Public disclaimer: This is an independent public teardown based solely on publicly available information retrieved on 2026-06-17. It is research support, not investment, legal, tax, or financial advice. Company-provided statements are labeled as claims, not verified facts. RunAnywhere did not participate in or review this report. Errors and omissions are possible; final decisions remain with the reader.
Two engineers. Ten thousand dollars of disclosed capital. And a stack most funded teams never finish: a cross-platform on-device AI inference SDK, a proprietary Apple-Silicon engine, a fleet control plane, and a second product line, all shipped out of a real YC W26 batch. The build density is on the public record — thousands of commits, a Delaware Form D, stable Swift and Kotlin targets. What carries the valuation is not. The engine's headline speed edge is the company's own benchmark, reproduced by no one, and the paid layer still has no price, no named customer, and no production fleet anyone can point to. Five partners spent a single day across 38 public sources, every claim cited and no founder framing waved through. The entire case turns on whether one thing the public record cannot yet see proves out. The teardown carries our answer. Read it before the next demo day reprices the round.
What The Company Does
company claim per the GitHub SDK repository.company claim — a claim that has not been independently reproduced. MetalRT is Apple Silicon arm64-only and carries a proprietary license source-backed fact per the MetalRT binaries repository. A separate HexagonRT engine for Qualcomm neural processing units (NPUs) is described as "coming soon" on the engines page company claim, with no shipped binary found in public repositories as of the retrieval date.company claim per the SDK page. Console has no public pricing, no publicly named customer, and no independently verified production deployment. The commercial motion is contact-driven, with a Google Form waitlist. A third product, Mirar, is a vision-agent platform in early access on a staging host.source-backed fact.Key Takeaways
Investor reaction
A two-person YC W26 team building at unusually high velocity in a real category window — but every load-bearing claim (benchmarks, prior-employer scale, production demand) is company-controlled or absent, the paid layer is entirely unvalidated, and the only proprietary asset is Apple-only and independently unreproduced.
Current verdict
hold, medium confidence.
Why investors might lean in
An evidenced build density of 3,438 SDK commits and roughly 1,887 GitHub contributions per year for the CTO is unusual for a two-person, minimally-capitalized pre-seed team. The category thesis is correctly timed: edge AI chipset revenue is projected to grow from $34.4B (2026) to $96B (2031) per ABI Research, and the non-obvious competitive insight — that a neutral, cross-platform fleet control plane occupies a white space that platform owners are structurally disincentivized to fill — is architecturally defensible. The pre-seed entry price is an appropriate check size for an option on an unmeasured ceiling.
Why investors might pull back
The popular layer (the open-source SDK) is free and depends on free backends (llama.cpp and ONNX Runtime) that the company neither owns nor controls. The paid layer (Console) has no public pricing, no named customer, and no independently verified production deployment. MetalRT, the only proprietary performance asset, is Apple Silicon-only and unreproduced by any third party. Founder prior-employer scale claims are unverified (public LinkedIn employment history was not retrievable from public data sources), and a three-way self-contradiction exists between the privacy policy, documentation, and SDK telemetry code.
Highest-leverage fix
Produce one named, retained, paying Console design partner at any price, with a reference willing to take a 15-minute investor call confirming that fleet management — not runtime speed — is why they pay.
Best next move
Schedule a first founder call. Arrive with the critical diligence trifecta: paying Console customers or committed design partners with price and fleet context; a MetalRT benchmark methodology disclosure enabling independent reproduction; and a reconciled privacy and telemetry statement. That call determines whether this is a first-check option or a pass.
What Makes This Potentially Fundable
The fundable version of RunAnywhere is the default cross-platform control plane for on-device AI model fleets — the infrastructure layer that every major smartphone original equipment manufacturer (OEM), enterprise mobile platform team, and edge device operator uses to deploy, update, and monitor on-device AI models across heterogeneous hardware. In that version, the open-source SDK is not the product; it is the free distribution funnel. The product is Console, priced per device per year, with OTA model delivery and policy routing creating fleet-level switching costs and an accumulating operational dataset that makes the platform harder to displace with each release cycle.
The non-obvious investment thesis is structural: Apple, Google, Meta, and Qualcomm each publish their own inference runtime (MLX, LiteRT, ExecuTorch, and Qualcomm Neural Network (QNN) respectively), but each is optimized for its own hardware and incentivized toward platform lock-in. None has an incentive to build a genuinely cross-platform fleet management layer, because cross-platform portability works against their strategic interest in keeping developers on their own silicon. RunAnywhere, as a neutral party, occupies the exact white space that no platform owner will voluntarily cede.
The evidence gap that keeps this honest is substantial. The fund-returning mechanism routes through a paid Console layer that is entirely unvalidated commercially. No named paying customer, no public pricing, and no independent production deployment exist in the public record. The category timing is right in general terms, but enterprise software buyers are estimated to be 18–36 months behind the hardware build-out, which means the company may be correctly positioned for a wave that has not yet arrived. That timing uncertainty is what separates this from a clear pursue.
The Four Holes To Close Before Fundraising
| Hole | Investor Fear | What To Bring |
|---|---|---|
| Zero independently verified paid demand: no named paying customer, no public pricing, no production fleet, no contract value | "This is a popular free tool that will never monetize. The paid layer is a hope, not a product." | At minimum two to three named design partners under NDA with signed letters of intent or paid pilots; a pricing page or price sheet with a documented per-device or SaaS tier; at least one customer citing Console fleet operations — not runtime speed — as the reason they pay |
| Three-way self-contradiction between the privacy policy, documentation, and SDK telemetry code | "The founders have a disclosure gap that is an enterprise procurement dealbreaker and a potential regulatory liability at scale." | A reconciled, scope-accurate privacy policy covering the SDK, Console, and iOS app; a published data-flow diagram showing exactly what telemetry transmits and to which endpoint; completed Apple App Store and Google Play data-safety declarations |
| MetalRT differentiation is company-claim-only and Apple Silicon-only; HexagonRT is unshipped | "The only defensible proprietary asset is a single-vendor claim with no external validation and no path to the Android total addressable market." | A published, methodology-disclosed third-party benchmark reproducing the MetalRT performance edge over Apple MLX and llama.cpp on independently selected hardware; shipped HexagonRT binaries on at least one Qualcomm device with a public benchmark |
| Neither founder has independently verified enterprise software background; zero enterprise go-to-market (GTM) hire is visible; Console commercial motion is contact-driven with no price | "Even if the product is real, this team cannot sell it to the enterprise buyer that the value-capture thesis requires." | A named founding-sales hire with a reference; or two or more enterprise design partners signed by the founders, demonstrating direct enterprise sales execution; alternatively, a credible plan with a named and committed GTM hire |
Decision Snapshot
- One-sentence company description: RunAnywhere is a pre-seed, YC W26 company building a cross-platform on-device AI SDK and a proprietary inference engine (MetalRT), with a fleet control plane (Console) as the intended commercial layer — but the paid surface is entirely unvalidated.
- Screen: hold
- Confidence: medium
- Suggested next action: Schedule a first founder call targeting the critical diligence trifecta — a paying Console design partner, a MetalRT methodology disclosure, and a reconciled privacy and telemetry statement — before committing any capital.
- Why this matters now: Edge AI hardware is deploying at scale today; the window to build the control-plane standard before platform owners are incentivized to act may be narrow; a pre-seed entry at the right company, if the paid thesis verifies, is better than a seed entry after the category is defined.
- Investor-readiness diagnosis: Developer distribution is unusually strong for pre-seed, but the commercial layer is not yet investor-grade — no pricing, no customers, no independent validation of the core performance claim, and a disclosed privacy contradiction that enterprise procurement teams will flag immediately.
- Best founder use of this report: Fix the telemetry contradiction and get one signed, paying Console pilot before the next investor conversation; the entire holding pattern resolves on that single datapoint.
IC Disagreement Map
The five partners independently reviewed the same Evidence dossier and submitted initial votes before any cross-reading. Each lens covers a distinct dimension: the power-law partner tests fund-returning upside and the falsifiable home-run mechanism; the prepared-mind partner tests category timing, competitive structure, and product insight; the founder-jockey partner tests founder-market fit, execution credibility, and team quality; the risk-reduction partner tests critical flaws, unsupported claims, and value-capture durability; and the long-horizon partner tests portfolio fit, compounding assets, ownership math, and incentive alignment across the full stakeholder map.
Power-law partnerThe fund-returning layer is not the inference runtime — which commoditizes — but the operational control plane: fleet management, OTA model delivery, policy routing, and inference observability priced per AI-enabled device. As AI silicon ships in roughly 1.25 billion phones per year plus more than 21 billion IoT devices, per-device operations spend on a compounding control plane could be a multi-billion-dollar ceiling. Stars are not the thesis; the stars are the free distribution funnel for a paid layer that is unmeasured, not disproven.
Would flip onOne retained, paying Console fleet at any price would move this partner toward pursue; a credible primary sizing of the on-device AI fleet-ops software layer at multi-billion-dollar scale, combined with at least one paying customer, would move them firmly to pursue. Confirmed structural inability to capture value above the free floor — meaning enterprises universally prefer self-hosting — would move them to pass.
