Tomofun Cuts AI Costs With AWS Inferentia2 for Pet Cams

Tomofun Taps AWS Inferentia2 to Power Smart Pet Monitoring

Tomofun, the Taiwan-headquartered pet-tech startup behind the popular Furbo Pet Camera, is leveraging AWS Inferentia2 chips to deploy vision-language models at a fraction of the cost of traditional GPU-based inference. By migrating to Amazon EC2 Inf2 instances, the company has found a cost-effective path to running sophisticated AI models that detect and interpret pet behavior in real time — a move that signals a broader shift in how startups approach AI infrastructure spending.

The deployment addresses one of the most pressing challenges facing AI-driven startups today: how to maintain model accuracy while dramatically reducing the compute bill. For Tomofun, which processes millions of video frames daily from pet cameras worldwide, even marginal cost savings per inference call translate into significant annual savings.

Key Takeaways at a Glance

• Tomofun deploys vision-language models on AWS Inferentia2 for real-time pet behavior detection
• EC2 Inf2 instances offer purpose-built AI acceleration at lower cost than traditional GPU instances
• The Furbo Pet Camera uses AI to identify barking, movement, and other pet activities remotely
• AWS Neuron SDK enables model compilation and optimization for Inferentia2 hardware
• Cost reduction is achieved without sacrificing inference accuracy or latency requirements
• The approach provides a replicable blueprint for other startups running vision-language workloads

Why Vision-Language Models Are Critical for Pet Tech

Vision-language models (VLMs) represent a significant leap beyond traditional computer vision. Unlike conventional image classifiers that simply label objects in a frame, VLMs combine visual understanding with natural language reasoning. This allows them to generate descriptive, context-aware outputs — such as telling a pet owner that 'your dog appears anxious and is pacing near the door' rather than simply flagging 'motion detected.'

For Tomofun, this capability is central to the Furbo Pet Camera's value proposition. Pet owners expect more than basic alerts. They want intelligent insights about their pet's emotional state, activity patterns, and potential health concerns. VLMs make this possible by interpreting visual data through a linguistic lens.

However, running these models at scale is expensive. VLMs are significantly larger and more compute-intensive than standard convolutional neural networks. A typical VLM can require billions of parameters, making inference on general-purpose GPUs like NVIDIA A100s or H100s prohibitively expensive for a consumer hardware startup operating on thin margins.

How AWS Inferentia2 Slashes Inference Costs

AWS Inferentia2 is Amazon's second-generation purpose-built chip designed specifically for machine learning inference workloads. Unlike general-purpose GPUs that handle training, gaming, and scientific computing, Inferentia2 is optimized exclusively for running trained models in production — which means it strips away unnecessary overhead and delivers better price-performance for inference tasks.

The key advantages of Inf2 instances for workloads like Tomofun's include:

• Up to 4x higher throughput compared to first-generation Inferentia chips
• Up to 10x lower latency for large model inference compared to comparable GPU instances
• Cost savings of up to 50% versus GPU-based alternatives for inference workloads
• 384 GB of accelerator memory across each Inf2 instance, supporting large VLMs
• Native support through the AWS Neuron SDK for frameworks like PyTorch and TensorFlow
• NeuronLink interconnect technology enabling efficient model sharding across chips

For Tomofun, this translates to the ability to serve millions of inference requests from Furbo cameras globally without the GPU price tag that would erode margins on a consumer product priced under $200.

The Technical Pipeline: From Model to Production

Deploying a vision-language model on Inferentia2 is not a simple lift-and-shift operation. The process involves several critical steps that Tomofun's engineering team navigated to bring their pet behavior detection system into production.

First, the pre-trained VLM must be compiled using the AWS Neuron SDK. This SDK converts standard PyTorch or TensorFlow models into optimized Neuron-compatible formats that can run efficiently on Inferentia2 hardware. The compilation step analyzes the model's computation graph and maps operations onto the chip's architecture for maximum throughput.

