Every time you ask an AI a question, a computer somewhere does the work. That work is called AI inference.
Most people have never heard the term. It is the reason data centers are going up across South Florida.
What AI Inference Actually Is
Building an AI model has two phases. Training teaches the model from huge piles of data. Inference is what happens afterward.
Inference is the model making decisions in real use. You type a prompt, the model computes an answer, and the answer comes back.
Training is a one-time, very expensive project. Inference happens every time anyone uses the product.
A simple comparison helps. Training is a chef spending years learning a cuisine. Inference is that chef plating a dinner order.
That repetition is why it matters. McKinsey expects inference to make up about half of all AI compute by 2030, according to MAIN’s Equinix coverage.
Everyday Miami examples are easy to find. Every call that Miami-based Heytruffle answers for a restaurant is inference. So is every document an AI reads for a lender.
Each answer uses chips, power and cooling. Cloud providers bill by the amount of text the model reads and writes, measured in tokens.
Gartner calls the result the inference paradox. Token prices are falling, but the cost of running AI agents is not.
Agents use far more tokens per task than a simple chatbot. MAIN covered the research in AI tokens are getting cheaper, running AI agents isn’t.
Gartner projects token costs to fall 95% by 2030. It also expects agent workflow costs to rise more than fivefold within two years.
For a business, that means the bill depends on how much the AI works, not what one answer costs.
Who Runs Inference in Miami
Inference rewards being close to users. A short trip for the data means a faster answer. Miami’s location suits that.
Equinix runs the NAP of the Americas downtown. The facility is a primary gateway for internet traffic between the US and Latin America.
More than half of its roughly 130 network carriers are based in Latin America and the Caribbean. That is the kind of reach inference customers want.
Equinix has struck a deal with Nvidia and Together AI called the Equinix Inference Exchange. It is due to launch in early 2027.
“The companies you already use are all running on us,” said Maryam Zand, an Equinix vice president.
Nvidia chief executive Jensen Huang said distributed systems let you be “both simultaneously close and far away.”
Metrobloks is building MIA-A1, a data center near Sweetwater designed for high-density AI inference. It is on track to open this year.
MAIN’s Metrobloks report covers the site. Miami is not building giant training campuses. Local projects target nearby businesses that need fast, local compute.
Miami Beach-based SOVAIR.AI takes a different route. It runs inference on-premise in air-gapped environments, with no cloud connection at all.
Its customers are utilities and other regulated organizations that cannot send sensitive data to a cloud AI service.
Brickell-based TrustedRouter sits one layer higher. It sends each request to whichever model fits the cost, speed and security need.
What Buyers Should Check
Inference is a service, not a building. A data center announcement does not mean usable capacity for your workload.
Percepture’s guide to deploying AI inference in Florida offers a useful checklist. Our summary of its main points:
Define the workload first. A customer chatbot and an overnight batch job need different tests.
Verify the carriers physically present in the building and who may order cross-connects. Treat a storm or connectivity failure as a test, not an assumption.
Measure from your own customers’ networks, not from a machine inside the facility. Separate resources already operating from those only planned.
The advice is generic, and Percepture is a marketing agency. Its checklist is a useful framework, not a Miami benchmark.
The short version for readers is simple. Inference is where AI meets customers. In Miami, a handful of operators are building the places where it happens.
Sources:
- MAIN: Equinix’s AI Boom Runs Through a Miami Landmark
- MAIN: Metrobloks’ MIA-A1 Is Quietly Becoming Miami’s Next AI Data Center
- MAIN: AI Tokens Are Getting Cheaper. Running AI Agents Isn’t
- MAIN: SOVAIR.AI, Miami Beach Air-Gapped AI for Energy Operators
- Percepture: AI Inference in Florida: Where to Deploy and What to Verify