Prepared-mind partnerThe company is marketing on the wrong battleground — the benchmark war for runtime speed is commoditizable. The defensible battleground is cross-platform model-fleet operations, an MDM-for-models layer that Apple, Google, Meta, and Qualcomm are each structurally disincentivized to build because each wants developers locked to its own runtime. On that battleground, incumbents will not contest a neutral cross-platform player; accumulated fleet telemetry creates a learning-advantage moat. The runtime and SDK are correctly read as distribution channels, not the product.
Would flip onMultiple paying enterprises managing mixed-device model fleets through Console with documented retention would move this partner to pursue. Continued absence of production or paid pull while peers capture the same category insight would move them to pass.
Founder-jockey partnerThe jockey case is the most underweighted signal. A two-person team building a proprietary low-level runtime, a stable five-target cross-platform SDK, a fleet control plane, and a second product on minimal disclosed capital is demonstrating an unusually high, evidenced build density. Teams that ship at this density are the ones that survive a wedge being commoditized by re-aiming quickly. One biographical verification would convert this to pursue.
Would flip onReference-grade confirmation of both founders' prior roles at the claimed scope — CEO's Intuit mobile SDK leadership and CTO's Azure/AWS infrastructure depth, plus MetalRT authorship confirmed by independent technical walkthrough — would move this partner to pursue. Unverifiable résumés, or MetalRT confirmed as a thin wrapper over Apple's free Metal API with no novel kernel work, would move them to pass.
Risk-reduction partnerThe proprietary surface is structurally too thin to capture fund-returning value. The SDK's inference runs on llama.cpp and ONNX Runtime — free software the company neither owns nor controls. MetalRT is confined to Apple Silicon, where Apple itself offers a free competing framework (MLX, approximately 27,000 GitHub stars). The company's differentiation lives entirely on the one surface where the platform owner has both the incentive and the capability to commoditize it for free. Layer on a privacy-policy-vs-telemetry contradiction visible in the company's own merged code, unpopulated app-store data-safety labels, and unverifiable founder scale claims, and the rational base case is a popular free runtime that never converts to a defensible paid layer.
Would flip onTwo or more named, paying Console customers with documented retention at a disclosed price, plus an independent MetalRT benchmark reproduction showing a durable edge over Apple MLX, plus verified founder employment at the claimed scope, would move this partner to pursue. Confirmation of any one of the five kill criteria — no paying customers after an active push, MetalRT parity with free alternatives, Apple policy noncompliance, a platform owner shipping first-party cross-platform fleet management, or an integrity break — would move them to pass.
Long-horizon partnerThe binding constraint is not technology or team — it is the ownership-adjusted value-capture math. The monetization surface is squeezed between free OSS below (including RunAnywhere's own SDK dependencies) and platform-owner runtimes above (all free and incentivized to give fleet tooling away). Even in the execution-bull case, the only durable monetization surface competes against platform incentives to commoditize it. Layer on an unconfirmed pre-seed ownership path: a below-target stake in a company whose paid layer is structurally contested may not be fund-returning even if the founders succeed.
Would flip onAccessible fund-returning ownership confirmed at a check inside the fund's target band, combined with at least one signed and paid Console design partner with documented retention, would move this partner to pursue. A round fully subscribed at terms that preclude a fund-returning ownership position, or contested capture confirmed from the ownership side, would move them to pass.
Scenario Range
| Scenario | What The Company Looks Like In 3-5 Years | Falsifiable Trigger To Watch | Earliest Evidence |
|---|---|---|---|
| Strikeout | RunAnywhere remains a widely-downloaded free SDK with no paid layer; Console stays a waitlist indefinitely; runway exhausts before any enterprise design partner converts; founders return to employment; the Apache-2.0 SDK becomes a community-maintained artifact without commercial backing | An active design-partner push of more than 90 days with zero paid or committed customers; MetalRT benchmark independently reproduced as equivalent to Apple MLX on the same silicon; Apple flags the OTA model-delivery mechanism as noncompliant under App Store Guideline 2.5.2 | Zero paying customers at first-call confirmation; a third-party benchmark showing MetalRT parity with MLX; static or declining GitHub star trajectory after the YC batch launch window |
| Base | A small but real SaaS control plane for on-device model fleets — Console at a modest per-device annual rate for mid-market mobile-AI teams; MetalRT shipping on Apple plus Qualcomm (HexagonRT); SDK install base converting at 2–5% to paid Console customers; a Series A at $5M–$15M; competitive with YC-backed peers; acquired by a larger infrastructure player at a modest multiple | A handful of paying Console customers sign and retain across at least two release cycles; HexagonRT ships within 9 months; the team adds one enterprise-GTM hire; a public pricing page goes live | A signed design-partner case study, even under NDA; public pricing with a defined tier; a HexagonRT GitHub release tag with an Android benchmark |
| Home run | RunAnywhere becomes the default cross-platform control plane for on-device AI fleet operations — the standard for OTA model delivery, policy routing, and inference analytics across iOS, Android, and IoT — with accumulating fleet-telemetry switching costs, a compounding learning advantage from deployment data, and a durable cross-platform performance moat (MetalRT and HexagonRT). Enterprise fleet management is priced per device per year at a recurring rate. The fund-returning mechanism routes through Console ARR multiplied by fleet lock-in and exit multiple, not star counts. | Three or more retained, paying enterprise fleets using Console OTA across heterogeneous device types, citing fleet operations — not runtime speed — as the buying reason; MetalRT edge independently reproduced; HexagonRT shipped with an Android benchmark; no first-party cross-platform fleet management visible on Apple, Google, or Qualcomm public roadmaps | A named paying enterprise customer (even on background) plus a published pricing page plus a HexagonRT GitHub release tag with an Android benchmark — these three signals together at the same time |
What Investors Will Test
| What We Looked For | Current Read | Investor Implication | Founder Prep Priority |
|---|---|---|---|
| A named, falsifiable paid surface with a credible path to fund-returning scale | Console is the stated paid layer; it has no pricing, no customers, and no production fleet in the public record | No investor can underwrite a home-run thesis that routes through a paid layer that does not yet demonstrably exist | Before the next investor conversation, commit at least one paying Console pilot at any price and publish a pricing page |
| Evidence beyond company-controlled surfaces for every load-bearing claim | MetalRT benchmarks, founder scale claims, production deployments, and design-partner count all trace exclusively to company-controlled pages; no independent corroboration was found for any of these claims | Every positive claim must be downgraded to company claim until independently corroborated; the bull case rests almost entirely on narrative | Obtain one third-party MetalRT benchmark reproduction and one employer reference for each founder confirming role and scope |
| A correct category thesis named before the investment meeting | The category-creation read — cross-platform fleet control plane as the defensible layer — is architecturally coherent and defensible, but the company's own public marketing leads with runtime speed, not fleet operations | Investors who understand the category will ask why the company is marketing on the wrong battleground; misaligned positioning undermines conviction in the company's own product insight | Reframe all external positioning to lead with "the cross-platform control plane for on-device model fleets" and retire the benchmark-first narrative |
| A team whose publicly evidenced competencies cover the three hardest required builds | Build velocity is evidenced; prior-employer scale is claimed without independent verification; enterprise GTM coverage is absent | The jockey thesis is the most convertible positive; one reference call per founder would move it from claimed to evidenced | Prepare one named employer reference for each founder and be ready to give a technical walkthrough of MetalRT kernel work |
| Value-capture alignment across the full stakeholder map | Enterprise buyers may prefer self-hosting the free stack; platform owners are incentivized against a cross-platform fleet layer; the pricing of the paid surface is unknown | The value-capture sandwich — free OSS below, platform-owner incentives above — is the single most structural risk; investors will probe whether any enterprise buyer will create a distinct budget line for a third-party fleet control plane | Prepare a concrete answer to "who creates the budget line for Console and what does it replace?" with at least one real enterprise reference to illustrate |
| Staged milestones with symmetric double-down and quit triggers | No milestone structure can be attached to specific ARR gates without pricing and design-partner count; the staged option logic is defined but not yet anchored to concrete numbers | Investors who believe in staged discipline will not commit reserves without at least a first gate defined by observable, falsifiable criteria | Publish pricing; define the 12-month milestone the company will hit to justify a follow-on conversation; state the quit condition spontaneously |
| Evidence that at least one compounding asset is already accruing | None of the three candidate compounding assets — MetalRT performance moat, installed SDK base switching costs, and fleet control-plane telemetry data — has any independent evidence of accruing | Underwriting a five-to-seven year hold requires a compounding asset that is at least beginning to form; "beginning to form" means at least one retained fleet using Console OTA across release cycles | Demonstrate at least one Console fleet that has received OTA model updates across two or more release cycles, showing switching-cost formation |
Claim Reconciliation: Inconsistencies Investors Will Catch
| Claim Or Metric | Where It Appears | Conflicting / Unreconciled Versions | Why Investors Flag It | How To Reconcile |
|---|---|---|---|---|
