Decision Lab | Metachips.ai

Economics engine

Fleet TCO and cost per million tokens

Combines accelerator acquisition, host and network cost, facility overhead, energy, utilization, useful life and model throughput.

AcceleratorsAccelerator price ($)Non-accelerator system cost / accelerator ($)Network + storage / accelerator ($)Average accelerator power (W)PUEElectricity ($/kWh)Utilization (%)Useful life (years)Output tokens/sec/acceleratorAnnual software + operations (% CapEx)Residual value (%)

Total CapEx–

Annual energy–

Lifetime TCO–

Cost / 1M output tokens–

Lifetime cost composition

Workload model

Inference performance and SLA envelope

Estimate concurrency, throughput, time-to-first-token and decode latency under simplified prefill/decode assumptions.

AcceleratorsReplicas / acceleratorRequests per secondInput tokens / requestOutput tokens / requestPrefill throughput / accelerator (tok/s)Decode throughput / accelerator (tok/s)Batching efficiency (%)Scheduler overhead (%)SLA TTFT target (ms)

Demand tokens/sec–

Effective capacity–

Estimated TTFT–

Estimated utilization–

–

Queue–

→

Prefill–

→

Decode–

Facility model

Rack power, campus demand and cooling load

Translate accelerator density into rack-level IT power, facility demand, heat rejection and annual electricity use.

Accelerators / rackAccelerator TDP (W)Non-accelerator overhead (%)Number of racksAverage load factor (%)PUELiquid cooling share (%)Cooling delta-T (°C)Electricity ($/kWh)

Rack IT power–

Campus facility demand–

Heat rejected–

Annual electricity–

Utility–

→

Facility–

→

IT–

→

Accelerators–

Approximate facility load breakdown

AI fabric designer

Network bandwidth and switch-count model

Size a simplified leaf-spine fabric using endpoint speed, oversubscription, switch radix and deployment scale.

AcceleratorsEndpoints / acceleratorEndpoint speed (Gb/s)Leaf switch portsLeaf downlink share (%)Spine switch portsOversubscription ratioPort reserve (%)Optical share of links (%)Transceiver ASP ($)

Leaf switches–

Spine switches–

Fabric bisection–

Optics value–

Spines

Leaves

Endpoints

Supplier economics

Switch, optics and cable attachment-rate model

Translate accelerator shipments into attached infrastructure units and market value using editable topology and pricing assumptions.

Annual accelerator shipmentsAccelerators / compute nodeCompute nodes / rackTop-of-rack switches / rackSpine switches / 32 racksScale-up switches / rackOptical modules / acceleratorHigh-speed cables / acceleratorAverage switch ASP ($)Average optical module ASP ($)Average cable ASP ($)Vendor share scenario (%)

Racks enabled–

Total switches attached–

Optical modules attached–

Addressable attached value–

Attached value by category

Memory system

Model weights, HBM fit and KV-cache capacity

Estimate memory needed for model weights and concurrent KV cache under common transformer assumptions.

Model parameters (billions)Weight precision (bits)HBM per accelerator (GB)Accelerators per replicaLayersKV headsHead dimensionKV precision (bytes)Context lengthMemory reserve (%)

Weight memory–

Usable HBM / replica–

KV / sequence–

Concurrent sequences–

Weights

KV pool

Reserve

Fleet planner

Capacity, demand growth and procurement timing

Estimate accelerators, racks and megawatts required under demand growth, utilization target and deployment buffer.

Current peak requests/secAnnual demand growth (%)Planning horizon (years)Tokens / requestUseful tokens/sec/acceleratorTarget utilization (%)Capacity buffer (%)Accelerators / rackRack IT power (kW)PUE

Future peak RPS–

Accelerators required–

Racks required–

Facility capacity–

Demand and capacity ramp

Scenario studio

Five-year infrastructure projection

Build a transparent scenario for accelerator deployments, rack density, power demand and attached network value.

Starting accelerators deployedAnnual deployment growth (%)Starting accelerators / rackAnnual density improvement (%)Starting rack power (kW)Annual rack power increase (%)Network value / accelerator ($)Annual network value change (%)Power price ($/kWh)PUE

Year-5 accelerators–

Year-5 racks–

Year-5 facility GW–

5-year network value–

Five-year system trajectory

This projection is generated only from the assumptions above and is intentionally not presented as a market forecast.

Decision architecture

Every model should end with a question.

Architecture

What topology, memory hierarchy, power chain or cooling loop fits the workload?

Economics

Which variable dominates lifetime cost, and what must be true for the investment to work?

Timing

Which transition is real now, which is emerging, and what needs qualification first?

Supplier strategy

Where does content attach, how sticky is the design win and what displaces it?

Custom modeling

Need a decision-grade model?

These public tools are intentionally transparent and directional. Metachips can build workload-specific architecture, cost, supplier and deployment models.

Start a modeling engagement