Legal name: Luma AI, Inc.
Luma AI Deep Dive: From 3D Capture and the Ray3 Reasoning Video Model to the Open Physical AI Lab, Mapping the World Model Strategy and Its Boundaries After a $900M Series C
Luma AI began with 3D capture and now commercialises Ray3 video models and Luma Agents. It launched an Open Physical AI Lab in June 2026, while its text-to-3D assets already feed the RoboCasa simulation pipeline.
Executive Summary
Luma AI, legally Luma AI, Inc., was founded in September 2021 in California. It began as a 3D capture and neural rendering company, entered generative video with Dream Machine in June 2024, released Ray3 in September 2025 as what it calls the world's first reasoning video model, raised a $900M Series C in November 2025 led by HUMAIN, the Saudi Public Investment Fund's AI company, and launched Luma Agents with the Ray and Uni model families in March 2026. In June 2026 it established the Open Physical AI Lab, an open science effort framed around generalization in physical AI. The company states its mission as building unified general intelligence that can generate, understand, and operate in the physical world.
This site's company research list has so far covered 74 companies centred on scientific and engineering simulation, surrogate models, digital twins, spatial intelligence and world models. Luma's revenue comes chiefly from creative generation tools, so its inclusion needs a stated boundary. Two verifiable links justify it. First, the official RoboCasa project page says its kitchen object assets are sourced from Objaverse 1.0, LightWheel AI and AI-generated assets from Luma AI; the original RoboCasa paper records that 1,592 of 2,509 high-quality object assets came from Luma.ai. That places Luma's generative models inside the data production chain of a robot simulation benchmark. Second, the June 2026 lab announcement makes generalization in physical AI a company-level research programme, backed by world model research and a planned 2-gigawatt compute cluster. Inclusion does not equate Luma with CAE surrogate vendors. Luma optimises visual and temporal consistency across generated frames, while most companies on this list optimise verifiable numerical prediction of physical fields; the objective functions, acceptance tests and buyer workflows are different. This report assembles the public record on team, technology, capital and commercial evidence, and rules item by item on the eight groups of conflicting public accounts recorded in DISCREPANCIES. Data is current as of 14 November 2026.
Company Overview
The official About page dates the founding to September 2021 and names Amit Jain as co-founder and chief executive and Alex Yu as co-founder and chief technology officer, with Jiaming Song joining as chief scientist. Jain previously worked on Apple's computer vision team and on product design for Apple Vision Pro; Yu is known for early work on neural radiance fields and real-time neural rendering; Song is the first author of DDIM. Chinese coverage of the Series C also records a September 2021 founding and notes Jain's Apple background. The corporate headquarters is given as Redwood City, California in the official FAQ, while the company's own careers page describes a team in Palo Alto and CNBC likewise uses Palo Alto; one funding database uses San Francisco. These are office, mailing and aggregator conventions rather than a substantive dispute, but they should not be silently merged.
Governance changed during 2026 and the public record is incomplete. A May 2026 press release from Reactor, a real-time generative video infrastructure startup, describes its co-founder and chief executive Alberto Taiuti as having previously co-founded Luma AI and served as its chief technology officer. An AI-generated wiki entry, citing Alex Yu's personal site as of October 2026, describes Yu as the former co-founder and chief technology officer of Luma and says two of three co-founders had moved on by that date; the same source reports that Chief Scientist Jiaming Song announced his departure in May 2026, attributing that to a Chinese technology outlet. None of these departures was the subject of a Luma announcement located in this round of research. The official About page continues to list only Jain and Yu as founders. CNBC's 2026 Disruptor 50 profile also lists two founders, Jain and Yu. The most defensible reading is that the official two-founder account is the company's current position, that Taiuti's co-founder and CTO role is supported by a primary press release from his new company, and that the exact allocation of the CTO title over time remains unconfirmed.