| Privacy and telemetry: "no analytics" vs "anonymous analytics by default" vs active telemetry code | App privacy policyRunAnywhere docsGitHub SDK repo | Three surfaces directly contradict each other; additionally, the privacy policy scope is limited to the Android sample app and does not cover the SDK, Console, or iOS app; app-store data-safety declarations are unpopulated | This is the kind of disclosure inconsistency that enterprise legal and procurement teams flag before integration; it raises the question of whether the founders optimize for narrative over accuracy under pressure; it is also a direct FTC and GDPR risk if SDK telemetry transmits data beyond the policy's stated scope | Publish a single, scope-accurate privacy policy covering the SDK, Console, and iOS app; disclose the telemetry data-flow diagram; complete app-store data-safety declarations on both platforms; and reconcile the "no analytics" and "anonymous analytics" language into a single, accurate statement |
| Headcount: YC team size 2 vs LinkedIn company profile showing 9 employees | YC profileLinkedIn company page | The discrepancy is unexplained and creates uncertainty about who owns the Console multi-tenant backend build and the enterprise GTM motion | Investors building a team thesis around a two-founder team will not discover nine employees as a positive surprise; they will ask why the discrepancy exists and whether the additional people are employees, contractors, advisors, or otherwise | Publish a current FTE list with roles and start dates; distinguish employees from contractors and advisors; confirm explicitly who owns Console backend engineering and enterprise sales |
| Open-source marketing vs. proprietary engine | The company website and homepage use "open source" framing prominently | The MetalRT binaries repo carries a "Proprietary" license with the licensor listed as founder@runanywhere.ai; no published end-user license agreement (EULA) or commercial use terms exist for MetalRT | Enterprise legal teams cannot approve integration of a runtime with undefined redistribution and commercial-use terms; the dual-license reality — free SDK, proprietary engine — is standard in open-core businesses, but needs to be stated clearly rather than obscured by "open source" branding | Publish a MetalRT EULA defining commercial use, redistribution rights, and pricing tiers; update public positioning to clearly describe the open-core model |
| "100% on-device" vs. hybrid routing with cloud destinations | RunAnywhere docs | RunAnywhere homepage and SDK page describe hybrid local/cloud routing by policy, with cloud destinations including OpenRouter and vLLM | Investors building an on-device privacy thesis will ask which claim governs when a real enterprise deploys | State clearly and consistently that on-device is the default mode and that cloud routing is an opt-in policy configuration under explicit user control; specify when — if ever — data leaves the device during hybrid routing |
Diligence Findings: Issues To Fix Before You Raise
2 deal-breaker5 high1 medium1 low
No independently verified paid demand for Console: no named paying customer, no public pricing, no production fleet, no contract or letter-of-intent value
deal-breakerFixSecure at minimum one signed, paid Console pilot at any price with a reference contact; publish pricing
Three-way self-contradiction between the privacy policy ("no analytics"), documentation ("anonymous analytics by default"), and SDK code (active telemetry pull requests); app-store data-safety declarations unpopulated
deal-breakerFixReconcile and republish a scope-accurate privacy policy; complete app-store data-safety declarations; prepare a data-flow diagram for enterprise review
MetalRT performance benchmark is company-claim-only; no independent reproduction found; MetalRT is Apple Silicon-only with a proprietary license and no published EULA; HexagonRT is unshipped
highFixDisclose benchmark methodology; publish a MetalRT EULA with commercial-use and redistribution terms; ship HexagonRT with a public Android benchmark
Founder prior-employer scale claims unverified: CEO "50M+ users" Intuit claim corroborated only at employment level (not scope or scale); CTO Azure/AWS and MetalRT authorship unverifiable; CTO identity yields an ambiguous press match
highFixOne named employer reference per founder prepared to confirm role, product, and scope on a 15-minute call; resolve the CTO identity question
No enterprise GTM coverage: neither founder shows a verified enterprise sales track record; zero public job postings; Console commercial motion is contact-driven with no pricing
highFixName the founding sales hire or demonstrate two or more enterprise design partners closed by the founders directly
Round structure and cap table unknown: $10K sold on a $1M Form D offering; YC deal terms not public; fund-returning ownership availability unconfirmed
highFixShare cap-table summary with share class, pre/post-money, and remaining allocation
Apple OTA policy risk unresolved: marketing states "manage models on thousands of devices without app-store releases"; no Apple confirmation or prior review approval found; MetalRT is entirely Apple-dependent
highFixObtain and share Apple developer documentation or prior review approval confirming OTA model-weight delivery is cleared under App Store Guideline 2.5.2
No SOC 2, ISO 27001, DPA, or enterprise security page found; Console targets enterprise fleet operations
mediumFixCommit to a SOC 2 Type I timeline; publish a security page describing Console infrastructure; prepare a data-processing agreement template for enterprise customers
No published Terms of Service for the SDK, Console, or Mirar; only the Android sample-app privacy policy is indexed
lowFixPublish a broadly-scoped Terms of Service covering the SDK API, Console, OTA model delivery, and the Mirar service
Data Room Readiness Checklist
- Procurement-readiness score:
not ready - Rationale: Console targets enterprise fleet operations buyers whose procurement gatekeepers will require a data-processing agreement, a SOC 2 or equivalent attestation, a subprocessor list, a security page, and completed app-store data-safety declarations before integration. None of these artifacts are publicly available; the only privacy document found is scoped to the Android sample app and directly contradicts the SDK's own code.
| Data Room Area | What Investors Expect | Current Read | Status | Priority |
|---|---|---|---|---|
| Company overview & corporate | One-pager, incorporation, good standing, entity structure | RunAnywhere, Inc. Delaware incorporation confirmed via SEC Form D; entity CIK and officers confirmed; corporate documents are private | partial | High |
| Financials | Monthly P&L (18–24 months), three-year model with assumptions, burn and runway | No revenues at Form D filing; no public pricing; monthly burn and runway are unknown; no financial model is public | missing | Critical |
| Cap table & funding history | Clean cap table, prior rounds, SAFEs or notes, 409A | Form D shows $1M offering with $10K sold and 2 investors; Crunchbase confirms pre-seed and YC investor; round terms, instruments, and cap table are private | partial | Critical |
| Legal & IP | Bylaws, board consents, IP assignments (founders and contractors), material contracts | No Terms of Service, no MetalRT EULA, no IP assignment evidence, no board consents visible in public sources | missing | High |
| Product & technology | Roadmap, architecture overview, security and compliance docs | SDK repo and docs are public; Console is login-gated with no architecture disclosure; no security page, no compliance attestation, no SOC 2; privacy policy contradicts code | partial | Critical |
| Team | Org chart, key employment/advisor agreements, vesting | YC team size 2; LinkedIn shows 9 employees with unresolved reconciliation; no org chart, no employment agreements, no vesting schedule public | missing | High |
| Customers & traction | Retention cohorts, ARR/MRR bridge, pipeline, 3–5 references | No named paying customers; no ARR; demo apps only; no reference available; GitHub stars are awareness proxy | missing | Critical |
| Market & competition | TAM/SAM/SOM, competitive landscape, pricing | Hardware TAM data available via public analyst summaries; software-layer SAM is a low-confidence inference; SOM unavailable (pricing unknown); competitive landscape documented | partial | Medium |
First-Call Agenda For The Startup
| Time | Topic | Founder Goal | Evidence To Bring |
|---|---|---|---|
| 0–5 minutes | Establish the fleet-control-plane thesis | State clearly that Console — not runtime speed — is the product that justifies a recurring paid subscription, and articulate in one sentence why enterprise buyers cannot get this from self-hosting llama.cpp or Apple MLX | A single positioning statement framing RunAnywhere as "the cross-platform control plane for on-device model fleets, not a faster runtime" |
| 5–20 minutes | Commercial traction: paying Console design partners | Demonstrate that at least one paying or committed enterprise customer exists and has chosen Console for fleet operations rather than benchmark performance | Signed letter of intent or pilot agreement; reference contact willing to speak to an investor; per-device or per-fleet pricing; brief use-case description naming OTA delivery, policy routing, or fleet analytics as the value driver |
| 20–30 minutes | Privacy and telemetry reconciliation | Show that the company's disclosure practices are enterprise-procurement-safe and reflect disciplined judgment | Updated, scope-accurate privacy policy; completed app-store data-safety declarations; a brief data-flow explanation suitable for enterprise legal review |
| 30–40 minutes | MetalRT differentiation and CTO biography | Disclose the benchmark methodology so an independent engineer can reproduce the test, confirm the CTO's authorship of the Metal kernels, and resolve the identity question | Hardware spec, model, and run configuration for the M4 Max benchmark; clear description of what the Metal kernels do that Apple MLX cannot; CTO willingness to give a technical walkthrough |
| 40–50 minutes | Team structure and trajectory | Resolve the headcount discrepancy and confirm that the two hardest builds — Console backend and enterprise GTM — have named, capable owners | Current FTE list with roles; explicit ownership assignment for Console backend engineering and enterprise sales; name of any enterprise-GTM hire in progress |
| 50–60 minutes | Cap table and round access | Confirm whether the fund can reach a meaningful ownership position before the Series A | YC deal terms; post-YC post-money; remaining round allocation; founders' next-financing plan and timeline |
Business Model And Revenue Signals
Open-core model: Apache-2.0 SDK (free) + proprietary MetalRT engine + Console SaaS fleet control plane
The open-core structure is clear; commercial pricing for Console and MetalRT license terms are not public
No revenues at SEC Form D filing date; $10K sold of $1M offering to 2 investors
SourceSEC Form DSnapshot at filing date; may have been updated; represents the only public revenue disclosure
Demo apps on App Store (iOS and Android) are free; no in-app purchase or subscription model visible
Demo apps only; no indication of commercial pricing layer in the app
Ollama Pro ($20/month) and Max ($100/month) cloud tiers as a pricing reference for the inference-runtime layer
SourceOllamaPricing comp for the inference layer, not the fleet control plane; the relevant pricing comparison for Console is MDM-style per-device-per-year, not per-user-per-month
Google Search and the RunAnywhere website were checked and showed no public Console pricing page. Login-gated pricing may exist; this absence carries moderate weight for investor benchmarking but does not disprove a paid layer.
