All case studies
    MODEL LAYER TRAPMarch 2026· 9 min

    Stability AI vs Midjourney: Why Open-Source L2 Couldn't Monetize

    Stability AI logoStability AI
    Midjourney logoMidjourney
    L2L7L8
    Verdict: L2a without L1b/L4a/L8c

    Stability AI

    Peak

    $1B (2022)

    Now

    Restructured (2024)

    ≈-90%

    Layer Scoring

    L-1
    Resources
    L0
    Infra
    L1
    Data
    L2
    Models
    L3
    Gates
    L4
    Access
    L5
    Execution
    L6
    Orchestration
    L7
    Surface
    L8
    Memory
    L2 Models
    Stability gave away its only asset. Midjourney kept it closed. The L2 layer is fragile when open *and* unprotected by adjacent layers.
    L4 Access
    Stability had no real distribution surface. Midjourney chose Discord, eccentric, sticky, and a real L4 in its own right.
    L7 Surface
    Stability shipped Clipdrop and Dreamstudio late. Midjourney made the surface itself a moat, every generation public, every user a marketer.
    L8 Memory
    The decisive layer. Midjourney's style refs, character refs, and mood-board memory of *your* taste are unreplicable. Stability never built it.

    Sublayer Impact Map

    Which of the 50 sublayers this case actually touches, and at what magnitude.

    L2 Models
    Models
    Stable Diffusion (open)
    plays here: Stability → everyone
    Owns
    Midjourney v6+ (closed)
    plays here: Midjourney
    Owns
    L4 Access
    Access
    Discord community
    plays here: Midjourney
    Owns
    L7 Surface
    Surface
    Dreamstudio/Clipdrop
    plays here: Stability (sub-scale)
    Touch
    Discord generation feed
    plays here: Midjourney
    Owns
    L8 Memory
    Memory
    Aesthetic memory / style refs
    plays here: Midjourney
    Owns
    Impact: Touch = enters · Share = meaningful · Owns = dominates· bars = magnitude

    Intelligence Cube · 2D

    Footprint across Functions × Verticals × Layers, the three axes that determine structural fate.

    Layers × Verticals

    6 cells · 3×2

    L-1
    L0
    L1
    L2
    L3
    L4
    L5
    L6
    L7
    L8
    FinTech
    EdTech
    Legal
    Health
    Travel
    eCom
    Media
    Gov
    SaaS
    Horizontal

    Layers × Functions

    6 cells · 3×2

    L-1
    L0
    L1
    L2
    L3
    L4
    L5
    L6
    L7
    L8
    Dev/Eng
    Design
    Product
    PM/Proj
    Ops
    Mktg
    Sales
    CustCare
    Strategy
    Finance

    Two 2D projections of the Intelligence Cube (Functions × Verticals × Layers). Filled cells = this move occupies that intersection.

    Timeline

    Aug 2022

    Stability releases Stable Diffusion as open source. Genuinely revolutionary, immediately ubiquitous.

    Oct 2022

    Stability raises at ~$1B. Midjourney quietly profitable on subscription, no funding round.

    2023

    Stable Diffusion becomes the de facto open model. Stability sees almost none of the resulting commercial value.

    Late 2023

    Midjourney v6 ships with style references and character consistency, the L8 wedge widens.

    2024

    Stability AI restructures. Founder departs. Sean Parker-led group invests on heavily revised terms. Valuation reported at a fraction of peak.

    2025–26

    Midjourney crosses $200M+ ARR, remains profitable. Stable Diffusion lineage continues open-source, but the company that birthed it is no longer the commercial vehicle.

    - Who Wins

    • Midjourney. Closed L2 + Discord L4 + compounding L8 aesthetic memory. The textbook closed-model + memory-moat play.
    • Every downstream product built on Stable Diffusion. Got a free industrial L2. Captured the commercial surface that Stability didn't.
    • Open-source as an ecosystem (vs. Stability the company). Stable Diffusion + ComfyUI + LoRA culture is one of the most generative open ecosystems in tech. The ecosystem won; the foundry didn't.

