**Author:** Darsh Shah | **Date:** 2026-02-24 **Purpose:** Remove the blocker on generation volume estimates to finalize MG commitments and move forward with provider contract signing. --- ## 1. The Problem We Need to Solve We need to commit to **Minimum Guarantee (MG) volumes** with our 7 model providers (Meshy, Rodin, Tripo3D, Hitem for 3D; Gemini Nano/Banana, Flux Kontext, Qwen Image for 2D) to sign contracts. Those MG commitments depend on a reliable forecast of **how many generations Fab AI will produce per month**. The original forecast assumed a **fab.com standalone interface** -- a full creative tool with visual editing, inpainting, multi-variation comparison, etc. That launch plan has changed: **Fab AI will now launch exclusively within EDA.** This fundamentally changes the usage model, and we need to re-forecast before signing. --- ## 2. What Changed: fab.com vs. EDA-Only ### 2.1 UX Implications | Dimension | fab.com (Original Plan) | EDA-Only (Current Plan) | | ------------------------ | ---------------------------------------------------------------- | ------------------------------------------------------------------ | | **Interaction mode** | Visual canvas + text prompts | Text-only chat | | **Image editing** | Inpainting, outpainting, style transfer, masking, upscaling | None -- no visual editing tools | | **Iteration method** | Paint a region, adjust sliders, visual A/B | Text-only refinement ("make it more worn", "change the color") | | **Multi-variation** | Side-by-side visual grid (4+ options) | Thumbnail previews in chat (limited comparison) | | **Image-to-3D pipeline** | Upload/generate image -> visually edit -> feed to 3D | Text-driven: describe what you want -> get 3D result | | **Texture workflow** | Generate PBR set, preview on 3D surface, tweak maps individually | "Generate a brick wall texture" -> get result, accept or re-prompt | | **Discovery path** | User goes to fab.com intending to create | User is working in UE, need surfaces naturally during development | | **Session intent** | Dedicated creative session | Asset need embedded in a broader development task | ### 2.2 What This Does to Generation Volume Assumptions **Things that DECREASE generation count:** - No visual editing means fewer iterative passes per asset (no inpaint -> refine -> inpaint cycle) - No canvas-based exploration means less "creative wandering" (fewer speculative generations) - EDA can surface Fab marketplace assets first, satisfying the need WITHOUT generation - The conversational UX adds friction vs. a "Generate" button (user has to describe, wait for agent response) - Users are mid-task in UE -- they want to get back to work, not spend 20 minutes iterating on a texture **Things that INCREASE generation count:** - Fab AI is embedded where developers work -- zero switching cost to try it (don't have to open a browser, navigate to fab.com, create an account) - EDA can proactively suggest generation ("I noticed you don't have a door texture -- want me to generate one?") - The bar for "good enough" may be lower in context -- a quick generated asset placed in-scene is more valuable than a perfect asset the user never bothered to go find - If EDA user base is large, even low per-user generation rate produces high absolute volume - Batch generation via programmatic tool calling (EDA's scripting capability) could produce bulk jobs **Things that CHANGE the distribution:** - The user profile shifts. fab.com would attract users who came specifically to generate. EDA-only means the users are developers who happen to need an asset. Their behavior pattern is different: - Fewer power users running hundreds of generations - More casual users running 1-5 generations per month - The "long tail" of light users becomes much larger - Heavy creative iteration (the "power" segment) nearly disappears without visual editing tools ### 2.3 Revised User Segment Distribution **Original assumption (fab.com):** | Segment | % of Users | Gens/Month | Description | | -------- | ---------- | ---------- | ------------------------------- | | Low | 30% | 