Revenue and profit are not publicly available for RunAnywhere beyond the no-revenues Form D disclosure. No revenue figure, ARR, or burn rate has been publicly confirmed. Any scenario estimate is an illustrative model, not a fact.
Revenue Quality Checklist
Funding And Ownership Context
| Item | Public Read | Evidence Label | Diligence Request |
|---|---|---|---|
| Total raised / rounds | Form D: $1M offering with $10K sold to 2 investors, first sale 2025-09-25; Crunchbase confirms pre-seed with YC as investor | source-backed fact via SEC Form D; Crunchbase | Total capital raised including YC deal; updated Form D if any subsequent offering has closed |
| Instruments (equity / SAFE / notes) | Form D Rule 506(b) private placement; instrument type (SAFE, note, or equity) not disclosed | source-backed fact for the offering type; unknown for the instrument | YC SAFE terms (MFN vs. post-money SAFE); terms of any other instruments |
| Investors | Y Combinator (confirmed via Crunchbase and YC profile); 2 investors as of Form D filing | source-backed fact via YC profile; Crunchbase | Whether any angels or non-YC investors participated; whether the round is still open |
| Post-money valuation | Not public | unknown | Post-YC close post-money valuation and cap table summary |
| Cap table / ownership | Not public; founder ownership likely pre-dilution at near-100% before YC; YC standard terms include equity or SAFE but specific terms are private | unknown | Cap table summary with share class, founder dilution, YC ownership, and remaining allocation |
| Burn & runway | Not public | unknown | Current monthly burn, runway in months, and next financing milestone |
| Next-round plan | Not public; YC batch implies Demo Day fundraising cycle in mid-2026 as the likely near-term trigger | inference | Founders' target raise amount, target valuation, and whether the round is actively open |
Founder, Team, And Related Entities
Founder And Team
Sanchit Monga
Co-Founder & CEO; named executive officer on Form D; active GitHub contributor to RunAnywhere organization repos
Evidencesource-backed fact for identity and role; company claim for prior-employer scale
Confidencehigh for identity; low for scale claim
Shubham Malhotra
Co-Founder & CTO; named officer on Form D; approximately 1,887 GitHub contributions in the past year heavily weighted to the RunAnywhere organization
Evidencesource-backed fact for identity, role, and build velocity; company claim for Azure/AWS and MetalRT authorship
Confidencehigh for identity and velocity; low for biography claims
LinkedIn employment history: both founders
Both founders' LinkedIn employment history fields were not populated in public dataset captures; employment claims on the About page and YC profile cannot be independently verified from this source alone
Evidencesource-backed fact for the absence of employment data
Confidencehigh (that the data is absent)
Inc42 Datalabs partial corroboration
A third-party employment aggregator (Inc42 Datalabs) independently corroborates Sanchit Monga's Intuit employment, listing software engineering roles from 2021 through 2025 and RIT education; the "50M+ users" scale or mobile-SDK-leadership claims are not corroborated by this source
Inc42 Datalabs (third-party aggregator, URL not publicly citable; source not openly accessible for independent link-out)Evidencesource-backed fact for Intuit employment at junior individual-contributor level; company claim remains for scale
Confidencemedium for employment; low for scale
CTO press identity: ambiguous match
A Business Insider profile of a "Shubham Malhotra" with an RIT educational background and Amazon/Microsoft employment history describes that individual as currently employed at Amazon as of December 2025, creating an unresolved identity question relative to full-time CTO status at RunAnywhere
Business Insider (third-party press, URL not publicly citable due to identity ambiguity)Evidenceinference (ambiguous match; identity not confirmed)
Confidencelow
YC team size 2 vs LinkedIn company showing 9 employees
YC profile lists team size as 2; LinkedIn company dataset shows 9 employees in the 2–10 band; the reconciliation between these signals is unknown
Evidencesource-backed fact (conflict)
Confidencehigh for the conflict; unknown for reconciliation
Founder Competency Coverage
Domain depth
- Sanchit Monga
- claimed
- Shubham Malhotra
- claimed
What The Evidence IsSanchit claims Intuit mobile SDK experience; Intuit employment corroborated at junior IC level by a third-party aggregator; "50M+ users" scale is not corroborated. Shubham claims Microsoft Azure and AWS EC2 Spot infrastructure experience; LinkedIn employment history not available; an identity-ambiguous press match does not confirm the claim.
Technical build capability
- Sanchit Monga
- evidenced
- Shubham Malhotra
- evidenced
What The Evidence IsGitHub repository history confirms active contributions from both founders at the RunAnywhere organization; Shubham's approximately 1,887 contributions in the past year on closed and open repositories is independently verifiable. This is evidence of build velocity, not of the specific depth required for MetalRT-level kernel work.
Product
- Sanchit Monga
- claimed
- Shubham Malhotra
- unknown
What The Evidence IsSanchit claims mobile-SDK product leadership at Intuit; individual scope unverified. No independent evidence of product ownership for Shubham beyond the RunAnywhere SDK itself.
GTM / sales
- Sanchit Monga
- absent
- Shubham Malhotra
- absent
What The Evidence IsNo verified enterprise sales background for either founder. Commercial motion is contact-driven with no pricing visible. No enterprise-GTM hire is visible in public job postings.
Leadership / hiring
- Sanchit Monga
- unknown
- Shubham Malhotra
- unknown
What The Evidence IsYC team size 2; headcount discrepancy unresolved; no org chart or evidence of team management visible in public sources.
Fundraising history
- Sanchit Monga
- evidenced
- Shubham Malhotra
- evidenced
What The Evidence IsYC W26 acceptance and Form D confirmed; both founders co-led the pre-seed raise successfully.
Prior founding outcomes
- Sanchit Monga
- absent
- Shubham Malhotra
- absent
What The Evidence IsNo prior ventures independently verified for either founder; Crunchbase and SEC records show no prior company under either name.
Low-level proprietary inference kernel (MetalRT) competitive with platform-owned free alternatives
HighTeam coverage todayShubham:
claimed— authorship unverified; no independent technical validation of Metal kernel depth exists outside the closed proprietary binaryWhat would close itIndependent MetalRT benchmark reproduction and CTO authorship confirmed via a technical walkthrough with a Metal GPU expert
Multi-tenant enterprise control plane (Console) with fleet OTA and analytics at production grade
CriticalTeam coverage todayBoth founders:
unknown— Console is login-gated with no architecture review or multi-tenant deployment evidence availableWhat would close itConsole architecture walkthrough; confirmation of multi-tenant production readiness serving independent enterprise fleets
Enterprise go-to-market and sales execution
HighTeam coverage todayNeither founder:
absent— contact-driven motion only; no enterprise-GTM hire visible in public job postingsWhat would close itA named founding-sales hire with a reference, or two or more enterprise design partners signed by the founders directly, demonstrating independent enterprise sales execution
Public Professional Footprint
These checks of public technical artifacts, professional social presence, and education and credential traces carry low decision weight by design — they corroborate or weaken the competency coverage above and sharpen the questions worth raising on a founder call; none of them alone changes the verdict.
Technical artifacts (repo level)
- Sanchit Monga
- GitHub profile (github.com/sanchitmonga22) shows RunAnywhere organization repos pinned alongside a personal project (AllInVault); SDK repo contributions confirmed through the organization history
- Shubham Malhotra
- GitHub profile (github.com/shubhammalhotra28) shows approximately 1,887 contributions in the past year, heavily weighted to RunAnywhere organization repos; a whisper-web fork is pinned; no original personal repo of sufficient depth independently confirms MetalRT-level Metal kernel work outside the closed-source binary
What The Public Record ShowsBoth founders are actively contributing at an unusual density for a two-person pre-seed team. The MetalRT proprietary binary exists on GitHub but is closed-source; authorship cannot be inspected from the public repository
Professional social content
- Sanchit Monga
- LinkedIn posts visible; posts reference RunAnywhere milestones and YC batch activity; the "50M+ users" scale claim originates on the company bio and YC profile, both authored by the founders — repetition across these surfaces is not independent validation
- Shubham Malhotra
- LinkedIn posts visible; a "5,000+ developers" hackathon adoption claim appears in a LinkedIn post by Shubham; no independent event-organizer or participant-count data found corroborating this figure; IEEE/IJARIIT publication list appears on the GitHub profile as a self-reported link
What The Public Record ShowsSelf-authored distribution of claims with no cross-venue contradictions in substance. Claims identified as originating in founder-controlled posts are labeled
company claimthroughout this report.Education / credentials
- Sanchit Monga
- Rochester Institute of Technology (RIT) affiliation corroborated via the Inc42 Datalabs profile cross-reference; degree completion not independently confirmed
- Shubham Malhotra
- RIT affiliation appears in the identity-ambiguous Business Insider profile match and the Inc42 cross-reference; degree completion not independently confirmed; the Business Insider match creates an unresolved identity question
What The Public Record ShowsEducation affiliation to RIT has multi-source corroboration; degree status is not independently confirmed for either founder
Publications / patents / certifications
- Sanchit Monga
- No patent filings or independently archived publications found in public record searches
- Shubham Malhotra
- IEEE/IJARIIT publication list cited on the GitHub profile; content not independently retrieved or verified; a Google Scholar profile link appears on GitHub but the target page was not retrieved
What The Public Record ShowsNo independently verified publications for either founder; Shubham's self-reported IEEE/IJARIIT citations remain
claimeduntil the DOI-resolved papers are reviewed
Net read: The footprint confirms two technically active founders building in public at an unusual velocity, but it does not independently verify the scale claims, the prior-employer biographical anchors, or the MetalRT authorship that anchor the jockey thesis. The verdict gate does not move on footprint evidence alone.
