Andromeda

1.1 Identity, footprint, and operating model

Andromeda presents itself as a two-sided market and infrastructure platform through which AI teams can buy and operate GPU capacity while infrastructure owners can sell capacity. Its official site says the platform sources, benchmarks, and prices compute across more than 100 providers, then standardizes contracts, service levels, invoicing, support, and observability; it also claims more than 1,000 completed transactions. Those are company-reported activity indicators rather than audited customer or revenue measures. Independent reporting traces the operation to a 2023 NFDG compute project, but the legal incorporation date remains unverified. A San Francisco office is reported, while a precise legal headquarters address is supported only by a low-reputation aggregator and should not be treated as confirmed.[CO001, CO002, CO003, CO011]

1.2 Leadership, founders, and governance exposure

Nat Friedman and Daniel Gross originated the project, and Gross recruited Wil Moushey to run it in late 2023. Moushey is the visible operating leader and chief executive: reporting attributes the spinout, commercial model, scale targets, and market-making thesis directly to him. Friedman and Gross later joined Meta and, according to Upstarts, no longer have an active relationship with Andromeda. That transition reduces direct founder involvement but increases dependence on Moushey, especially because no public board roster, succession plan, or broader executive bench was found in the assigned evidence. A company-issued recruitment post for a founding marketing leader says the narrative, launch, customer-proof, and feedback-loop functions are still being built from scratch, signaling an early organizational layer despite reported revenue scale.[CO003, CO004, CO006, CO013]

1.3 Capital, reported scale, and stakeholder alignment

Upstarts reported that Andromeda became a standalone company backed by NFDG and Paradigm, with Paradigm's total investment reaching $60 million at a $1.5 billion valuation in March 2026. The same report says 2025 revenue run rate exceeded $100 million, up from more than $50 million in 2024, and says the company had operated profitably since its start. Revenue run rate is not ARR, and reported profitability is not evidence of cash generation. A separate aggregator labels $160 million as total funding across a $100 million seed and $60 million Series A, but Upstarts describes NFDG's earlier amount as an upfront commitment associated with a GPU stockpile rather than clearly disclosed equity financing. Total raised, dilution, liquidity, debt, and available cash therefore remain unresolved rather than safely collapsing into one headline number.[CO007, CO008, CO009, CO010, CO014, CO015]

1.4 Milestones and diligence implications

The operating sequence is unusually compressed: an NFDG-backed capacity project in 2023, Moushey's recruitment that November, a reported stockpile above 4,000 GPUs by early 2024, revenue-run-rate expansion through 2025, and a standalone financing at a unicorn valuation in March 2026. By that financing report, a roughly 20-person company managed tens of thousands of GPUs and targeted as many as 100,000 during 2026. The resulting scale per employee could reflect a capital-light brokerage model, but it also concentrates execution, counterparty, and service risk. One customer told Upstarts that larger buyers may eventually internalize cluster capabilities, while the article noted that a durable marketplace requires enough buyers and sellers. No assigned source established the absence of lawsuits, outages, sanctions, or other incidents; the adverse search returning no retained result is a diligence gap, not exculpatory evidence.[CO004, CO005, CO006, CO007, CO008, CO009, CO011, CO012, CO016]

1.5 Exhibits

2.1 Market boundary and substitutes

The narrowest supportable market boundary is externally sourced GPU compute for AI training and inference, coupled with tooling to buy, sell, route, and operate that capacity. Andromeda's company-authored recruiting copy places the platform between model builders and a fragmented supply base of hyperscalers, neoclouds, and independent data centers. Product descriptions add orchestration through Slurm, Kubernetes, or SSH plus storage and operational support. That boundary excludes general-purpose cloud spend, owned private clusters, chips sold directly by OEMs, and capital deployed into data centers unless Andromeda intermediates the resulting capacity. The status-quo substitutes are single-provider cloud procurement, direct provider contracts, internal infrastructure teams, and bespoke brokered deals; the assigned evidence does not quantify their relative shares.[CM001, CM002, CM003, CM008, CM009]