1-5 | Tried it, occasional use | | Low-Med | 25% | 6-15 | Regular light use | | Med | 20% | 16-40 | Active creative workflow | | Med-High | 15% | 41-100 | Heavy creative user | | High | 7% | 101-300 | Power user, daily use | | Power | 3% | 300+ | Pipeline integration, batch gen | **Revised assumption (EDA-only):** | Segment | % of Users | Gens/Month | Description | |---|---|---|---| | Low | 40% | 1-3 | Tried it once or twice, sporadic use | | Low-Med | 30% | 4-10 | Uses it when they hit an asset gap during development | | Med | 20% | 11-25 | Regular use, relies on it for quick assets | | Med-High | 7% | 26-60 | Active user, uses it for most placeholder/draft assets | | High | 2.5% | 61-150 | Heavy user, batch gen via EDA scripting | | Power | 0.5% | 150+ | Pipeline integration, automated batch workflows | **Key shift:** The center of gravity moves left. Average generations per user per month drops significantly because the creative exploration loop (edit -> refine -> edit) is removed. But the user base may be larger because there's zero friction to first use. ### 2.4 Revised Iteration Count per Asset | Metric | fab.com Estimate | EDA-Only Estimate | Why | |---|---|---|---| | **Avg iterations per 2D asset** | 4-8 (generate, inpaint, adjust, upscale) | 1.5-3 (generate, maybe re-prompt once or twice) | No visual editing tools means fewer refinement steps | | **Avg iterations per 3D asset** | 3-6 (generate, retexture, refine geometry) | 1.5-2.5 (generate, maybe regenerate or retexture via text) | Text-only refinement is coarser -- users accept or regenerate, don't fine-tune | | **Avg iterations for image-to-3D** | 5-10 (generate image, edit image, generate 3D, refine) | 2-4 (describe -> get 3D, maybe refine once) | The image editing middle step is largely eliminated | | **Effective gens per "asset need"** | 4-8 | 1.5-3 | This is the critical multiplier for total generation volume | --- ## 3. The Fab Marketplace + Fab AI Funnel ### 3.1 The User Journey in EDA In EDA, the user journey for asset needs is NOT "open Fab AI and start generating." It's: ``` User working in UE Editor | v Asset need surfaces ("I need a rock texture for this cliff") | v EDA agent interprets the need | +---> Search Fab Marketplace first | (does a suitable asset already exist?) | | | +---> Good match found --> User downloads/purchases | | (NO generation needed) | | | +---> Partial match found --> User may still generate | | ("close but not quite what I need") | | | +---> No match / poor results --> Falls through to generation | +---> Generate with Fab AI (text-to-3D or text-to-image-to-3D) | v Result delivered | +---> Accept --> Import to project +---> Refine --> Re-prompt (text-only) --> New result +---> Reject --> Try marketplace instead, or abandon ``` **This funnel means Fab marketplace search is the FIRST filter.** Generation only happens when the marketplace doesn't satisfy the need. The generation volume estimate depends heavily on **what percentage of asset needs fall through marketplace search to generation.** ### 3.2 Search Success Metrics We Need For the marketplace search step, we need analogs of standard search product metrics: | Metric | Definition | Why It Matters | |---|---|---| | **Query Success Rate (QSR)** | % of searches where the user engages with a result (clicks, previews, downloads) | If QSR is high (e.g., 70%), most asset needs are served by marketplace -- generation volume is low | | **Zero-Result Rate (ZRR)** | % of searches that return no results at all | These are guaranteed fall-throughs to generation (or abandonment) | | **Search-to-Purchase Conversion** | % of searches that result in a download or purchase | Direct measure of marketplace solving the need | | **Search-to-Generate Fall-Through Rate** | % of searches where user proceeds to Fab AI generation instead | THIS IS THE KEY NUMBER -- it directly drives generation volume | | **Abandonment Rate** | % of asset needs where user does neither (gives up, uses a placeholder, finds another solution) | Reduces both marketplace and generation numbers | **The critical