Hardest Part

Translating volatile application metrics into precise hardware constraints with zero tolerance for under-provisioning to prevent catastrophic downtime.

Min Viable Scope

Build a read-only recommendation engine exclusively for stateless Kubernetes workloads on AWS. Deliberately leave out stateful database capacity planning, automated infrastructure deployment, and multi-cloud support.

Cold Start Problem

The model lacks baseline utilization and traffic spike data to make accurate predictions for new architectures. Break this by offering a backward-looking historical analyzer that ingests the past 90 days of Datadog and AWS metrics from seed design partners.

Time To First Value

Same-day via read-only historical API pulls from existing cloud monitoring tools.

Data Moat Available

true

Technical Difficulty

High

Wedge

The beachhead targets mid-tier contract manufacturers in the consumer electronics sector. These firms face extreme seasonal demand spikes and acute component shortages, making their capacity pain highly visible and financially punishing. Once the system owns this high-volatility niche by proving throughput increases, it expands laterally into automotive parts and industrial equipment manufacturing where production cycles are longer but constraints are structurally identical.

Timing

Large language models and multimodal agents now reliably parse unstructured data like PDF supplier schedules, email updates, and irregular ERP exports without brittle robotic process automation setups. This eliminates the data-entry bottleneck that previously kept capacity planning trapped in manual spreadsheets.

Why This ICP

Mid-market manufacturers experience the exact same complex demand-supply volatility as enterprise firms but lack the budget to implement multi-million dollar monolithic supply chain deployments. They are highly motivated buyers seeking enterprise-grade optimization at a fraction of the cost and implementation time.

Size Of Prize

There are ~30,000 mid-market manufacturing and logistics firms in the US. Multiplying this entity count by an average annual spend of ~$40,000 on dedicated capacity planning labor and legacy software licenses yields a total addressable prize of roughly $1.2B.

Gap Narrative

Mid-market manufacturers and logistics providers rely on siloed, static spreadsheets to balance demand forecasts against machine hours, labor shifts, and material availability. This manual reconciliation process causes stockouts during demand spikes and idled floors during lulls, costing millions in missed revenue and wasted labor. They require a system that continuously ingests these variable inputs to dynamically generate actionable production schedules and identify capacity constraints in real time.

Defensibility

The product builds defensibility through deep workflow lock-in and historical data compounding. As the agent ingests months of planned versus actual production data, it builds a proprietary, localized model of the factory's specific throughput quirks, machine downtime patterns, and seasonal yield variations. Switching to a competitor requires abandoning this highly tuned operational model and starting over with generic baseline assumptions.

Why This Thesis

An agentic software approach fits perfectly because capacity planning is a continuous constraint-satisfaction problem fed by messy, asynchronous inputs. Agents handle the unstructured data ingestion from supplier emails and PDFs while algorithmic solvers handle the optimization math, directly replacing the manual labor of a human planner copying data between systems.

SAM

~$100M-200M Tier 2 cloud providers and specialized GPU hosting services

SOM

~$10M-25M

TAM

~3,000 global cloud and colocation providers × ~$150k-200k/yr ≈ ~$450M-600M

Growth Rate

~25-35%/yr, driven by rapid GPU cluster deployments and the high capital cost of stranded AI hardware

Paid Comparable Spend

~$100k-300k/yr spent on legacy DCIM software licenses and internal data engineering labor for custom spreadsheet models

Kill Thresholds

• Time-to-first-cluster-ingestion > 14 days
• D30 user retention < 40 percent
• Recommendation acceptance rate < 50 percent
• POC conversion rate to paid < 20 percent
• Identified stranded capacity < 5 percent per customer

Leading Metrics

• Time-to-first-cluster-ingestion
• Percentage of stranded GPU capacity identified
• Weekly active planner logins per account
• Recommendation acceptance rate vs manual overrides
• Time spent editing capacity constraints in-app

What Proves Right

Tier 2 cloud providers and GPU hosting services integrate their Kubernetes schedulers and hypervisors with the software within the first week of onboarding. Cohorts retain at over 85 percent after three months by reclaiming at least 15 percent of their stranded GPU capacity. These infrastructure operators accept a $150,000 annual price point because the system directly offsets the capital expense of purchasing redundant hardware.

What Proves Wrong

Target customers refuse to grant read-only access to their cluster metrics due to strict data compliance policies. Infrastructure operators generate one-off static reports during onboarding but immediately revert to manual spreadsheet models for actual deployment decisions. The software consistently produces allocation recommendations that engineers override because the logic fails to account for physical data center power and networking constraints.