Founder-Market Fit Read
- What is promising: The CTO's GitHub contribution density (~1,887 GitHub contributions per year in the past year) is independently observable and is the most evidenced signal in the founder portfolio. The CEO's Intuit employment is now independently corroborated at the employment level (though not at the scale claimed). Building a closed low-level runtime, a stable five-target cross-platform SDK, a fleet control plane, and a second product on minimal disclosed capital is the kind of velocity that survives a wedge being commoditized — teams that ship this fast can re-aim.
- What is missing: The three most load-bearing competency cells — CEO product scale at Intuit ("50M+ users"), CTO MetalRT authorship and Azure/AWS infrastructure depth, and enterprise GTM coverage for either founder — remain unverified. The hardest required build (multi-tenant enterprise control plane) has no named owner with independently verified experience. Enterprise sales competency is entirely absent from the public record.
- What to prepare: One named employer reference per founder, prepared for a 15-minute investor call confirming role, specific product, and an honest description of individual contribution scope. An explicit answer to "who owns Console backend engineering and enterprise sales" with a named person or hire. A technical walkthrough of MetalRT kernel work that a Metal GPU expert can evaluate as novel versus orchestration-level.
Related Entities And Founder-Associated Companies
RunAnywhere, Inc. (Delaware; CIK 0002092245)
Operating entity of the company; sole legal entity found in public records
Evidencesource-backed fact
Confidencehigh
Follow-upNo follow-up needed; entity confirmed
Mirar (vision agent)
Second product under the RunAnywhere brand, in early access on a Vercel staging host; uses a design-partner onboarding model
Confidencehigh
Follow-upEngineering focus and monetization intent are unclear; concurrent active development of Mirar alongside Console is a product-focus risk; investors should ask whether Mirar competes with Console for the CTO's engineering attention
AllInVault
Personal GitHub project pinned by Sanchit Monga; not a separate legal entity
Evidencesource-backed fact
Confidencemedium
Follow-upEngineering-scope indicator only; not a separate company
Traction
Customer Status Table
No named customers appear anywhere in the public evidence — no company names, no case studies, no App Store references to production apps using the RunAnywhere SDK, and no customer-controlled public documentation. The strongest signals are developer community awareness proxies (GitHub stars and HN launch), not commercial traction.
| Customer | Strongest Evidence | Customer-Controlled? | Status | Paid? | Seats | Renewal | Confidence |
|---|
The RunAnywhere website, Apple App Store, Google Play, and Crunchbase were checked and showed no named paying customer, pilot, design partner, or production deployment. These absences carry high weight for commercial traction because enterprise buyers and investors expect at least one customer-controlled reference at this stage.
Traction Signals
~10,300 GitHub stars on the open-source SDK; 360 forks; 3,438 commits; Apache-2.0 license
SourceGitHub SDK repoStars are developer awareness, not production adoption; automation accounts inflate contributor count; production integrations by non-RunAnywhere developers are unknown
HN YC W26 launch: 240 points, 153 comments; community controversy including spam and outreach-ethics allegations; HN moderator clarified launch placement
Community interest is genuine; the controversy is a trust signal risk in the developer community that should be addressed proactively
iOS sample app: 3 ratings at 5.0 stars; free; Developer Tools category; version 0.30
SourceApple App Store listingSample/demo app only; not a production product; minimal ratings
Android sample app: approximately 956 US downloads; 3 reviews; free; Tools category
SourceGoogle Play listingUS downloads only; sample app; provider-estimated count; not a production deployment
Website traffic: approximately 5,220 monthly visits; approximately −61% month-over-month change (Semrush estimate, April 2026)
SourceCrunchbaseStale estimate (April 2026); directional only; the decline may reflect post-YC-launch fade
Google Search (US) AI assistant visibility: ChatGPT recommends ONNX Runtime, ExecuTorch, Firebase AI Logic, MediaPipe, and Core ML as primary choices for on-device hybrid AI SDK; RunAnywhere blog cited in sources but not in primary recommendations
SourceGoogle Search (US) and ChatGPT answer snapshots — snippet-grade, not in Source LogT4 AI-visibility signal only; does not prove category standing; cited in sources means awareness, not recommendation
Apple App Store and Google Play were checked and showed no production apps beyond RunAnywhere's own sample apps crediting the SDK. These absences carry little weight for early-stage infrastructure companies where adoption may be embedded in private enterprise builds.
Hiring And Org Momentum
LinkedIn company page
SourceLinkedIn companyHeadcount growth signal, but unreconciled with YC team size of 2; FTE vs contractor vs advisor breakdown unknown
Indeed, Glassdoor, LinkedIn Jobs, and the RunAnywhere careers page were checked and showed no matching listing. These absences carry little weight since job boards are often incomplete, and employee sentiment remains unavailable.
- Role-mix read: No public postings exist to interpret role mix; the only public signal is the LinkedIn headcount figure of 9, which cannot be broken down by role from available public data.
- Momentum read:
unknown— headcount may be growing (LinkedIn feed shows hire activity) but no public postings confirm scale, role mix, or velocity. The absence of any public job postings at the same time as a LinkedIn count of 9 is unexplained.
Traction Quality Read
| Traction Dimension | Current Status | Good Enough For First Call? | Needed For Deep Diligence |
|---|---|---|---|
| Developer adoption (open-source SDK) | Approximately 10,300 GitHub stars and a 240-point HN launch; community enthusiasm is genuine and unusual for a pre-seed two-person team | Yes — developer awareness is real and above average for stage and team size | Evidence of non-sample-app SDK integrations in production; confirmation that stars translate into developer community pull, not just awareness |
| Commercial traction (paid Console) | Zero: no pricing, no named customer, no production fleet, and no contract value in any public source | No — the paid layer is entirely unvalidated | One named, paying Console design partner with a reference call available, a price, and a retention signal |
| App store and consumer signals | Three iOS ratings and approximately 956 US Android downloads — both for sample/demo apps, not a production product | No — sample-app metrics only | A production app (not the sample app) in any app store with meaningful usage metrics that attributes itself to the RunAnywhere SDK |
| Hiring and org momentum | Zero public job postings across LinkedIn Jobs, Indeed, Glassdoor, and the company careers page | No — no observable hiring scale signal | At least one public job posting for an enterprise-GTM or Console-backend engineering role would signal commercial intent and execution capacity |
Competitive Landscape
| Segment | Examples | Customer Alternative | Pressure On Company |
|---|---|---|---|
| Platform-owned free runtimes | Apple MLX (~27,100 stars), ExecuTorch (Meta/PyTorch, 12+ hardware backends), Google LiteRT (TFLite successor) | Each provides a free, actively maintained runtime optimized for its own hardware platform | Platform owners are directly incentivized to maintain these as free alternatives; Apple MLX competes directly with MetalRT on the same Apple Silicon hardware and has approximately 2.6× more GitHub stars; the strongest bear case in this report rests on this segment |
| Open-source inference community | llama.cpp (~117,000 stars, MIT license), ONNX Runtime (cross-platform, enterprise logo wall) | Free, cross-platform, large active communities; RunAnywhere's SDK uses both as backends | RunAnywhere depends on these tools for its SDK functionality; enterprises can self-host the same free stack without paying for Console; llama.cpp has approximately 11× more GitHub stars than the RunAnywhere SDK |
| Hybrid on-device/cloud peers | Cactus Compute (YC W26, ~5,300 stars self-reported, cross-platform SDK + hybrid routing) | A direct peer from the same YC batch with a similar value proposition | Two YC W26 companies with similar positioning narrows the time window before one is established as the category reference; Cactus's hybrid routing story may appeal to the same enterprise buyer RunAnywhere targets |
| Cloud AI gateway layer | LiteLLM (YC-backed, 100+ LLM providers, OSS and enterprise), Portkey (acquired by Palo Alto Networks) | Cloud-side routing and observability for teams that have not committed to on-device inference | Portkey's acquisition by Palo Alto Networks signals consolidation in the adjacent cloud AI gateway layer; LiteLLM's scale (self-reported 240M+ Docker pulls) shows the cloud routing category is maturing |
Category Visibility Snapshot
Google Search (US)on-device AI inference SDK competitors
- Captured page-one results (2026-06-17)
- RunAnywhere blog post ranks #1; Cactus Compute ranks #2; infrabase.ai ranks #3
- Was the company present?