equation:** ``` Monthly Fab AI generations = EDA monthly active users x Asset needs per user per month x (1 - Marketplace QSR) <-- needs that marketplace doesn't satisfy x (1 - Abandonment Rate) <-- needs that actually proceed to generation x Iterations per generation <-- re-prompts per asset ``` ### 3.3 Estimating the Funnel (Assumptions We Need to Validate) | Funnel Step | Optimistic | Base Case | Conservative | Need from Product Director | |---|---|---|---|---| | EDA MAU | [from Jay] | [from Jay] | [from Jay] | Jay's EDA usage model | | Asset needs per user per month | 8-12 | 5-8 | 2-4 | What % of EDA sessions involve asset work? | | % needs that trigger search | 80% | 60% | 40% | Will EDA default to searching Fab first? | | Marketplace QSR | 50% | 35% | 20% | How deep is current Fab marketplace inventory for game-ready assets? | | Search-to-generate fall-through | 60% | 45% | 30% | Of unsatisfied searches, what % will try generation vs. abandon? | | Iterations per generation | 3 | 2 | 1.5 | With text-only editing, how many re-prompts before accept/abandon? | | Effective generation multiplier | 3 | 2 | 1.5 | Combining iteration count with occasional regeneration | **Base case example with placeholder numbers:** ``` 50,000 EDA MAU x 6 asset needs/user/month = 300,000 asset needs x 60% trigger search = 180,000 searches x (1 - 35% QSR) = 65% unsatisfied = 117,000 unsatisfied needs x 45% proceed to generation = 52,650 generation requests x 2 iterations per request = 105,300 total generations/month Plus: direct generation (user skips search, asks to generate directly) ~20% of asset needs go straight to generation: 300,000 x 20% x 2 iterations = 120,000 generations/month TOTAL: ~225,000 generations/month (base case) ``` *These numbers are illustrative. Every assumption needs validation.* --- ## 4. MG Commitment Implications ### 4.1 What We Need to Commit To For each provider, the MG is typically structured as: ``` MG = Minimum monthly spend (or minimum monthly API calls) over a contract term (typically 12-24 months) ``` We need to forecast: - **Total generations per month** (derived from Section 3) - **Split by type:** 2D vs. 3D (determines which providers are called) - **Split by quality tier:** draft vs. standard vs. high (determines cost per generation) - **Growth curve:** month-over-month ramp tied to EDA rollout timeline - **Provider routing split:** what % goes to each of the 4 (3D) / 3 (2D) providers ### 4.2 The EDA-Only Impact on MG Sizing | Factor | fab.com MG Basis | EDA-Only MG Basis | Direction | |---|---|---|---| | User base size | Fab.com visitors who opt into AI gen | EDA MAU (all UE devs using EDA) | Potentially LARGER user base | | Gens per user | High (dedicated creative tool, visual editing) | Low (text-only, casual use) | SMALLER per-user volume | | Iteration multiplier | 4-8x (visual editing loops) | 1.5-3x (text re-prompts) | SMALLER multiplier | | Marketplace cannibalization | None (fab.com Fab AI is separate from search) | Significant (EDA searches marketplace first) | REDUCES generation volume | | Time to ramp | Immediate (fab.com launch = day 1 traffic) | Tied to EDA rollout schedule | SLOWER ramp | | Net effect on total volume | ? | Likely **30-60% lower** than fab.com estimates | Depends on EDA user base size | ### 4.3 Risk Ranges for MG Given the uncertainty, we should model three scenarios: | Scenario | Monthly Gens (Steady State) | MG Recommendation | |---|---|---| | **Conservative** | 75,000 - 125,000 | Commit to this level. Low risk of under-delivery. | | **Base Case** | 150,000 - 300,000 | Target this for planning. Some risk. | | **Optimistic** | 400,000 - 750,000 | Do NOT commit to this. Use as upside scenario only. | **Recommendation:** Set MG at the conservative scenario. Negotiate contract terms that allow us to scale up commitments (with better pricing) as actual usage data comes in. See Section 6 on contract structure. --- ## 5. Questions for [Fab Product Director] These are ordered by priority -- the first group directly unblocks MG commitment sizing. ### 5.1 CRITICAL -- Unblocks MG Commitment **User Base & Reach:** 1. What is the projected EDA MAU timeline? (Month 1, Month 6, Month 12 after Fab AI ships in EDA.) We need this from Jay or the EDA team if you don't have it. 2. What percentage of EDA users will have access to Fab AI? Is it all EDA users, or gated behind a tier/entitlement? 