Headcount has only one official data point. The November 2025 Series C announcement says Luma had grown to more than 130 researchers, engineers, designers, artists and operators. A commercial database lists a band of 101 to 500 employees in 2026, which is an inferred range and is marked MEDIA and UNVERIFIED here; it cannot be read as a current headcount. Expansion is easier to verify than staffing: Dream Lab LA opened in July 2025, a London office in December 2025, and a Riyadh office in February 2026, the last two tied to European, UK and Saudi partnerships. For a private company without audited financials, registry and office records, official timelines and named-partner announcements are the higher-grade sources; third-party aggregator fields are retained only as leads.
Product and Technology
Luma's technical lineage starts with 3D capture. The original product used neural radiance fields and Gaussian splatting to turn phone video into explorable 3D scenes, and the Luma App still offers 3D capture workflows. Genie, its text-to-3D model, is described in Alex Yu's own research summary as one of his Luma contributions. A third-party review states that the Genie text-to-3D feature was sunset on 1 January 2026; that is a single non-primary source and is treated as unconfirmed. The company's current emphasis is elsewhere, and the official FAQ asks that Dream Machine no longer be treated as the forward product.
The video line is the best documented. Ray 1 shipped with Dream Machine in June 2024; Ray 2 arrived in January 2025 with a new multimodal architecture; Ray 3 was released on 18 September 2025 and was presented as the first reasoning video model, adding native 1080p, 16-bit high dynamic range output in ACES2065-1 EXR, character references, keyframes and a draft mode. Ray 3 Modify followed in December 2025 for hybrid live-action workflows, Ray 3.14 in January 2026 with faster and cheaper generation, and Ray 3.2 on 9 June 2026 with frame-level control of up to 16 keyframes, 1080p clips up to 20 seconds, HDR with 16-bit EXR export, performance tracking for up to eight faces, reframing tools and a developer API. The official site lists an accompanying Ray3 evaluation report. A media comparison quotes Jain as saying Ray3 benchmarks above OpenAI's Sora 2 and around the level of Google's Veo 3; that is a company-side competitive claim, not an independent benchmark.
On the image and multimodal side, Uni-1 and the Uni-1.1 API of 5 May 2026 are built on what Luma calls Unified Intelligence: the model carries style references, character references and visual grounding across a campaign so that brand context survives from one asset to the next. The API accepts up to nine reference images and launched with creative-tool partners. Luma Agents, introduced in March 2026, wraps these models in an agent workspace that plans, generates and revises across video, image, audio and text. Later features include Luma Skills for saving repeatable workflows, Layers for generating independently editable image layers, and Scenes for turning a brief into an editable storyboard. The workspace also exposes third-party models, including Google Veo, ByteDance Seedance and Seedream, Kling, MiniMax and GPT Image, so Luma is simultaneously a model lab and an orchestration layer over competitors' models. The research page lists work on inductive moment matching for pre-training and terminal velocity matching for inference-time scaling. Enterprise documentation states SOC 2 compliance, ISO certification, private-by-default handling and a commitment not to train on enterprise customer content.
Two boundaries matter for this list. Every metric above measures perceptual quality, temporal continuity, controllability or workflow throughput. None measures conservation error, convergence, contact forces or agreement with a physical test. The models are production-grade creative infrastructure, and the official research is credible, but the technical evidence for physical-world prediction is a separate and much thinner record.
The Physical AI Turn and the Simulation Pipeline
On 1 June 2026 Luma published the announcement of the Open Physical AI Lab. Its argument is explicit: physical AI and robotics remain in a pre-AI-scale era; robots are trained on small examples of specific tasks in specific environments and replay them, which creates a generalization crisis; the default remedy of scaling teleoperation across every task combination is physically and economically impossible. Luma says four years of training frontier foundation models across 3D, image, video and unified language-visual generation produced infrastructure for making sense of internet-scale raw multimodal data, and that empirical evidence from its Unified Models turned a suspicion into a conviction that generalization is not unique to text. The lab will research and scale methods for generalization, develop world models for understanding and interaction with the physical world, collaborate with peer labs and academia, build evaluations to measure progress and assess safety, and partner with industry on chips, hardware and physical agent systems. The announcement frames openness as a governance choice, arguing that control of the world's intelligence and physical infrastructure by one or a few companies would be a profound risk. In an interview reported by Semafor, Jain said Luma was not building robot hardware and that outside robotics teams could build their own systems on top of its technology.