2.2 Evidence-constrained sizing

No retained source provides an independently constructed market total, growth rate, geographic split, or denominator from which a defensible TAM, SAM, or SOM can be calculated. Company-authored job copy says that "trillions of dollars of compute" could flow through Andromeda in five years, but it is an aspiration about platform throughput rather than a measured market forecast. It must not be treated as revenue, GMV, SAM, or available spend. A credible sizing model would separately estimate accelerator capacity available from third parties, workloads willing to use external supply, Andromeda-compatible regions and hardware, and the platform's achievable take rate. None of those denominators is present in the assigned pool, so this chapter presents sizing layers without numeric scale.[CM012, CM013, CM015, CM016]

2.3 Buyers, users, payers, and adoption path

The evidence points to technical infrastructure leaders as the primary users and evaluators, while founders, commercial leaders, or enterprise budget owners likely approve spend. Andromeda's sales posting targets frontier labs, AI-native startups, and enterprises whose requirements have outgrown one provider's fleet. It describes a consultative sales cycle spanning architecture discussions, technical validation, commercial negotiation, and parallel engagement of technical, commercial, and executive stakeholders. On the supply side, GPU cluster providers are counterparties that need demand, utilization, and potentially financing, rather than end users of compute. VC funds and accelerators are described as channels into portfolio-company demand. These roles support a marketplace adoption path, but the evidence supplies no conversion rates, sales-cycle length, contract duration, retention, or budget data.[CM004, CM005, CM006, CM007, CM010, CM011]

2.4 Drivers, constraints, and valuation relevance

The central demand driver is the company's thesis that AI workloads will require more reliable high-performance compute than a small number of vertically integrated providers can supply. The matching supply driver is fragmented GPU capacity seeking customers and utilization. Those are company claims, not independent market validation. Adoption still depends on technical compatibility, capacity reliability, price transparency, workload portability, security, data-transfer economics, procurement approval, and confidence that a marketplace can manage provider and operational risk. The assigned evidence does not measure any of those constraints or compare marketplace economics with direct cloud contracts. Valuation therefore cannot be anchored to broad AI-infrastructure growth alone; diligence needs verified transaction volume, take rate, gross margin, repeat usage, concentration, and supply utilization before a market-share or revenue-multiple argument is credible.[CM002, CM003, CM007, CM013, CM015]

2.5 Exhibits

3.1 Landscape and buyer alternatives

The relevant competitive boundary is wider than GPU brokers. Tracxn lists 340 active competitors for Andromeda and names CoreWeave, Together AI, FluidStack, Crusoe, Aethir, exaBITS, Yotta, and Neysa among the ranked set, showing that a buyer can choose direct GPU clouds, distributed supply networks, broader AI platforms, or general cloud providers rather than a broker. CompuX adds a financing-led alternative at the API layer, while hyperscalers and internal procurement remain substitutes for teams willing to accept one-provider exposure or operate their own vendor portfolio. Andromeda's company-issued sales posting targets frontier labs, AI-native startups, and enterprises whose demand exceeds a single provider's fleet, narrowing the most plausible buyer wedge to sophisticated, capacity-constrained teams rather than the full cloud market.[CP005, CP007, CP008, CP009, CP010]

3.2 Capability, pricing, and delivery model

Retained evidence places Andromeda at the procurement and operations layer. Its own recruiting material describes a platform to buy, sell, and operate compute, with a console where teams find, buy, and run capacity across global supply. Tracxn reports access through Slurm, Kubernetes, or SSH, plus storage, DevOps support, service-level agreements, and a third-party GPU marketplace. A competitor comparison adds claims of more than 100 providers, one contract, one invoice, a single SLA, provider certification, and a live pricing index; these details are useful directional evidence but are not independent verification. CompuX instead combines financing with OpenAI-compatible routing. Public list prices, realized discounts, contract minimums, SLA remedies, and like-for-like benchmark results remain unavailable, so the chapter does not claim a verified cost or performance lead.[CP001, CP003, CP004, CP005, CP006, CP013]