3. Is there a phased rollout plan (e.g., beta -> GA, or region-by-region)? **Generation Volume:** 4. What is your current estimate of asset needs per EDA user per month? (How often does a typical UE developer need a new asset while working in the editor?) 5. In the EDA-only UX, will the agent default to searching Fab marketplace before suggesting generation? Or will the user explicitly choose "search" vs. "generate"? 6. What is the current Fab marketplace search success rate for common game-dev asset queries? (e.g., if someone searches "medieval shield", how often do they find something usable?) This tells us how much generation volume gets absorbed by marketplace. 7. What is a reasonable search-to-generate fall-through rate? Of users whose search doesn't return a good match, what percentage do you expect will try AI generation vs. abandoning (using a placeholder, going to another tool, etc.)? 8. With text-only interaction (no visual editing), how many iteration cycles (re-prompts) do you expect per asset? Our estimate is 1.5-3, down from 4-8 with a visual editor. 9. What is the expected split between 2D and 3D generation requests? This determines how we allocate MG across the 2D providers (Gemini, Flux, Qwen) vs. 3D providers (Meshy, Rodin, Tripo3D, Hitem). 10. Is batch generation (via EDA's programmatic tool calling) in scope for v1, or is it a later phase? Batch gen could significantly spike volume for Studio-tier users. **Pricing & Tiers:** 11. Is Fab AI generation included in an existing EDA subscription, or is it a separate paid add-on? This directly affects adoption rate and generation volume. 12. If tiered: what generation limits per tier are you considering? (This caps the upside of our volume forecast.) 13. Are there any "unlimited" tiers being considered? (These create MG risk if a small number of heavy users drive disproportionate volume.) ### 5.2 IMPORTANT -- Informs Forecast Accuracy **User Behavior:** 14. Do you have any data from EDA beta/preview users on how often asset-related conversations come up? Even rough anecdotal data helps. 15. What is the expected session length for an EDA conversation that involves asset generation? (Short = "generate me a barrel" -> done. Long = extended back-and-forth refining multiple assets.) 16. Will users be able to reference previously generated assets in new conversations? ("Make another barrel like the one I generated last week, but blue.") This affects whether generations are truly independent or build on each other. 17. What happens when generation fails or produces poor quality? Does the user get free retries, or does each attempt count against their limit? **Marketplace Interaction:** 18. When EDA searches Fab marketplace and finds results, how are they presented to the user? (Thumbnails in chat? Links to fab.com? In-editor preview?) The presentation quality affects whether users accept marketplace results or fall through to generation. 19. Will Fab AI-generated assets be listable on the Fab marketplace? If so, does that create a flywheel where generated assets improve future search results? 20. Is there a concept of "generate something LIKE this marketplace asset but with modifications"? (e.g., "I like this rock texture but make it snow-covered.") This hybrid flow would count as a generation but is guided by a marketplace asset. ### 5.3 IMPORTANT -- Informs Contract Structure **Timeline:** 21. What is the target launch date for Fab AI in EDA? We need to align MG ramp schedules with the actual launch. 22. What is the EDA rollout timeline relative to Fab AI launch? (Is EDA already widely deployed, or are both launching together?) 