Next, the compiled model is deployed on EC2 Inf2 instances, which come in several configurations ranging from the inf2.xlarge (with 1 Inferentia2 chip) to the inf2.48xlarge (with 12 chips). Tomofun can scale horizontally by adding more instances or vertically by selecting larger instance types depending on traffic patterns — such as evening hours when pet owners are most likely to check on their animals.

Model Optimization Strategies

Beyond basic compilation, several optimization techniques help squeeze additional performance from Inferentia2 hardware:

• Dynamic batching groups multiple inference requests together to maximize chip utilization
• Model quantization reduces parameter precision from FP32 to BF16 or INT8 without meaningful accuracy loss
• Tensor parallelism distributes model layers across multiple Neuron cores for faster processing
• Caching mechanisms store frequently accessed model components in on-chip memory

These optimizations are particularly important for VLMs, which process both image tokens and text tokens simultaneously and require careful memory management to avoid bottlenecks.

Industry Context: The Rise of Purpose-Built AI Silicon

Tomofun's migration to Inferentia2 reflects a broader industry trend away from general-purpose GPU dependence. As AI inference costs become the dominant expense for production ML systems — often exceeding training costs over a model's lifetime — companies are increasingly seeking specialized hardware alternatives.

NVIDIA still dominates the AI chip market with an estimated 80% market share in data center GPUs. However, purpose-built inference chips from AWS, Google (with its TPUs), and startups like Groq and Cerebras are carving out significant niches. AWS has been particularly aggressive, offering Inferentia2 at price points that undercut GPU instances by 40-60% for equivalent inference workloads.

This matters especially for startups like Tomofun that operate in the consumer electronics space. Unlike enterprise SaaS companies that can pass infrastructure costs to customers through subscription pricing, consumer hardware companies must absorb cloud compute costs within fixed product margins. Every dollar saved on inference directly improves unit economics.

Compared to running the same VLM workload on EC2 P4d instances powered by NVIDIA A100 GPUs, Inf2 instances can deliver comparable throughput at roughly half the cost — a difference that compounds dramatically at Tomofun's scale of millions of daily inference calls.

What This Means for Developers and AI Startups

Tomofun's deployment offers a practical blueprint for any team running vision-language models in production. The key lesson is clear: not every AI workload requires the most powerful (and expensive) GPU available.

For developers evaluating their own infrastructure options, the decision framework comes down to several factors. If the workload is inference-only (no training), if latency requirements are in the tens-of-milliseconds range rather than single-digit milliseconds, and if cost optimization is a priority, then purpose-built inference chips like Inferentia2 deserve serious evaluation.

The AWS Neuron SDK has matured significantly since its initial release, now supporting popular model architectures including transformers, diffusion models, and multimodal VLMs. This reduces the engineering friction that previously made migration to custom silicon a daunting proposition.

Startups in adjacent verticals — home security, elder care, retail analytics, and agricultural monitoring — can apply the same approach to their own computer vision and VLM workloads. The pattern of compiling a model with Neuron, deploying on Inf2, and optimizing with quantization and batching is transferable across domains.

Looking Ahead: Custom Silicon Reshapes AI Economics

Tomofun's successful deployment on Inferentia2 is likely just the beginning. AWS continues to invest heavily in its custom silicon roadmap, with Trainium2 chips expected to further blur the line between training and inference optimization. As these chips mature, the cost advantage over general-purpose GPUs will likely widen.

The pet-tech market itself is projected to exceed $30 billion globally by 2027, with AI-powered devices driving much of that growth. Companies that solve the inference cost equation early — as Tomofun has done — will hold a significant competitive advantage as they scale.

For the broader AI industry, this case study reinforces a critical point: the future of AI deployment is not just about building better models. It is equally about deploying existing models more efficiently. Purpose-built silicon, smart compilation pipelines, and disciplined optimization practices will separate the startups that scale profitably from those that burn through cloud budgets chasing accuracy they could achieve at half the cost.

Tomofun's journey from GPU-based inference to Inferentia2 is a reminder that in production AI, the smartest engineering often happens after the model is trained.