AdAI
Definition

Inference is what happens when a trained AI model gets used. You type a question into ChatGPT, the model generates an answer. That is inference. A churn-prediction model scores a customer record. That is inference. A camera classifies an image as cat or dog. That is inference. Inference is the production-phase work of AI: the part that runs every time the model is asked to do something useful, as opposed to training, which happens once.

Key Takeaways

  • Inference is using a trained model. Training is teaching it. Two different jobs, two different cost profiles.
  • Inference is the part of AI that costs ongoing money. For every dollar a company spends training a model, it can spend 15 to 20 dollars running inference on that model over its lifetime.
  • The cost per inference has dropped about 50-fold since late 2022. GPT-4-class capability now costs about $0.40 per million tokens, down from around $20 per million tokens at GPT-4's launch.
  • For SMBs, inference is most often invisible: it is the cost the AI vendor (OpenAI, Anthropic, Google, HubSpot AI, etc) absorbs and recovers through subscription or API pricing.
  • When building custom AI agents, choosing the right model size for the job is where most inference-cost overruns get fixed. A smaller cheaper model usually does most work fine.
50x
cost drop for GPT-4-class inference between late 2022 and early 2026: from about $20 per million tokens to about $0.40
Source: Epoch AI / GPUnex Inference Economics Report, 2026
~2/3
of all AI compute now goes to inference rather than training, up from about 1/3 in 2023
Source: GPUnex AI Inference Economics Report, 2026
15-20x
multiplier on inference cost over a model lifetime, relative to its one-time training cost
Source: GPUnex / IDC FutureScape 2026

In Simple Terms

Training an AI model is like teaching someone a skill. It takes a long time, costs a lot of money up front, and happens once. Inference is what the trained person does for a living afterwards: answering questions, making predictions, generating content. Every interaction with the model is an inference. Every output is the result of inference running.

For an SMB owner, inference is the part of AI you actually pay for over time. Training GPT-4 cost OpenAI an estimated $100 million-plus. You will never pay anything close to that. But every time someone in your business uses ChatGPT, an inference call runs, and the cost of those calls is what your subscription is covering.

Inference is also where most of the recent cost improvements have happened. The same task that cost $20 in late 2022 costs about $0.40 in early 2026, a roughly 50-fold drop in three years. That is why AI features have become economically reasonable to embed in everyday SMB tools (HubSpot AI, Microsoft Copilot, Notion AI, Intercom Fin) when they were not in 2022.

Why Inference Costs Have Fallen So Fast

Three drivers operating at once.

Better hardware. NVIDIA's H100 and Blackwell GPUs, Google's TPU v5 and Trillium, AMD's MI300 and MI350, and specialised inference chips like Groq's LPUs have each cut the cost-per-token of running a model. NVIDIA H100 cloud pricing alone has fallen from $8-10 per GPU-hour at peak to under $3 per hour by early 2026.

Better software. Model distillation produces smaller versions of large models that keep most of the capability at a fraction of the inference cost. Quantisation reduces the precision of model weights with little accuracy loss. Speculative decoding predicts the next few tokens before the model finishes the current one. Each technique trims compute per inference by a meaningful percentage; combined they compound.

Pure competition. OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, and a half-dozen others are all selling inference. DeepSeek's late-2024 entry at roughly 90% lower pricing than incumbents triggered another round of frontier-model price cuts. The dynamics look like the AWS-vs-Azure cloud wars of the early 2010s, just on a faster clock.

What This Means for SMBs Choosing AI Tools

For most SMB cases, you do not see inference costs directly. You see a monthly subscription or an API bill, and the vendor handles inference under the hood. The practical things to watch:

Model choice in custom builds. If you are commissioning a custom AI agent or using a tool that lets you pick the model (n8n, Make, custom apps via API), the difference between using GPT-4o ($2.50 per million input tokens, roughly) and GPT-4o-mini ($0.15) is 16x on cost. For most internal automation, the smaller model is fine. Pay the premium only where the task genuinely needs the frontier model's reasoning.

Customer-facing volume. An internal team using ChatGPT 50 times a day is a fixed-price problem. A customer-facing chatbot answering thousands of conversations a day is an inference-cost problem. Build the cost model before you launch a high-volume AI feature.

Latency vs capability. The more capable the model, the slower the inference. Real-time use cases (voice agents, instant chat) often need smaller models or specialised inference hardware. Background use cases (overnight email triage, weekly report generation) can use the slowest, most capable model without anyone noticing.

The trend is your friend. Whatever you build today will be cheaper to run tomorrow. The inference cost curve has consistently surprised even AI insiders on the downside. Designing AI features that assume costs will keep falling is a safer assumption than designing on today's prices.

Frequently Asked Questions

What is the difference between training and inference?
Training is teaching the model. Inference is using it. Training happens once (or every time the model is updated). Inference happens every time someone asks the model a question or runs new data through it. Training takes weeks and costs millions for the largest models. Inference is fast (milliseconds) and cheap per call, but adds up to far more total cost over a model's lifetime.
Why has inference become so much cheaper?
Three reasons. Better hardware: NVIDIA's H100 and Blackwell, plus Google TPUs, AMD MI series, and specialised chips like Groq's LPUs, all cut cost per token. Better algorithms: model distillation, speculative decoding, and quantisation make models run more efficiently without losing much accuracy. Pure competition: when OpenAI, Anthropic, Google, Meta, and now DeepSeek are all racing for the same customers, prices fall.
What does inference cost an SMB in practice?
For most SMB use cases, very little. A small business running ChatGPT Team ($25 per user per month) gets unlimited inference at a fixed cost. A custom AI agent using Claude or GPT-4-class models via API typically costs a few cents to a few dollars per day for a typical SMB volume. The high-cost cases are large-volume customer-facing applications: a chatbot handling 10,000 conversations a day can run into hundreds of dollars per month in inference.
Why do AI agents seem slow sometimes?
Inference latency. Generating a response token by token takes real time, especially for the larger, more capable models. Smaller models (GPT-4o-mini, Claude Haiku) are faster but less capable. Specialised inference hardware like Groq's LPUs can hit 300 tokens per second on smaller models, making real-time voice agents possible. The latency-versus-capability tradeoff is one of the main design decisions when building an AI agent.
What should an SMB watch for on inference costs?
Three things. The unit cost per call or per million tokens: this should be in your AI vendor's published pricing. The volume: how often the AI runs (per visitor, per ticket, per email). The model choice: a smaller cheaper model often does most jobs at a fraction of the cost of a frontier model. Most SMB inference bills get bloated because someone defaulted to GPT-4 or Claude Opus when GPT-4o-mini or Claude Haiku would have been fine.

Related Glossary Terms & Resources

Join 5,000+ SMB owners getting weekly AI agent insights

Subscribe Free