Normalized claim
Time: 2 x increase
Epoch cycles became 2x faster.
GCPEvidence: Medium50/100
Resemble AI case study – Gemini/Vertex AI data labeling and deepfake detection on Hypercomputer
Use case typeAI platformUpdated Jun 13, 2026
Resemble AI builds enterprise voice and audio models for real-time cloning, multilingual dubbing, deepfake detection, invisible watermarking, and media identification. As its datasets grew beyond 60 TB, the company struggled with slow data preparation, inconsistent version control, and stalled training pipelines that left accelerators waiting on data.
- Organization
- Resemble AI
- Industry
- Tech & Comms
- Location
- United States
- Published
- January 2025
Reported outcomes
Time: 2×
Time & speed
Deepfake detection fine-tuning time: −99%
Catalog median for time & speed deployments: +59% across 138 reported metrics. Compare benchmarks →
Planned next steps
- The source says the organization aims to achieve: Improved deepfake detection coverage.
Why do we believe this?Outcome claims, sources, and evidence checks
Normalized claim
Deepfake detection fine-tuning time: 99% decrease
Deepfake detection fine-tuning time was reduced by 99%, from 7 days to 1 hour.
Correction recorded Jul 26, 2026 · Previously: 99% decrease · Replaced endpoint-only extraction with one source-verified duration transition.
Normalized claim
Deepfake detection coverage update time: 60 minutes increase
The team could update deepfake detection coverage in under 60 minutes and scale to more than 100 inference requests per second with sub-250 ms latency.
Correction recorded Jul 26, 2026 · Previously: 7 days decrease · Removed a duplicate fine-tuning endpoint after preserving the complete reduction claim.
Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks
- Customer
- Resemble AI
- Provider
- GCP
- Maturity
- Unknown
- Linked source
- Google Cloud Customer Stories
No explicit deployment-stage evidence found.
Customer identity supportedSource describes one deploymentMaturity evidence evaluated
Primary read
Use case focus
Showing 3 of 3
- 1AI model training acceleration
- 2Data labeling automation
- 3Deepfake detection
- Resemble AI moved from ML Engine to Vertex AI for fine-tuning jobs and scaled its training on Google Cloud AI Hypercomputer.
- The company used Hyperdisk ML with A3 VMs to provide high-throughput access to large static datasets, while N2 VMs handled upstream cleaning and transformation.
- It used Google Compute Engine, Google Kubernetes Engine, Cloud Storage, and Cloud Storage FUSE to support training and serving workflows.
- Resemble also adopted Gemini and Gemma to accelerate data labeling for deepfake detection workflows.
- Epoch cycles became 2x faster.
- Deepfake detection fine-tuning time was reduced by 99%, from 7 days to 1 hour.
Architecture
Resemble AI used Google Cloud AI Hypercomputer components including A3 VMs and Hyperdisk ML. N2 VMs handled data cleaning and transformation, Cloud Storage stored model weights, and Cloud Storage FUSE loaded them for serving. Vertex AI handled fine-tuning jobs, with GKE and Compute Engine supporting development and production workloads.
Sources & evidence1
Evidence: Medium50/100Evidence strength
- Customer explicitly identified
- Primary source available
- Quantified outcome available
- Technical implementation details available
Type: Customer StoryPublished: Jan 1, 2025Publisher: Google Cloud Customer StoriesEvidence: PrimaryConfidence: High
AI-generated summary. Verify important details with the linked sources before relying on this case.
Explore related AI use cases
Customer evidence
Provider evidence
AI patterns
Browse the catalog
Was this useful?
Community
Comments
No published comments yet.