Small devices can run useful AI locally, with less delay and more privacy. That can replace some server requests. It does not remove large-model training, demanding shared services, or the need to examine a proposed facility’s long-term financial risks.
Training creates a model; inference uses it. Moving selected inference to a device changes where that work happens.
Conceptual lifecycle, not a measured traffic or energy split. Sources: Google Gemma 3 technical report and Gemma 3n model card, listed below. The diagram does not imply every local request contacts a server.
Data centers support training, large-model inference, storage and shared services. Edge computing can reduce selected cloud workloads, while growth in users and applications can increase others. Historical efficiency gains have allowed much more computing without proportional energy growth; that history does not settle future demand or justify every proposed campus.
Peosta’s August 2026 concept describes 500-850 MW of on-site natural-gas generation, not a verified computing load. It leaves final equipment and water figures to later operator requirements. A small-device demonstration cannot establish this campus’s viability, and national AI growth cannot establish it either. Ask for the committed operator, phased demand and enforceable exit obligations. See the city-posted concept in the sources below.
A pocket-sized device does roughly 70 trillion operations a second, so it can replace a data center.
No. Running a small model does not give an 8 GB device the capacity of a data center. TOPS counts elementary arithmetic, not complete answers or model capability. Memory capacity, the speed of moving data and connections between processors are decisive. The large model in the fact card cannot fit in that device’s memory. A collection of small devices also lacks a rack’s tightly connected GPU system.
If both models answer questions, they do the same job.
AI is a category, not one interchangeable program. A compact model may recognize speech or summarize a short message; a more capable model may handle harder reasoning, longer documents or complex coding. Parameters are learned numerical settings: 8B means eight billion. More parameters need more storage, but size alone does not determine quality. Training, model design and the time spent answering also matter. Compare success on the same task, not whether both can display a chatbot reply.
An open-weight model should be easy to run at home.
Open weights means the learned model data is available under its license; it does not mean the model is small or cheap to run. Kimi K3 uses a mixture of experts: only part of the model works on each step, but the full weights still need storage. If the needed weights do not fit in fast memory, fetching them from slower storage can add delay. Running large models at home is possible with sufficient hardware, but it is a different proposition from using an 8 GB device. Closed weights describes access, not a guaranteed size or quality advantage; undisclosed sizes cannot support a memory comparison.
Gemma runs on a device, so AI no longer needs data centers.
Inference means using an already-trained model. Some local tasks can work offline, without a cloud request. Google’s Gemma 3n was trained on TPU systems; Gemma 3 documents distillation, where larger models help teach smaller ones. New model versions still require development and distribution. Local recognition and quick responses complement larger models and shared services serving many users. Not every device task needs the cloud.
The small device uses no water, so distributing AI removes its energy and cooling footprint.
A low-power device can release heat to room air. A rack concentrates much more heat in one place. Distributing work changes where heat is released; it does not erase electricity use. Total energy can rise or fall with hardware, utilization and workload. Liquid cooling does not necessarily consume water through evaporation. The final heat-rejection design and electricity supply determine water impacts.
It only costs about $2 to run.
For illustration, a constant 25 W for 30 days uses 18 kWh. At an assumed $0.15/kWh, that is $2.70. This is arithmetic, not a measured Jetson bill or a Peosta tariff; complete-system power can differ. It excludes purchase cost and model development. A low operating bill for one small workload says little about the cost of serving many users or training large models.
The hardware will be outdated in two to four years, leaving an empty building.
Servers are replaced or repurposed as needs change. Meta’s accounts distinguish roughly five-year server lives from decades for buildings. Buildings, power distribution, cooling and fiber connections can support successive equipment generations, although upgrades may be costly and suitability is not guaranteed. Equinix reports facility impairment and restoration obligations. Neither a chip refresh nor a long depreciation schedule proves whether a particular site will retain a tenant.
More efficient computing means total demand must fall.
In the global 2010-2018 study, compute instances grew 550% while electricity rose 6%: efficiency substantially restrained energy growth. Separately, LBNL estimates U.S. data-center electricity grew from 58 TWh in 2014 to 176 TWh in 2023. Cheaper or more efficient tasks can enable more use, but these trends do not prove that lower cost caused growth, or that future AI demand must rise.
Residents should ask who pays if the tenant leaves.
A specialized facility can lose a tenant or require expensive conversion even if data centers remain useful overall. A funded decommissioning or reclamation agreement is a reasonable local request. East Manchester Township, Pennsylvania, requires a plan and financial security at 110% of estimated removal cost, excluding salvage offsets. That is an example, not Iowa law. Peosta should assess enforceability, responsible parties and costs for both campus and power plant.
Without these inputs, a project-specific verdict is incomplete. Treat missing evidence as an open question.
One Jetson Orin Nano Super: memory, advertised compute and module power modes
Manufacturer specifications, not independent application benchmarks. Module power excludes the complete system and power-supply losses.Sparse and dense INT8 throughput for Jetson Orin Nano Super
Peak throughput depends on precision and sparsity; TOPS does not measure model quality or finished requests per second.One liquid-cooled rack: 72 GPUs, GPU memory and NVLink interconnect
Manufacturer system specifications. Aggregated GPU memory is distributed across connected GPUs; it is not one device memory chip.Approximately 120 kW at full load including rack nodes and components
Rack power is not total facility power and is not directly comparable to a Jetson module power mode.Kimi K3: 2.8 trillion total parameters; 104 billion activated per token
Published model architecture, not independent performance verification. Weight storage is a lower-bound calculation, not a complete hardware budget.Small models intended for low-resource devices; training hardware and limitations
On-device inference does not imply on-device original training. Model capabilities and deployment requirements vary.Gemma 3 training, knowledge distillation and model family
A disclosed example of making smaller models, not proof that every edge model uses the same training process.Heat removal, cooling loops and water efficiency in data centers
A liquid cooling loop does not by itself establish evaporative water consumption or a particular campus water budget.Global modeled data-center compute instances and electricity, 2010-2018
Historical modeled estimates, not a controlled test of efficiency-induced demand or a forecast for AI.Company accounting useful lives for servers, network assets and buildings
Depreciation assumptions are neither mandatory replacement schedules nor guarantees of physical or economic life.One operator portfolio; facility impairment, asset lives and restoration obligations
Company-specific evidence that sites can underperform. It does not estimate Peosta vacancy risk or guarantee reuse.Local decommissioning plans and financial security for data centers
A Pennsylvania example, not Iowa law or evidence that Peosta has adopted these protections.August 2026 preliminary campus and on-site generation concept posted by the city
Developer discussion estimates, not approved capacity, metered demand, an executed tenant lease or an independently verified operating plan.United States; historical estimates through 2023 and scenarios through 2028
Modeled national estimates with limited facility-level public data. Location-based emissions and indirect water do not include facility-specific contracts or behind-the-meter supply.