Revenue activation OS for AI factories to turn phased GPU deliveries into tenant-ready capacity without overselling launch dates.

1. GPU access is consolidating into multi-year campus-scale allocations rather than ad hoc spot buying.
2. Batam capacity is planned to land in phases across 2027 and 2028, so operators need to know exactly what can be sold and activated at each step.
3. The project's headline revenue claims already rely on committed customer deals, making commercialization and delivery execution a board-level issue before the site is fully live.
4. Smaller firms are explicitly part of the access thesis, which creates demand for standardized sub-cluster products rather than only whale-sized contracts.
5. Investors and operators are already treating GPU access as the product itself, not just the real estate wrapper around a data center.

Catalyst. Firmus' Batam project couples access to up to 170,000 Nvidia accelerators with claims of committed customer deals and an explicit goal of widening access for smaller firms, creating an urgent need to package phased supply into trustworthy smaller offerings.

The product becomes the system of record between commercial promises and physical cluster readiness. It ingests GPU delivery schedules, rack commission states, power-block energization, network sign-off, and tenant requirements from CRM, DCIM, cluster managers, and internal spreadsheets, then exposes only the capacity blocks that are truly sellable. Operators use it to create standard reservation SKUs, staged onboarding plans, entitlement maps, and acceptance checklists for each tenant instead of bespoke SRE projects. Over time it learns which dependencies most often break launch dates and becomes the default layer for capacity planning, tenant provisioning, and expansion launches across AI factories.

What's different. DCIM tools know which racks are energized, CRMs know which customers were promised capacity, and cloud schedulers know how to allocate live clusters, but none of them turns phased infrastructure build-out into contract-safe inventory. This company owns the narrow seam between commercial promise and infrastructure readiness, where early AI factory operators are most exposed. Its moat compounds through unique data on how delivery milestones, tenant requirements, and acceptance failures translate into revenue leakage or launch risk across multiple campuses.

Jobs to be done

flowchart LR
  Buyer[AI factory operator] --> Pain[Phased GPU supply and tenant promises drift apart]
  Pain --> Product[Capacity activation OS]
  Product --> Outcome[Sellable GPU blocks go live on time]

Executive takeaways

• The wedge is real because GPU supply is arriving in phased campus deals, while commercialization still happens through bespoke coordination across commercial, infrastructure, and provisioning teams.
• The near-term market is narrow but urgent: a small number of Southeast Asian AI-factory operators are making high-stakes revenue promises before every rack, power block, and cooling loop is production-ready.
• Competition is indirect rather than direct; the startup mainly fights DCIM suites, public-cloud reservation products, and turnkey AI clouds, none of which cleanly own sellable-capacity readiness for neutral operators.
• The biggest execution risk is data quality and category timing: operators must trust a new system of record before enough multi-tenant AI halls exist to create repeatable demand.

Market definition

Software that converts phased AI data-center build-out into contract-safe, tenant-ready GPU inventory by linking physical readiness, entitlement logic, and customer acceptance workflows.

Customer and buyer

The practical buyer is a COO, CCO, or head of capacity delivery at an independent AI-factory operator in Southeast Asia. Champions sit across commercial operations, data-center operations, and provisioning teams that must translate physical hall readiness into date-certain customer commitments.

Buying triggers

• Operators start signing committed offtake or anchor-tenant deals before all racks, power blocks, and network domains are fully ready, so revenue is exposed to readiness drift. [1][2][3][4]
• Regional AI-ready campuses are filling fast: AirTrunk says its Johor platform is almost fully contracted even as new phases are still being built. [24][25][15]
• Data-residency, efficiency, and sustainability rules make site choice and launch sequencing more consequential than generic cloud procurement. [7][10][11][12][13][14]

Willingness to pay

Budget exists inside major infrastructure programs, not innovation slush funds. Buyers are already committing billions to AI-ready campuses and cloud regions, and pre-leasing pressure means a tool that reduces oversell risk or accelerates billable go-live can justify enterprise-level pricing long before it becomes a line item called “capacity activation software.” [15][24][25][30][31][32][33]

Category dynamics

Growth signal 18.41% CAGR

Validation signals

• Firmus is already tying a 170,000-GPU deployment to large committed offtake expectations before the Batam campus is fully live.
• AirTrunk says its Johor AI-ready campuses are almost fully contracted even as it keeps adding new phases.
• AWS and Google are exposing reserveable GPU windows as products, confirming that guaranteed AI-capacity access has become a buyable workflow.
• AWS, Microsoft, and Google are continuing to invest heavily in Southeast Asian cloud and AI infrastructure, increasing the number of sophisticated infrastructure buyers in the region.