    - Who Loses

    • Stability AI (the company). Open L2, no L1, no L4, no L8, no L3. The cleanest example of layer-architecture failure in the AI cycle.
    • L2-only startups generally. Without an adjacent layer to capture value, your model is either a science project (open) or a vendor in a price war (closed).
    • The 'open source is automatically a moat' thesis. Open source is a *distribution* strategy, not a moat. You still need to own one of the other nine layers, or you'll watch the value flow past you.

    - Steelman: The Counter-Thesis

    The counter is that Stability's contribution is best understood as a *strategic* gift to the open ecosystem, not a failed business, and that the post-restructuring company can re-emerge as a focused enterprise vendor (fine-tunes, custom models, licensed weights) for buyers who specifically want non-OpenAI/non-Anthropic optionality. There is a real niche there, plausibly a $50–150M ARR business over time. But that's a 5–10× smaller outcome than the $1B valuation implied, which is the structural read the market has priced in.

    Stability AI is the most important structural cautionary tale about open-sourcing L2 without owning anything above or below it.

    What Stability built. Stable Diffusion, a genuinely revolutionary text-to-image model, open-sourced under a permissive license in mid-2022. Within months it was running on consumer GPUs, in ComfyUI, in Automatic1111, in every AI-image startup's backend, and inside every other company's product. Stability's L2 became infrastructure for an entire industry.

    What Stability captured. Almost none of the value the model created. Compute costs grew with usage they didn't monetize. Enterprise revenue stayed thin. There was no L1 (no proprietary training-data advantage), no L4 (no distribution surface of their own), no L8 (no per-user memory), and the L2 itself was, by design, available to every competitor for free. By 2024 the company had restructured, the founder had departed, and the valuation collapsed roughly 90% from the $1B peak.

    What Midjourney did instead.
    • L2, kept the model fully closed. No weights, no API for years, no fine-tuning leakage.
    • L7, chose Discord as the surface. Eccentric, sticky, community-native. Every generation is visible by default, a public aesthetic feed that made every user a marketer.
    • L8, the decisive layer. Style references, character references, mood boards, personalization tokens, Midjourney built a memory of your aesthetic that no other model can replicate without your usage history.
    • Pricing, subscription, not API. Captured value at the surface where the user actually was.

    Result: $200M+ annual revenue, profitable from year one, $10B+ in implied valuation, and a model that competitors cannot fully replicate even when their underlying L2 is technically comparable. Because the moat is no longer the model, it's the L8 wrapper around it.

    Law I, intelligence commoditizes downward. Stability accelerated this on themselves by open-sourcing their own L2. Midjourney accepted that L2 would commoditize eventually and built L7+L8 as the durable layers from day one.

    Law III, value migrates to the scarcest layer. Once Stable Diffusion was open, the scarce layer in AI image generation was not the model, it was the taste-and-memory layer that knew what looked good and remembered your prior work. Midjourney owns that. Stability gave it away.

    The generalizable lesson. Open-sourcing L2 is a defensible strategy only if you own one of the other nine layers. Meta open-sources Llama because they own L4 (Facebook, Instagram, WhatsApp distribution at 3B+ users). DeepSeek open-sources because they own L0 (a national-strategy compute relationship). Stability open-sourced and owned nothing else, which is why the open model became the entire industry's gain and Stability's structural loss.

    Public reporting; valuations approximate.

    What This Means for You

    Product Leader

    If your roadmap depends on 'we'll open-source our model and capture downstream value,' name the L1, L4, or L8 layer you own that the open model funnels users into. If none exist, you are not Meta, you are Stability.

    Investor

    L2-only startups (closed or open) without an adjacent moat layer are structurally exposed. The closed-vs-open debate is a distraction; the layer-architecture question is the actual decision.

    Operator

    For image generation, Midjourney is the L8 play (long-term consistency for a brand's aesthetic) and Stable Diffusion is the L0 play (cheap, owned, on-prem when you need it). They solve different problems, don't conflate them.

    AA

    Anand Arivukkarasu

    Ex-Meta product leader. Creator of Supply Chain of Intelligence™. Writes about where AI value accrues, and who can fire your product. LinkedIn

    Get the next teardown in your inbox.

    One issue when something structurally important happens, usually weekly. No spam, no filler, unsubscribe anytime.

    Worth sharing? Pull-quote: "Open-source L2 is a defensible strategy *only if* you own one of the other nine layers. Stability owned none."