Compute is the other half of the claim. The Series C announcement pairs the financing with Project Halo, a planned 2-gigawatt AI supercluster in Saudi Arabia built with HUMAIN, with deployment starting in the first quarter of 2026 and completion expected in 2028 or 2029. The official About page says access to Project Halo supports training large-scale world models from video, audio and language data through the lab. A 2-gigawatt build-out is a multi-year infrastructure programme with capital, energy, siting and export-control dependencies, so it should be read as a planned capability rather than installed capacity as of this report.
The most concrete physical-AI evidence is older and comes from outside Luma. The official RoboCasa project page states that its repository of more than 3,200 objects across 150-plus categories is sourced from Objaverse 1.0, LightWheel AI and assets generated by Luma AI. The original RoboCasa paper, published at RSS 2024, records 2,509 high-quality assets across 153 categories, of which 1,592 came from Luma.ai. RoboCasa365, the ICLR 2026 release, scales to 365 tasks across 2,500 kitchen environments with more than 600 hours of human demonstrations and 1,600 hours of synthetic demonstrations, and is used to benchmark policies including Diffusion Policy, pi-0 and GR00T. In other words, Luma's text-to-3D output was already a production input to simulation-ready robot training assets two years before the lab was announced. That is a real link, with a real limit: generated meshes supply geometry and appearance, while collision geometry, mass, friction, joint structure and semantic labels are added downstream by RoboCasa. The asset contribution is to visual and geometric diversity for domain randomisation, not to physical fidelity.
The gap between the two should be stated plainly. A video model that keeps motion, lighting, occlusion and object persistence coherent across frames is learning useful statistical regularities about the world. A simulator or neural surrogate that predicts stress, flow, heat transfer or contact must be checked against measurements and must conserve or respect the governing equations. Luma's own announcement acknowledges the data and generalization problem but does not publish a physical benchmark. As of 14 November 2026, this round of research located no released model, evaluation suite, technical report or named robotics deployment from the Open Physical AI Lab. Bloomberg, Semafor and trade coverage located so far describe the announcement, the open-science framing and the strategic intent, not a shipped physical-AI system.
Applications, Customers and Business Model
The paying business is creative production. The official app page lists cinematic VFX, product visuals, social advertising, packaging mockups, brand identity, model photography, animation, infographics and storyboards as target outputs, and names enterprise buyers as the primary organisational customer. Named engagements in public sources include Publicis Groupe as a preferred generative-AI creative partner in the Middle East and North Africa, Serviceplan Group for global creative deployment, Dentsu, Adidas, and a Mazda MX-5 heritage commercial produced by the Johannesburg agency Boundless with Luma Agents in under two weeks. Wonder Project and Luma launched Innovative Dreams in April 2026, a production services company backed and supported by Amazon Web Services that combines performance capture, virtual production and generative AI under the label Realtime Hybrid Filmmaking; its first project was produced with Amazon MGM Studios for a Prime Video companion series. HUMAIN Create is a separate sovereign-AI workstream aimed at Arabic and regional data. In its Series C announcement Luma says adoption by Fortune 500 companies in media, entertainment and advertising is accelerating, without naming them or giving contract values.
Pricing is published and unusually transparent for a frontier model lab. The Luma App lists Plus at $30 per month or $300 yearly with 10,000 credits and commercial use, Pro at $90 per month or $900 yearly with 40,000 credits and four times the Agents usage, and Ultra at $300 per month or $3,000 yearly with 150,000 credits and fifteen times the Agents usage. Team and Enterprise plans are quoted on contact and add member management, shared projects, usage analytics, shared credits, single sign-on, training and custom fine-tuning. A separate API is pay as you go: third-party documentation records Uni-1 image generation at about $0.0404 per output and Ray 3.2 draft video from about $0.06 for five seconds, with rates varying by resolution, duration and mode. Credits, app plans and API billing are separate wallets, an important detail for procurement. The company's careers page has cited more than 30 million users worldwide; that is a community or registration figure, not a paid-subscriber or active-user count.
