Skild AI

Phase A — Understand the business

Company Overview

What it is. Skild AI is a Pittsburgh-based robotics-AI lab building "the Skild Brain" — a single, robot-agnostic ("omni-bodied") foundation model that controls any robot body (quadruped, humanoid, tabletop arm, mobile manipulator) without prior knowledge of its exact morphology, and without per-robot retraining. Founded May 2023 by two former Carnegie Mellon Robotics Institute professors, Deepak Pathak (CEO) and Abhinav Gupta (President).

The deliberate strategic choice — no hardware. Unlike Figure, 1X, or Tesla Optimus, Skild builds no proprietary robot. It sells the intelligence layer to run on hardware made by anyone. This is the central bet: that in robotics, as in LLMs, value migrates from the body to the brain. The robotics commercial layer frames this exactly — Skild sits in the "Software-stack players" row ("General-purpose robot brain") next to Physical Intelligence, NVIDIA GR00T, and Covariant, distinct from the builder/integrator row (Tesla, Figure, Agility, Unitree).

Business model. B2B software — foundation-model licensing via cloud APIs, plus fine-tuning/post-training services and vertical software modules. As of late 2024 monetization was still "in development" with no reported customers; by the Series C the company reported "live revenue grew from zero to about $30M in just a few months in 2025… with multiple customers". The model is pre-trained on simulation + internet human-action video, then post-trained on targeted real-world robot data to deliver working customer solutions.

Customers / deployment sectors (named). Security, construction, delivery, data centers, warehouses, factory assembly. The marquee deployment: with Foxconn, the Skild Brain is being put on robotic assembly lines that build NVIDIA's Blackwell GPU server systems in Houston, Texas. NVIDIA is also brokering integrations with ABB Robotics and Universal Robots. A June 2026 MOU with VinDynamics (a Vingroup robotics arm) targets humanoid integration of the Skild Brain.

Key payment-term / contract structure. Not publicly disclosed — n/a — private, not disclosed. The implied shape is per-robot/per-deployment software licensing + post-training services; no take-or-pay or backlog figures are public.

One-line: A CMU-pedigree, no-hardware "robot brain" lab betting the entire robotics margin pool migrates from the body to the model — and that one model can rule every body.

Supply Chain

Skild is a software/model layer, so its "supply chain" is (a) the inputs to train the brain and (b) the hardware bodies + channel it must ride to reach an end customer. Mapping it against the robotics chain:

Upstream inputs (what feeds the Skild Brain):

• Compute — Skild explicitly partnered with Hewlett Packard Enterprise (HPE) for the AI infrastructure to train the model. Onboard inference rides NVIDIA Jetson / Thor, the concentration chokepoint for the whole sector. NVIDIA is simultaneously a Skild investor (NVentures), a compute supplier, a channel-maker (ABB/UR/Foxconn intros), and a competitor (GR00T) — an unusually dense four-way entanglement.
• Training data — three sources: large-scale simulation/synthetic data, internet human-action video, and targeted real-world robot data for post-training. This is the moat candidate the sector flags as "teleop data / fleet hours" — Skild's twist is to lean on internet video + sim to sidestep the fleet-ownership disadvantage that a no-hardware player otherwise faces.

Midstream (Skild) → downstream (bodies + buyers):

• Robot bodies it must run on: ABB, Universal Robots, Foxconn's line robots, VinDynamics humanoids. Skild does NOT control these — it is a guest on someone else's actuators+sensors stack.
• End buyers: Foxconn (for NVIDIA's own Blackwell assembly), plus the named verticals (warehousing, security, construction, data centers).

Chokepoints / single-source dependencies:

• Compute dependence on NVIDIA/HPE + AWS-class cloud — Contrary explicitly flags reliance on third-party compute (AWS) as a cost/dependency risk.
• Channel dependence on NVIDIA's convening power — the Foxconn/ABB/UR deals came through NVIDIA. That is leverage Skild does not own.
• Body fragmentation — every new robot platform = new actuator/sensor stack = real-world post-training cost. The "omni-bodied" claim is precisely the attempt to neutralize this chokepoint; whether it holds at deployment scale is the unresolved question (see Lens 13).

