4. RESULTS & IMPACT

From a GPU Resource Challenge to a More Flexible AI Infrastructure

Following the implementation, Pikarobot gained greater flexibility in accessing and using GPU resources for AI workloads while improving infrastructure cost efficiency.

~10% Reduction in GPU Costs

GPU infrastructure costs were reduced by approximately 10% compared with Local Cloud, helping Pikarobot improve the cost efficiency of its AI infrastructure.

RTX 6000 on AWS

Pikarobot uses RTX 6000 on Amazon EC2 for AI workloads and model experimentation.

~1,000 Thumbnails Generated in ~30 Minutes

For thumbnail generation, Pikarobot could process a batch of approximately 1,000 thumbnails, with the entire generation process completed in around 30 minutes.

Compared with API-based generation, the self-hosted approach delivered more than 90% cost savings for this generation workload.

High-Volume Educational Content Production Within a Day

Beyond thumbnail generation, the GPU infrastructure was also used to support high-volume educational content production workflows.

Within approximately one day, the infrastructure could support the production of:

• 4 music videos

• 10 animated episodes

• 2 intro/outro videos

• 20 songs

All of these assets were created to support children's learning and educational experiences, demonstrating that the GPU infrastructure could support not only AI model experimentation but also real-world, high-volume AI-generated media workloads.

~2 Weeks to Deploy

The GPU infrastructure was deployed in approximately two weeks, allowing Pikarobot to quickly move the new environment into experimentation and practical use.

1. CUSTOMER PROFILE

Pikarobot – An AI-powered EdTech Platform

Pikarobot is an EdTech product developed by Step Up Education, focusing on using technology to support English learning and conversational practice for children in Vietnam.

Pika Robot is an AI-powered learning companion that provides an interactive environment for children to practice English at home. The platform combines multiple AI technologies to support a more interactive and personalized learning experience.

As the product continued to evolve, the Pikarobot team needed to experiment with different AI models and build workflows that required GPU resources. In addition to model experimentation, the team also needed to handle high-volume AI-generated educational content workflows.

This created a need for infrastructure that could provide sufficient computing power while remaining flexible and cost-efficient based on actual usage.

2. THE CHALLENGE

More GPU Resources Were Needed for AI Experimentation, but Hardware Was Not Always Fully Utilized

As Pikarobot developed new AI capabilities, the team needed GPU resources to experiment with models and run multiple GPU-intensive workflows.

However, several challenges emerged:

Limited Hardware Resources for Experimentation

The team did not have dedicated GPU infrastructure that could be flexibly used whenever they needed to test new models or deploy new tools.

GPU Resources Were Not Continuously Utilized

GPU resources were required for different workloads, but not necessarily on a continuous basis. Paying for dedicated hardware on a monthly basis could therefore result in resources being underutilized.

Relatively High GPU Rental Costs

Monthly GPU rental options could be costly, while Pikarobot's actual usage varied depending on the workload and stage of experimentation.

Limitations of Hourly GPU Rental Services

Some hourly GPU rental options came with network performance limitations or shared/community environments, creating additional concerns around environment control and data protection.

Pikarobot therefore needed more than simply access to a powerful GPU. The team needed a flexible GPU environment that could support their workloads, enable self-hosted models, and optimize costs based on actual usage.

3. THE SOLUTION

Building Flexible GPU Infrastructure on Amazon EC2

Genextech Solutions worked with Pikarobot to deploy GPU infrastructure on Amazon EC2, using RTX 6000 on AWS for AI workloads and model experimentation.

The solution was designed around the team's actual usage requirements rather than requiring Pikarobot to maintain dedicated GPU hardware continuously.

Flexible Access to GPU Resources

With GPU compute on Amazon EC2, Pikarobot could access RTX 6000 resources for model experimentation and GPU-intensive workloads.

This gave the team greater flexibility to test models, evaluate performance, and deploy new AI workflows without investing in physical GPU hardware.

Cost Optimization with Spot Instances

Genextech Solutions incorporated Amazon EC2 Spot Instances to help Pikarobot optimize compute costs.

Instead of maintaining fixed GPU resources, the team could take advantage of suitable compute capacity based on workload requirements and improve cost efficiency for flexible workloads and experimentation.

Self-hosted Models for AI-generated Media

Some of Pikarobot's workloads involved AI-generated media.

Rather than relying entirely on external APIs, Pikarobot could self-host models on its GPU infrastructure, giving the team greater control over the model environment while significantly reducing costs for high-volume workloads.

The infrastructure was deployed in approximately two weeks.

5. BUSINESS IMPACT

Infrastructure That Supports More Flexible AI Development

For an AI-powered EdTech product like Pikarobot, flexible access to GPU resources directly supports the team's ability to experiment with and develop new AI capabilities.

Instead of relying on fixed hardware or GPU rental services with operational limitations, Pikarobot now has a cloud-based environment that can be adapted to different workloads.

By combining Amazon EC2, RTX 6000, Spot Instances, and self-hosted models, the team can:

• Experiment more flexibly with AI models and GPU workloads

• Maintain greater control over the model environment

• Optimize costs for AI workloads

• Access GPU resources based on actual requirements

• Support high-volume AI-generated content workflows

• Build a more flexible infrastructure foundation for AI product development

From GPU Infrastructure to AI-powered Content Production

For Pikarobot, GPU infrastructure supports more than model experimentation. It also provides a foundation for high-volume educational content production workflows.

The ability to process different types of AI-generated media within a short period gives the Pikarobot team greater flexibility to develop and scale content for children, including thumbnails, videos, animations, and music.

Within approximately one day, the infrastructure could support the production of 4 music videos, 10 animated episodes, 2 intro/outro videos, and 20 songs, demonstrating how GPU infrastructure can extend from experimentation to real-world production workloads.

Pikarobot's case demonstrates that GPU infrastructure is not simply about choosing the right hardware. It is also about designing an infrastructure approach that aligns with workload requirements, usage patterns, and the cost objectives of an AI product.

AWS SOLUTION

Customer: Pikarobot
Industry: EdTech
AWS Solution: Amazon EC2 – GPU Compute
GPU: RTX 6000
Cost Optimization: Amazon EC2 Spot Instances
AI Approach: Self-hosted AI models
Deployment Time: ~2 weeks

GENEXTECH SOLUTIONS

Empowering businesses with innovative cloud technology solutions

© 2025. All rights reserved.

Address

17B, Street 12C, Long Binh Ward
HCMC, Vietnam

Contacts

+84 38 569 4657
innovate@genextech-solutions.com

Subscribe to our newsletter

Privacy Policy