B300 compute / Proposed deployment
Our proposed initial fleet brings 64 B300 GPUs together in 8 complete Supermicro systems at Mead, Oklahoma. The campus concept places direct-liquid-cooled systems in Buildings A and C. Final equipment, installed costs and engineering acceptance determine procurement and service availability.
64 B300 GPUs in 8 complete Supermicro systems. The A/C campus split is illustrative; procurement and commissioning are ahead.
Shared inference uses nodes serving multiple customers; dedicated GPU servers are intended for a single tenant. Isolation, scheduling and performance require validation.
The direct-liquid campus design requires exact OEM SKU confirmation, installed cooling scope and acceptance tests. Those costs are not included in the air-cooled, one-building baseline in model 14 September management target · v6.1.1 cost basis.
AI hardware concept · Not installed inventory or an OEM configuration drawing.
Choose a topic, or open the 3D model.
Illustrative hall from the homepage. The current proposal is 64 B300 GPUs in 8 Supermicro systems; the scene is not an OEM layout or installed inventory.
8 complete Supermicro systems
Each proposed system contains 8 B300 GPUs. Management describes complete systems with CPUs, NVSwitch fabric, networking, storage, software and support. Exact configuration, included cluster infrastructure and support terms remain subject to supplier verification.
External networking, storage, installation and ongoing service obligations must be reconciled with the full deployment requirements.
Proposed allocation
The four-GPU reserve is a financial allocation within eight complete systems. It is not a complete standby server or a validated failover design.
Scheduling, maintenance and node-failure recovery need validation before we offer a service level. These GPUs are not currently bookable inventory.
Commercial stage. Planned SmartTec-owned and operated systems; not a purchased or commissioned fleet. Management reports signed three-year marketplace access; no signed customer contracts or guaranteed minimum receipts have been established. Payment is only for rented GPU-hours. Customer availability follows procurement, workload validation and facility and network acceptance.
Planned service
Planned inference on SmartTec-owned GPU nodes shared by multiple customers. Scheduling, tenant controls, supported models and performance targets will be validated before service.
Planned service
Planned SmartTec-owned GPU servers for one tenant per node. The proposal will define isolation, GPU allocation, software access and support.
Planned service
Planned space, power, cooling and connectivity for servers you own, subject to a separate equipment and facility allocation review.
SmartTec-owned hosting has two service modes: shared inference on nodes serving multiple customers, and dedicated GPU servers intended for one tenant per node. Both use the same proposed fleet. The 60-GPU financial allocation does not itself establish how complete eight-GPU servers will be divided between them; deployment design and customer terms must settle that allocation before service.
Tell us the application, model size, memory needs, storage throughput and intended software stack. We assess the required GPU allocation, networking, security, support and service targets against a representative workload.
Inference, rendering and suitable fine-tuning require different software and performance tests. A GPU count alone does not establish throughput, latency or a customer outcome.
The workload planner helps you prepare the model, tenancy, storage and service requirements for a deployment review. Hardware allocation and performance follow validation of that brief.
Published model 14 September management target · v6.1.1 cost basis retains an air-cooled, one-building financial baseline. The approved campus experience illustrates a proposed direct-liquid design across Buildings A and C; it requires updated equipment quotes, engineering and repricing before it is included financially. Financial scenarios remain in the protected investor area.
The campus concept uses direct liquid cooling for B300 systems in Buildings A and C, with a proposed chiller yard behind B. The owner specifies a closed-loop Daikin plant at 208V; the exact model and compatible operating conditions remain to be confirmed.
A coolant distribution unit (CDU) exchanges heat between separate facility-water and server-coolant circuits. Overhead supply and return pipes feed rack manifolds and server cold plates. Residual room cooling, maintenance access and local-weather performance remain part of the complete design.
Owner-reported Dobson handoff in Building B. The external provider route and internal campus fiber paths remain unverified. The reported handoff location does not establish an accepted or live service.
The campus electrical commissioning milestone is an 8 October 2026 target. Planned Dobson Telephone Company fiber has an 8 October–5 November 2026 target window. These management targets are distinct from completed customer service acceptance.
Explore the latest campus concept, then trace its cooling, electrical and fiber connections. Equipment placement and routing illustrate the proposal; they do not establish installed capacity.
We review platform fit, capacity, timing and support needs, then develop a proposal with the configuration and commercial terms. You can also discuss colocation for hardware you already own.