3.3 Distribution, switching, and control points

Andromeda's distribution thesis depends on aggregating fragmented provider supply and reaching buyers through direct enterprise sales and ecosystem partnerships. Current recruiting evidence shows both motions are still being built: an account executive is expected to bring an existing network and convert it into pipeline, while a new partnerships leader is expected to create the lifecycle across GPU providers, AI labs, OEMs, venture funds, and accelerators. That can produce two-sided liquidity, but it also gives providers and channel partners bargaining power until demand concentration and repeat usage are demonstrated. Switching costs appear more likely to arise from contracting, workflow integration, support, and operational familiarity than from proprietary hardware. The retained pool has no customer-level evidence on multi-homing, renewal, migration effort, provider concentration, or SLA performance, preventing a stronger lock-in conclusion.[CP002, CP010, CP012, CP016]

3.4 Moat durability and displacement risk

The potential moat is market liquidity plus operational standardization: more supply can improve availability and price discovery, while more buyer demand can improve provider utilization. The risk is that these benefits are contestable. Direct GPU clouds and hyperscalers control more of the physical fleet and runtime stack, financing-led aggregators can bundle capital with access, and internal procurement can replicate multi-vendor sourcing for the largest buyers. Company hiring language also says customer proof, win-loss analysis, competitive intelligence, and partnership playbooks are being established rather than already institutionalized. The resulting positioning is credible but not yet proven durable: Andromeda may win on breadth and procurement simplicity, yet the evidence does not establish exclusive supply, proprietary performance advantages, verified savings, or high switching costs.[CP011, CP012, CP014, CP015, CP016]

3.5 Exhibits

4.1 Revenue model and public traction

Andromeda appears to monetize access to fragmented GPU supply rather than sell a conventional software subscription. The public proposition combines on-demand capacity, provider vetting, multi-provider procurement, consolidated billing, and pricing intelligence. SiliconANGLE reported that customers can avoid long-term contracts, while a separate job-directory profile describes flexible usage with no minimum duration and no ingress or egress fees. Those terms describe customer packaging, not realized pricing or recognized revenue. The strongest traction evidence is third-party reporting that annualized revenue run rate increased from $50 million in 2024 to $100 million in 2025 and that the business has been profitable since launch. Neither metric is audited in the retained evidence, and revenue run rate must not be treated as ARR or cash generation.[CI001, CI002, CI005, CI006, CI007, CI008, CI009, CI015]

4.2 Go-to-market and delivery economics

Current hiring materials point to a relationship-led, technically intensive sales motion. An account-executive posting asks candidates to bring an existing network, own the full cycle through technical validation and negotiation, and sell to AI-native buyers whose needs may exceed one provider's fleet. Partnership hiring adds provider distribution and venture-portfolio channels, while an infrastructure role is charged with matching demand to capacity, increasing utilization, onboarding suppliers, and negotiating supplier cost. This suggests that sales efficiency depends on deal sourcing, technical evaluation, and capacity matching rather than low-touch self-service alone. Public materials provide no sales-cycle duration, quota attainment, CAC, payback, provider take rate, support cost, or utilization history, so the cost to produce each dollar of reported run rate remains unquantified.[CI010, CI011, CI012, CI013, CI016]

4.3 Capital adequacy and margin path

Reporting places the latest financing at a $1.5 billion valuation, but SiliconANGLE says the round's exact dollar amount was unclear. That prevents a reliable bridge from funding headlines to cash currently available. The retained public sources do not disclose cash on hand, burn, runway, debt, guarantees, working-capital exposure, or contractual commitments to providers. The operating model could be relatively asset-light when capacity is brokered, yet job materials also emphasize supplier negotiation, utilization, reliability, and operational complexity, all of which can create margin and working-capital risk. A low-reputation market profile additionally lists a compute-for-equity program, but no contract terms were available, so that item is a diligence lead rather than an underwritten revenue stream.[CI003, CI004, CI012, CI014, CI016]

4.4 Financial verdict and underwriting blockers

The public evidence supports demand and transaction activity, but not revenue quality. A doubling of reported annualized run rate and reported profitability are positive signals; more than 100 providers and more than 1,000 transactions also indicate that the platform has moved beyond a single-cluster experiment. However, none of those facts establishes gross margin, recurring retention, cash conversion, or resilience to GPU-price compression. The central underwriting question is whether Andromeda earns a durable spread for aggregation and operations without assuming inventory, credit, uptime, or minimum-purchase obligations that consume the apparent margin. Investment-grade diligence requires a recognized-revenue bridge, transaction-level take rates and contribution margins, concentration schedules, provider commitments, and a current liquidity and debt schedule.[CI001, CI002, CI006, CI007, CI012, CI016]

4.5 Exhibits

5.1 Product surface and customer workflow

Andromeda is not simply a directory of GPU inventory. Its public privacy policy describes a business-to-business marketplace in which customers submit GPU type, quantity, timing, and duration requirements while providers submit configuration, pricing, availability, networking, and storage data. Company-issued recruiting material adds sourcing, benchmarking, contracting, billing, and a single-invoice experience. Together, those disclosures indicate a commercial control plane spanning discovery through settlement. The customer-facing product appears to include capacity and pricing tables, usage and billing views, cluster and job dashboards, onboarding, documentation, and terminal-adjacent workflows, although a current public feature inventory or live documentation set was not available in the assigned evidence.[CE001, CE002, CE003, CE004, CE013, CE014]

5.2 Architecture, deployment, and operating model

The disclosed architecture separates the marketplace from an infrastructure control plane designed to abstract virtual machines, Kubernetes, bare metal, and schedulers. At the data layer, Andromeda uses WEKA NeuralMesh and its Kubernetes Operator for deployment and lifecycle management across heterogeneous providers. Supplier-originated reporting says the integration can use either dedicated storage or GPU-native Axon configurations and reports a roughly ten-minute bare-metal deployment path. This is a coherent portability strategy: normalize orchestration and storage above fragmented capacity. It also creates dependencies on provider configuration quality, WEKA software, and Andromeda's ability to certify and continuously operate unlike clusters to one service standard.[CE005, CE006, CE008, CE009]

5.3 Maturity and differentiation

Evidence of production use exists, but its provenance limits confidence. A WEKA-originated announcement carried by StorageReview says Andromeda operates across more than 50 providers, supports more than 100 customers, and has clusters sustaining more than 400 GB/s and six million IOPS; it also reports environment startup falling from about three minutes to 30 seconds. These figures are useful indicators of scale and technical ambition, not independent benchmarks. The likely differentiation is the operating layer that combines market access, multi-provider abstraction, certification, and portable storage rather than a proprietary accelerator. Public recruiting signals also show core product design, customer-proof, and capacity-planning functions are still being built, consistent with a maturing rather than fully standardized platform.[CE007, CE008, CE010, CE011, CE014, CE015, CE016]

5.4 Trust, reliability, and roadmap risk

The privacy policy states that Andromeda uses access controls, encryption in transit and at rest, and ongoing monitoring, and that it offers a data processing agreement on request for GDPR or UK GDPR customers. Those are useful baseline commitments, but they are not substitutes for an independent security audit, named certification, penetration-test summary, incident history, or service-level performance record. The same policy confirms that the platform processes commercially sensitive workload, infrastructure, financial, billing, and technical data and may share data between marketplace counterparties as needed to transact. Before underwriting enterprise-grade maturity, diligence should therefore test tenant isolation, provider access boundaries, failover behavior, incident response, subprocessors, uptime, and the division of support responsibility among Andromeda, capacity providers, and WEKA.[CE004, CE006, CE012, CE013]

5.5 Exhibits

6.1 A two-sided customer and capacity network

Andromeda addresses technical compute buyers rather than a single homogeneous customer class. Its public materials describe AI builders using one relationship to access capacity across providers, while recruiting materials identify AI-native founders, infrastructure leads, machine-learning teams, research labs, inference providers, and enterprises as target demand. The supply side includes hyperscalers, neoclouds, independent data centers, and GPU cluster operators. This creates two distinct commercial motions: winning sophisticated workload owners and recruiting capacity partners whose fleets can satisfy those workloads. The public positioning is strongest for customers whose requirements exceed one provider's fleet or who need training and inference capacity quickly. It is not possible to size revenue contribution by segment, geography, workload, or buyer type from the retained evidence.[CU004, CU006, CU007, CU013, CU014]

