Skild AI

1.1 Identity, product, and operating footprint

Skild AI was founded in 2023 and an independent January 2026 profile placed its operations in Pittsburgh, the San Francisco Bay Area, and Bengaluru. The company describes the Skild Brain as one robotics foundation model intended to generalize across tasks and hardware forms, while its homepage presents lower-level mobile-manipulation skills such as grasping, handover, and navigation through an API. That supports a software/control-layer identity, but the assigned evidence does not disclose pricing, contract structure, or the split between software, services, and hardware-linked revenue. The operating footprint is therefore supportable; the detailed business model is not.[CO001, CO002, CO003, CO004]

1.2 Founder leadership and governance visibility

Public reporting identifies Deepak Pathak as cofounder and CEO and Abhinav Gupta as cofounder and president. Their public comments span the technical thesis, model training, commercial testing, and deployment plans, indicating that the two founders remain central to both product direction and market communication. The same source set does not identify a board roster, independent directors, succession arrangements, executive bench, voting rights, or material leadership changes. This is evidence of concentrated visible leadership, not proof that broader governance infrastructure is absent; those records require direct diligence.[CO005, CO006]

1.3 Capitalization and stakeholder concentration

On January 14, 2026, Skild AI announced a $1.4 billion Series C led by SoftBank and said the financing valued the company above $14 billion. Independent reporting repeated those figures, while Technical.ly later reported more than $2 billion raised in total. The latter is retained as a third-party total rather than reconstructed from unavailable round data. The round brought together financial sponsors, strategic technology investors, and potential enterprise-channel stakeholders, but participation alone does not establish board control, commercial commitments, or customer status. No assigned source resolves ownership percentages, liquidation preferences, secondaries, debt, or credit facilities.[CO007, CO008, CO009]

1.4 Commercial scale, milestone sequence, and evidence limits

The company said live revenue rose from zero to about $30 million within a few months in 2025. By September 2026, NVIDIA described a $100 million annual revenue run rate ten months after first commercial deployment and more than 60 deployment partnerships; Technical.ly separately called the $100 million figure ARR. Those labels are not interchangeable, so this chapter preserves the conflict and does not convert either into audited revenue. NVIDIA also described S1's August launch and Skild's internal task results, while Technical.ly reported customer testing and a 2027 deployment target. The careers page displayed zero current openings at access, but that is neither headcount nor evidence of no hiring. No assigned source establishes the absence of legal, regulatory, outage, or other adverse events.[CO010, CO011, CO012, CO013, CO014, CO015, CO016]

1.5 Exhibits

2.1 Market Boundary, Taxonomy, and Structural Shifts

The addressable market for Skild AI occupies the horizontal physical intelligence layer of robotics automation, distinct from traditional robotics hardware OEMs, single-purpose industrial robotics cells, and purely digital multimodal language models. Historically, robotic automation required bespoke hardware engineering, hand-crafted control policies, and custom programming for every machine geometry and task. Skild AI defines its market around universal robotics foundation models that act as an operating system or software intelligence layer capable of controlling diverse morphologies including quadrupeds, mobile manipulators, humanoid systems, and tabletop robotic arms. The addressable software spend includes robotics operating software licenses, foundation model API usage, and vertical solution orchestration modules. Spend excluded from Skild AI's direct model revenue includes raw mechanical hardware, structural castings, and standard actuator manufacturing. Incumbent status-quo substitutes consist of fixed-function robotic automation, manual human labor across warehousing and manufacturing stations, and proprietary vertically integrated autonomy stacks developed by single-form-factor OEMs. By acquiring Zebra Technologies' robotics automation business, Skild AI absorbed established autonomous mobile robot (AMR) orchestration to directly access brownfield logistics and warehouse fulfillment operations.[CM001, CM002, CM003, CM007]

