GCP Vertex AI to AWS Bedrock + Claude, enterprise-grade.

Business challenges

Five concrete pain points shaped the brief.

• Insufficient flexibility in model selection. The existing GCP platform put hard limits on integrating and switching between frontier base models. Model swaps took 1–2 months — completely out of step with how fast the game product roadmap moved.
• Model capability bottlenecks. The GCP stack relied heavily on Vertex AI and Gemini. In game logic code generation specifically, Gemini satisfaction sat under 60%, with clear blockers in multi-model integration, platform integration and resource management.
• Limited support for complex AI scenarios. The existing architecture struggled with long-term memory, advanced reasoning and multi-model orchestration — covering only about 30% of the complex AI scenarios the business wanted to ship. It couldn't take the team into the next stage of AI applications or support intelligent game operations and personalised content production.
• Cross-cloud managed-service adaptation. The business had been deployed on GCP managed services for years. Subtle differences in functional support, interface specifications and operational conventions between GCP and AWS meant the partner had to deliver full-process technical support — adapting 13 categories of core GCP managed services to AWS, with feasibility studies and migration-method validation for each.
• Global expansion technical foundation. Core business is hosted in the target AWS region serving 6M+ APAC users. Looking ahead — overseas markets, AWS's product depth and partner ecosystem — Cloud Alliance wanted the 30-strong operations and R&D team to come out of the migration fluent in AWS, ready for Europe, the US and the Middle East.

Revenue & performance results

• Flawless cross-cloud adaptation, massive AI platform uplift. Precise adaptation of all 13 categories of GCP core services to AWS. Post-migration, the core business and AI platform run on AWS managed services — DB and cache availability at 99.99%, roughly 92% more stable than the previous GCP self-built setup. Annual incidents dropped from ~15 to under 1, single-incident recovery time fell from 2.5 hours to 12 minutes, fully closing the self-built stability risk. The AI platform — now Bedrock-centric — saw +90% model selection flexibility, code-gen satisfaction climbing from 60% to 95%, AI content generation efficiency up +70%, and full coverage of complex AI application scenarios.
• Operating costs sharply down, productivity released. After moving self-built to AWS managed: ops-labour cost down 62%, core data-services operating cost down 38%, overall cloud-resource cost 29% lower than the original GCP solution. With the team freed from infrastructure babysitting, R&D / ops productivity rose +42%. Game version iteration cycle compressed from 1 month to 18 days, AI feature iteration from 2 weeks to 5 days — enough headroom to ship 4–5 more game versions and 3–4 more AI iterations every year.
• Team fluent in AWS and the AI platform within 1 month. Customised training and on-site support brought the 30-engineer team to independent operation of AWS core products, AI platform management and model adaptation inside a month of cutover. AWS operational proficiency rose +82%, AI platform ops capability +85% — capability upgrade achieved, no recurring external training cost.
• Standardised global technology base, ready for overseas. The architecture follows global deployment specifications and reserves standardised interfaces and compliance hooks. New regional deployment is now under 1.5 weeks with 52% lower deployment cost. The first European/US test region is live, serving 600,000+ users, with AI service cross-region response latency held under 120 ms — the foundation for technology rollout and innovation in overseas markets is in place.