Revenue is not disclosed. Sacra, a private-markets research firm, estimated about $8 million in annualised revenue as of December 2024, shortly after Dream Machine subscriptions launched, and a 2026 aggregator still displayed the same figure. That estimate is two years old, single-source and unverified; it is recorded as MEDIA and UNVERIFIED and is not used here as a current revenue figure. The honest summary is that Luma has shipped products, published prices, signed named enterprise and media partnerships and built a large user base, while giving no audited financials, no revenue, no paid-seat count and no retention or unit-economics data. For a company valued above $4 billion by media reports, that is the single largest commercial information gap.
Competitive Landscape
In generative video, Luma competes with Runway, Google Veo, OpenAI Sora, Kling, MiniMax and ByteDance Seedance, and it also distributes several of those models inside its own workspace. Its stated differentiators are native HDR output, frame-level directorial control through keyframes, and brand consistency carried across a campaign by Unified Intelligence. Third-party benchmarks rank different leaders by different criteria, so a single ordering would not be well founded. One funding tracker states that OpenAI shut down Sora in April 2026 after heavy losses; that is a single secondary source and is not used as a factual anchor here.
In world models and physical AI the comparison set on this site is different. World Labs works on spatial intelligence and explorable large scenes and was the subject of an announced all-stock acquisition by AMD at about $8.2 billion in September 2026, still subject to approvals at the time of that reporting. NVIDIA supplies Cosmos world foundation models and the Isaac and Omniverse simulation stack. Physical Intelligence pursues generalist vision-language-action policies, AMI Labs pursues JEPA-style world models, and Veeda AI, TensorWorld and Physicl represent robot foundation world models, physics-native pretraining and physically accurate synthetic 3D data respectively. Luma's position is the representation and generation layer trained on internet-scale multimodal data, with an explicit statement that it will not build robot hardware. A Chinese self-media essay reports that Jain argued at a Stanford talk that spatial intelligence as a separate category does not exist, in contrast to Fei-Fei Li's framing; no primary recording or transcript was located, so this is recorded as an unverified viewpoint rather than a technical conclusion.
In 3D generation, the existing World Labs report on this site groups Luma AI with Tripo, Meshy and VAST as companies generating individual objects or small scenes and commercialising them as creator tools and game assets. That classification was accurate for Luma's 2024 era and remains the basis of the RoboCasa asset link, but it no longer describes the company's own product emphasis. Luma's relative strengths are a shipped multi-generation model family, a published pricing and API surface, named enterprise deployments, a research programme with released technical titles, and financing large enough to fund substantial compute. Its relative constraints in a physical-AI context are the absence of a released physical model or benchmark, a revenue base concentrated in creative tooling, and a dependence on a single sovereign partner for the planned compute expansion. These are differences in layer and evidence type, and this report does not rank the companies by technical merit.
Financing and Capital Structure
The official timeline gives three priced rounds. A $20 million Series A in March 2023 was led by Amplify Partners with General Catalyst, NVIDIA Ventures, Matrix Partners and South Park Commons participating. A $43 million Series B in January 2024 included Andreessen Horowitz among returning and new backers. The $900 million Series C in November 2025 was led by HUMAIN with significant participation from AMD, and with deepened participation from Andreessen Horowitz, Omniva, Amplify Partners and Matrix Partners. The official FAQ says Luma has raised more than $1.1 billion across five funding rounds since 2021; the three priced rounds above total $963 million, so the remaining amount sits in earlier or unpriced rounds that the company does not itemise. CNBC reported the Series C on 19 November 2025 at the US-Saudi Investment Forum and confirmed a valuation upwards of $4 billion; its May 2026 Disruptor 50 profile cited PitchBook at $1 billion raised and a $4 billion valuation. Chinese coverage on 20 November 2025 gave roughly RMB 6.397 billion for the round and about $4 billion in valuation, attributed to people familiar with the matter. No company announcement located in this research states a valuation.
