, Terraform on Google Cloud

*Learn more:*
→ How Imgix processes 8 billion images daily with G4 VMs powered by NVIDIA Blackwell https://cloud.google.com/blog/products/infrastructure/how-imgix-processes-8-billion-images-daily-with-g4-vms-powered-by-nvidia-blackwell