Related daily report
Firebird’s Armenia project makes the terminology worth understanding
The August 9 briefing covers Firebird’s newly opened Armenian AI facility and its much larger planned expansion. This explainer separates the infrastructure concept from the headline GPU numbers.
Open the August 9 reportThe direct answer
An AI factory is a computing environment optimized around AI workloads, but the phrase is not a universal engineering standard.
A general-purpose data center can already host GPUs, train models and serve AI applications. What people increasingly call an AI factory is usually a facility or computing environment designed much more aggressively around accelerated computing, high-speed interconnects, large data flows, dense power delivery, advanced cooling and software that keeps expensive AI hardware busy.
The term became popular because frontier-model training and large-scale inference no longer behave like ordinary office IT. Thousands of accelerators may have to work together as one distributed system. A weak network, insufficient cooling or poor workload scheduling can leave expensive chips waiting rather than computing.
A data center describes the place and computing infrastructure broadly. “AI factory” usually describes an AI-specialized operating model built to turn that infrastructure into sustained model training, fine-tuning or inference capacity.
Why “AI factory” does not have one fixed definition
NVIDIA defines an AI factory as specialized computing infrastructure spanning the AI lifecycle from data ingestion through training, fine-tuning and high-volume inference. Its commercial framing treats token throughput, cost per token, utilization and performance per watt as factory-like production measures.
The European Commission uses the same phrase more broadly. Its EuroHPC AI Factories combine supercomputing resources with data, researchers, universities, startups, industry and support services. In that policy context, an AI Factory is not merely a room full of accelerators; it is an ecosystem built around access to AI compute.
Those definitions overlap, but they are not identical. This is why the phrase should never be treated like a precise legal or engineering category. When a company or government announces an “AI factory,” readers still have to ask what equipment exists, what workloads it will run, who can access it, how much power is commissioned and which parts remain future plans.
The system underneath the label
GPUs are the visible part; the factory is the whole stack
Power
Accelerators, networking, storage and cooling require continuous electricity. Large clusters increasingly force operators to plan at utility scale.
Accelerated compute
GPUs and other AI accelerators perform the highly parallel numerical work behind training and inference.
Network fabric
High-bandwidth, low-latency links move model parameters, gradients, activations and data among many machines.
Storage and data pipelines
Training systems must feed enormous datasets quickly enough that compute does not sit idle waiting for data.
Cooling
Dense accelerator racks produce heat at levels that can require liquid cooling and carefully engineered heat rejection.
Software and orchestration
Schedulers, distributed-training software, inference systems, monitoring and failure recovery determine how efficiently the physical system is used.
Thinking in layers prevents a common reporting mistake: treating a GPU purchase as if it were already usable AI capacity. Hardware becomes productive only when the electrical, thermal, network, storage and software layers are ready at the same time.
Why thousands of fast GPUs can still form a slow system
Large models are usually too large, or too computationally demanding, to train efficiently on one accelerator. The work is divided across many devices. Those devices must repeatedly exchange information while the job runs.
This makes networking part of the computation rather than an accessory to it. If accelerators spend too long waiting for data from other accelerators, adding more GPUs can produce disappointing gains. At scale, engineers care not only about raw GPU performance but also about bandwidth, latency, congestion, topology and how communication is scheduled.
NVIDIA’s current AI-factory material emphasizes high-bandwidth interconnects and tightly integrated networking for this reason. The broader point does not depend on one vendor: distributed AI is a systems problem. The fastest component does not determine the speed of the whole cluster when another layer becomes the bottleneck.
Training and inference stress the infrastructure differently
Training
Training repeatedly processes huge datasets while coordinating many accelerators. The job can run for days or weeks and is sensitive to network performance, storage throughput and failures.
Fine-tuning
Fine-tuning usually starts from an existing model and adapts it. The jobs can be much smaller than frontier training but still benefit from accelerated infrastructure.
Online inference
Inference serves user or application requests. Latency, throughput, model loading, caching and cost per response become central.
Batch inference
Large offline workloads can process documents, video, scientific data or synthetic-data jobs where throughput may matter more than immediate response time.
That difference matters when evaluating claims about capacity. A cluster optimized for enormous training runs is not automatically the cheapest way to serve millions of small requests, and a highly efficient inference fleet may not be appropriate for training a frontier model from scratch.
Why AI projects are increasingly described in megawatts
GPU counts are intuitive but unstable. A newer accelerator can consume more power and deliver much more compute than an older one. Projects can also mix generations and configurations. Megawatts tell a different story: how much electrical capacity the site is designed to deliver to computing and supporting systems.
Power figures still require care. Planned capacity, energized capacity and actual operating load are not the same number. A project may announce 300 MW years before the grid connection, substations, cooling plant and equipment needed to use that power are complete.
This is why serious infrastructure reporting should record both compute and power while attaching a status to each figure: operating now, under construction, contracted, approved or simply proposed.
The expensive question is not how many GPUs you own, but how productively you use them
An accelerator generates no economic value while it waits for data, sits between jobs or is unavailable because another part of the system failed. That is why operators monitor utilization, uptime, throughput, energy efficiency and workload scheduling.
NVIDIA’s factory metaphor highlights tokens per second, tokens per watt and cost per token. Those are useful measures for large inference operations, but they are not the only possible output. An AI facility may also produce trained models, scientific simulations, embeddings, images, video or internal enterprise workloads.
The broader lesson is that installed hardware is an input. Productive compute is the output. A smaller well-utilized cluster can be economically stronger than a larger system that cannot secure enough customers, power or reliable operations.
Firebird shows why installed capacity and announced ambition must be separated
Firebird’s August 2026 opening in Armenia is a useful current example. The project has a real first phase, while much larger GPU and power figures describe expansion plans through 2027. NVIDIA and Armenian government sources also publish different end-2027 capacity figures.
That does not make the project unreal. It makes status labels essential. A responsible description distinguishes what has opened, what the first phase is planned to contain and what later phases are expected to add.
The same discipline should be applied to every “gigawatt AI factory” announcement. Treat construction schedules, equipment orders, financing, power availability and contracted customers as separate evidence rather than one giant headline number.
For the policy side of the same infrastructure story, see What Is Sovereign AI?
A reader’s checklist
Nine questions to ask when someone announces an AI factory
- 1What is operating today?
Separate installed equipment from the eventual plan.
- 2Which accelerators are involved?
Model generation and configuration matter more than a generic “GPU” count.
- 3How much power is actually energized?
Do not equate a planned grid connection with current load.
- 4What workloads will run there?
Training, inference and scientific computing have different requirements.
- 5Who are the customers?
Demand determines whether an ambitious facility can be utilized.
- 6How is the cluster networked?
Accelerator performance is limited if communication becomes the bottleneck.
- 7How is it cooled?
Dense systems require substantial thermal engineering.
- 8What is funded versus aspirational?
Approved investment and long-range ambition should not share the same certainty label.
- 9How will utilization be measured?
Installed capacity only becomes useful when workloads can use it reliably.
Sources