2.2 Market Sizing Lenses, Growth Drivers, and Near-Term Reality

Evaluating the addressable market for general-purpose physical AI requires contrasting long-range macroeconomic labor projections against concrete deployment metrics. Macroeconomic thesis drivers cite structural labor scarcity, including projections that 2.1 million U.S. manufacturing positions could remain unfilled by 2030 with potential economic output impacts exceeding $1 trillion. Long-horizon analyst estimates forecast the global humanoid robot market to reach $38 billion by 2035 and speculative evaluations reach $5 trillion by 2050. However, near-term shipment data highlights the early stage of physical deployment: global humanoid robot shipments reached approximately 13,300 units in 2025 with forecasts targeting 250,000 units by 2030. Consequently, Skild AI's near-term serviceable market is concentrated in non-humanoid industrial form factors, specifically mobile manipulation, quadrupedal security/inspection, and warehouse material transport. Commercial revenue traction expanded rapidly from zero to approximately $30 million during 2025, with company announcements stating annual recurring revenue crossed $100 million in September 2026 across 60+ enterprise customers. Of this recognized revenue base, manipulation tasks account for roughly 90%, while mobility solutions account for 10% (including 4% from acquired AMR solutions). This distribution illustrates that customer willingness to pay is currently driven by dexterous manipulation bottlenecks rather than standalone wheeled locomotion.[CM004, CM005, CM006, CM008, CM009, CM010]

2.3 Buyer Segmentation, Procurement Paths, and Adoption Constraints

Adoption of physical foundation models spans three primary customer profiles: robot hardware OEMs, industrial enterprise operators, and automated system integrators. Robot manufacturers license the Skild Brain to eliminate multi-year in-house AI development cycles, converting capital expenditure into cloud-based API licensing and edge runtime fees. Enterprise buyers in manufacturing, logistics, and critical infrastructure evaluate deployments based on cycle-time parity, station throughput, and error recovery rather than passive benchmark demonstrations. Commercial deployments across Foxconn Blackwell server assembly, Sumitomo Wiring Systems, and Mitsui commercial kitchen pilots reveal that industrial adoption triggers center on labor availability and retooling speed. However, critical constraints govern deployment velocity. Deployments require cycle-time parity with human line workers, where a robot operating with high accuracy but at one-tenth human speed cannot be placed into production. Furthermore, physical foundation models face impending regulatory hurdles: the European Union's Machinery Regulation takes effect on January 20, 2027, introducing mandatory third-party notified body assessments for machinery demonstrating self-evolving or out-of-distribution operational behaviors. Hardware integration friction, sim-to-real physical discrepancies, and liability risks for physical property damage remain persistent adoption bottlenecks.[CM011, CM012, CM013, CM014, CM015]

2.4 Exhibits

3.1 Competitive landscape and buyer alternatives

The relevant landscape is broader than direct robotics-foundation-model startups. Independent analysis groups Covariant, Physical Intelligence, and Genesis AI as horizontal peers, while Alphabet's Intrinsic and NVIDIA's Isaac represent platform alternatives with compute, ecosystem, and enterprise-distribution advantages. A separate threat comes from vertically integrated developers such as Figure, Tesla, and Apptronik, whose hardware-software deployments can produce proprietary operating data and reduce demand for a third-party intelligence layer. Internal development and continued use of task-specific automation are also buyer alternatives, but the assigned evidence does not quantify their cost or adoption share. The table therefore presents a representative, evidence-bounded landscape rather than an exhaustive market census.[CP001, CP002, CP003, CP004, CP012, CP014]

3.2 Capability positioning and technical differentiation

Skild's central differentiation claim is breadth across robot morphologies combined with rapid task adaptation. The company describes one brain spanning quadrupeds, humanoids, tabletop arms, and mobile manipulators, and says S1 can execute a new task from one video without task-specific fine-tuning. Its internal comparison reports 66% average cumulative per-step success on unseen long-horizon tasks at 100,000 hours of pre-training, versus 9% for a language-prompted VLA baseline. Those results are useful indicators of the intended moat, not independent proof: the benchmark, task suite, grading intervention, and model comparison are company-issued. Hardware breadth may lower initial integration friction, while proprietary protocols and fragmented hardware can recreate integration costs and lock-in at deployment.[CP004, CP005, CP006, CP007]