Names or it didn't happen: HPE (compute), NVIDIA Jetson/Thor (inference) + NVentures (capital) + GR00T (rival), Foxconn (line + buyer), ABB + Universal Robots (body channels), VinDynamics/Vingroup (humanoid MOU), AWS-class cloud (training).

Competitive Advantages (moats)

The bull's moat stack:

1. Capital + syndicate as a moat. ~$1.8–2.0B raised across 4 rounds is the largest war chest among pure model-labs — roughly 3× Physical Intelligence's ~$600M-led Series B base and a multiple of Covariant/Sanctuary. More important than the dollars is the strategic syndicate: SoftBank (lead), NVIDIA/NVentures, Bezos, Samsung, LG, Schneider Electric, Salesforce Ventures, plus IQT (In-Q-Tel — a US national-security tell). Strategics are distribution, not just money.
2. Data approach as a moat-substitute. A no-hardware player normally loses the "fleet hours" race. Skild's sim + internet-video + post-training recipe is an attempt to manufacture the data moat without owning a fleet — "a millennium of experience within days". If real, it converts the sector's #1 moat from a hardware-ownership problem into a research/compute problem — which suits Skild's strengths.
3. Founder IP / talent density. Pathak (artificial-curiosity paper, ~4K cites) and Gupta (~75K cites) are genuinely top-tier robot-learning researchers; the team is ex-CMU + ex-Meta FAIR. Automate.org's own headline frames it as "a ton of talent". In a field where the bottleneck is research breakthroughs, talent concentration is a real (if leaky) moat.
4. Switching costs (prospective). The sector note: teleop data + the policy model don't transfer across stacks, so once a customer's robots run on Skild-tuned policies, re-stacking forces re-training. This cuts both ways — it's Skild's lock-in once embedded, but it's also why a vertically-integrated builder may prefer its own in-house brain.

Bargaining power — who needs whom?

• vs. body-makers (ABB, UR, Foxconn lines): roughly balanced today, brokered by NVIDIA. Skild needs bodies to demonstrate; bodies need a brain they don't have to build. The danger: if the brain commoditizes, body-makers hold the customer relationship.
• vs. NVIDIA: Skild is the weaker party. NVIDIA supplies compute, capital, and channel — and ships GR00T. That is a partner who can become a competitor at will.
• vs. customers (Foxconn/NVIDIA-Blackwell): weak until proven — this is a test, not a moat, until the line throughput data exists.

Verdict on the moat: Real on capital + syndicate + research talent; aspirational on the data/omni-body claim that is the whole thesis. The durable-moat question reduces to one thing: does one model generalizing across bodies create enough lock-in before open-source (GR00T is partly open) + Physical Intelligence + builders' in-house brains commoditize the layer? — unproven.

Segments

n/a — private, not disclosed. There is no segment-level revenue/EBITDA/geography breakout for a private company at this stage, and our figures is header-only. The only disclosed revenue datum is the aggregate ~$30M "live revenue," 2025, with no split by sector, customer, or geography. By deployment vertical (a qualitative proxy, not revenue): security, construction, delivery, data centers, warehouses, factory assembly — with factory assembly (Foxconn/Blackwell) the lighthouse. Treat any finer breakdown as unsourced.

---

Phase B — Measure performance (+private: funding/valuation trajectory replaces earnings)

Funding & Valuation Trajectory (+private swap for "Earnings Result")

The defining feature is velocity: $1.5B → $14B+ in ~18 months.

Cumulative raised: ~$1.8B (PitchBook/Tracxn report $1.81B over 4 rounds, 30 investors). Some sector trackers say "total funding now exceeds $2B" — minor conflict; treat ~$1.8B as the disciplined figure.