- Yes — a RunAnywhere-authored blog post about its own competitors holds the top organic position; self-reinforcing share of voice
Google Search (US)on-device AI SDK hybrid routing local cloud
- Captured page-one results (2026-06-17)
- Firebase tutorials, Microsoft framework tutorials, and a service named Xybrid appear prominently; RunAnywhere absent from the top 10 results
- Was the company present?
- No — RunAnywhere does not appear when the query frames the category as hybrid routing
ChatGPT AI assistant (US)best on-device hybrid mobile AI SDK
- Captured page-one results (2026-06-17)
- Primary recommendations: ONNX Runtime, ExecuTorch, Firebase AI Logic, MediaPipe, Core ML; RunAnywhere blog cited in sources section but not featured in the primary recommendation list
- Was the company present?
- Partial — appears in sources, not in recommendations
- Visibility read (
inference, low confidence): RunAnywhere owns the SERP for its own framing ("on-device AI inference SDK") through a self-authored blog post, but is absent when the query frames the category as hybrid routing — which is the framing the company's own marketing sometimes uses. AI-assistant recommendation engines currently favor established platform tools (ONNX Runtime, ExecuTorch, Firebase) over RunAnywhere. This is consistent with early-stage awareness: real within the YC and developer-curiosity orbit, not yet established as the category reference. Visibility is not market share; this read carries low confidence.
Pricing And Competitive Benchmark
| Alternative | What It Offers | Public Price Signal | Price Vs. This Company | Evidence Label |
|---|---|---|---|---|
| llama.cpp | Free, MIT-licensed LLM inference; RunAnywhere SDK uses it as the GGUF backend | Free | RunAnywhere SDK is free; the only pricing comparison will be Console vs. self-hosting llama.cpp for free | source-backed fact via llama.cpp GitHub |
| Apple MLX | Free, Apple-research-maintained array framework; Apple Silicon only | Free | MetalRT claims a performance edge over MLX; the price comparison is a free alternative vs. a proprietary binary with no published license terms or pricing | source-backed fact via Apple MLX GitHub |
| Ollama | Local LLM inference; developer-facing | Free (local); Pro $20/month; Max $100/month (cloud-enhanced tiers) | The Ollama commercial tier benchmarks the inference-layer pricing ceiling; Console's per-device-per-year fleet-management pricing would need to justify a different value axis (operational scale vs. individual-user inference) | source-backed fact via Ollama |
| ExecuTorch | Free, Meta/PyTorch sponsored; 12+ hardware backends | Free | Platform-sponsored free substitute targeting the same cross-platform inference market | company claim for production-scale claims via ExecuTorch |
| Google LiteRT | Free, Google-maintained TFLite successor; cross-platform | Free | Platform-owned free substitute deployed at scale across Google's Android ecosystem | company claim for deployment scale claims via Google LiteRT |
| ONNX Runtime | Free, cross-platform; enterprise logo wall | Free (open-source) | RunAnywhere SDK uses ONNX Runtime for speech-to-text and text-to-speech workloads; enterprise teams can use ONNX Runtime directly for free | source-backed fact via ONNX Runtime |
- Price positioning read: RunAnywhere has no public price for Console, so direct price positioning is currently unknown. The entire competitive benchmark is "free vs. free vs. no price." The competitive pricing argument only becomes relevant once Console pricing is published; without that, investors and enterprise buyers cannot assess value relative to self-hosting the same free stack. Publishing a pricing page is one of the highest-leverage investor-readiness steps the company can take.
- Price claims to correct or substantiate: The MetalRT benchmark on the engines page claims specific throughput numbers compared with llama.cpp, MLX, and Ollama. These claims are
company claimand have not been independently reproduced; if a third party reproduces them and finds them inaccurate or methodology-dependent, the pricing premium narrative collapses. The company should disclose the benchmark methodology before any investor or enterprise buyer reproduces it independently and finds a discrepancy.
Competitive Wedge
inference, not source-backed fact.The runtime performance wedge (MetalRT) is more fragile: it is Apple-only, unreproduced, and directly contested by a platform-owned free alternative (Apple MLX) on the same hardware. The cross-platform SDK abstraction wedge is the thinnest of the three — a thin convenience layer over free backends that any developer can self-assemble.
If the fleet-control-plane white space validates through enterprise buyer interviews, it is the one non-copyable asset: neutrality is structural and cannot be replicated by any platform owner without abandoning its platform lock-in strategy. If it does not validate — if enterprise buyers build their own fleet management on top of MDM tooling or if a platform owner ships cross-platform fleet management when enterprise demand materializes — the wedge does not exist.
Risks And Open Questions
| Risk | Severity | Evidence | What To Ask |
|---|---|---|---|
| Enterprise buyers self-host the free stack and no Console budget line forms | deal-breaker | SAM of $100M–$1.0B is a low-confidence inference; zero paying customers in public record | On the founder call: what enterprise has evaluated Console and committed, even verbally, to paying if it reaches a specific milestone? |
| Privacy/telemetry self-contradiction and unpopulated app-store data-safety labels | deal-breaker | source-backed fact — three-way contradiction between privacy policy, documentation, and SDK code | On the founder call: walk through the exact data-flow and reconcile each surface; what is the timeline for updated declarations? |
| MetalRT moat limited to Apple Silicon and unreproduced; HexagonRT unshipped | high | source-backed fact — MetalRT is Apple-only and proprietary; HexagonRT binary does not exist in public repositories | In technical diligence: reproduce the MetalRT benchmark on independently selected hardware; obtain a dated HexagonRT milestone commitment |
| Apple platform dependence: MetalRT, OTA policy, and Metal API access all at Apple's discretion | high | source-backed fact — MetalRT requires Apple Silicon; OTA model delivery invites App Store Guideline 2.5.2 scrutiny | On the founder call: what is the fallback if Apple restricts Metal access or changes OTA policy? |
| Founder prior-employer scale claims unverified; CTO identity-ambiguous press match | high | source-backed fact for the LinkedIn employment absence and Business Insider identity question; company claim for the scale and MetalRT authorship | On the founder call: reference calls for both founders; CTO identity resolution |
| Value-capture sandwich: free OSS below (including the company's own dependencies) and platform-owner runtimes above | high | source-backed fact — SDK uses llama.cpp and ONNX Runtime; platform-owned free substitutes confirmed | In enterprise buyer interviews: are buyers willing to fund a third-party fleet control plane, or will they self-host the free stack? |
| Cactus Compute (same YC batch, similar positioning) may occupy the category before RunAnywhere establishes Console traction | medium | source-backed fact for Cactus's existence and YC backing via Cactus Compute | On the founder call: what is the positioning differentiation from Cactus, and has RunAnywhere signed any design partner that Cactus is also pursuing? |
| Product focus diluted across MetalRT, SDK, Console, Mirar, and a technical newsletter at two founders | medium | source-backed fact for multiple simultaneous products via GitHub and website; inference for focus risk | On the founder call: which product gets engineering focus this quarter; what has been paused or deprioritized? |
Pre-Mortem: The Most Likely Obituary
- Cause of death (one sentence): RunAnywhere builds a widely-used free SDK that enterprises self-host forever, the Console paid layer never signs a paying customer, and the company runs out of runway before the enterprise software timing cycle matures.
- The causal chain (3-5 steps from today to the shutdown): First, enterprise interest focuses on the SDK benchmarks and the YC pedigree; developer teams integrate the free SDK. Second, DevOps teams evaluate Console but conclude they can wrap llama.cpp, Apple MLX, and ONNX Runtime themselves using existing MDM tooling at near-zero incremental cost. Third, Apple MLX achieves performance parity with MetalRT on Apple Silicon as Apple continues investing in free developer tooling, eroding the performance differentiation narrative. Fourth, HexagonRT ships too late — after Cactus or a platform-owned alternative captures the Android fleet-ops story — and RunAnywhere remains Apple-only into a market that has already defined its preferences. Fifth, Console waitlist converts zero paying customers after an active design-partner push, the founders cannot demonstrate commercial viability to raise a Series A, and the company shuts down or becomes a community-maintained open-source project without a commercial entity behind it.
- The earliest observable warning sign: An active design-partner push of more than 90 days with zero paid or committed customers — this is not a timing problem, it is structural evidence that no enterprise budget line for a third-party fleet control plane will form.
- The question that defuses this chain today: Who is the first enterprise willing to sign a paid Console pilot, at any price, and can they articulate why fleet operations — not runtime speed — justify paying above the free self-hosted stack?