3.3 Commercial scale, distribution, and packaging

Skild reports crossing $100 million in ARR ten months after its first commercial deployment and serving more than 60 paying customers. It also reports that about 90% of revenue comes from manipulation and 10% from mobility, including 4% from acquired Zebra/Fetch Robotics AMR solutions. These disclosures suggest that deployment access, integration capability, and acquired distribution are becoming competitive assets, rather than the model alone determining outcomes. However, TNW reported that the company declined to name the customers described in its coverage. AI Business separately reported $30 million of 2025 revenue; that historical accounting-period figure is not interchangeable with the later ARR claim. No retained source provides public list pricing, realized pricing, discounting, contract duration, or competitor packaging on a comparable basis.[CP008, CP009, CP010, CP011, CP012, CP013]

3.4 Moat durability and displacement risk

The moat depends on whether cross-embodiment learning and deployment data improve faster than integrated competitors' proprietary fleets and big-tech platforms' ecosystems. Skild's reported funding provides substantial capacity, but capital scale does not demonstrate benchmark leadership, pricing power, or low switching cost. Independent analysis explicitly warns that vertically integrated players may shrink the addressable market for a horizontal provider, while hardware fragmentation can constrain portability. Regulatory trust will also matter: TNW reports that Europe's Machinery Regulation will cover AI-based safety functions and self-evolving machinery from 20 January 2027 and require notified-body assessment. The fetched careers page did not establish current hiring scale, so organizational capacity relative to larger platforms remains unverified. The investment conclusion is therefore differentiation with material proof risk, not a demonstrated durable monopoly.[CP002, CP003, CP004, CP012, CP014, CP015, CP016]

3.5 Exhibits

4.1 Commercial Model and Revenue Structure

Skild AI commercializes its general-purpose robotics foundation model through enterprise software licensing and integrated automation deployments. In September 2026, the company disclosed reaching $100M in annual recurring revenue (ARR) run rate across more than 60 paying enterprise customers, representing a sharp ramp from approximately $30M in calendar 2025 revenue. Commercial deployments span industrial manipulation, site inspection, security, food preparation, logistics, and factory operations. The revenue mix is heavily weighted toward robotic manipulation, which accounts for roughly 90% of recurring revenue, while mobility and autonomous mobile robot (AMR) workflows constitute the remaining 10%, including a 4% contribution from the acquired Zebra Technologies automation business. Pricing structures remain opaque and private, typically packaged as enterprise platform licensing tiers and compute-indexed deployment agreements rather than standardized public list prices.[CI001, CI002, CI003, CI004, CI005, CI006]

4.2 Operating Cost Drivers and Margin Dynamics

The economic profile of physical AI differs sharply from asset-light SaaS due to continuous real-world inference, multi-embodiment data collection, and physical fleet servicing. For every dollar Skild AI spends on collecting robotics data, it spends three dollars on quality control, screening data points for low-level precision and task coherence. Furthermore, the company's April 2026 acquisition of Zebra Technologies' robotics automation division transitioned Skild AI away from a purely virtual foundation model provider into managing physical warehouse mobile robots and fulfillment orchestration software. While this direct asset integration short-circuits enterprise hardware deployment friction, it introduces ongoing maintenance, field depot support, and fleet operational overhead that depress gross margins relative to standard pure-play software multiples.[CI007, CI008, CI009, CI010]

4.3 Capital Runway and Underwriting Verdict

Skild AI secured $1.4B in Series C capitalization in January 2026 led by SoftBank Group, with strategic participation from NVentures, Samsung, and Jeff Bezos, establishing a post-money valuation of $14B. This followed earlier rounds including a $300M Series A at a $1.5B valuation in July 2024 and a $100M Series B at a $4.5B valuation in June 2025. With cumulative equity capital exceeding $1.8B, Skild AI possesses substantial multi-year liquidity to absorb massive accelerated computing training runs on NVIDIA infrastructure. However, trading at an extreme valuation multiple of 140x forward ARR run rate (and ~467x against trailing 2025 revenue), underwriting Skild AI requires accepting high execution risk. Key financial unknowns include unreleased audited cash burn, unknown gross margins across hardware vs. software streams, and regulatory compliance exposure under Europe's Machinery Regulation governing self-evolving machinery.[CI011, CI012, CI013, CI014, CI015, CI016]

4.4 Exhibits

5.1 Product Definition and Omni-Bodied Operating Model

Skild AI defines its core product offering around general-purpose physical artificial intelligence designed to decouple robotic software capabilities from fixed machine morphologies [CE001]. Rather than constructing narrow behavioral policies tailored to specific mechanical arms or proprietary mobile bases, the company develops the Skild Brain, an action-centric foundation model engineered to control heterogeneous robotic form factors [CE001, CE002]. The core operating model structures execution through a two-tier hierarchical policy [CE002]. A low-frequency, high-level navigation and manipulation policy interprets task goals and environmental state to generate spatial action plans [CE002]. This high-level command sequence is subsequently fed into a high-frequency, low-level action policy that directly outputs joint angles and motor torques governing the physical robot hardware [CE002]. By abstracting hardware control into a unified policy architecture, Skild AI enables robotic platforms ranging from quadrupeds and bipedal humanoids to table-top manipulators and mobile industrial bases to run on a shared model foundation [CE001, CE002]. Furthermore, the company reported crossing $100 million in annual recurring revenue ten months after its initial commercial deployment, with manipulation applications accounting for approximately 90% of total revenue and mobility solutions generating 10% [CE006]. Commercial deployments span diverse industrial workflows, including dual-arm manipulation for NVIDIA Blackwell system assembly with Foxconn, wire harness manufacturing automation with Sumitomo Wiring Systems, and commercial kitchen food preparation pilots with Mitsui & Co. [CE006].

5.2 S1 Architecture, In-Context Learning, and Pre-Training Pipeline

The technical centerpiece of Skild AI's manipulation capabilities is S1, a flagship foundation model built from the ground up as an in-context learner for robotics [CE003]. In contrast to conventional vision-language-action models that require collecting tens to hundreds of hours of post-training demonstration data and fine-tuning model weights for every novel deployment condition, S1 adapts to unseen tasks entirely through visual video prompting without altering neural network weights [CE003]. Under this paradigm, a user specifies a target task by supplying a single egocentric visual demonstration video; S1 processes the demonstration prompt within its context window, infers the demonstrator's intent, resolves functional correspondences across embodiment viewpoints, and translates the prompt into autonomous motor commands [CE003]. The company's controlled internal benchmark evaluations report that on long-horizon unseen tasks (spanning 4 to 8 minutes across dozens of manipulation primitives), scaling pre-training data to 100,000 hours enabled S1 to achieve a 66% cumulative per-step success rate, compared to 9% for language-prompted VLA baselines, demonstrating a 7x performance advantage on out-of-distribution manipulation [CE004]. Skild AI estimates that a single video demonstration in context achieves task performance comparable to approximately 380 post-training teleoperation demonstrations, eliminating an estimated 50 to 100 hours of manual teleoperation per new task [CE005]. Nevertheless, real-world deployment requires balancing accuracy with execution cycle time to prevent factory line bottlenecks, demanding continued distillation and hardware-specific latency optimizations [CE006].

5.3 Training Data Flywheel, Simulation Stack, and Technical Dependencies

Acquiring actionable physical data at scale represents the fundamental bottleneck in robotics foundation model development [CE002, CE007]. Skild AI addresses this data scarcity through a multi-pronged data engine that combines large-scale synthetic physics simulation, web-scale human video datasets, and targeted real-world robot teleoperation [CE002, CE007]. The company's computational and simulation stack is heavily dependent on NVIDIA AI infrastructure, utilizing Isaac Lab for scalable reinforcement learning scenarios and Cosmos Transfer to multiply simulation datasets across varying environmental lighting, textures, and camera parameters via text prompts [CE007, CE008]. This simulation-first pipeline allows thousands of virtual robot instances across multiple morphologies to experience millions of edge-case failure modes safely, enabling robots to recover from physical disruptions such as jammed wheels within 2 to 3 seconds or continue operating after limb failures through in-context behavioral adaptation [CE007, CE008]. To complement synthetic physics data, Skild AI mines internet-scale video repositories to extract functional manipulation affordances by treating humans as biological robots [CE007]. Real-world production deployments feed into a recursive self-improvement flywheel (physical RSI), where specialized task edge cases encountered across 60+ commercial customer sites are distilled back into the generalist base model to enhance universal manipulation capabilities [CE006]. On the software engineering layer, the runtime stack relies heavily on production-level C++ and Python modules integrated with the ROS/ROS2 robotics middleware ecosystem for motion planning, perception, SLAM, and high-level behavioral control [CE010, CE011].

5.4 Differentiation, Operational Reliability, and Technical Limitations

Skild AI's primary technical differentiation lies in true cross-morphology generalization and demonstration-driven in-context adaptation, contrasting sharply with conventional robotic systems that require fixed kinematic calibrations or proprietary hardware lock-in [CE001, CE002, CE007]. By establishing hardware-agnostic control, Skild AI enables commercial partners to deploy robotic intelligence on commodity hardware priced between $4,000 and $15,000, compared to traditional specialized industrial robotics requiring upwards of $250,000 [CE007]. In field locomotion testing conducted in Pittsburgh urban environments, Skild's vision-driven models achieved 60% to 80% task performance within hours of data collection across unstructured terrain, outdoor parks, and fire escapes without pre-mapped environments [CE007]. Despite these demonstrated strengths, substantial technical and diligence risks persist [CE012]. The company has published limited peer-reviewed, independently audited benchmarks on sustained long-horizon precision and mean time between failures (MTBF) under high-speed manufacturing cycle times [CE006, CE012]. Furthermore, third-party industry analyses note that generalized robotics models face significant physical AI data bottlenecks and unproven safety verification standards when transitioning from piloted proof-of-concept cells to mission-critical industrial assembly lines [CE008, CE012]. Software engineering recruiting profiles indicate ongoing development across SLAM, low-latency motion planning, and hardware safety interlocks, underscoring that commercial scaling remains constrained by edge-case physical uncertainties and dependency on external compute stacks [CE009, CE010, CE011].

5.5 Exhibits

6.1 Enterprise and OEM Customer Segmentation

Skild AI segments its commercial revenue across two primary functional modalities: robotic manipulation and robotic mobility. According to disclosures from company leadership upon crossing the $100M annual recurring revenue threshold, manipulation tasks constitute approximately 90% of total revenue, while mobility applications account for the remaining 10%. Within mobility, autonomous mobile robot (AMR) orchestration solutions acquired from Zebra Technologies represent roughly 4% of total ARR, or approximately $4M. The company's customer base is structured primarily around high-value enterprise original equipment manufacturers (OEMs) and industrial conglomerates rather than fragmented end-user accounts. Target buyers consist of Vice Presidents of Manufacturing, Plant Automation Directors, and Supply Chain Executives across electronics manufacturing, automotive components, logistics fulfillment, and commercial food preparation. Skild AI delivers its foundation intelligence via software subscription agreements with terms varying across customer accounts, enabling enterprise customers to upgrade robot capabilities dynamically as foundation model iterations advance without requiring complete robotic hardware overhauls.[CU001, CU002, CU008, CU011]

6.2 Commercial Trajectory and Revenue Run-Rate Acceleration

Skild AI's commercial adoption trajectory has expanded rapidly throughout 2025 and 2026, advancing from zero commercial revenue in early 2025 to approximately $30M in live revenue late in the year, and ultimately surpassing $100M in annual recurring revenue ten months after its initial commercial deployment. Over this same ten-month operating window, the company reported scaling from eight commercial customer accounts to more than 60 paying enterprise customers. In addition to the $100M recurring run-rate, the company recognized over $50M in cumulative revenue since beginning deployments. Operational footprints span facility inspections, point-to-point delivery, security, commercial kitchens, manufacturing facilities, and data centers. The rapid velocity of deployment is facilitated by the S1 foundation model's in-context learning architecture, which reports a 96% success rate on seen manipulation tasks and 66% on novel, unseen tasks from a single video demonstration, reducing deployment setup from weeks of specialized programming to rapid in-situ prompting.[CU001, CU002, CU010, CU011]

6.3 Named Customer Deployments and Strategic Integration

Skild AI's public customer footprint is anchored by high-visibility deployments with major strategic partners and investors. On NVIDIA Blackwell GPU assembly lines operated in collaboration with Foxconn in Houston, Skild Brain powers dual-arm robotic manipulators executing high-precision tasks including busbar insertion, screw fastening, and component placement. In automotive manufacturing, the company is conducting advanced integration work with Sumitomo Wiring Systems to automate delicate wire harness assembly operations previously deemed intractable for mechanical automation. In commercial food service, Skild AI has deployed S1 in commercial kitchen pilots with Mitsui & Co., whose distribution operations supply 1.4 million meals daily across Japan. Furthermore, the company is collaborating with LG CNS to deploy foundation models across partner logistics facilities, citing Market Kurly as a key operational target. However, independent diligence highlights that nearly all publicly named customer references represent corporate shareholders or strategic go-to-market partners, leaving third-party commercial traction outside partner circles largely unverified.[CU003, CU004, CU005, CU006, CU007]

6.4 Retention Durability, Channel Dependence, and Concentration Risks

Despite reporting top-line ARR acceleration, Skild AI exhibits notable customer concentration and durability vulnerabilities. The company relies heavily on indirect sales channels through legacy industrial robot manufacturers, notably ABB, Universal Robots, and Mobile Industrial Robots, rather than an extensive direct enterprise sales organization. This indirect go-to-market structure limits direct operational feedback from frontline factory floors and exposes Skild AI to partner channel friction. Furthermore, Skild AI has disclosed zero audited metrics regarding net revenue retention (NRR), gross retention (GRR), annual contract renewal rates, or cohort-level churn. Additional churn risk stems from the April 2026 acquisition of Zebra Technologies' robotics division: taking over an autonomous mobile robot business that Zebra had actively slated for downsizing in late 2025 threatens customer retention across inherited Fetch Robotics accounts. Internationally, pending European regulatory scrutiny under the Machinery Regulation in January 2027 introduces compliance barriers for self-evolving AI machinery, creating potential friction for European customer expansion.[CU007, CU009, CU012, CU013, CU014]

6.5 Exhibits

7.1 Regulatory, legal, and safety exposure

The clearest identified regulatory issue is Europe's Machinery Regulation, which The Next Web reports will cover AI-based safety functions and machinery with self-evolving behavior from January 20, 2027, with notified-body assessment rather than only manufacturer declaration. Applicability to Skild AI is conditional: the same report says the company had not disclosed whether it sells into Europe. The assigned evidence contains no regulator filing, conformity plan, safety certification, litigation search result, incident register, privacy assessment, export-control classification, or insurance disclosure. The investment implication is therefore not a finding of noncompliance; it is a readiness gap. Before underwriting European expansion or safety-critical deployments, diligence should verify intended jurisdictions, legal entity responsibilities, applicable standards, independent testing, incident reporting, and contractual allocation of product-liability risk.[CR001, CR002]

7.2 Operational reliability, quality, and security

Skild AI itself identifies the hard part of robotics deployment as the final reliability margin: its egg task took one week to work once and another two months to work reliably across eggs and setups. It also says a robot that is 99.9% accurate but ten times too slow is not deployable, because one slow station constrains an entire production line. Those observations are more decision-useful than demo footage, but they do not provide fleet uptime, intervention frequency, mean time between failures, safety-event rates, or customer-level cycle times. S1's published results use internal benchmark suites, and the described evaluation used human intervention mainly to recover the VLA baseline. Kill criteria should therefore focus on audited production telemetry: repeated safety incidents, intervention rates that erase labor savings, or cycle time below contracted throughput after an agreed remediation window.[CR006, CR007, CR008, CR009]

7.3 Partner, platform, and integration dependencies

Skild AI depends on external infrastructure and counterparties even though its product thesis is hardware-agnostic. S1 is built on NVIDIA AI infrastructure, while the company's data strategy combines sources with different tradeoffs: teleoperation is close to deployment hardware but scales poorly, and egocentric video scales better but has a larger robot-domain gap. The Zebra Robotics Automation acquisition adds the Symmetry Fulfillment orchestration platform and creates an end-to-end warehouse proposition, but the available release does not disclose purchase price, transferred liabilities, customer retention, service obligations, or integration milestones. Public customer breadth also does not reveal revenue concentration: Skild AI reports more than 60 paying customers, while independent reporting says customers were largely unnamed. Investment diligence should test compute portability, acquisition integration, renewal quality, and concentration rather than treating a broad customer count as proof of diversified economics.[CR004, CR010, CR012, CR013]

7.4 Financial-model risk and thesis-break triggers

The public financial signals are promising but not interchangeable. Skild AI says it crossed $100 million in annual recurring revenue and recognized $50 million of revenue during its first ten months of deployment; the latter is not ARR, and neither figure discloses gross margin, contract duration, collections, or concentration. A separate company-issued acquisition release described approximately $30 million of 2025 revenue. Independent reporting says the company raised $1.4 billion in January 2026 at a $14 billion valuation, but funding raised is not evidence of cash currently available. The valuation leaves execution sensitive to slower conversion, integration costs, or weak unit economics. Thesis-break triggers are failure to reconcile ARR to audited revenue, customer cohorts and cash receipts; runway below the investment plan once burn is verified; or material dependence on a small number of counterparties without contractual protection.[CR003, CR005, CR014, CR015]

7.5 Exhibits

8.1 Recommendation and price discipline

The recommendation is research-more, with medium confidence, high risk, and a stretched valuation stance. Forge and Silicon Valley Investclub both place the January 2026 post-money valuation at roughly $14 billion, and the financing was reported at approximately $1.4 billion. Skild AI's September 2026 claim of $100 million in ARR makes the simple post-money valuation-to-ARR reference 140x; this is neither an enterprise-value multiple nor an audited revenue multiple. The company separately described $50 million as recognized since deployment began, so funding, ARR, and recognized revenue must not be treated as interchangeable liquidity or accounting measures. Entry at the latest mark is not supported by the retained public evidence without audited revenue quality, margin, retention, burn, cash, and customer-concentration proof.[CV001, CV002, CV006, CV007, CV009, CV016]

8.2 Thesis, anti-thesis, and comparable context

The thesis is that a hardware-agnostic robot foundation model can compound deployment data across embodiments, and Skild reports both technical progress and early commercial scale. Its S1 evaluation reports 66% success on unseen long-horizon tasks from one demonstration, while HPE describes dedicated private infrastructure for training and inference. The anti-thesis is evidence quality: those technical results are company-issued, the commercial metrics are unaudited, and an Exa directory remained materially stale after the latest financing. Retained market-data pages identify CoreWeave and LightOn only as broad public yardsticks and disclose no usable like-for-like revenue multiple. Competitor comparisons therefore do not provide an independent valuation benchmark, and the current financing mark is the only defensible price anchor.[CV011, CV012, CV013, CV014, CV016]

8.3 Scenario framing and downside triggers

Scenario values are bounded by observed financing marks rather than invented multiples. A bull case can defend $14 billion or more only after the $100 million ARR claim is audited, customer references establish durability, and margins and retention support software-like economics. The base case lies somewhere between the May 2025 $4.7 billion Series B mark and the January 2026 $14 billion Series C mark, but public evidence cannot locate a fair value within that interval. The bear case falls to or below the prior mark if ARR proves services-heavy, concentrated, or non-recurring, or if regulatory deployment friction rises. These are decision bands and triggers, not probability-weighted fair-value estimates; no numerical probabilities are supportable from the retained sources.[CV001, CV003, CV006, CV008, CV009, CV010, CV015, CV016]

8.4 Financing terms, exit readiness, and diligence asks

Forge's records indicate a 1.0x liquidation preference, a 6% non-cumulative dividend, and non-participating Series C economics, while Silicon Valley Investclub estimates roughly 10% dilution from the round; neither source is a complete fully diluted cap table. The retained evidence does not establish an IPO timetable, a credible strategic-acquirer set, or a supportable target return, hold period, or exit window. Before underwriting price, obtain audited revenue recognition and ARR schedules, gross margin by deployment, retention and concentration cohorts, burn and cash, the complete preference stack and option pool, customer reference calls, and jurisdiction-specific machinery compliance. Thesis-break triggers are failed revenue verification, weakening renewal quality, hidden preference seniority, or a deployment regime that materially slows commercialization.[CV004, CV005, CV008, CV015, CV016]

8.5 Exhibits