Valuation step-ups: A→B ≈ 3.0–3.1×; B→C ≈ 3.0× — i.e. the price roughly tripled at each of the last two rounds in <12 months between B and C. Crunchbase's framing: "tripling valuation to $14B in just 7 months".

Burn / traction signal. Revenue 0 → ~$30M in months in 2025. Against a >$14B post-money, that is a ~470× EV/run-rate-revenue mark. Even on a generous forward ARR, this is a pure option value price on owning the robot-software layer — not a multiple any operating metric supports. Headcount is small and capital-light by hardware standards: ~50–64 employees (Tracxn: 64 as of 2026-03-31; PitchBook/Datanyze: ~50). Implication: the ~$1.8B raised is almost entirely compute + data + research talent, not plant — consistent with the no-hardware thesis.

⚠ Conflict to surface (provenance discipline): Sacra labels the latest round "Series B, March 2026, $1.4B at $15B" and adds a 2023 seed of $14.5M; the company's own blog + BusinessWire + TechCrunch call it Series C, Jan 2026, $1.4B at >$14B. I weight the primary-source ($14B+ Series C, Jan 2026) and flag Sacra's $15B/round-label as a secondary discrepancy, not silently averaged.

Founder Interviews / Narrative Trend (+private swap for "Earnings Calls")

No earnings calls exist. The narrative surface is founder interviews + launch posts:

• Davos / press (Jan 2026): Pathak frames the bet as "one AI brain for every robot" — intelligence, not hardware, is the bottleneck. Consistent, confident, single-message discipline.
• "First look" (Jul 2025): the pivot from stealth-claims to a shown brain with emergent behaviors (adapting to limb loss, jammed wheels, new payloads without retraining).
• Tone shift over time: 2024 = "general-purpose brain, no customers yet, monetization in development" → 2025 = "here's the brain + emergent robustness" → 2026 = "live revenue, multiple customers, Foxconn/Blackwell line." The arc is claim → demonstration → commercial proof-point, which is the right sequence. The recurring phrase ("omni-bodied," "any robot, any task, any environment") is now load-bearing marketing — and the thing skeptics test against (Lens 13).
• What they stopped saying: the early "no customers / monetization TBD" candor has been replaced by traction language — natural for the stage, but worth watching that disclosure doesn't get less specific as the valuation climbs (see Form-D gap, Lens 10).

Cap Table & Secondary Marks (+private swap for "Comps")

Syndicate quality — the IPO-proximity tell. The Series C cap table is unusually strong on crossover / strategic capital:

• Crossover / late-stage public-market-adjacent: SoftBank (lead, multi-round), Macquarie Capital, Mirae Asset, Coatue, Mirae, Andra/Alpha Square/Destiny growth vehicles. A Coatue + SoftBank + Macquarie + Mirae presence is the kind of late-stage book that precedes an IPO/secondary market.
• Strategics (= distribution): NVIDIA (NVentures), Samsung, LG, Schneider Electric, Salesforce Ventures, CommonSpirit (healthcare system).
• National-security signal: IQT (In-Q-Tel) participation — a tell that the security/inspection vertical has US-government interest.
• Founders + CMU: Carnegie Mellon University is on the cap table (Amazon Industrial Innovation Fund / Alexa Fund per Sacra at earlier rounds).

Mechanism/structure comps (the right comparison — by layer, not P/E):

Secondary marks: No disclosed mutual-fund markups/markdowns or tender pricing — n/a. Read: Skild is the best-capitalized, most strategically-syndicated pure model-lab, priced at ~2.5× Physical Intelligence's last mark. That premium is the bet that Skild's omni-body + Foxconn proof-point makes it the category winner. There is no traditional multiple to anchor it — write multiples as n/a — pre-revenue-scale.

Funding / Product Catalysts (events that re-rated the company) (+private swap for stock-price catalysts)

The "tape" for a private is the markup ladder + product/deal milestones:

• Jul 2024 — emergence from stealth, $300M Series A @ $1.5B: established Skild as a serious model-lab on day one.
• Mar 2025 — HPE compute partnership: solved the training-infrastructure input.
• May 2025 — $500M Series B @ ~$4.5B (SoftBank-led): 3× re-rate; SoftBank conviction signal.
• Jul 2025 — "first look" at the Skild Brain + emergent robustness: claim → demonstration.
• Dec 2025 — SoftBank + NVIDIA talks leak at $14B: the re-rate telegraphed.
• Jan 2026 — $1.4B Series C @ >$14B closes: ~3× re-rate, NVentures in.
• 2026 — Foxconn/NVIDIA-Blackwell line + ABB + Universal Robots integrations: the first real commercial proof-point — and the single most important catalyst ahead.
• Jun 2026 — VinDynamics MOU (humanoid): geographic + form-factor expansion.

Pattern: the market (private investors) re-rates Skild on (1) SoftBank/NVIDIA conviction events and (2) credibility milestones (HPE compute, the brain demo, the Foxconn line). What it has NOT yet re-rated on is audited deployment economics — because none exist. The next re-rate (up or down) hinges on whether the Foxconn line ships throughput numbers.

---

Phase C — Judge people & books

Management

Deepak Pathak (Co-founder/CEO). IIT (gold medal 2014) → PhD AI, UC Berkeley → co-founded VisageMap (acquired by FaceFirst, 2015) → Facebook AI Research (2019) → CMU Robotics faculty. Authored the seminal "artificial curiosity" intrinsic-motivation paper (~4K cites). Track record: genuine research leadership + one prior (small) startup exit. First time running a company at this scale/capital — the open question is operator-vs-researcher (see red flags).

Abhinav Gupta (Co-founder/President). IIT → PhD 2004 → CMU Robotics Institute professor (2009) → founding member of Meta FAIR (2018). ~75K citations across manipulation/locomotion/navigation — one of the most-cited robot-learning researchers alive. They jointly won Best Robotic System Award at CoRL (2021–22) for large-scale adaptive sim2real — the literal technical seed of Skild. They reportedly discussed founding a company together for "nearly a decade" before launching.

Skin in the game / tenure. Founders since May 2023; founder equity meaningful but undisclosed. CMU itself is on the cap table (alumni/IP alignment). No insider-transaction data exists for a private — our figures absent.

Capital allocation. Two years in, the decision set is: (a) no hardware — capital goes to compute/data/talent, not plant; (b) raise aggressively and bank optionality (~$1.8B); (c) buy credibility via strategics (NVIDIA/Samsung/LG/Schneider as distribution). For a model-lab this is coherent capital allocation — concentrate spend where the moat (research + data + compute) actually is. The risk is over-raising into a valuation that pre-commits them to a winner-take-all outcome they may not achieve.

Red flags (management-specific):

• Operator inexperience at scale — both are first-time founders of a $14B-valued, 50-person company being asked to become a commercial enterprise. Research brilliance ≠ go-to-market execution; the field is littered with lab-strong/commerce-weak robotics companies.
• Disclosure opacity — the Form-D gap (Lens 10) is partly a management/governance choice and deserves a direct question (Lens 14).
• Founder archetype: classic technical-founder/researcher pair — implication: world-class at the model, unproven at the P&L. At this stage that's acceptable if they hire a commercial bench; watch for a COO/CRO appointment.

Forensic Red Flags (+private: web-only, unaudited)

No audited financials exist — so "forensic accounting" reduces to disclosure quality + governance signals, all, unaudited.

The headline red flag — the missing Form D. Technical.ly's analysis found that Skild's announced raises are not backed by Form D filings under the Skild name with the SEC. The only "Skild" Form D is a special-purpose vehicle (May filing) logging just over $1M raised — against the $814M total PitchBook reported at the time, and despite the $500M (Apr 2025) and $300M (Jul 2024) rounds. Startups are required to file Form D (Reg D exemption), though they may file under an alias.

• How to read it (balanced): Form D non-filing/under-an-alias is common and not per se illegal; large startups frequently file under an LLC/SPV name or rely on a different exemption. It is not evidence of fraud. But it is a real transparency gap that makes the announced round sizes/valuations unverifiable against a primary regulatory source — which is exactly why the $14B mark must be treated as not [fact]. It belongs at the top of the diligence list (Lens 14, Q-priority).

Revenue-claim verifiability. The "0 → ~$30M live revenue in months" figure is company-sourced (the Series C blog), uncorroborated by any filing. Standard for a private — but combined with the Form-D gap, the prudent stance is to discount self-reported traction until a strategic partner or filing corroborates it.

Regulatory / legal findings (required sub-section).

• SEC (EDGAR LR + AAER): None. Per the Stage-1 pre-fetched file: "Skild AI has no CIK — it is private and not required to file with the SEC. No EDGAR enforcement search is possible." total_sec_findings: 0.
• Non-SEC (FTC/DOJ/FDA/CFPB): Web search "Skild AI" (FTC OR DOJ OR lawsuit OR litigation OR fine OR penalty OR investigation) returned no Skild-specific enforcement or litigation — only general AI-enforcement-trend coverage. No consent decrees, fines, or settlements found.
• Item 3 (Legal Proceedings): n/a — no 10-KForm 10-KA company’s audited annual report to the US regulator. The most complete thing it publishes. exists (private).
• Conclusion: No material regulatory or legal findings — verified via SEC EDGAR EFTS (LR, AAER, none/no-CIK), web search (no hits), and 10-K Item 3 (n/a, private) as of 2026-06-18. The governance flag (Form-D gap) is a transparency concern, not an enforcement action.

---

Phase D — Project & stress-test

IPO-Readiness & Path-to-Tradeable (+private swap for "Forward Projection")

Note: research/private-watch.json has no Skild entry, so there is no pre-graded stage/ipo_readiness/catalyst to anchor — this is my against the schema's 1–5 readiness scale (1 early → 5 S-1 filed/IPO imminent).

Estimated IPO-readiness: 3 / 5 (late-stage), trending toward 4. Reasoning:

• Capital/syndicate say 4 — SoftBank/NVIDIA/Macquarie/Coatue/Mirae is a pre-IPO-grade book, and $14B is public-IPO scale.
• Fundamentals say 2–3 — ~$30M unaudited revenue, ~50 employees, first commercial deployment in progress not proven, and a Form-D/disclosure gap that would have to be cleaned up before any S-1. No company files an S-1 on $30M of self-reported revenue at $14B.

Milestones that unlock an S-1 (the gating path):

1. Foxconn/Blackwell line ships measurable throughput (uptime, cycle-time, defect rate vs. human/conventional automation) — the single proof that the brain works in production, not just demos.
2. Revenue scales to a credibly-public number (call it ~$200–400M ARR, audited) — i.e. several Foxconn-scale deployments, not one.
3. Disclosure/governance cleanup — proper audited financials + the Form-D ambiguity resolved.
4. A durable-moat proof vs. Physical Intelligence + GR00T (retention/expansion within deployed accounts).

Estimated window: No realistic IPO before 2028–2029. Before then, the tradeable event is far more likely a structured secondary / tender (the syndicate quality supports one) or — a real tail risk — acquisition by a strategic (NVIDIA, SoftBank-orchestrated, or a body-maker buying the brain). Covariant's acqui-hire into Amazon is the cautionary template for how a model-lab can get absorbed rather than IPO.

Brier forecast (would-log, but suppressed in --watchlist): "Skild AI files an S-1 or completes an IPO by 2028-12-31 — p ≈ 0.25". Per --watchlist rules and the strict wave boundary, no our model create is run in this unattended sweep.

Bull vs Bear

Bull case. Value in robotics migrates from the body to the brain, exactly as it did from PC hardware to Windows and from handsets to iOS. Skild is the best-capitalized pure model-lab with the strongest strategic syndicate (NVIDIA + SoftBank + Samsung + LG + Schneider as distribution), a genuinely differentiated omni-body + internet-video-data recipe that sidesteps the fleet-ownership moat trap, top-1% robot-learning founders, and — critically — the first real production deployment (NVIDIA's own Blackwell line via Foxconn) that no other pure model-lab can claim. If one model can truly run any body, Skild becomes the toll road for the entire $22.8B-by-2034 physical-AI layer, and $14B looks early. The secular tailwind (labor automation, 41.2% sector CAGR, 50K+ foundation-model-trained robots by 2026) is as strong as any in tech.

Bear case (permanent-impairment risks).

1. The omni-body claim doesn't hold at deployment scale. "Adapts to any body" is easy in a sizzle reel and brutal on a real Foxconn line with safety, cycle-time, and reliability constraints. If the brain needs heavy per-deployment post-training, Skild is a services company, not a platform — and the multiple collapses.
2. Commoditization from above and below. NVIDIA ships GR00T (partly open) — the same partner brokering Skild's deals can substitute its own brain; Physical Intelligence ($5.6B, well-funded) + Google DeepMind + open-source push the model layer toward zero-margin. The robot-software layer may be a race to commoditized infrastructure, not a winner-take-all monopoly.
3. The price. ~470× run-rate revenue bakes in category dominance. Any disappointment — Foxconn line underperforms, a builder defects to its own brain, a down-round at a peer — re-rates the whole private robot-software complex.

Pre-mortem (18 months out, thesis broke): The Foxconn/Blackwell deployment underdelivered on throughput; NVIDIA quietly favored GR00T for its own ecosystem; body-makers (ABB/UR/Figure-class) concluded the brain is better built in-house or bought cheaply from a commoditizing field; Skild's ~$30M revenue stalled while burn climbed; a Physical Intelligence or a Chinese model-lab matched the capability at a fraction of the mark; and the Form-D opacity turned into a governance headline during the next raise. The $14B mark became the high-water print.

Are multiples too high? There is no multiple — it's an option price. At $14B on ~$30M unaudited revenue, you are not buying a business; you are buying a call option on owning the robot-OS layer. That can be the right bet — but it is a venture bet priced like a near-certainty.

Contrarian view (what the market refuses to see): The consensus treats "robot foundation model" as the obvious next platform monopoly. The thing being under-weighted: in robotics the body-maker may keep the customer, and the brain may end up the commoditized, NVIDIA-subsidized layer — the inverse of the software-eats-hardware thesis. The most dangerous scenario for Skild isn't a competitor out-modeling them; it's the layer itself not being defensible.

Devil's Advocate (short-seller)

Dismantling the bull case:

• Where revenue is concentrated: ~$30M, unaudited, with the lighthouse account being NVIDIA's own Blackwell line (via Foxconn) — i.e. revenue concentrated in deployments brokered by Skild's own investor. Strip out the NVIDIA-orchestrated demand and ask: what's the organically-won book? Not disclosed. If the flagship customer is effectively your lead investor showcasing its portfolio company, that's a circular-credibility risk, not a market validation.
• Why the moat is weaker than bulls think: The one moat the sector says matters — fleet/teleop data — is the one a no-hardware company structurally lacks. Skild's answer (sim + internet video) is clever but unproven to close the real-world gap at production reliability. Meanwhile switching costs cut against Skild whenever a body-maker prefers in-house control of its policy stack.
• The most dangerous competitor bulls underestimate: NVIDIA itself. GR00T is partly open, NVIDIA owns the inference silicon every robot uses, and it's inside Skild's cap table and deal flow. NVIDIA can commoditize the brain to sell more Jetson/Thor + Blackwell — its incentive is a cheap, ubiquitous robot brain, which is the opposite of Skild's $14B-monopoly thesis. Physical Intelligence (Alphabet/CapitalG-backed) is the #2 threat.
• Worst governance/capital-allocation signals: the Form-D gap (raises unverifiable against SEC filings; only a ~$1M SPV on record vs. $800M+ announced) + self-reported revenue + a valuation that tripled twice in a year on conviction, not audited metrics. None of this is illegal; all of it is un-auditable, which is precisely the short's edge.
• Assumptions that must hold for $14B: (1) one model truly generalizes across bodies in production; (2) the brain layer is defensible vs. open-source + NVIDIA + Physical Intelligence; (3) body-makers stay as channel, not as in-house competitors; (4) revenue compounds from $30M toward hundreds of millions within ~2 yrs. If growth disappoints by 20–30%, there's no earnings floor to catch the fall — a private re-rate could be brutal and is the base case in a robotics-funding-winter.
• The single scenario that permanently impairs the business: the robot-software layer proves non-defensible (commoditized by NVIDIA/open-source) while body-makers retain the customer — leaving Skild a well-funded research lab with no durable margin pool. Plausibility: moderate and rising, and the thing to underwrite against hardest.

Management Questions (ordered by information value)

1. The Form D: Why are the $300M and $500M rounds not on file as Form Ds under Skild AI, with only a ~$1M SPV on record? Walk through the exact legal structure and exemption used for each round. (Highest info value — it gates the credibility of every other number.)
2. Foxconn/Blackwell line: What are the measured production metrics to date — uptime, cycle-time, defect/intervention rate vs. the human or conventional-automation baseline it replaced?
3. Of the ~$30M 2025 revenue, what share is from NVIDIA-orchestrated deployments (Foxconn/Blackwell, ABB/UR intros) vs. independently-won customers? What's net revenue retention on the earliest accounts?
4. Per-deployment economics: For a new robot platform, how much targeted real-world post-training is required before production reliability — measured in engineer-weeks and real-robot-hours? (This decides platform-vs-services.)
5. On NVIDIA as partner and competitor (GR00T): What contractually prevents NVIDIA from favoring its own brain on its own silicon/ecosystem? What's your moat the day GR00T is good enough and free?
6. Generalization evidence: Show the quantitative cross-body transfer result — success rate on a body the model never trained on, at production tolerances, not a demo.
7. Defensibility vs. open-source + Physical Intelligence: Where, specifically, is the durable moat — data, weights, distribution, or switching cost — and how do you measure it widening?
8. Path to a public-grade number: What audited revenue/ARR and how many Foxconn-scale deployments do you need before an S-1 is credible, and on what timeline?
9. Burn & runway: What's the monthly burn against the ~$1.8B raised, and what does the next 24 months of compute+data spend buy in capability terms?
10. Commercial bench: Who runs go-to-market? What's the plan to convert two researcher-founders' company into a sales-led enterprise (COO/CRO hires)?
11. Customer ownership risk: When you ride ABB/UR/Foxconn bodies, who owns the end-customer relationship and the renewal — you or the body-maker?
12. IQT / national-security vertical: What does In-Q-Tel's involvement unlock (procurement, classified deployments), and how large is the security/inspection opportunity?
13. China & export controls: The VinDynamics/Vingroup MOU and the "US falling behind China on hardware deployment" critique — how do you think about Chinese low-cost bodies + export-control exposure for your model?
14. Capital-allocation discipline: Given $14B prices a near-monopoly, what's the honest probability you don't win the category, and what's the plan (and the down-side for employees/investors) if the layer commoditizes?
15. Acquisition vs. independence: Covariant got absorbed into Amazon. Under what conditions would you sell the brain to a strategic rather than stay independent to an IPO?

---