Decision-Critical Unknowns
| Unknown | Why It Is Decision-Critical | Best Evidence | Decision Effect |
|---|---|---|---|
| Whether the absence of a paying Console customer is a timing problem or a structural problem | This is the binary IC question: if timing, a pre-seed option check is appropriate; if structural, the value-capture route is closed regardless of product quality | Enterprise buyer interviews (3–5 calls with mobile platform or DevOps leads at mid-market and large companies); or at least one enterprise that evaluated Console and would sign at a specific milestone | Resolves the hold toward either pursue or pass; no other single piece of evidence has this impact |
| Whether MetalRT's performance edge over Apple MLX is independently reproducible and durable | The differentiation moat rests entirely on this claim; if unreproducible, the proprietary surface collapses to a commodity | A third-party benchmark on independently selected Apple Silicon hardware, disclosed model and methodology, measuring the gap over Apple MLX and llama.cpp on the same hardware | If the edge is durable, it supports the paid Console thesis; if it is equivalent to MLX, the bear case is confirmed |
| Whether the fund can reach 8–15% ownership at a pre-seed-compatible check size | Even if the company executes perfectly, a below-target stake in a contested capture scenario may not produce a fund-returning return | Cap-table summary from the founders: post-YC post-money, remaining allocation, and pro-forma at the fund's target check size | If accessible, the option check is financeable; if the round is fully subscribed or priced out of fund-returning ownership, the decision is pass regardless of company quality |
| What the SDK telemetry transmits and whether it is consistent with any reasonable reading of the privacy policy | A material undisclosed data transmission is an integrity disqualifier per the fund's stated criteria | A clear data-flow diagram showing exactly what fields are collected, where they are sent, and what consent mechanism governs them; a reconciled, scope-accurate privacy policy | If the telemetry is benign and the contradiction is a documentation oversight, it is fixable; if the telemetry transmits identifiable or behavioral data without disclosed consent, it is a pass condition |
Diligence Questions
First Call
| Question | Why It Matters | Good Evidence |
|---|---|---|
| Who are your current paying Console customers or committed design partners — name, use case, device count, and the price you are charging? | The single question that unlocks all five investor perspectives simultaneously | Named customer (even on background), signed pilot agreement or letter of intent, per-device pricing, and a reference contact available for a 15-minute call |
| Reconcile the telemetry and privacy contradiction: what does the SDK transmit by default, to which endpoint, and under what consent mechanism? | Enterprise procurement teams will block integration until this is resolved; it is also a potential regulatory liability | A plain-language data-flow description; a reconciled, scope-accurate privacy policy; completed app-store data-safety declarations live on both platforms |
| Sanchit — what specific Intuit product did you personally ship and what portion of the "50M+ users" scale is attributable to your individual work versus the team? | The CEO's domain-depth anchor is the foundation of the founder-market-fit thesis | Named product, an honest description of individual vs. team contribution, and a manager or peer reference willing to confirm the scope |
| Shubham — what did you build at Microsoft Azure and AWS EC2 Spot, and did you personally write the MetalRT Metal kernels, or does MetalRT primarily call Apple's Metal API on top of the free backends the SDK already uses? | The CTO's technical depth and the differentiation moat both rest on this answer | Named Azure or AWS system, a description of what the Metal kernels do that Apple MLX cannot, and willingness to give a technical walkthrough |
| How many production applications beyond the sample apps use the SDK today? | GitHub stars are awareness, not adoption; one confirmed production app changes the traction narrative | An App Store identifier or app name, even under NDA; approximate active inference calls per day or active users |
Follow-Up
| Question | Why It Matters | Good Evidence |
|---|---|---|
| Commission an independent MetalRT benchmark: provide hardware, model, and run configuration so a third party can reproduce the test on independently selected Apple Silicon hardware | The differentiation moat rests entirely on a company-controlled claim; independent reproduction is the minimum bar for technical diligence | Third-party benchmark showing a statistically significant, methodology-reproducible edge over Apple MLX and llama.cpp on the same hardware |
| Enterprise buyer interviews: three to five calls with mobile platform or DevOps leads at mid-market and enterprise companies building on-device AI — do they have a budget line for fleet ops, or do they self-host the free stack? | This answers the binary timing-vs-structural question that the committee cannot resolve from public evidence alone | Two or more of five interviewed buyers describe a concrete budget line for on-device fleet management separate from existing MDM spend; at least one willing to sign an LOI |
| Cap-table and round-structure review: post-YC post-money, remaining allocation, YC instrument type, and pro-forma at the fund's target check size | The fund cannot model a fund-returning outcome without ownership data | Founders share cap-table export; fund confirms it can reach 8–15% at a check inside $0.5M–$5M |
| Console architecture walkthrough: is Console production-grade multi-tenant or a demo-grade single-tenant prototype? | The difference between "waitlist demo" and "production-grade multi-tenant" is the difference between a vision and a product | A multi-tenant architecture serving at least 10 independent fleets without data commingling; a credible path to enterprise-grade service-level agreements |
| Founder employment reference calls: one named reference per founder, confirming role, specific product, and scope | Converts the jockey thesis from claimed to evidenced; the most flip-convertible positive in the report | A manager or peer reference for Sanchit confirming the named Intuit product and individual contribution scope; identity and departure date resolution for Shubham |
Kill Criteria
| Kill Criterion | Evidence That Would Trigger It |
|---|---|
| Active design-partner push of 90+ days with zero paying or committed customers and no credible price or packaging — structural value-capture failure, not timing | Founders confirm Console has no paying customers and no committed design partners after an active push, with no draft pricing and no enterprise buyer expressing willingness to pay at any milestone |
| MetalRT performance edge confirmed as equivalent to Apple MLX on independently selected Apple Silicon hardware | A third-party benchmark on disclosed hardware and methodology shows MetalRT and Apple MLX within measurement noise |
| SDK telemetry confirmed to transmit user-identifiable or behavioral data inconsistently with the privacy policy without disclosed consent | Technical review of the telemetry registry reveals data beyond "anonymous and aggregate" being transmitted to company endpoints without user-facing consent |
| Apple notifies RunAnywhere or publicly signals that OTA model delivery under the current mechanism is noncompliant with App Store Guideline 2.5.2 | An App Store rejection citing this mechanism; public Apple developer documentation excluding model-weight delivery from the data/code exemption |
| A platform owner (Apple, Google, or Qualcomm) announces first-party cross-platform fleet management or OTA model delivery | Any public roadmap announcement or developer documentation from Apple, Google, or Qualcomm describing a cross-platform on-device model fleet management tool |
Double-Down Criteria
| Double-Down Criterion | Evidence That Would Justify More Diligence |
|---|---|
| One named, retained, paying Console design partner confirmed — any price, any fleet size — with a reference call available | A signed pilot agreement or letter of intent; a reference contact willing to speak on the buying reason (fleet operations, not runtime speed); at least one OTA model update delivered to the fleet |
| Independent MetalRT benchmark reproduction showing a durable, methodology-reproducible performance edge over Apple MLX and llama.cpp on disclosed Apple Silicon hardware | A third-party benchmark with disclosed hardware, model, and methodology that reproduces the performance edge and confirms it is not within methodology noise |
| Cap-table path to 8–15% initial ownership at a check inside $0.5M–$5M confirmed by founders | Cap-table summary showing remaining allocation and a post-YC post-money valuation at which the fund's target stake is achievable without adverse dilution dynamics |
| HexagonRT ships with a public Android benchmark within 6 months | GitHub release tag with binary plus a benchmark post with disclosed hardware (named Qualcomm Snapdragon device) and model, showing meaningful performance relative to ONNX Runtime |
Founder Action Plan
| Timeframe | Action | Output |
|---|---|---|
| Before next investor conversation | Commit one existing waitlist contact to a signed, paid Console pilot at any price — even $500 per month for a 3-month pilot covering a real device fleet | Signed letter of intent or pilot agreement; reference contact willing to speak to an investor on background or on record; a brief use-case description naming fleet operations as the value driver |
| Before next investor conversation | Reconcile the telemetry/privacy contradiction: publish a corrected, scope-accurate privacy policy covering the SDK, Console, and iOS app; complete the Apple App Store and Google Play data-safety declarations | Live, accurate privacy policy URL; completed data-safety declarations visible in both store listings |
| Within 30 days | Publish a Console pricing page or pricing sheet with at least a "starting from $X per device per month" figure | A public or shared pricing document; a price entered into the company's pipeline tracking |
| Within 60 days | Reframe all external positioning to lead with "the cross-platform control plane for on-device model fleets" and retire "fastest on-device runtime" as the primary marketing claim | Updated homepage and YC profile; a single, consistent positioning statement the founders deliver without prompting |
| Within 60 days | Publish a MetalRT EULA or developer license with commercial-use terms, redistribution policy, and pricing tiers | A live, accessible MetalRT EULA URL linked from the GitHub repository and the product website |
| Within 90 days | Ship HexagonRT first binary for at least one Qualcomm Snapdragon device with a reproducible, methodology-disclosed benchmark compared with ONNX Runtime on the same hardware | Public GitHub release tag with binary; a benchmark post with disclosed hardware and model |
| Within 90 days | Obtain one external employer reference for each founder confirming role, specific product, and scope | Named reference contacts for each founder, prepared for a 15-minute investor call |
| Within 90 days | Pause Mirar active development and concentrate engineering on Console production-hardening and HexagonRT | A public or shared roadmap showing the focus decision; reduced Mirar surface on the website |
| Within 6 months | Achieve three signed, paying Console design partners; publish public pricing | A pipeline dashboard; a pricing page with defined tiers; a reference-ready customer case study, even under NDA |
Decision
- Screen: hold
- Confidence: medium
- Rationale: No confirmed fatal flaw exists in the public record — the category is real, the build velocity is evidenced, the YC W26 pedigree is a weak positive signal, and the cross-platform fleet control-plane thesis is architecturally defensible. Equally, no confirmed foundation for a first check exists: every load-bearing positive claim (benchmarks, founder biography at scale, production demand, the Console paid layer) rests on company-controlled surfaces or is entirely absent from the public record. The hold is a cheap pre-seed option, not a conviction bet. It preserves the first-call option without committing reserves.