Project Halo is the strategic core of the round. HUMAIN is owned by Saudi Arabia's Public Investment Fund and led by Tareq Amin; the partnership covers a 2-gigawatt supercluster in Saudi Arabia, with the Series C announcement stating deployment from the first quarter of 2026 and completion in 2028 or 2029, and a parallel HUMAIN Create initiative for Arabic and regional models. One industry dataset estimates that the $900 million round accounts for roughly a quarter of all tracked funding across 131 AI video deals, which illustrates concentration but is a database calculation rather than an official statistic. The London and Riyadh offices, the Publicis Groupe partnership in MENA and the regional model work show that the sovereign capital came with a regional commercial agenda, not only a compute commitment.
The capital structure has two readable consequences. It gives Luma a research runway and a claim on frontier-scale compute that few independent model labs can match. It also concentrates financing, compute and a substantial early commercial programme in one counterparty and one geography, with a build-out that will not finish until late in the decade. For diligence, signed transaction documents, a company statement reconciling the round count and total, and disclosed Project Halo milestones would carry more weight than database totals or press valuations.
DISCREPANCIES
1. Total raised, round count and the 2024 round label. Account A: the official FAQ states more than $1.1 billion across five funding rounds, and the official timeline lists a $20M Series A in March 2023, a $43M Series B in January 2024 and the $900M Series C in November 2025. Account B: a funding tracker gives $1.07B across six rounds and calls a $90M December 2024 financing a prior Series C including Amazon; CNBC and PitchBook give about $1 billion; a market database gives $1.3B or more. Ruling: use the official three priced rounds and the official total, record the $90M December 2024 round as an unverified database item, and mark the exact round count and pre-Series-A amounts as unconfirmed. Source grade: OFFICIAL versus DATABASE and MEDIA.
2. Valuation and its provenance. Account A: CNBC said it had confirmed a valuation upwards of $4 billion in November 2025, and its 2026 profile cited PitchBook at $4 billion. Account B: Chinese coverage attributed about $4 billion to people familiar with the matter; a funding tracker says $4B-plus. Account C: neither the Series C announcement nor the official FAQ states a valuation. Ruling: report $4 billion as a media-confirmed and database-corroborated figure, note that the company has not disclosed it, and do not treat it as an official valuation. Source grade: MEDIA and DATABASE versus OFFICIAL silence.
3. Headquarters. Account A: the official FAQ says Redwood City, California. Account B: the company's own careers page says Palo Alto, and CNBC uses Palo Alto. Account C: one funding database says San Francisco. Ruling: use Redwood City as the official corporate description, record Palo Alto as the team and press location, and mark San Francisco as an aggregator convention; these are not merged into one address. Source grade: OFFICIAL versus OFFICIAL and DATABASE.
4. Number of co-founders and the CTO title. Account A: the official About page and the CNBC Disruptor 50 list two founders, Amit Jain as CEO and Alex Yu as CTO, with Jiaming Song as chief scientist. Account B: Reactor's May 2026 primary press release describes Alberto Taiuti as a co-founder and former CTO of Luma, and an AI-generated wiki says two of three co-founders had left by October 2026 and that Song departed in May 2026. Account C: Yu's personal site reportedly describes him as former co-founder and CTO as of October 2026. Ruling: keep the official two-founder account as the company position, record Taiuti's role as supported by a primary external press release, and mark the allocation of the CTO title over time, the third co-founder claim and the departures as unconfirmed by any Luma announcement. Source grade: OFFICIAL and MEDIA versus OFFICIAL press release and UNVERIFIED aggregator.
