Generative media · USA
Luma AI
A company growing on a video model that gets cheaper with every release and on shareable output, investing its $900 million round in compute; the product evidence is strong, and there is not a single official revenue figure.
Verdict: Perception level
Updated: 2026-10-06
01
Profile
- What it does
- Generative video and image models: the Dream Machine product; the Ray2, Ray3, Ray3.14 and Ray3.2 video models; the Uni-1.1 image API; workflow features such as Modify Video, Scenes, Skills and Layers; an API [2].
- Founded
- 2021, San Francisco (secondary data). The company describes itself as the world leader in "multimodal artificial generative intelligence" for the creative industry [4].
- Founders
- Amit Jain (CEO) and Alex Yu (secondary data, not opened in a primary source)
- Pricing model
- Credits: Plus $30 a month for 10,000 credits, Pro $90 for 40,000, Ultra $300 for 150,000; Team and Enterprise through sales. Credits for failed generations are returned [6].
02
Growth timeline
2025-05-15
HUMAIN partnership: compute, go-to-market, investment
"Multimodal AI is the biggest disruptive force in how media is created, how companies advertise their message"
[5]2025-09-18
Ray3: in the company's words the first reasoning video model and the first 16-bit HDR generation
[2]2025-11-19
$900M Series C; Project Halo 2 GW supercluster; team of more than 130
"begin deploying starting Q1 2026 and finish by 2028-29"
[1]2026-01-26
Ray3.14: 3x cheaper and 4x faster than Ray3 (720p)
"3× cheaper", "4× faster"
[3]2026-02-11
Riyadh office; preferred partnership with Publicis Groupe Middle East
"Saudi Arabia is the natural home for our regional headquarters in MENA."
[7]
03
Growth loop
- 1New model: cheaper, faster
- 2More generation
- 3Output shared on social
- 4New users, agencies, capital
- 5Compute capacity: Project Halo
- Leak: revenue node: no disclosure
The company's thesis: owning compute is a cost advantage, and a cost advantage is distribution. The red arrow marks what the loop is missing: revenue has never been disclosed.
The drawing shows the shape of the loop; each node is explained in the list below. The red arrow marks value leaking out of the loop.
04
Mechanism
- Acquisition channels
- A viral product launch: the claim of "1 million users in 96 hours with zero marketing budget" at Dream Machine's 2024 launch is a news report only and could not be verified on an official page (secondary data) [8]. Documented channels: a new model or feature announcement nearly every month in 2025–2026 [2], the Adobe Firefly integration [2], the agency channel through Publicis and the Dream Brief competition [7] [2], the developer channel through the Ray3.2 and Uni-1.1 APIs [2].
- Loop
- A new, cheaper model (Ray3.14, 3x cheaper than Ray3) [3], more generation, visible output on social, new users and agency interest, capital, compute capacity (the 2 GW Project Halo) [1], a cheaper model again. The $900M announcement foregrounds the 2 GW supercluster [1]; the inference that most of it goes to infrastructure is ours (opinion).
- Activation
- The first free generation; returning credits on failed generations lowers friction [6]. No activation rate is disclosed.
- Retention evidence
- No claim. The credit subscription is consumption-driven; no cohort or churn data has been disclosed.
- Role of pricing
- Three credit tiers (10,000, 40,000, 150,000) and a credit multiplier for agent usage [6]. As the model gets cheaper, generations per credit rise; protecting average revenue depends on the upper tiers (opinion).
- Distribution moves
- A regional market through HUMAIN (Saudi Arabia, an Arabic interface) [5] [7], a preferred partnership with Publicis Groupe Middle East across 8 markets [7], a London office with a plan for 200 roles in 2026 [4], and the Dream Brief competition with $1M for a Cannes Lions gold winner [2].
05
Real reading
The five questions of the Growth Claim Check, applied to this company's public claims.
- 01
What was measured, and by whom?
no
No user or revenue figure in official sources; "30M users" and "1M in 96 hours" are news reports.
- 02
Perception or outcome?
unclear
Launch speed and model headlines are perception; the product-side evidence (releases, price drops) is concrete.
- 03
Is there a stability or quality metric?
no
No quality or stability metric.
- 04
Is there a counterfactual or cohort?
no
No cohort.
- 05
Nominal or real?
unclear
A large part of the $900M is infrastructure investment; its relation to revenue is not disclosed.
Verdict
Perception level
Strong evidence on the product and model side, none on the revenue side. The only thing that can be read as growth is release cadence and partnerships; no business outcome has been disclosed.
Test your own claim with the same five questions: Growth Claim Check →
06
Shadow side
- Revenue is not disclosed; a capital-intensive plan with a 2 GW supercluster commitment; unit economics unknown (opinion) [1].
- Strategic dependence on one large partner: compute, market and investment from the same source, HUMAIN [1] [5] [7].
- User figures do not appear on official pages; the figures in news reports should be verified against a primary source before publication [8].
07
What it teaches a founder in Türkiye
This section is opinion; a reading, not a source.
- 01
(opinion) In a product whose output can be shared, the product is the ad. The first days come from shareable output rather than a marketing budget; but revenue evidence is a separate job.
- 02
(opinion) When model cost falls with every release, rebuild the price tiers and the credit definition; otherwise revenue falls while usage rises.
- 03
(opinion) A compute-ownership strategy takes capital. For a founder paying dollar GPU costs and earning lira, what can be copied is the release cadence and the agency partnership.
- 04
(opinion) Market entry through a regional partnership is a concrete roadmap for a founder in Turkey looking to expand into the Middle East.
A question for you: remove this company's logo and drop the loop onto your own product. Which node stays empty?
08
Sources
Numbers are the references in the text. An 'opened' source was read by the researcher; 'secondary' ones are verified against their primary source before publication.
- [1]Luma AI raises $900M Series C led by HUMAIN
Luma (official announcement) · 2025-11-19 · opened
“2GW compute supercluster – Project Halo. This colossal infrastructure will begin deploying starting Q1 2026 and finish by 2028-29”
- [2]Luma news and press list
Luma · accessed 2026-10-05 · opened
“the world's first reasoning video model and the first to generate high fidelity 16 bit HDR”
- [3]
- [4]Luma AI opens for international business with former WPP executive Jason Day leading new London office
Luma (official announcement) · 2025-12-02 · opened
“we have the capital and capacity to bring world-scale AI to creatives everywhere”
- [5]Luma joins forces with HUMAIN
Luma (official announcement) · 2025-05-15 · opened
“Multimodal AI is the biggest disruptive force in how media is created, how companies advertise their message”
- [6]Dream Machine pricing page
Luma · accessed 2026-10-05 · opened
“Any credits used for failed generations will be automatically returned to your account.”
- [7]Luma AI to open Riyadh office to accelerate HUMAIN Create and partner with Publicis Groupe
Luma (official announcement) · 2026-02-11 · opened
“Saudi Arabia is the natural home for our regional headquarters in MENA.”
- [8]Luma AI Series C press release (30M users, $4B valuation, "1M in 96 hours")
BusinessWire · 2025-11-19 · secondary
Note: could not be opened (403)
09
Open questions
- 30M registered users, 1M users in 96 hours and the $4B valuation were not found in official text.
- No revenue or ARR claim.
- Founders and founding year not verified in a primary source.
- No Amit Jain podcast or interview transcript was found.