- What would move this to pursue: At least one retained, paying Console design partner confirmed at a disclosed price — any price, any fleet size — with a reference call available; a MetalRT methodology disclosure enabling independent benchmark reproduction; and a cap-table path to 8–15% initial ownership at a fund-compatible check size. All three together in a single founder call would move the IC toward pursue.
- What would move this to pass: Confirmation of any one of the five kill criteria — no paying customers after an active design-partner push (structural value-capture failure), MetalRT benchmark reproduced as equivalent to Apple MLX, Apple policy noncompliance on OTA model delivery, a platform owner shipping first-party cross-platform fleet management, or an integrity break in which telemetry transmits identifiable data inconsistently with the privacy policy.
- Recommended next step: Schedule a first founder call, targeted within 30 days, carrying the critical diligence trifecta: a paying Console design partner or a credible near-term commitment; a MetalRT methodology disclosure; and a reconciled telemetry/privacy statement. That call is the gate between a first-check option and a pass.
- Founder preparation standard: The company should walk into the next investor conversation with three artifacts prepared: (1) a signed pilot agreement or letter of intent from a paying Console customer, even under NDA, with a reference contact; (2) a live, reconciled privacy policy and completed app-store data-safety declarations; and (3) a MetalRT benchmark methodology document that an engineer could use to reproduce the test. Without these three, every investor conversation will return to the same unanswerable questions.
How We Would Miss This One
- The miss scenario: The cross-platform fleet control-plane thesis validates faster than anticipated — enterprise buyers do create a distinct, recurring budget line for on-device model fleet management, the MetalRT and HexagonRT performance moat reproduces independently, and the CTO's Metal kernel work turns out to be a genuinely novel contribution rather than an orchestration layer. The evidenced build density, which the committee's structural analysis may have over-weighted relative to the market-structure concerns, turns out to be the leading signal of a team that can re-aim quickly when the wedge matures. RunAnywhere becomes the category default before the committee schedules a follow-on call.
- Flip conditions: One retained, paying Console enterprise fleet confirmed with a reference call available; an independent MetalRT benchmark showing a durable edge over Apple MLX; a cap-table path to 8–15% ownership at a first check inside $0.5M–$5M. If all three arrive in a single founder call, the hold flips to pursue.
- Revisit trigger / date: First founder call within 30 days; revisit immediately on any of the following external signals: a paying Console customer announcement (even under NDA to the fund), any independent MetalRT benchmark result (positive or negative), a HexagonRT GitHub release tag, a platform owner announcing first-party cross-platform fleet management (flip to pass), or a financing event that resolves the ownership path question.
Source Log
RunAnywhere homepage
runanywhere.aiProduct positioning, OTA claim, routing diagram
RunAnywhere About page
runanywhere.aiFounder bios, employer claims, YC batch
RunAnywhere Inference Engines page
runanywhere.aiMetalRT benchmarks (company claim), HexagonRT roadmap claim
RunAnywhere SDK page
runanywhere.aiSDK scope, Console features, OTA claim, absence of pricing
RunAnywhere Documentation
runanywhere.ai"100% on-device" claim, "anonymous analytics" claim
RunAnywhere app privacy policy
runanywhere.ai"No analytics" claim; scope limited to Android sample app
RunAnywhere GitHub SDK repository
github.comStars, commits, Apache-2.0 license, backend dependencies (llama.cpp, ONNX), telemetry pull requests
RunAnywhere MetalRT binaries repository
github.comProprietary license, Apple Silicon arm64-only, closed-source
YC RunAnywhere company profile
ycombinator.comYC W26 batch, team size 2, stage, launch narrative
SEC Form D — RunAnywhere, Inc.
sec.govDelaware incorporation, officers (Monga/Malhotra), $1M offering, $10K sold, 2 investors, no revenues
Crunchbase — RunAnywhere
crunchbase.comPre-seed stage, YC investor, Redwood City, 1–10 employees, traffic estimate
Apple App Store — RunAnywhere
apps.apple.com3 ratings, free, Developer Tools, iOS 17/M1 requirement, privacy label fields empty
Google Play — RunAnywhere
play.google.com~956 US downloads, 3 reviews, free, Qualcomm Genie/QNN note in changelog, Data Safety N/A
Hacker News — RunAnywhere YC W26 launch
news.ycombinator.com240 points, 153 comments, spam/outreach-ethics allegations, community reception
Sanchit Monga — GitHub profile
github.comPinned RunAnywhere repos, org affiliation
Shubham Malhotra — GitHub profile
github.com~1,887 contributions/year, RunAnywhere org repos
RunAnywhere — LinkedIn company page
linkedin.com9 employees (2–10 band), followers, hire activity
Sanchit Monga — LinkedIn profile
linkedin.comPosts, education hints, honors; employment history not populated
Shubham Malhotra — LinkedIn profile
linkedin.comPublications list, posts, honors; employment history not populated; "5,000+ developers" claim origin
Mirar landing page (Vercel staging)
mirar-beryl.vercel.appVision-agent early access, WIP demo, design-partner onboarding
ABI Research — edge AI market trends
abiresearch.comEdge AI chipset revenue $34.4B → $96B (2026–2031); 711.3M chipset shipments (2026)
IDTechEx — AI chips for edge applications
idtechex.comEdge AI chips >$80B by 2036; every major OEM had AI-enabled flagship phones by Jan 2026
IDC — smartphone market share tracker
idc.com1.25B smartphone shipments (2025 forecast); $578.9B market value (2026)
GSMA Mobile Economy 2026
gsma.com~5.8B unique mobile subscribers (2025); 80% 5G by 2030
IoT Analytics — connected IoT devices
iot-analytics.com21.1B IoT devices end-2025; 39B by 2030
NVIDIA FY2026 Form 10-K
sec.govRevenue $215.9B; gross margin 71.1%; AI infrastructure economics benchmark
Qualcomm FY2025 Form 10-K
sec.govQCT IoT $6.617B; QTL 72% EBT margin; on-device AI commentary
Arm FY2025 Form 20-F
sec.govRevenue $4.007B; 97% gross profit; royalty $2.168B; IP/royalty-layer economics benchmark
Apple FY2025 Form 10-K
sec.govServices GM 75.4% vs Products GM 36.8%; platform-gatekeeper incentive benchmark
llama.cpp GitHub repository
github.com~117k GitHub stars; MIT license; primary free substitute; RunAnywhere SDK uses it as GGUF LLM backend
Apple MLX GitHub repository
github.com~27.1k GitHub stars; Apple Silicon array framework; free Apple-research alternative to MetalRT
Ollama homepage
ollama.comFree local inference; Pro $20/month, Max $100/month commercial tiers
ExecuTorch homepage
executorch.aiMeta/PyTorch-sponsored free substitute; 12+ hardware backends; cross-platform
Google LiteRT documentation
ai.google.devTFLite successor; Google-maintained; cross-platform
ONNX Runtime homepage
onnxruntime.aiCross-platform OSS inference runtime; RunAnywhere SDK uses it for STT/TTS
Cactus Compute homepage
cactuscompute.comHybrid on-device/cloud routing; cross-platform SDK; YC-backed; 5.3k+ stars (self-reported)
LiteLLM homepage
litellm.ai100+ LLM providers; OSS and enterprise; YC-backed
Portkey homepage
portkey.aiPalo Alto Networks acquisition banner; AI gateway product
Access limitations: Several sources central to this report's most important claims could not be retrieved or were inaccessible to public research. The LinkedIn employment history fields for both founders returned no data in public dataset captures, which means the core prior-employer scale claims — CEO's Intuit "50M+ users" and CTO's Microsoft Azure/AWS and MetalRT authorship — cannot be independently verified and remain labeled as company claims throughout. The Console fleet control plane is login-gated; no architecture review, multi-tenant evidence, or customer list was accessible, so every claim about Console's capabilities comes from the company's own SDK page. The MetalRT binary is closed-source; the performance benchmarks on the engines page are company-controlled and no independent reproduction was found in any indexed publication. X (formerly Twitter) profiles for RunAnywhere and both founders were inaccessible. The primary analyst market-sizing data for the on-device AI fleet-management software category specifically (as distinct from the broader edge AI hardware market) does not exist as a public research product; the SAM figures in this report are derived estimates with low confidence. The YC deal terms and cap-table detail are not public; all ownership and round-structure items are marked unknown. Three total sources in the research were blocked and could not be cited; they are omitted from this log. Absence-of-record checks for job postings (LinkedIn Jobs, Indeed, Glassdoor) and for pricing (site:runanywhere.ai search) are documented as named-source absence findings in the relevant sections.
Public disclaimer: This is an independent public teardown based solely on publicly available information retrieved on 2026-06-17. It is research support, not investment, legal, tax, or financial advice. Company-provided statements are labeled as claims, not verified facts. RunAnywhere did not participate in or review this report. Errors and omissions are possible; final decisions remain with the reader.