5. Headcount and user count. Account A: the Series C announcement says more than 130 staff as of November 2025, and the careers page cites more than 30 million users worldwide. Account B: a commercial database lists 101 to 500 employees in 2026; a funding tracker says more than 30 million Dream Machine users, while an earlier company figure was 25 million registered users in November 2024. Ruling: use 130-plus as a dated official headcount, treat the database band as an inferred range marked MEDIA and UNVERIFIED, and describe 30 million-plus as a community or registration figure rather than paid or active users. Source grade: OFFICIAL versus DATABASE and MEDIA.
6. Revenue and commercial traction. Account A: Luma states that adoption by Fortune 500 companies in media, entertainment and advertising is accelerating and publishes subscription and API prices, but discloses no revenue, paid seats or contract values. Account B: Sacra estimated about $8 million in annualised revenue as of December 2024, a figure still displayed by an aggregator in October 2026. Ruling: record named partnerships and published pricing as verified commercial evidence, mark the $8 million estimate as a two-year-old single-source unverified figure, and state that current revenue is undisclosed. Source grade: OFFICIAL versus MEDIA and UNVERIFIED.
7. Current product line and model naming. Account A: the official FAQ says the current production models are Ray for video and UNI for image and multimodal work, that Photon has been retired, and that Dream Machine remains available but is no longer the forward model or work canvas; the About timeline dates the Ray and Uni families and the Agent canvas to March 2026. Account B: much media and database coverage still describes Dream Machine and Ray3 of September 2025 as the flagship, and one third-party review lists a separate Dream Machine price ladder in 2026. Account C: an aggregator says a company information page updated in July 2026 asks that the Dream Machine name no longer be used. Ruling: follow the official FAQ for the current line, record Ray3 and Dream Machine as dated milestones, and mark parallel pricing pages and the naming instruction as secondary and partially conflicting. Source grade: OFFICIAL versus MEDIA and DATABASE.
8. Genie text-to-3D status and the physical-AI classification. Account A: the RoboCasa project page says part of its object assets were AI-generated by Luma AI, the RSS 2024 paper attributes 1,592 of 2,509 assets to Luma.ai, and Alex Yu's research summary lists Genie as a Luma contribution. Account B: a third-party review says the Genie text-to-3D feature was sunset on 1 January 2026, and the official FAQ no longer lists 3D generation among current production models. Account C: the World Labs report on this site classifies Luma as a 3D generation company, while the 2026 lab announcement frames it as a physical-AI and world-model lab. Ruling: record the RoboCasa asset link as a verified historical contribution to simulation data, mark the Genie sunset as single-source and unconfirmed, and classify Luma in this list as an adjacent world-model and synthetic-data company whose physical-AI programme has not yet published a model or benchmark. Source grade: OFFICIAL and PAPER versus MEDIA and site-internal classification.
Risks and Due Diligence Gaps
The first gap is the distance between the physical-AI claim and product evidence. The lab announcement is a research programme with a clear diagnosis of the generalization problem, a stated open-science governance position and a planned compute base. This round of research found no released world model, evaluation suite, technical report, benchmark result or named robotics deployment attributable to the lab. Until such artefacts exist with independent baselines, the physical-AI position should be read as a strategic direction supported by adjacent capabilities, not as a validated product line.
The second is the technical boundary between visual plausibility and physical correctness. Coherent motion, lighting and object persistence across generated frames are useful learned regularities, but they do not establish force, torque, contact, friction, deformation, conservation behaviour or long-horizon rollout stability. A video model can depict a hand moving across a table without modelling grip or safety. Luma's own framing of the generalization crisis is about data coverage and transfer; engineering acceptance additionally requires numerical accuracy, uncertainty estimates and comparison with physical measurement. Diligence should ask for exactly those artefacts, and should treat the RoboCasa asset link as evidence of geometric and visual diversity for domain randomisation rather than evidence of simulation fidelity.
The third gap is financial transparency. There are no audited financials, no revenue, no paid-seat count, no retention and no disclosed unit economics, while compute obligations are large and the only public revenue estimate is two years old and unverified. Subscription credits, Agents usage multipliers and API billing are separate wallets, which makes average revenue per customer hard to reconstruct from published prices. For a company valued above $4 billion by media reports, the ratio between disclosed commercial evidence and disclosed financial evidence is unusually wide.