23. Are there any planned marketing moments or events (GDC, Unreal Fest, etc.) that would create usage spikes? **Strategic:** 24. Is the EDA-only decision final, or is fab.com still a possibility for a later phase? If fab.com is coming in 6-12 months, our MG commitments should account for that volume increase. 25. Are any of the 7 providers also being used by other Epic teams? If so, can we negotiate umbrella contracts with combined volume for better rates? 26. What is the competitive intelligence on Unity Muse's usage numbers or pricing? Helps calibrate our estimates against a comparable product. --- ## 6. Contract Signing Strategy ### 6.1 The Problem Contracts were being negotiated assuming a **fab.com launch**, which implied: - Higher per-user generation volume (visual editing loops) - A dedicated user base coming to fab.com to generate - Faster ramp to steady-state volume The **EDA-only pivot** means: - Lower per-user generation volume (text-only, fewer iterations) - User base is the EDA installed base (potentially larger, but generation rate is lower) - Ramp is tied to EDA adoption, not Fab AI marketing - Total steady-state volume is likely **30-60% lower** than original estimates Signing contracts with MGs based on the original fab.com estimates creates risk of **paying for volume we don't use.** ### 6.2 Recommended Contract Structure **Principle: Commit low, negotiate scale-up terms.** ``` +--------------------------------------------------------------+ | CONTRACT STRUCTURE RECOMMENDATION | +--------------------------------------------------------------+ | | | TERM: 12 months (not 24) | | Shorter term = less risk given launch pivot | | Renegotiate after 6-9 months with real data | | | | MG LEVEL: Set at Conservative scenario | | (75K-125K gens/month at steady state) | | with a 3-month ramp period at reduced MG | | | | RAMP: Month 1-3: 25% of steady-state MG | | Month 4-6: 50% of steady-state MG | | Month 7-12: 100% of steady-state MG | | | | OVERAGE: Negotiate a committed rate for volume above | | MG (e.g., same per-unit price up to 2x MG, | | discounted rate above 2x MG) | | | | SCALE-UP If actual usage exceeds 150% of MG for 2 | | CLAUSE: consecutive months, trigger renegotiation | | for higher MG at better per-unit pricing | | | | PROVIDER Don't commit 100% of volume to any single | | SPLIT: provider. Structure as: | | - Primary provider: 40-50% of MG | | - Secondary: 25-30% | | - Tertiary: 15-20% | | - Flex: 5-10% uncommitted (use for A/B | | testing, new providers, failover) | | | +--------------------------------------------------------------+ ``` ### 6.3 Negotiation Talking Points for Provider Conversations **Framing the pivot:** The pivot from fab.com to EDA-only should be positioned as a **distribution advantage**, not a scaling-down: - "We're embedding Fab AI directly into Unreal Editor via EDA, which gives us access to [X] million active UE developers. This is a larger addressable market than a standalone web tool." - "The per-user generation rate is lower, but the user base is significantly larger and more engaged -- these are professional developers in their daily workflow." - "We're starting with conservative MGs to prove the model, with a commitment to scale up as adoption grows." **What to ask for:** | Ask | Rationale | |---|---| | Lower initial MG with ramp schedule | We don't have usage data yet -- reduce upfront risk | | 12-month term (not 24) | Renegotiate with real data after launch | | Unused MG rollover | If Month 3 is below MG, carry credits to Month 4 | | No take-or-pay penalty in ramp period (Month 1-3) | Launch timing is uncertain; don't penalize for late start | | Volume-based pricing tiers (not flat rate) | As we scale, per-unit cost should decrease | | Quality SLA with credits | If provider has downtime or quality degradation, MG obligation is reduced | | Benchmark/evaluation period | 30-60 day evaluation period before MG kicks in (use for A/B quality testing) | ### 6.4 Provider-Specific Considerations **3D Providers (Meshy, Rodin, Tripo3D, Hitem):** - We should A/B test all 4 during the ramp period before committing to a primary/secondary split. - Quality benchmarking on game-ready metrics (topology, UV quality, PBR material accuracy, poly count adherence) should determine routing, not just subjective visual quality. - Ask: Can we run a 60-day benchmark period where we send identical prompts to all 4 and evaluate results before finalizing the MG split? **2D Providers (Gemini Nano/Banana, Flux Kontext, Qwen Image):** - In an EDA-only world, the 2D generation role may be smaller than originally planned. 