TERAFAB provides free AI infrastructure tools and calculators - training cost, GPU clusters, semiconductor fab, datacentre power, RAG vs fine-tune vs train, plus a token counter and AI glossary. No sign-up required.
Yes. All tools and calculators are free. We use published benchmarks and scaling laws; no paywall or sign-up.
We use industry benchmarks and first-principles models (e.g. Chinchilla scaling, GPU throughput). Results are planning estimates-actual costs vary with vendor, region, and utilisation.
Training costs for frontier models can exceed $100M+, driven by massive GPU usage, data scale, and engineering complexity. Use our AI Training Cost calculator to estimate.
It's increasingly power, not chips. Datacentres require enormous electricity, and energy availability is becoming the main constraint on AI growth.
Beyond hardware, costs include cooling, networking, storage, and low utilisation. A cluster's true cost of ownership is often far higher than just the GPUs.
It depends on scale. APIs are cheaper at low usage, but self-hosting becomes more cost-effective at high volume, especially for inference-heavy applications.
RAG retrieves external data at query time to augment answers. Fine-tuning adapts a pre-trained model on new data. Full training builds a model from scratch. Each has different cost, latency, and complexity-use our RAG vs Fine-tune vs Train calculator to compare.
A scaling law from DeepMind research: optimal training uses roughly 20 tokens per parameter. Training longer often beats scaling up model size. Our AI Training Cost calculator applies Chinchilla when you enable the option.
A unit of text that AI models process. Roughly 4 characters or 0.75 words in English. Input and output are both counted in tokens-API pricing is typically per token. Use our Token Counter to estimate.
Training is the expensive process of learning weights from data. Inference is running the trained model to generate outputs-much cheaper per token.
Some researchers believe high-quality training data could become a bottleneck, pushing models toward synthetic data. Others argue data is still abundant. Useful, high-quality data is getting harder to find.
Power Usage Effectiveness: the ratio of total facility power to IT power. A PUE of 1.2 means 20% overhead for cooling, lighting, and other systems. Lower PUE is better for cost and carbon. Our Datacentre Power Budget calculator includes PUE.
Use our AI Training Cost calculator. Enter model parameters (e.g. 70B), training tokens, GPU type, cluster size, and facility markup. It applies Chinchilla scaling and industry benchmarks to estimate total cost and training time.
RAG is best when you need up-to-date or domain-specific data at query time. Fine-tuning suits adapting a model to a specific style or task. Full training is for building a new model from scratch. Use our RAG vs Fine-tune vs Train calculator to compare costs and trade-offs.
It estimates CO₂ emissions from AI model training based on model size, training tokens, region (grid carbon intensity), and energy mix. You can optionally include embodied carbon from GPU manufacturing. Useful for sustainability reporting and comparing regions.
Use our Token Counter. Paste your text and select the token model (GPT/Claude, code-heavy, or dense). It estimates tokens, characters, and words. Roughly 4 characters or 0.75 words per token for English. Different models tokenize slightly differently.
Given your available megawatts at the grid connection, it calculates the maximum GPU capacity you can support. Power is often the binding constraint for AI scale. It factors in PUE, cooling type, GPU TDP, and reserve headroom. Useful for planning new builds or retrofits.
Contact Us
Questions or feedback? Get in touch via our contact form.
FAQ
Visit our FAQ Center for more information and commonly asked questions.