The fourth is concentration. HUMAIN is simultaneously lead investor, compute partner, regional commercial channel and the sponsor of a sovereign model programme; Project Halo will not be complete until 2028 or 2029 and carries energy, siting, supply-chain and export-control dependencies. A 2-gigawatt plan is a roadmap, not installed capacity. Regional data and Arabic model work also introduce jurisdiction-specific content, privacy and procurement requirements that differ from Luma's North American creative market.
The fifth is team and narrative continuity. Public sources describe the departure of two co-founders and the chief scientist during 2026, a CTO title that appears in conflicting accounts, and a product line in which Dream Machine and Photon have been retired and 3D generation may have been sunset. None of the personnel changes was confirmed by a company announcement located in this research. Key-person retention, the division of research ownership between the creative and physical-AI programmes, and the identity of the team executing the lab's roadmap are all open questions.
The sixth is legal and reputational exposure around training data. CNBC reported that a viral 2024 Dream Machine demo drew comparisons to a Pixar film and that Jain said the incident came from a user-uploaded image rather than scraping, adding that Luma had built detection systems. The company's enterprise terms state that customer content is not used for training. For a business selling into Hollywood and large advertisers, provenance, indemnification and model-output safeguards are commercial requirements, and any future dispute would matter more than a benchmark point.
The seventh is classification risk for this list. Luma's shipped products, named customers and published prices are in creative media; its physical-AI contribution so far is an announced lab, a planned cluster and a historical 3D asset pipeline. If the lab produces nothing verifiable, Luma will remain an adjacent generative-media company with a world-model narrative. If it produces a released model with independent physical evaluations and named robotics deployments, its classification here would need to be revised upward. This report therefore records the boundary explicitly rather than treating the company as a CAE or neural-simulation vendor.
Overall Assessment
Technical strength scores 8.0. Luma has shipped several generations of video and image models, published research on pre-training and inference-time scaling, built an agent workspace with frame-level control and HDR output, and retains an unusual 3D capture and neural rendering lineage. The deduction reflects the absence of any released physical-AI model, benchmark or evaluation, and the fact that creative-generation metrics do not measure physical accuracy.
Commercial capability scores 7.0. There is a published price ladder, a pay-as-you-go API, named enterprise and agency deployments, a joint production venture with AWS support, a large user base and an enterprise security posture. The deduction reflects undisclosed revenue and paid seats, an unverified and dated revenue estimate, and a buyer base concentrated in media and advertising rather than engineering.
Financing scores 9.5. A $900 million Series C led by a sovereign investor, an official total above $1.1 billion and a committed 2-gigawatt compute programme place Luma among the best-funded private companies on this list. The deduction reflects an undisclosed official valuation, a round count that does not reconcile across sources and dependence on a single counterparty for the compute expansion.
Product maturity scores 8.0. Multiple released model generations, a documented API, tiered subscriptions, team and enterprise controls, compliance attestations and three international offices are all shipping product evidence. The deduction reflects a product line in transition, with Dream Machine and Photon retired, Genie's status unconfirmed, third-party models inside the workspace, and no physical-AI deliverable from the lab.
Market position scores 7.5. Luma is a recognised leader in generative video with a distinct HDR and directorial-control proposition and a credible bridge into world models through multimodal pretraining. The deduction reflects intense competition from better-capitalised platform labs, a physical-AI field where World Labs, NVIDIA Cosmos and the robot foundation model companies have earlier and more specific footprints, and a revenue mix that is still overwhelmingly creative.
The composite is 8.0. Luma AI is a production-grade generative media company making a deliberate, well-funded turn toward physical AI, with one verified historical contribution to robot simulation data and a research agenda aimed squarely at generalization. Its principal assets are a shipped model family, published commercial terms and an exceptional compute partnership; its principal liabilities are the absence of a released physical model or benchmark, undisclosed financials, and eight groups of conflicting public accounts. The signals worth tracking, in order, are a first released model and evaluation suite from the Open Physical AI Lab, a named robotics deployment with independent results, a company statement reconciling financing and valuation, the first verifiable revenue disclosure, and an official clarification of product and leadership changes.
