Reference Index — Sector 4: The Labs and the Buyers of Physical-World Data

Companion reference file for the thesis (parts I–VI). Every entry gives the entity's status, its current market cap or last disclosed valuation with an as-of date, its latest reported financials, the specific thesis claims that rest on it, and working links. Compiled 2026-07-14.

OpenAI

What it is: The largest single buyer of AI compute and, since 2026, a first-party collector of physical-world data — it restarted an in-house robotics program that teleoperates its own arms rather than buying demonstration data from vendors, which is the thesis's exhibit for "the labs stopped trusting the label layer." Status: Private (OpenAI Group PBC, controlled by the OpenAI Foundation). Microsoft holds ~27% of the PBC. Market cap / valuation: $852B post-money, announced 2026-03-31 on the close of a $122B round — the largest private funding round on record. Anchors: Amazon up to $50B, Nvidia $30B, SoftBank $30B; also a16z, D.E. Shaw Ventures, Microsoft (size undisclosed), and ~$3B raised from retail/individual investors through bank channels. Latest financials: ~$2B in revenue per month (~$24B annualized run-rate) as stated at the round's close, March 2026; 900M+ weekly ChatGPT users, of whom 50M+ are paying subscribers. Still lossmaking on a heavy compute-commitment base (~$300B Oracle cloud contract). Cited for: (a) the $852B/$122B round as the top of the private-capital stack — I.1 (OUTLINE §"the private stack"; ARGUMENT §I.1); (b) circular financing — Nvidia's up-to-$100B commitment to OpenAI, OpenAI's ~$300B of Oracle cloud, Oracle buying Nvidia chips — ARGUMENT §I.1 move 2; (c) the robotics restart: in-house collectors teleoperating Franka arms via GELLO rigs, hiring across actuator design, DAQ stations, and simulation — I.3 (the "~10 buyers" of physical-world data); (d) OpenAI's wind-down of Scale AI work after the Meta deal — I.4, the neutrality law. Links:

Anthropic

What it is: The second pole of the frontier-lab capital stack and, in the thesis, the reference point for how much a lab will pay for *environments* rather than labels — the reported ">$1B on RL environments" line that anchors Part II's argument that the label layer is being replaced by interactive data. Status: Private (Anthropic PBC). Market cap / valuation: $965B post-money, Series H closed 2026-05-28, raising $65B (led by Altimeter, Dragoneer, Greenoaks, Sequoia; co-led by Capital Group, Coatue, D1, GIC, ICONIQ, XN). This *supersedes* the $380B Series G ($30B, 2026-02-12, led by GIC and Coatue) that the thesis currently cites, and puts Anthropic above OpenAI's $852B as the most valuable private AI company. Latest financials: Run-rate revenue crossed $47B in May 2026 (disclosed with the Series H), up from $14B at the Series G in February 2026 and ~$9B at end-2025; Claude Code alone passed $2.5B run-rate by Feb 2026. Compute commitments disclosed at the Series H: up to 5 GW with Amazon, 5 GW of TPU capacity with Google/Broadcom, plus GPU capacity from SpaceX's Colossus 1 and 2. Cited for: (a) the private-capital stack table — I.1 (OUTLINE §"the private stack" lists "$30B Series G at $380B"); (b) Anthropic leaders reportedly discussing spending >$1B on RL environments — II.1/II.2 (OUTLINE part-2 lines 56, 85; traceable to The Information's reporting); (c) as one of the ~10 buyers whose demand sets the clearing price for training data. Links:

Alphabet / Google DeepMind

What it is: The hyperscaler that is also a first-party robotics-data collector — Gemini Robotics is trained on data captured in-house (ALOHA-2 teleop) *and* streamed continuously from Apptronik's "Robot Park" and from Hyundai/Boston Dynamics factory floors; the thesis uses it as the cleanest example of a lab that owns its physical-data pipeline end-to-end. Status: Public (NASDAQ: GOOGL / GOOG). Google DeepMind is a wholly-owned Alphabet division. Market cap / valuation: ~$4.3T as of 2026-07-10 (Alphabet crossed $4T on 2026-01-12 and traded in the $4.28–4.34T range through mid-July 2026). Latest financials: Q1 2026 (reported 2026-04-29): net income $62.6B, up 81% y/y; EPS $5.11, up 82%; Google Cloud revenue grew 63% to over $20B, cloud backlog nearly doubled q/q to over $460B. Capex $35.7B in Q1 2026 alone (~60% servers). Full-year 2026 capex guidance raised to $180–190B (from the $175–185B given in February 2026, itself roughly double the ~$91B spent in 2025), with 2027 capex to "significantly increase." Cited for: (a) the ~$725B aggregate 2026 hyperscaler capex figure and the per-company line "Alphabet ~$175–185B (from ~$91B in 2025)" — I.1 (OUTLINE lines 29–32; ARGUMENT §I.1 "the measured base"); (b) Google DeepMind as one of the ~10 buyers of physical-world data — the Apptronik "Robot Park" capture facility in Austin (~90,000 sq ft, expanded facility announced 2026-06-30; Apollo 2 fleets collecting continuously; further Robot Parks at DeepMind itself and at customers Mercedes-Benz and GXO), feeding the Gemini Robotics models — I.3 (OUTLINE lines 227–231; ARGUMENT §I.3); (c) Google as Scale AI's largest customer (~$200M planned annual spend) that split within days of the Meta deal — I.4, the neutrality law. Note for drafting: the thesis's "Alphabet ~$175–185B" is the February 2026 guidance; the current guidance is $180–190B. Update or date-stamp the figure. Links:

NVIDIA

What it is: Simultaneously the toll-taker on every lab's compute bill, an equity investor in nearly every lab it sells to, and — through GR00T/Cosmos/Isaac — the vendor of the robot foundation model itself; it also supplies the thesis's single most load-bearing empirical result, the EgoScale scaling law that says egocentric *human* video buys robot dexterity predictably. Status: Public (NASDAQ: NVDA). Market cap / valuation: ~$4.9T as of June 2026 — the most valuable listed company in the world. Latest financials: Q1 FY2027 (quarter ended 2026-04-26, reported 2026-05-20): revenue $81.6B, up 85% y/y; Data Center revenue $75.2B, up 92% y/y and 21% sequentially, on Blackwell 300 adoption. Announced an $80B additional buyback authorization and raised the quarterly dividend from $0.01 to $0.25/share. FY2026 (ended Jan 2026) full-year revenue ~$215.9B. Cited for: (a) the Sequoia "$600B question" arithmetic — take Nvidia's run-rate data-center revenue, double it for total cost of ownership, add margin, and ask where the revenue is — ARGUMENT §I.1 (lines 53–54); (b) circular financing: Nvidia commits up to $100B to OpenAI and is an equity investor in OpenAI, xAI, SSI, Figure, Skild, Physical Intelligence and Thinking Machines while remaining their chip supplier — ARGUMENT §I.1 moves 1–2, and the "neutrality law" tension in I.4 (lines 388–391: the defence is that Nvidia sells fungible compute, not a capture-able dataset); (c) EgoScale: a log-linear scaling law, R²=0.9983, no saturation, from pretraining on 20,854 hours of egocentric human video — 1k→20k hours more than doubles average task completion, and a human-pretrained policy transferred to a Unitree G1 gave a +30 percentage-point absolute success-rate improvement over G1-only training. This is the empirical backbone of Part I's "the missing data is human data" claim and of GR00T N1.7 (June 2026) — I.2/I.3 (OUTLINE lines 130, 186–193); (d) depreciation: hyperscalers depreciate Nvidia hardware over 5–6 years against a shorter real economic life (Burry bear case) — ARGUMENT §I.1 move 3; (e) NVIDIA Halos (ANAB-accredited ISO/IEC certification anchor, Jun 2026) and Isaac's adoption of the MCAP log format — II.4 and IV (OUTLINE part-2 line 246; part-4 lines 436, 444, 472–474). Links:

Meta Platforms

What it is: The only buyer with an *ambient* egocentric-capture funnel already in consumers' hands — a ~9–10M-unit smart-glasses installed base — paired with a dedicated humanoid program ("Metabot") whose stated strategy is to be the Android of robotics; it is also the antagonist of the thesis's neutrality law, having bought 49% of Scale AI and destroyed its neutrality. Status: Public (NASDAQ: META). Market cap / valuation: ~$1.69T as of 2026-07-10 (companiesmarketcap/stockanalysis). The stock fell 12.8% between the Q1 filing on 2026-04-29 ($668.53) and 2026-07-02 ($582.90) on capex anxiety. Latest financials: Q1 2026 (reported 2026-04-29): revenue $56.3B, up 33% y/y; operating income $22.9B, up 30%; net income $26.8B, up 61%. Full-year 2026 capex guidance raised to $125–145B (from $115–135B), nearly double 2025 and more than 2025 and 2024 combined; the revision added ~$107B in new contractual commitments. Meta depreciates AI hardware over 5.5 years. Cited for: (a) the aggregate hyperscaler capex marker, "Meta ~$125–145B" — I.1 (OUTLINE lines 29–32; ARGUMENT line 375); (b) depreciation-schedule extension (Meta 5.5 yrs) as a component of the Burry bear case — ARGUMENT §I.1 move 3; (c) the ~9–10M-unit smart-glasses installed base as an ambient egocentric-capture funnel — ARGUMENT §I.3 (line 153). Sourcing: EssilorLuxottica sold over 7M AI glasses in 2025 alone (Ray-Ban + Oakley), on top of ~2M across 2023–24 combined — i.e. ~9M cumulative; Meta holds ~82% of global smart-glasses shipments, and Meta/EssilorLuxottica were discussing doubling capacity to 20M units/yr by end-2026; (d) "Metabot" — a humanoid team inside Reality Labs led by Marc Whitten (ex-Cruise CEO), with MIT's Sangbae Kim hired in, pursuing a platform/licensing ("Android of robotics") strategy with world-model support from Meta Superintelligence Labs — I.3 (OUTLINE lines 262–267); (e) the Scale AI acquisition — ~$14.3B for a 49% non-voting stake, June 2025, implying ~$29B — and the resulting customer flight, which is the natural experiment of I.4 (OUTLINE lines 122–123, 322–333); (f) the mid-2025 talent raid: Ruoming Pang (ex-Apple) reported above $200M, the disputed Tulloch offer — I.1 (OUTLINE lines 75–81, 98–99; flag as *reported offers*, not paid comp). Links:

Microsoft

What it is: The largest single capex line in the hyperscaler stack and OpenAI's landlord-turned-shareholder (~27% of OpenAI Group PBC); in the thesis it appears not as a physical-data collector but as (i) a component of the capex base, (ii) an early mover on depreciation-schedule extension, and (iii) one of the labs that walked away from Scale AI the week Meta bought in. Status: Public (NASDAQ: MSFT). Market cap / valuation: ~$3.07T as of May 2026. Latest financials: Q3 FY2026 (quarter ended 2026-03-31, reported 2026-04-29): revenue $82.9B, up 18% y/y; operating income $38.4B, up 20%; Azure and other cloud services grew 40% y/y (fastest in over a year); Microsoft Cloud revenue $54.5B in the quarter (>$200B annualized run-rate); the AI business hit a $37B annual run-rate, up 123% y/y. Capex $31.9B in Q3 with Q4 guided above $40B. Microsoft has since called for ~$190B of calendar-2026 capital spending — about $35B above the $154.6B analyst consensus, with ~$25B of the increase attributed to higher memory/component prices. Cited for: (a) the aggregate hyperscaler capex marker — the thesis's per-company line is "Microsoft ~$110–120B" for 2026 (OUTLINE I.1 line 32), which is now badly stale against the company's own ~$190B calendar-2026 figure; (b) depreciation: Microsoft extended server depreciation from 4 to 6 years, a core plank of the Burry bear case — ARGUMENT §I.1 move 3 (line 81); (c) Microsoft's pullback from Scale AI after the Meta deal — I.4 (OUTLINE lines 332–333, 376). Note for drafting: if the ~$725B aggregate is retained, it must be recomputed — Microsoft ~$190B, Alphabet $180–190B, Meta $125–145B, Amazon ~$200B now sums to roughly $695–735B, so the total survives but the Microsoft component does not. Links:

Amazon

What it is: The largest 2026 capex line, an anchor investor in OpenAI (up to $50B) and Anthropic's compute counterparty (up to 5 GW) — and, separately, the owner of the Delivery Service Partner franchise chassis that Part V reads structurally as the template for financing a robot-operator network, plus the biggest single deployer of Agility's Digit. Status: Public (NASDAQ: AMZN). Market cap / valuation: ~$2.6–2.7T as of 2026-07-13. Latest financials: Q1 2026 (reported 2026-04-29): revenue $181.5B, up 17% y/y; operating income $23.9B, up 30% (13.1% margin); AWS revenue $37.6B, up 28% — its fastest growth in fifteen quarters; AWS backlog $364B, excluding a ~$100B Anthropic deal. Capex $44.2B in Q1 alone, up 77% y/y, tracking to ~$200B for full-year 2026 (vs ~$123B in 2025 and ~$77B in 2024). Trailing-twelve-month free cash flow fell to $1.2B, down ~95% y/y, on AI investment. Cited for: (a) "Amazon ~$200B" in the 2026 hyperscaler capex line — I.1 (OUTLINE lines 29–32); (b) Amazon as an anchor of OpenAI's $122B round (up to $50B) — I.1; (c) the DSP franchise economics that Part V reads as the operator chassis: $10–30k operator entry, 20–40 vans, financing through Element Fleet, an Amazon scorecard (Fantastic/Great/Fair/Poor) gating route allocation, $1M–$4.5M/yr revenue at 7–10% net margin, and Amazon-guided owner earnings of $75k–$300k/yr — V.3 and VI.3 (part-5 OUTLINE lines 205–226; part-4 line 167; part-6 line 59); (d) Amazon as the largest site in Agility Digit's >100-unit deployed base, and as a Foxglove customer — VI.1/VI.2. Links:

Apple

What it is: The weakest physical-world-data signal among the megacaps — no robot foundation model, no capture fleet — but a 2027 tabletop robot and a 2026 home hub put it on the board as a latent buyer; separately it is the private-sector counterparty in the MP Materials magnet structure that Part III leans on. Status: Public (NASDAQ: AAPL). Market cap / valuation: ~$4.65T as of July 2026 — the world's second most valuable company. Latest financials: Q2 FY2026 (March quarter, reported 2026-04-30): revenue $111.2B, up 17% y/y — the largest non-holiday quarter in Apple's history; diluted EPS $2.01, up 22%; Products $80.2B, Services an all-time-high $31.0B (+16%); iPhone $57.99B, up 22%. Operating cash flow >$28B in the quarter; dividend raised 4% to $0.27; a further $100B buyback authorized. Cited for: (a) the "weaker signal" robotics read — a home smart hub targeted for 2026 and a tabletop robot (an iPad on a motorized arm) targeted for 2027, both on a new "Charismatic" OS — I.3 (OUTLINE lines 268–271); (b) Apple's $500M offtake with MP Materials for recycled NdFeB magnets from Fort Worth, starting 2027, as the private half of the state-backed demand template — III.2 and VI.4 (part-3 OUTLINE lines 203, 211; part-6 lines 64, 114); (c) Ruoming Pang, Apple's ex-head of foundation models, as the highest-profile loss in Meta's 2025 talent raid — I.1 (line 77). Links:

ByteDance (Seed Robotics)