6.2 Scale signals are substantial but mostly issuer-reported

The clearest adoption signals are scale markers rather than customer-level operating disclosures. WEKA reports that Andromeda serves more than 100 AI customers across more than 50 capacity providers, and the company says it has supported more than one billion GPU-hours. An independent trade publication repeats the customer, provider, and GPU-hour figures, but its article is based on the same WEKA announcement and therefore is not fully independent corroboration. Company job materials also say the original managed cluster filled almost immediately. Together, these signals suggest real demand and meaningful workload throughput, but none supplies a cohort denominator, customer start date, active-account definition, paid-versus-pilot split, or revenue concentration bridge. The figures should therefore be treated as company and partner-reported scale, not audited retention evidence.[CU001, CU002, CU003, CU011]

6.3 Deployment outcomes exist, but customer identities are withheld

The WEKA integration provides the most concrete customer-use evidence. One research lab reportedly overcame a supplied-storage bottleneck without adding infrastructure; AI teams reportedly reduced environment startup time from about three minutes to 30 seconds; and one biotech workload reportedly moved metadata-heavy transfers from hours to minutes for directories containing roughly one billion files. These examples indicate production-scale technical usage, yet the customers are unnamed and the disclosures do not identify contract value, workload duration, renewal status, baseline method, or whether the outcomes were independently measured. Across the retained source set, no paying buyer is named together with verified deployment scope and contract status. The examples support product utility, but they do not establish repeatable customer economics or reference quality.[CU008, CU009, CU010, CU015]

6.4 Expansion mechanics are visible; durability metrics are not

Public recruiting materials outline an expansion model built around relationship-led sales, technical validation, supply-demand matching, and distribution partnerships. The planned partnership motion would channel portfolio-company demand through venture funds and accelerators while helping GPU providers find customers, which could expand both sides of the marketplace. However, the founding-marketing role explicitly calls for building case studies, reference customers, and customer feedback loops, signaling that institutionalized customer proof remains a work in progress. No retained source discloses gross or net revenue retention, churn, renewal rates, contract length, customer acquisition cost, top-customer share, or channel-attributed revenue. Without those measures, the more-than-100-customer headline cannot show whether adoption is durable, diversified, or economically attractive. Underwriting should require account-level cohort and concentration data before treating reported scale as recurring demand.[CU012, CU013, CU014, CU016]

6.5 Exhibits

7.1 Ranked residual exposure

The highest residual risks are multi-tenant security, provider reliability, and operating execution. ClusterMAX reports that Andromeda can deploy Slurm on Kubernetes with light namespace isolation between tenants and procures capacity from multiple neoclouds, while Airframe describes a broker responsible for consolidating commercial terms, billing, and observability across providers. Those facts establish a broad control surface, but the assigned sources do not disclose control testing, incident history, SLA attainment, or provider concentration. A second tier comprises founder-ecosystem continuity, early commercial systems, cap-table opacity, and the operational challenge embedded in a stated ambition to reach as many as 100,000 GPUs. Severity and likelihood below are analyst assessments of the cited exposure, not ratings published by the sources.[CR001, CR002, CR003, CR005, CR006, CR013, CR014, CR016]

7.2 Regulatory and legal risk

Public materials in the assigned pool describe global, multi-provider compute brokerage, standardized commercial terms, and multi-tenant orchestration, but they do not include a regulator record, court record, DPA, privacy control matrix, export-control policy, or certification report. This is an evidence limitation, not proof that controls or incidents are absent. The most consequential legal diligence therefore concerns cross-tenant data handling, sanctions and export screening across providers and geographies, allocation of outage and performance liability, and corporate ownership. The latter deserves focused review because Silicon Valley Investclub says the full shareholder register is not public and its tear sheet mixes named investors, SPVs, and unattributed entries.[CR002, CR006, CR008, CR012, CR014, CR015]

7.3 Operational, quality, and security risk

Andromeda's marketplace model can reduce sourcing friction, yet it also makes service quality depend on heterogeneous upstream fleets, networking, storage, orchestration, and commercial interfaces. ClusterMAX's light-isolation description is the sharpest technical risk signal in the assigned evidence; it should be tested rather than extrapolated into an incident claim. Company-issued job material separately acknowledges capital, risk, and operational complexity, and a reported 100,000-GPU aspiration would increase the importance of automated qualification, capacity provenance, change management, and escalation discipline. Public staffing and customer-proof signals are inconsistent or thin, so mitigation maturity should remain unscored until private evidence is reviewed.[CR002, CR005, CR007, CR009, CR010, CR011, CR016]

7.4 Dependencies, financial model, and execution

The company sits between fragmented GPU suppliers and demanding AI buyers, making both sides of the marketplace potential concentration points. ClusterMAX ties the original model to NFDG portfolio companies and questions continuity after Nat Friedman and Daniel Gross joined Meta; Andromeda's own sales posting says an existing relationship network is the fastest route to pipeline and that the sales motion is still being built. A reported $60 million round at a $1.5 billion valuation and aggregate funding claims indicate access to capital, not cash availability, margin quality, or runway. Investment approval should therefore require provider and customer concentration schedules, recognized-revenue and take-rate bridges, cash commitments, SLA history, and named owners for remediation.[CR001, CR003, CR004, CR008, CR009, CR013, CR014, CR016]

7.5 Exhibits

8.1 Financing context and entry discipline

Multiple secondary sources report a $60 million Paradigm-led Series A at a $1.5 billion post-money valuation. One report also places annualized revenue run rate above $100 million and says the company has been profitable since launch; those are reported operating claims, not audited revenue, ARR, cash generation, or cash available. Dividing the financing mark by the stated revenue-run-rate floor produces a headline multiple of no more than about 15x. That is an orientation point, not an underwriting multiple, because take rate, gross margin, revenue recognition, working capital, and concentration are undisclosed. Public funding totals also conflict, while the full shareholder register is unavailable. Entry discipline should therefore require primary financial and cap-table evidence before accepting the current mark.[CV001, CV002, CV003, CV004, CV005, CV015, CV016]

8.2 Thesis, anti-thesis, and comparability

The thesis is that Andromeda can become a trusted transaction and operations layer across fragmented GPU supply. Its own account describes benchmarking, certification, standardized contracts, matching, and operations across providers, and reports more than one billion GPU-hours. The anti-thesis is that scale claims have not yet been translated into independently verified margin quality or durable customer economics. Company-authored recruiting material says customer proof, feedback loops, win-loss analysis, the sales motion, and the design function are still being built. Third-party profiles also blend current marketplace positioning with older compute-for-equity and cluster narratives. The named competitor set is useful for strategic context, but it is not an independent valuation benchmark and contains no sourced comparable multiples.[CV006, CV007, CV008, CV009, CV010, CV011, CV012, CV013, CV014]

8.3 Bull, base, and bear cases

A quantified valuation range is not supportable from the assigned public evidence, so the scenarios are framed as underwriting states rather than false-precision price targets. The bull case requires audited conversion of marketplace activity into durable net revenue, attractive gross margin, diversified customers and providers, and repeatable cash generation. The base case assumes reported activity is directionally real but that margin and concentration remain unproven, leaving the financing mark ahead of diligence quality. The bear case is triggered by weak take rate, revenue reversals, concentrated counterparties, preference or debt overhang, or failure to convert a founder-led network into repeatable go-to-market. Each state should be re-priced after primary diligence rather than assigned an unsupported probability today.[CV002, CV003, CV008, CV010, CV014, CV015, CV016]

8.4 Recommendation, exit readiness, and diligence asks

Recommendation: track, with medium confidence, high risk, and a stretched valuation stance. The current mark can become investable if the company substantiates recognized revenue, take rate, gross margin, cash conversion, retention, counterparty concentration, service performance, and cap-table terms. Until then, a target return, hold period, and exit value cannot be responsibly calculated. Exit readiness is likewise unproven: the public record lacks audited financials, a disclosed shareholder register, independently verified customer proof, and a comparable transaction framework. Thesis-break triggers are material revenue reclassification, sustained negative cash conversion, top-customer or provider dependence, punitive preferences, and failure to build repeatable sales and customer proof. A lower entry price can compensate for some execution risk, but not for missing ownership or revenue-quality evidence.[CV001, CV003, CV007, CV009, CV010, CV012, CV014, CV015, CV016]

8.5 Exhibits