2D was heavily tied to the editing/inpainting workflow on fab.com. - Remaining 2D use cases: concept reference for image-to-3D, texture generation, concept thumbnails. - Consider: Do we need all 3 2D providers at launch, or can we start with 1-2 and add the third later? - MG for 2D providers should be proportionally lower than 3D providers. ### 6.5 Scenario-Based MG Table | Provider Type | Conservative MG (Monthly) | Base Case MG | Notes | |---|---|---|---| | **3D Primary** (e.g., Meshy) | 25,000-35,000 gens | 50,000-70,000 | Largest share; best all-rounder | | **3D Secondary** (e.g., Rodin) | 15,000-20,000 gens | 30,000-45,000 | High-quality tier routing | | **3D Tertiary** (e.g., Tripo3D) | 10,000-15,000 gens | 20,000-30,000 | Fast/draft tier routing | | **3D Flex** (e.g., Hitem) | 5,000-10,000 gens | 10,000-20,000 | Evaluation, failover, A/B | | **2D Primary** (e.g., Flux) | 10,000-15,000 gens | 25,000-40,000 | Texture gen, concept-to-3D input | | **2D Secondary** (e.g., Gemini) | 5,000-10,000 gens | 15,000-25,000 | General image gen | | **2D Tertiary** (e.g., Qwen) | 3,000-5,000 gens | 10,000-15,000 | Diversity, cost optimization | | **TOTAL** | **73,000-110,000** | **160,000-245,000** | | *Primary/secondary/tertiary assignments are illustrative. Actual assignment should follow the benchmark period.* --- ## 7. Open Risks | Risk | Impact | Action Needed | |---|---|---| | **EDA MAU is lower than projected** | Generation volume misses MG | Get firm EDA projections from Jay. Set MG below conservative if uncertain. | | **Marketplace search is too good** | Very few users fall through to generation | Good for Fab marketplace revenue, bad for Fab AI generation volume. Model both outcomes. | | **Marketplace search is too poor** | Users skip Fab entirely, generation overload | Need marketplace inventory health data from Product Director. | | **EDA-only pivot reverts** | fab.com launch added later, volumes spike past committed capacity | Negotiate overage terms and scale-up clauses now. | | **Text-only UX limits quality perception** | Users try once, get mediocre result, never return | Invest in prompt enhancement (agent adds detail/context to user's short prompt). Track first-generation quality scores. | | **Provider quality diverges over time** | One provider improves faster than others, breaking routing assumptions | Build routing flexibility into contracts. No exclusive volume commitments. | | **Contract timing pressure** | Signing now with uncertain volume = overpaying or under-committing | Push for evaluation periods. Accept slightly worse per-unit pricing in exchange for lower MG and shorter terms. | --- ## 8. Immediate Next Steps | # | Action | Owner | Unblocks | | --- | -------------------------------------------------------------------------------- | ------------------------ | -------------------------------------- | | 1 | Get EDA MAU projections from Jay (users, sessions, tokens) | Darsh | Generation volume forecast | | 2 | Schedule meeting with [Fab Product Director] to walk through Section 5 questions | Darsh | Funnel assumptions | | 3 | Request Fab marketplace search success rate data (QSR, ZRR) from Fab search team | Darsh / Product Director | Fall-through rate estimate | | 4 | Run sensitivity analysis: MG at conservative/base/optimistic with 3D/2D split | Darsh | Contract negotiation range | | 5 | Draft revised contract term sheets per Section 6.2 structure | Darsh + Legal/BizDev | Contract signing | | 6 | Propose 60-day benchmark period to providers for A/B quality evaluation | BizDev | Provider routing split | | 7 | Confirm whether EDA-only is final or fab.com is a later phase | Product Director | Contract term length (12 vs 24 months) |