Regulatory & technical constraints

• Indonesia PDP compliance, cross-border data-transfer sequencing, and the still-forming DPA mean operators need auditable evidence around where customer data and operational artifacts live.
• Singapore-style approval thresholds on efficiency and sustainability reinforce that AI-capacity sales are coupled to facility performance, not just chip availability.
• High-density GPU halls now depend on liquid-cooling and tightly managed flow, pressure, and power envelopes; poor commissioning can throttle sellable capacity.

There is no obvious category king for this exact workflow. The incumbent set is fragmented: DCIM vendors model power and space, public clouds sell reserved GPU windows, and AI-native clouds offer turnkey alternatives. The gap is the commercial-readiness control point that tells an operator what fraction of a half-built hall is truly safe to sell, to whom, and for which date.

Why incumbents do not win by default

• Cloud platforms. AWS and Google already productize reserved GPU windows, but they do so inside their own cloud footprints rather than helping neutral operators commercialize partially commissioned halls.
• DCIM suites. Schneider and Sunbird can model capacity, reservations, and infrastructure telemetry, yet they stop short of tenant-safe entitlement logic, commercial packaging, and acceptance evidence by customer.
• AI-native clouds. CoreWeave and Crusoe reduce the need for neutral operators by selling turnkey AI capacity directly, but that is a substitute business model, not an operator control plane.
• Regional operator stacks. Operators such as AirTrunk, DCI, and Nxera can build internal workflows, but current public positioning centers on capacity, cooling, and resilience rather than a reusable commercial-readiness product.
• In-house tooling. The default alternative remains spreadsheets, tickets, DCIM, and scripts because the workflow spans multiple systems of record and there is not yet a universally adopted product layer.

Independent AI factory operators in Southeast Asia are signing committed anchor-tenant capacity deals before every rack, power block, and cooling loop is production-ready, creating simultaneous risks of overselling unavailable inventory and leaving sellable GPU blocks stranded for want of a reliable readiness signal. No existing product owns the seam between commercial promises and physical hall readiness: DCIM suites track infrastructure, CRMs track contracts, and public-cloud schedulers allocate live clusters, but none converts a phased GPU build-out into contract-safe, tenant-ready inventory for a neutral operator. This company builds the capacity activation OS—a system of record that ingests rack commission states, power-block energization, cooling sign-off, and tenant requirements from operator CRM, DCIM, and provisioning systems, then surfaces only the GPU blocks that are truly sellable at each phase. The beachhead is narrow but urgent: 3-5 Southeast Asian AI factory operators (starting in the Singapore-Johor-Batam corridor) launching their inaugural 2,000-8,000 GPU halls with signed go-live dates across 2027-2028. The Firmus Batam project, committing up to 170,000 Nvidia accelerators in phased deliveries alongside public revenue commitments already contingent on committed customer deals, is the clearest signal that the buyer moment is real and time-constrained. Venture scale follows from compounding hall-level readiness data into the default commercialization layer for multi-campus AI factory operations, tenant brokerage, and cross-site workload portability as the regional GPU market matures.

Problem

• AI factory operators secure giant Nvidia allocations before a clean system exists to stage, sell, and activate that capacity; DCIM, CRM, and cluster tools remain disconnected, leaving commercial promises untrackable against physical readiness milestones.
• Phased GPU deliveries across 12-18 month campus builds force operators to make contractual go-live commitments before all racks, power blocks, and cooling loops are commissioned, creating twin failure modes—oversold capacity that cannot be activated and sellable inventory that stays unmonetized because commercial teams cannot confirm its readiness state.
• Non-hyperscale buyers needing 32-256 GPU reserved blocks require standardized reservation SKUs, entitlement maps, and acceptance checklists; today each tenant onboarding is a bespoke SRE project rather than a repeatable workflow, raising per-launch cost and delaying time-to-bill.

Solution

• A capacity activation control plane that ingests GPU delivery schedules, rack commission states, power-block energization, network sign-off, and tenant requirements from CRM, DCIM, cluster managers, and spreadsheets, then gates and surfaces only truly sellable GPU blocks with evidence links attached to each block.
• Standard reservation SKU builder and tenant entitlement generator that converts gated capacity blocks into contractible 32-256 GPU products, staged onboarding plans, and automated acceptance checklists—eliminating bespoke SRE runbooks per tenant launch.
• Readiness risk analytics dashboard that shows operators their go-live exposure across all in-flight tenant commitments so launch-date risk is visible at board level before it becomes a contractual breach.

Why we win

• No incumbent product bridges the commercial-readiness seam; DCIM suites (Schneider, Sunbird) stop at infrastructure telemetry, while CRMs and public-cloud reservation products do not model partial-commissioning states for neutral-operator campuses.
• Each hall deployment generates a readiness graph uniquely calibrated to that operator's infrastructure, supplier, and tenant mix; this dataset compounds with every launch and is structurally difficult for a generic workflow tool to replicate.
• The buying trigger is a board-level revenue risk event—overselling committed anchor tenants—not a productivity optimization, which shortens sales cycles and raises willingness to pay relative to the annual software subscription cost.
• Entering during the first-hall launch creates a high-trust expansion path: the operator naturally grants access to later halls, additional sites, and downstream billing and brokerage workflows once the first deployment proves ROI.

Milestones

flowchart LR
  Wedge[Beachhead: first-hall launch risk] --> MVP[MVP: readiness dashboard + evidence links]
  MVP --> Proof[Proof: tenants live on contracted date, zero oversell]
  Proof --> Expand[Expand: second hall same operator]
  Expand --> MultiSite[Multi-site control plane]
  MultiSite --> Platform[Platform: billing, brokerage, workload portability]

Founding team

Experiment roadmap

Risk assessment

What must be true

• At least 5 independent Southeast Asian AI factory operators will have signed committed anchor-tenant deals for 32-256 GPU reserved blocks before their first hall is fully commissioned by end of 2027.
• The economic cost of a single oversell event—breach of anchor-tenant contract or delayed go-live—exceeds $500K, making a $0.5-0.8M annual subscription justifiable on prevention value alone.
• DCIM, CRM, and commissioning systems at target operators expose enough structured data via API or export to build a hall-level readiness graph without a 12-month custom integration project.
• No Nvidia-supplied tooling, hyperscaler reservation product, or DCIM vendor will release a neutral-operator capacity activation workflow before the startup reaches 2-3 paid reference customers.
• The operator's COO or CCO has budget authority for $0.5-1M enterprise software investments inside major GPU hall launch programs without requiring board approval for each deal.

Open diligence questions

• How many Southeast Asian AI factory operators are currently signing committed anchor-tenant capacity deals ahead of partially commissioned GPU halls—can you name three and describe their current readiness gap?
• What is the typical contractual penalty or revenue exposure when a go-live date is missed for a 64-256 GPU anchor tenant, and how is that tracked today?
• Will Schneider Electric, Sunbird, or a major DCIM vendor release a capacity activation workflow covering commercial entitlements and tenant acceptance within 18 months?
• How long does a typical DCIM and CRM integration take at a 100-500 person AI factory operator, and what data quality can be expected from day one?
• Is the Firmus Batam project representative of the near-term beachhead, or are most operators in the corridor still 18-24 months away from first-tenant go-live commitments?
• Have any of the 5 most likely customers already scoped or started a homegrown version of this readiness workflow—what did they build and why did it fall short?

Model sanity

• Revenue engine. Base-case revenue comes from moving from 1 live hall in Y1 to 4 live halls by Y3 at roughly $700K of blended annual revenue per hall.
• Must go right. The first design-partner deployment has to convert into a reference subscription quickly enough to support a second live hall by Q2Y2 and a third by Q4Y2.
• Model breaks if. If pricing compresses toward $650K or sales cycles slip by a quarter, Y3 revenue falls by roughly $180-260K and the seed cash buffer shrinks meaningfully.
• Next-round proof. The Series A story is 3 live paid halls, $1.5-2.1M ARR, and 2 written ROI-backed reference deployments that show the platform prevented oversell or accelerated go-live.

Scenarios

Sensitivity

flowchart LR
  QualifiedOperators --> PaidHalls
  PaidHalls --> SubscriptionRevenue
  SubscriptionRevenue --> GrossProfit
  GrossProfit --> Cash
  ReferenceDeployments --> QualifiedOperators

Flags: The base case assumes the company wins 4 live halls from a beachhead the research sizes at only about 18 near-term halls, so logo-count validation is the biggest top-line risk. · Implementation and launch-support economics are normalized into blended ARPU instead of shown as a separate services line, which smooths quarterly volatility but can understate cash lumpiness. · Even in the base case the company is not fully EBITDA-positive by Y3, so the next round still depends more on ARR quality and reference evidence than on profitability. · Customer concentration stays high: losing one hall in Y3 would remove about $700K of exit ARR and meaningfully compress the cash buffer.

• Internal-build bias. Early AI factory operators may try to stitch together CRM, DCIM, and cluster tooling themselves instead of buying a new category. Mitigation: Land with a narrowly scoped launch-revenue workflow that shows time-to-bill and oversell-risk savings inside one hall before asking to replace broader systems.
• Channel compression. Nvidia, hyperscalers, or major cloud brokers could centralize capacity distribution and shrink the independent-operator segment. Mitigation: Focus first on neutral regional operators and become the interoperability layer across multiple supply sources rather than depending on any single upstream channel.
• Messy source-of-truth data. Construction, infrastructure, and commercial systems may disagree on what capacity is actually launch-ready, undermining trust in the product. Mitigation: Start read-only, reconcile multiple systems into an auditable readiness view, and attach evidence to every sellable capacity block before automating downstream actions.