References
- Luma AI official About page, including the company timeline and FAQ: founding date and founders, headquarters, priced rounds, official total raised, current models, the retirement of Dream Machine and Photon, and the Open Physical AI Lab.
- Luma AI official Series C announcement, "AGI is multimodal and reality is the dataset of AGI" (November 2025): the $900M round, HUMAIN and the investor list, Project Halo's 2GW scale and deployment window, headcount above 130, and the Ray3 reasoning video model claim.
- Luma AI official research announcement, "Luma Announces The Open Physical AI Lab" (1 June 2026): the generalization crisis, the open science programme, world models, evaluations, and the governance argument.
- Luma AI official home page and Luma App pricing page: mission statement, Ray3.2, Uni-1, Luma Agents, research titles, subscription tiers, credits, enterprise controls and API availability.
- CNBC, "Luma AI raises $900 million in funding round led by Saudi AI firm Humain" (19 November 2025) and the CNBC Disruptor 50 profile (19 May 2026): valuation confirmation, founders, headquarters, PitchBook figures, HUMAIN and Project Halo, and the copyright controversy.
- Zhidongxi via Tencent News, Chinese coverage of the Series C (20 November 2025): RMB conversion, valuation attributed to people familiar with the matter, founding date, Amit Jain's Apple background and North American market focus.
- RoboCasa official project page, the RoboCasa paper (RSS 2024) and the RoboCasa365 citation (ICLR 2026): object asset sources including Luma AI, the 1,592 of 2,509 asset figure, and the 365-task, 2,500-scene benchmark scale.
- Semafor reporting, Startup Fortune and trade coverage of the Open Physical AI Lab, together with Bloomberg-derived coverage: the decision not to build robot hardware, the invitation to outside robotics teams, and the absence of a released physical-AI system.
- Reactor's stealth launch press release via PRNewswire and Yahoo Finance (28 May 2026) and a Value Add VC company profile (October 2026): Alberto Taiuti's prior Luma co-founder and CTO role, and the related leadership timeline.
- Secondary databases and industry references used only as leads: PitchBook figures via CNBC, Sacra revenue estimate via an AI-generated wiki, AI Market Watch, AI Funding Tracker, Bot Memo, GTM Directory, aiproviderindex, Fast.io, AITOP and a Chinese self-media essay. Valuation, revenue, user counts, model dates, pricing details and the reported Stanford remarks are marked MEDIA, DATABASE or UNVERIFIED in the text.
This report is compiled from public sources and does not constitute investment advice. Data is current as of 14 November 2026. Luma AI is a private, unlisted company with no audited financials and no disclosed revenue; financing totals, valuation, headquarters, founder and CTO roles, headcount, user counts, product status and the classification of its physical-AI programme carry eight groups of conflicting public accounts (see DISCREPANCIES). Uncertainties are flagged throughout; please rely on the company's primary announcements, signed transaction documents, peer-reviewed papers and independent benchmarks.
Related tools
Related reports
- World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B
- NVIDIA Deep Dive: Building the Full-Stack Physical-AI Infrastructure with PhysicsNeMo, Warp, Cosmos World Foundation Models and Omniverse
- Physical Intelligence Deep Dive: A $2.1B Bet on the GPT Moment for Robotics, With pi0.7 Self-Reporting 60-80 Percent Success on Unseen Tasks and Neither Pricing nor Revenue Disclosed for the Pi API
- AMI Labs: A USD 1.03B Seed Bet on World Models, LeCun's Non-Generative Path and Its Blank Scorecard
- Physicl Deep Dive: A Physical-AI Synthetic-Training-Data Infrastructure Spun from E-commerce 3D Rendering, Closing the Sim-to-Real Gap with Physics-Accurate Edge-Case Data, on a $10M Inovia-led Seed
This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.