What it is: The one buyer that owns both the robot and the consumer headset that captures the human data to train it — Seed Robotics' GR-3 model is fine-tuned on human trajectories captured through ByteDance's own VR hardware, which is the thesis's proof that a consumer device can be a data-acquisition instrument. Status: Private (ByteDance Ltd., Cayman/Beijing). PICO is a wholly-owned ByteDance subsidiary. Market cap / valuation: ~$550B implied by a General Atlantic secondary stake sale in early 2026, up from ~$480B in a November 2025 share auction and $330B at the September 2025 employee buyback. No primary-issuance round; these are secondary marks. Latest financials: Not a filer — ByteDance publishes no audited statements. Reporting indicates ByteDance surpassed Meta in total annual sales in 2025, with 2025 annual profit projected around $48B. Treat both as press estimates, not disclosed figures. Cited for: (a) ByteDance as one of the ~10 buyers of physical-world data, running in-house egocentric capture through its own PICO headsets to fine-tune GR-3 — I.3 (OUTLINE lines 239–243; ARGUMENT line 158). The GR-3 technical report is the primary: it describes co-training on web-scale VLM data plus efficient fine-tuning from human trajectory data collected via VR devices (egocentric video + hand trajectories), paired with the ByteMini 22-DoF dual-arm mobile robot; (b) the reported ByteDance robotics spend figure ("80B", from a single Chinese-language source) — flagged in OUTLINE line 289 as unverified, and it remains so: could not source that figure to a primary link as of 2026-07-14. Caveat for drafting: the GR-3 report says "VR devices," not "PICO 4" specifically, and gives no "~450 [units/hours]" figure. The PICO attribution is a reasonable inference (PICO is ByteDance's own headset line) but is not stated in the primary. Either soften to "its own VR headsets" or source the PICO 4 claim separately. Links:

Physical Intelligence (π)

What it is: The purest instance of the thesis's central move — a lab that self-collects its own robot demonstrations rather than buying them, and open-sources the model while keeping the data. Status: Private. Founded by Karol Hausman, Sergey Levine and Chelsea Finn; San Francisco. Market cap / valuation: $5.6B post-money — Series B, November 2025 — is the last *confirmed, closed* mark. A raise of ~$1B at a valuation above $11B was reported by Bloomberg on 2026-03-27 (Founders Fund and Lightspeed in talks; existing backers Thrive, Lux) and was still being reported as "in talks / targeting" as recently as 2026-07-13. It has not been announced as closed. Latest financials: No revenue disclosed; pre-commercial. Nvidia is an investor. Cited for: (a) ~10K hours of self-collected robot demonstrations behind π0/π0.5, and PI's status as both a self-collector and a buyer — I.3 (OUTLINE lines 217–221; ARGUMENT line 199); (b) the "PI $11B" valuation in the [NEW] figures list — I.3 (OUTLINE line 284); (c) Nvidia's equity stake in PI, part of the circular-financing pattern — ARGUMENT §I.1 (line 77). Correction for drafting: the thesis states PI $11B as fact. As of 2026-07-14 the $11B is a *reported round in talks*, not a closed round; the last confirmed valuation is $5.6B (Series B, Nov 2025). State it as "reported to be raising at >$11B" or cite $5.6B. Links:

Skild AI

What it is: The sim-first counter-case to Physical Intelligence — Skild Brain is pretrained largely in simulation (Isaac Lab + Cosmos) with real data reserved for post-training, and the thesis uses the pair to frame the open question of how much real-world data a robot foundation model actually needs. Status: Private. Founded 2023 (Pittsburgh/CMU lineage). Market cap / valuation: >$14B post-money, Series C of ~$1.4B announced 2026-01-14, led by SoftBank Group, with NVentures (NVIDIA), Macquarie Capital, Jeff Bezos (Bezos Expeditions), Disruptive and 1789 Capital. That is roughly 3× the $4.5B valuation of its previous round seven months earlier (summer 2025). Latest financials: No revenue disclosed; pre-commercial foundation-model company. Cited for: (a) Skild as one of the ~10 buyers, with the sim-first strategy (Isaac Lab + Cosmos, real data for post-training) — I.3 (OUTLINE lines 222–226); (b) the "$1.4B raise, ~$14–15B valuation, ~3× in seven months" figure in the [NEW] list — I.3 (line 284); (c) Nvidia's equity stake, again part of the circular-financing pattern — ARGUMENT §I.1 (line 77). The omni-bodied claim — Skild Brain controlling robots it never trained on and adapting in real time to lost limbs, jammed wheels, or added payload without retraining — is the company's own framing in the Series C announcement. Links:

xAI

What it is: A frontier lab that, as of 2026, is no longer a standalone private company — it was absorbed into SpaceX, which then executed the largest IPO in history; in the thesis it is one of the private-stack valuations and one of the labs that fled Scale AI after the Meta deal. Status: Now a wholly-owned subsidiary of Space Exploration Technologies Corp. (NASDAQ: SPCX) following an all-stock triangular merger announced 2026-02-02 (share exchange: 1 xAI share → 0.1433 SpaceX shares). Market cap / valuation: Standalone: $20B Series E at a ~$230B valuation, January 2026 (Valor, Nvidia, Cisco Investments, StepStone, Fidelity, QIA, MGX, Baron; Saudi HUMAIN $3B reported). In the merger, xAI was valued at $250B against SpaceX's $1T, for a combined $1.25T. Post-IPO: SPCX listed 2026-06-12 at $135, closed day one at $160.95 (+19%) for a market cap around $2.1T, having raised ~$75B; the stock has since fallen back toward the $135 IPO price (2026-07-13). Latest financials: xAI does not report separately; its results are now consolidated into SPCX. Merger rationale as stated by Musk: vertically integrating AI, rockets and satellite internet to build orbital data centers. Cited for: (a) the private-capital stack: "xAI — $20B Series E, Jan 2026, ~$230B valuation … then folded into an all-stock SpaceX–xAI merger valuing the combined entity at ~$1.25T (SpaceX $1T + xAI $250B), Feb 2026" — I.1 (OUTLINE lines 44–47); (b) xAI's withdrawal from Scale AI after the Meta deal — I.4 (OUTLINE lines 332, 376; ARGUMENT line 184); (c) Nvidia's equity stake in xAI — ARGUMENT §I.1 (line 77). Correction for drafting: the thesis lists xAI in the *private* stack. Since 2026-06-12 its parent trades publicly as NASDAQ: SPCX; the "$230B/$250B" marks are now historical, and the live number is SPCX's public market cap (~$1.8–2.1T range since listing). Anthropic's compute deal for GPU capacity in Colossus 1 and 2 also now runs to this entity. Links:

Mistral AI

What it is: The European entry in the private-capital stack, and — separately — the model family whose benchmark scores drop most visibly under decontamination, which is why it appears in Part II's contamination argument. Status: Private (Paris). Largest shareholder: ASML, which bought an 11% stake for €1.3B leading the Series C. Market cap / valuation: €11.7B (~$14B) post-money, Series C of €1.7B, September 2025 (ASML leading). In June 2026 Bloomberg reported Mistral in early talks to raise ~€3B at a ~€20B (~$23B) valuation — nearly double — still at an early stage and not closed as of 2026-07-14. Total raised to date ~$4B (PitchBook), against OpenAI's ~$186B and Anthropic's ~$161B. Latest financials: Revenue not disclosed. Financing detail the thesis uses: $830M of debt financing raised in March 2026 for data-centre expansion (reported as backing a purchase of ~13,800 Nvidia chips) — a European lab tapping credit rather than equity for compute. Stated goal: 1 GW of own compute capacity by 2030. Cited for: (a) the private stack — "Mistral AI — €1.7B Series C at ~€11.7B/$14B, Sept 2025 (ASML leading)" and the €3B/€20B rumour — I.1 (OUTLINE lines 56–60); (b) the debt-for-compute move (~$830M, ~13,800 Nvidia chips) — I.1; (c) benchmark contamination: swapping contaminated items for clean mirrors drops scores by up to ~13 points, with Mistral and Phi families showing ~10–13-point drops versus inflated scores — II.5 (OUTLINE part-2 lines 409–411). Links:

Thinking Machines Lab

What it is: The clearest marker of where private AI capital's price discipline broke and then reasserted itself — a $12B seed valuation with no product, followed by a collapsed attempt to raise at $50B+. Status: Private. Founded by Mira Murati (ex-OpenAI CTO). Market cap / valuation: $12B post-money on a $2B seed, July 2025 — the largest seed round in AI history (investors: Nvidia, Accel, ServiceNow, Cisco, AMD, Jane Street). Bloomberg reported talks in November 2025 at $50–60B; those talks collapsed without a deal by January 2026. As of mid-2026 the company still carries the $12B mark — the last actual round. Latest financials: No revenue disclosed. ~140–170 employees; two shipped products (including Tinker); billions in infrastructure commitments from Nvidia and Google. Cited for: (a) the private stack — "Thinking Machines Lab — $2B seed at a $12B valuation, July 2025 (Nvidia, Accel…)" — I.1 (OUTLINE lines 52–55, 94); (b) the Andrew Tulloch offer from Meta (reported at $1.5B, disputed) — I.1 (OUTLINE lines 78, 98–99: state as a *reported offer*, not paid comp; Tulloch subsequently left TML for Meta); (c) Nvidia's equity stake in TML — ARGUMENT §I.1 (line 77). Correction for drafting: if the thesis anywhere implies TML's valuation kept climbing, it did not — the $50B round failed and $12B stands. That failure is arguably *better* evidence for Part I's argument about the limits of the capital flow than a successful markup would have been. Links:

Safe Superintelligence (SSI)

What it is: The limit case of the capital flow — the highest-valued AI lab in the world with no commercially available product — which is the exhibit the thesis uses to show that the money is pricing a research bet, not a revenue line. Status: Private. Founded June 2024 by Ilya Sutskever, Daniel Gross (since departed to Meta) and Daniel Levy. Market cap / valuation: ~$32B, reported April 2025 on a $2B raise; total raised ~$6B since June 2024 (first round: $1B at $5B, September 2024). Backers include Alphabet, Nvidia, Greenoaks, Sequoia, a16z, Lightspeed. No newer mark has been announced as of 2026-07-14. Latest financials: Zero revenue — no product. That is the point of the citation. Cited for: (a) the private stack — "Safe Superintelligence (SSI) — ~$32B valuation on ~$6B raised since June 2024, with no product (Greenoaks-led $2B extension; Alphabet, Nvidia backing)" — I.1 (OUTLINE lines 48–51, 94); (b) Nvidia's equity stake in a lab it also sells chips to — ARGUMENT §I.1 (line 77). Note: Sutskever's public framing (2026) — that the 2020–2025 "scaling era" is over and a new research era begins in 2026 in which algorithmic innovation rather than compute drives progress — is directly useful to Part I's "the gravity is bending" argument and to Part II's claim that the binding constraint has moved from compute to data. Links:

World Labs

What it is: Fei-Fei Li's spatial-intelligence lab — the pure-play bet that the missing capability is 3D world modelling, and the clearest commercial instance of *synthetic* physical-world data (its product, Marble, generates persistent 3D environments) standing in for captured data. Status: Private. Co-founded September 2024 by Fei-Fei Li and Justin Johnson. Market cap / valuation: $1B at emergence from stealth (Sept 2024, on $230M). In February 2026 it raised $1B more from NVIDIA, AMD, Autodesk (which anchored with $200M, its largest-ever startup investment, plus a strategic advisor seat), Emerson Collective, Fidelity and Sea. The company did not disclose a post-money valuation; press reporting ahead of the round estimated ~$5B. Total raised: ~$1.23B. Latest financials: Revenue not disclosed. Marble, launched November 2025, is its first commercial product — generating and editing persistent 3D environments from text, image, video or 3D-layout prompts, with free and paid tiers and mesh/video export. Cited for: World Labs is not named anywhere in the current OUTLINE or ARGUMENT files (verified by grep across all six parts, 2026-07-14). It belongs to the sector but the thesis does not yet rest a claim on it. Where it *would* attach: Part II's "synthetic data" vector (OUTLINE part-2 line 75, "capture-supply explosion via smart glasses, synthetic…") and Part I.3's buyer list — World Labs is the counter-argument to the whole thesis, the claim that you can generate the physical-world data rather than collect it. Its valuation being undisclosed, and roughly an order of magnitude below Skild's, is itself a data point about how the market prices synthetic against captured. Links:

Wayve

What it is: The end-to-end driving-foundation-model company — the one buyer whose physical-world data comes from vehicle fleets rather than robot arms or headsets, and a named customer of Foxglove, the data/observability layer Part VI identifies as the best-capitalised claimant on the record substrate. Status: Private (London). Market cap / valuation: $8.6B post-money, Series D, announced 2026-02-24. Round size reported as $1.2B (CNBC/TechCrunch) and $1.5B by Wayve itself, the difference being up to $300M of milestone-based capital from Uber. Led by Eclipse, Balderton and SoftBank Vision Fund 2, with NVIDIA, Microsoft, Uber and automakers Mercedes-Benz, Nissan and Stellantis, plus Ontario Teachers', Baillie Gifford, British Business Bank and Schroders Capital. On 2026-06-30 Wayve opened an $85M employee tender at an $8.5B valuation — a secondary mark essentially flat to the Series D. Latest financials: Revenue not disclosed. Commercial commitments: a Nissan partnership to ship Wayve's AI in driver-assistance systems from 2027, and public robotaxi trials with Uber beginning in London in 2026, targeting 10+ markets. Cited for: Wayve appears in the thesis as one of Foxglove's named customers (alongside NVIDIA, Amazon, Anduril and Dexterity) in the argument that the multimodal robotics data/observability layer is where the record substrate is being claimed — VI.2 (part-6 OUTLINE line 43). It is otherwise not load-bearing; its Series D is not currently cited. Links:

Hyundai Motor Group (and Boston Dynamics / RMAC)

What it is: The closed loop the thesis points at — Hyundai owns Boston Dynamics, owns the factories, owns the actuator supply chain (Hyundai Mobis), and has committed *every* 2026 Atlas unit to its own Robotics Metaplant Application Center and to Google DeepMind; the factory data trains the model that runs the robot that works in the factory. Status: Public (KRX: 005380, Hyundai Motor Co.). Boston Dynamics is a Hyundai Motor Group subsidiary (acquired from SoftBank in 2021; SoftBank retained a minority stake). Market cap / valuation: Hyundai Motor ~₩109.75T (~$84.1B) as of June 2026. Latest financials: Hyundai has announced a $26B investment in its US operations, including a new robotics plant with capacity for 30,000 robots per year. Boston Dynamics does not report separately. Cited for: (a) the Boston Dynamics + Hyundai + Google DeepMind closed loop — electric Atlas production launched at CES 2026; all 2026 Atlas units committed to RMAC and Google DeepMind; Gemini Robotics trained on Hyundai automotive assembly demonstrations; Hyundai factory data feeds RMAC; ~$26B robotics/AI-inclusive US investment — I.3 (OUTLINE lines 255–261; ARGUMENT lines 155–156); (b) the count of buyers — Hyundai/BD is one of the additions that pushes the "~10 buyers" framing toward 12–13 (OUTLINE line 286). Supporting detail (all primary-sourced): Atlas deployment at Hyundai Metaplant America in Savannah, Georgia by 2028, starting with parts sequencing; Hyundai Mobis supplies Atlas's actuators; Boston Dynamics announced the Google DeepMind foundation-model partnership for Atlas cognition; Atlas won CNET's "Best Robot" at CES 2026. Links:

Samsung Electronics / Rainbow Robotics

What it is: The Korean mirror of the Hyundai loop — a megacap that bought control of a humanoid maker outright rather than partnering with one; the thesis flags it as an adjacent buyer that was surfaced but not profiled, and it belongs in the buyer count. Status: Samsung Electronics is public (KRX: 005930). Rainbow Robotics is public (KOSDAQ: 277810) and, since 2024-12-31, a consolidated Samsung subsidiary — Samsung exercised a call option to take its stake to 35.0%, becoming the largest shareholder. Market cap / valuation: Samsung Electronics ~$1.108T (₩1,611T) as of 2026-07-13 — it crossed $1T on the semiconductor cycle. Rainbow Robotics ~₩12.99T (~$9–10B) as of April 2026, the fourth-largest stock on KOSDAQ. Latest financials: Samsung's stake purchases: ₩59B (~$40M) for 10.2% in January 2022, raised to 14.99% in March 2023, then ~$181M to reach 35.0% on 2024-12-31. Samsung has stood up a Future Robotics Office and targets commercialization of an advanced humanoid in 2028, combining Rainbow's hardware with Samsung's AI/software. Cited for: OUTLINE I.3 line 295 explicitly flags "Samsung / Rainbow Robotics and XPeng (IRON) surfaced adjacent but were not profiled this pass" — this entry closes that gap. The structural point the thesis can now make: Samsung did what Meta did to Scale (bought control) but in the *hardware* layer, where the neutrality law does not bite the same way — nobody stops buying Samsung memory because Samsung owns a robot company. That asymmetry between owning a data vendor and owning a hardware vendor is a live argument for I.4. Links: