Decart

1.1 Identity, operating model, and product surface

Decart describes itself as a fully vertically integrated frontier AI research lab founded in 2023 that builds state-of-the-art real-time world models together with the ultra-optimized infrastructure to run them. The company's own May 2026 funding release frames the business as three product lines that share one stack: Lucy, a real-time world model for immersive video and interactive experiences; Oasis, a real-time world model targeted at Physical AI and robotics simulation; and DOS, the Decart Optimization Stack that compiles model weights down to NVIDIA GPUs, Google TPUs, and Amazon Trainium silicon. Decart sells API access on a pay-as-you-go basis through platform.decart.ai, with realtime models such as Lucy 2.1 and Oasis 3 Preview priced at roughly $0.02 per second of active generation and image edits at single-cent unit prices, and offers enterprise pricing on top through a sales contact. The company anchors its operations in Tel Aviv, with disclosed offices in northern Israel, San Francisco, and New York and a 2025-opened San Francisco R&D center led by Dr. Kfir Aberman. The pre-2026 product surface also included Mirage, the company's livestream diffusion model that Decart positioned as the predecessor to the Lucy family of real-time video models.[CO001, CO002, CO003, CO004, CO005, CO006]

1.2 Founders, leadership, governance, and key-person dependence

Decart was founded in late 2023 by Dr. Dean Leitersdorf (CEO) and Moshe Shalev (CPO), who met as Unit 8200 reservists in Israel's military intelligence corps. Leitersdorf earned three computer-science degrees and a doctorate from the Technion before turning twenty-four, while Shalev grew up in Israel's ultra-Orthodox community in Bnei Brak before serving in Unit 8200 and joining Decart as its operations-and-product lead. Public materials also identify Dr. Kfir Aberman, a co-creator of the DreamBooth diffusion-tuning technique and a former senior researcher at Snap and Google, as the head of the company's San Francisco R&D center from the August 2025 Series B onward. Decart's own May 2026 announcements continue to treat Leitersdorf as the public-facing CEO and primary spokesperson, and the company has not published an independent board composition or governance disclosures in any reviewed source. That pattern, paired with the small executive surface area named in coverage, supports a high key-person dependence on Leitersdorf and a still-thin governance picture even after the $300 million Radical-led round. Decart is a Delaware C-corp marketed as Decart.AI, Inc. through investor and trade releases, and the public record does not yet describe outside directors, an audit committee, or any independent board chair.[CO013, CO014, CO015, CO016, CO017, CO018]

1.3 Capital base, valuation path, and stakeholder map

Decart's funding history is unusually compressed. The company emerged from stealth on October 31, 2024 with a $21 million round led by Sequoia and joined by Oren Zeev's Zeev Ventures, alongside the viral Oasis Minecraft-style demo. Ynet's August 2025 interview reports that two early back-to-back rounds totaled $53 million at a roughly $500 million valuation before the company crossed unicorn status. On August 7, 2025 Decart announced a $100 million Series B at a $3.1 billion post-money valuation, bringing cumulative capital to $153 million, with Sequoia, Benchmark, and Zeev Ventures rolling over and Aleph VC entering as a new investor. On May 18, 2026 Decart disclosed a $300 million round led by Radical Ventures at a roughly $4 billion valuation, with NVIDIA, eBay Ventures, Adobe Ventures, Toyota Ventures, Atreides Management, and Valor Equity Partners joining alongside returning Sequoia, Benchmark, and Zeev Ventures; private backers in the round include OpenAI co-founder Andrej Karpathy, former Disney CEO Michael Eisner, members of the Nintendo founding family, and gaming investor Moritz Baier-Lentz. Decart itself states the company has raised over $450 million to date, and Amazon is publicly described as a strategic customer rather than a disclosed equity investor. Leitersdorf told Ynet in August 2025 that Decart had used less than $10 million of its $153 million in capital, a self-reported burn that has not been independently audited.[CO023, CO024, CO025, CO026, CO027, CO028]

1.4 Milestone chronology, scale signals, and evidence gaps

The public chronology runs from the late-2023 founding, through the October 31, 2024 stealth exit and Oasis viral demo, into a 2025 build-out (Mirage launch, $100 million Series B, San Francisco R&D center, growth from roughly fifteen to about sixty employees) and a 2026 strategic-infrastructure push (DOS 2.0 announce, AWS Trainium3 partnership for Lucy2, Amazon strategic-customer designation, $300 million Radical-led round at a roughly $4 billion valuation, and the June 10, 2026 Oasis 3 Preview launch on a public API at $0.02 per second). Decart publicly claims that Lucy operates at sub-30-millisecond response time and that the DOS stack delivers more than 100-fps full-HD inference and over 1,600 tokens-per-second agentic throughput, and the company says it serves a developer community exceeding 100,000 users. Adverse and uncertain signals coexist with this growth story: TechCrunch's October 2024 review of Oasis flagged unresolved copyright questions because the model was trained on Minecraft footage without disclosed permission from Microsoft; TechCrunch's June 2026 review of Oasis 3 reported that long generations degrade, that cars drive through other cars, and that physics consistency is an open research problem; and audited revenue, an exact current headcount, an itemized cap table, and an independent board composition are still not in the public record. The chapter therefore preserves Decart as a fast-scaling, well-capitalized infrastructure-leaning world-models lab whose key diligence variables remain founder concentration, undisclosed financial detail, and the gap between hyper-real demos and production-grade physics.[CO038, CO039, CO040, CO041, CO042, CO043]

1.5 Exhibits

2.1 Market boundary, included spend, and status-quo substitutes

Decart's addressable market sits at the intersection of three loosely linked spend pools rather than one mature category. The first is generative AI in gaming, defined by independent analyst firms as revenues from procedural content generation, level/world building, NPC behavior, narrative generation, adaptive personalization, and AI-driven QA inside game studios and creator-economy platforms. The second is real-time generative video and interactive video, where Decart's Lucy 2.1 transforms or generates a continuous stream and competes with edit-pass video models (Runway, Luma, Adobe Firefly Video) and with the in-game video pipelines of UGC platforms (Fortnite, Roblox, Minecraft). The third is physical-AI / world-model simulation, where Oasis 3 Preview targets autonomous-vehicle data generation alongside Waymo's in-house simulator, NVIDIA's Cosmos/DRIVE Sim stack, World Labs Marble, Google DeepMind Genie 3, and Tesla's and Wayve's proprietary AV simulators. Status-quo substitutes are decisive at the buyer level: hand-built game content (artists, designers, cinematics teams), traditional driving-sim engines (Unreal/CARLA), and conventional non-real-time video models (Sora, Veo, Runway Gen) that ship batch outputs rather than sub-30-millisecond interactive frames. Excluded spend includes regulated commercial gambling (the $51B/month AGA series is a different pool), GPU hardware capex (an input cost rather than software revenue), and broadcaster/film VFX spend that is not real-time. Decart's $0.02-per-second list price prices its API at the boundary between developer-tools spend and inference compute, which means the company's market boundary is best read as a horizontal real-time generative-AI infrastructure category rather than as a gaming-only line item.[CM001, CM002, CM003, CM004, CM005, CM006]

2.2 Sizing lenses, ranges, and methodology limits

No single public sizing captures Decart's full opportunity, so three lenses anchor the chapter. Lens 1 (generative AI in gaming, the closest pure-play): The Business Research Company sizes the market at $1.79 billion in 2025 growing to $2.21 billion in 2026 and $5.09 billion by 2030 at 23.2% CAGR, with Asia-Pacific the largest region in 2025; Research and Markets republishes a near-identical 2025-2030 forecast and extends the series to 2035 with deterministic and nondeterministic segments. Lens 2 (cloud gaming as a proxy for low-latency, server-rendered interactive entertainment): BCG's 2026 Video Gaming Report puts cloud gaming at roughly $1.4 billion in 2025 expanding to about $18.3 billion in 2030 at a compound annual growth rate above 50%, and notes that 60% of surveyed players have tried cloud gaming with 80% reporting a positive experience. Lens 3 (UGC creator economy, the platform layer Decart's tools enable): BCG reports Fortnite and Roblox UGC creator payouts exceeding $1.5 billion in 2025 and 40% of gamers consuming more UGC than a year earlier. None of these lenses size the physical-AI / world-model simulation segment that Oasis 3 explicitly targets, and McKinsey's State of AI report - the most-cited cross-industry benchmark - was access-denied during this run. The chapter therefore preserves three quoted ranges side-by-side, treats the generative-AI-in-gaming series as the most directly relevant SAM lens, and flags physical-AI simulation as an evidence-constrained TAM whose independent sizing is not yet published.[CM010, CM011, CM012, CM013, CM014, CM015]

2.3 Buyers, users, payers, and adoption path

Four buyer cohorts pay for or use Decart's stack today. (a) Game studios and engine teams, where BCG estimates roughly 50% of studios already use AI of some kind and notes that around 20% of new Steam titles disclosed AI use by mid-2025; budget owners are CTOs and head-of-content, with adoption triggered when generative tools demonstrably shorten content cycles for AAA titles whose budgets can reach $300 million per game. (b) UGC platforms and creators, where the $1.5 billion creator-payout pool attracts diffusion- and world-model vendors as picks-and-shovels suppliers; the payer is the platform (Roblox, Fortnite, Minecraft) and the user is the creator. (c) Generative-AI infrastructure customers, where cloud providers (Amazon/AWS is a named Decart strategic customer) license DOS to optimize their own model serving across NVIDIA, TPU, and Trainium silicon; the payer is the cloud platform, the user is its enterprise tenant, and the adoption trigger is GPU/Trainium efficiency. (d) Physical-AI and AV simulation buyers (Toyota Ventures and the Nintendo family appear on Decart's cap table as adjacent signals), where Oasis 3 Preview is positioned against in-house simulators at Waymo, Tesla, Wayve, and NVIDIA's DRIVE Sim; budget owner is the autonomy program head, and adoption is gated by physics fidelity and safety-case acceptability. Across all four cohorts, the published self-serve adoption path is Decart's API at platform.decart.ai - a $0.02-per-second realtime price tier plus cent-level image-edit calls - layered with enterprise contracts on top. The 100,000-developer community Decart reported to TechCrunch in June 2026 is the early funnel for paid usage.[CM021, CM022, CM023, CM024, CM025, CM026]

2.4 Growth drivers, adoption constraints, and contradictions to preserve

The tailwinds are concrete. Latency-sensitive workloads (real-time gameplay, AV scene generation, livestream video transformation) are an underserved gap left by batch-oriented video and image generators; cloud gaming's >50% CAGR and the doubling of AI-disclosing Steam titles between 2024 and 2025 both indicate a real-time generative-AI infrastructure wedge. The headwinds are equally concrete. The European Parliament's March 2026 own-initiative report (rapporteur Axel Voss, 17-3 vote) demands itemised transparency over all copyright-protected training data, proposes a rebuttable presumption of infringement absent transparency, and floats a 5-7% global-turnover flat-rate copyright fee on generative-AI providers; Osborne Clarke characterizes this as a "watershed moment" for the EU regulatory regime. The EU AI Act's General Purpose AI obligations have been in force since 2 August 2025, with the Digital Omnibus delaying the high-risk obligations originally scheduled for August 2026. In the US, the California AI Transparency Act and GenAI Training Data Transparency Act took effect on 1 January 2026, the Texas Responsible AI Governance Act took effect the same day, and the Trump administration's 11 December 2025 executive order directed DOJ to challenge state AI laws on preemption grounds. The US Copyright Office's ongoing report series remains the primary federal framing for AI and copyright. Contradictions to preserve: (i) cloud-gaming forecasts disagree with generative-AI-in-gaming forecasts on scale and unit (player-hours vs. studio software spend); (ii) physical-AI simulation has no published TAM yet Decart's investors (Toyota Ventures, NVIDIA, Radical) bet that it becomes Decart's largest segment; (iii) studio adoption of AI is rising but BCG's own survey shows player resistance to AI in art/animation, suggesting consumer-side adoption risk; (iv) the GDC 2026 State of the Game Industry Report - the canonical developer survey - is paywalled in this run and not directly cited in the figures here.[CM031, CM032, CM033, CM034, CM035, CM036]

2.5 Exhibits

3.1 World-model landscape in 2026: direct labs, incumbents, adjacents, and internal-build substitutes

Decart does not compete in a well-bounded software category. Its Lucy model for immersive experiences competes with video-generation platforms such as Runway and Luma AI that are themselves pivoting toward world models. Its Oasis model for physical-AI simulation competes with both research-grade systems (Google DeepMind's Genie 3, released as a research preview in August 2025) and large internal build programs at Waymo, Tesla, NVIDIA, and Wayve. Its DOS inference stack competes with hyperscaler GPU clouds and dedicated inference optimization vendors. The common thread across all three product lines is real-time, interactive generation at sub-30-millisecond latency — a capability that, as of mid-2026, no competitor has matched in a commercially available API. Genie 3 generates frames at 24 fps at 720p but remains a research preview without a public paid API. Runway's latest Gen 4.5 produces high-definition video with audio and character consistency, but targets post-production and creative workflows rather than live interactive simulation. World Labs' Marble product creates downloadable 3D environments for design and entertainment, not real-time driving simulation. Luma AI raised $900 million in November 2025 at a multi-billion valuation and is expanding into world models, but has no publicly disclosed real-time interactive API. The two classes of competitor that most directly undercut Decart's long-term thesis are: (1) internal builds by the top-tier AV players who have Google-grade research budgets and proprietary sensor data, and (2) future commercial releases by the same general-purpose video labs already raising capital at comparable valuations. Scenario and Inworld AI represent adjacent substitutes focused on gaming tooling and AI characters, respectively, but do not directly address real-time driving simulation.[CP001, CP004, CP005, CP006, CP007, CP019]

3.2 Competitor profiles: Runway, World Labs, DeepMind/Genie 3, Luma AI, Inworld, and Rosebud

Runway raised $315 million in a Series E in February 2026, nearly doubling its valuation to $5.3 billion. The company is best known for physics-aware video generation (Gen 4.5) and is explicitly pre-training the "next generation of world models" for products and industries beyond media. Runway already has partnerships with Adobe and is expanding into gaming and robotics, making it the most direct frontier peer to Decart's Lucy model in creative and media workflows. It has a CoreWeave compute partnership and a team of approximately 140 people in February 2026. World Labs raised $1 billion in February 2026, including $200 million from Autodesk, at a valuation reportedly targeting $5 billion. Founded by Fei-Fei Li, its Marble product generates editable, downloadable 3D environments from images, video, or text. The Autodesk partnership focuses on media and entertainment use cases initially and positions World Labs as the closest spatial-AI peer to Decart's Oasis world for 3D content creation workflows, though not for real-time driving simulation. Google DeepMind's Genie 3, released in research preview in August 2025, generates interactive worlds at 24 fps and 720p, supports promptable world events, and already powers Waymo's internal simulation architecture. It is not available as a public commercial API, which gives Decart a temporal window before Google potentially monetizes this capability externally. Luma AI raised $900 million in November 2025 with a 2-gigawatt AI supercluster partnership in Saudi Arabia. It is primarily known for Dream Machine and video generation, with world models as a stated future direction. Inworld AI pivoted to realtime voice AI for consumer companion apps, achieving 1 million users in 19 days on a $500 million valuation, and is no longer a direct simulation competitor. Rosebud AI focuses on text-to-3D-game creation and represents a development-tooling substitute rather than an inference-infrastructure competitor.[CP002, CP003, CP004, CP005, CP007, CP008]

3.3 Decart differentiation: real-time latency, DOS cost moat, and developer-ecosystem strategy

Decart's core differentiation rests on three pillars that competitors have not matched in production systems as of mid-2026. First, real-time inference latency: Lucy runs at under 30 milliseconds time-to-first-frame, generating up to 100 frames per second via the DOS stack optimized for Amazon Trainium3, versus Genie 3's 24-fps research-only pipeline. Second, cost efficiency: Decart claims a 100x improvement in cost efficiency over comparable systems, enabled by DOS, which achieves over 80% Model FLOPS Utilization on Trainium3 hardware. Third, the developer ecosystem: by pricing Oasis 3 at $0.02 per second via public API from day one, Decart is building the same developer ecosystem flywheel that made OpenAI's text API sticky — the startup already has more than 100,000 developers building on Lucy. The DOS stack is architecturally differentiated in that it runs on any hardware (NVIDIA GPU, Amazon Trainium, Google TPU) rather than being tied to one cloud provider. Oasis 3 generates three synchronized camera feeds (front-facing plus two side-facing), matching the perception setup of most camera-first AV stacks. Decart's CEO stated the company had burned "drastically less" than $100 million in its lifetime despite raising over $450 million, which if accurate suggests DOS cost efficiency applies to internal inference economics as well as external sales. However, this is a company-claimed metric without independent audit corroboration. The key strategic bet is that mid-tier AV companies, robotics startups, and drone programs that cannot afford Google-DeepMind-scale research budgets will prefer to rent Oasis simulation at $0.02/sec rather than build internal systems — an addressable segment the CEO estimates at dozens of programs globally.[CP009, CP010, CP011, CP012, CP021, CP022]

3.4 Moat durability, switching costs, and commoditization risk

Decart's moat is real but narrow and time-bounded. The switch to Oasis 3 is sticky while it is the only production-ready API for real-time driving simulation: once AV training pipelines integrate Oasis-generated scenarios, the switching cost includes re-collecting or re-labeling synthetic data from a different format. However, the moat has at least three structural vulnerabilities. First, the two firms most capable of building a competing real-time world model — Google (via DeepMind) and NVIDIA (which is an investor in Decart but runs its own DRIVE simulation stack) — have R&D budgets orders of magnitude larger than Decart's. Waymo has already implemented Genie 3 internally, setting a precedent for the top AV tier self-sourcing. Second, Runway, World Labs, and Luma AI are all explicitly investing in world models and have raised comparable or larger financing; if any of them ships a competitive real-time API, Decart's first-mover developer ecosystem advantage faces direct pressure. Third, the distribution channel through AWS Bedrock, while positive for reach, makes Decart dependent on Amazon's go-to-market and pricing architecture. Scenario's aggregator model, which offers 550+ models from 50+ providers, illustrates how commodity infrastructure markets evolve: once multiple world-model vendors exist, an aggregator could undercut any single vendor's pricing moat. For now that risk is prospective, but it frames the 3–5 year competitive durability question correctly.[CP009, CP024, CP030, CP031, CP032, CP034]

3.5 Adverse dynamics: physics-consistency limitations, internal builds, and buyer caution

Independent testing by TechCrunch at launch found that Oasis 3's environment thematic integrity degrades rapidly during extended interaction — a New York City street prompt produced a strong initial scene that quickly became a generic Western urban environment, and the model did not preserve geographic landmarks across navigation. More critically, vehicles in Oasis 3 drive through each other, indicating the model does not yet simulate rigid-body physics correctly. Decart's CEO acknowledged this as "a major research problem we're cracking now," attributing it to data imbalance between normal driving and collision scenarios. These limitations matter because AV companies evaluating Oasis 3 for edge-case synthetic data generation will need to validate that generated scenarios transfer to real-world behavior before committing training budgets. Synthetic data that does not respect physics may introduce distributional artifacts into trained perception or planning models. On the internal-build dimension, Waymo, Tesla, NVIDIA, and Wayve are all building comparable world-model simulation in-house. Decart's real addressable market is therefore the second and third tiers of AV programs, robotics labs, and drone startups that cannot self-fund a world-model research effort. This concentration of addressable customers at the mid-tier reduces the near-term revenue ceiling and creates concentration risk if large customers are few in number.[CP013, CP014, CP015, CP023, CP031, CP032]

3.6 Exhibits

4.1 Revenue model and pricing architecture

Decart operates three publicly documented revenue streams as of mid-2026. The first — and historically earliest — is DOS licensing to major cloud providers, AI laboratories, and hyperscale infrastructure companies through multi-million-dollar commercial contracts. The company's optimization stack was profitable enough on this basis for the CEO to claim in August 2025 that Decart had spent less than $10 million of its $153 million in capital raised, funding its tens of millions in GPU compute costs through licensing revenue. The second stream is API pay-per-second: Lucy 2.1 (realtime video editing, 720p) is priced at $0.02 per generated second, Lucy Restyle 2 (style transfer) at $0.01 per second, Oasis 3 Preview (physical AI and driving simulation) at $0.02 per second for self-serve access, and the prior-generation Lucy Clip at $0.15 per second. Enterprise pricing for Oasis 3 and custom deployments depends on use case and is not publicly disclosed. The third stream is enterprise distribution via AWS Bedrock, announced as part of the $300M Series C in May 2026, which enables Amazon's enterprise customer base to access Decart technology for media, commerce, advertising, and physical AI use cases under custom commercial terms. Revenue recognition dynamics differ across streams: DOS licensing contracts likely follow completion-of-delivery or ratable-license recognition; pay-per-second API revenue is recognized as consumed; enterprise Bedrock contracts would follow negotiated multi-year delivery schedules. The mix between these streams and the relative sizes of each segment are not publicly disclosed, constituting the primary revenue-quality diligence gap. [CI001, CI003, CI004, CI006, CI007, CI014]

4.2 GTM motion and developer flywheel

Decart's go-to-market strategy as of mid-2026 combines a self-serve developer API with direct enterprise sales through strategic partnerships. The developer-led path has produced a community of more than 100,000 developers building primarily on the Lucy API for e-commerce virtual try-on and livestreaming transformation use cases, with Oasis 3 launched in June 2026 targeting autonomous vehicle simulation developers as a second vertical. The AWS Bedrock integration is the enterprise distribution anchor, giving Decart access to Amazon's sales force and cloud marketplace credibility without the cost of building a direct enterprise sales organization from scratch. This reduces upfront customer acquisition cost but cedes margin and control to Amazon's commercial terms. Strategic investors Toyota, Adobe, and eBay are described by CEO Leitersdorf as "all potential customers," creating a captive initial enterprise pipeline that does not reflect arms-length commercial conversion. NVIDIA's dual role as investor and hardware partner further blurs the distinction between financial support and commercial demand. No customer count by revenue tier, net revenue retention, or sales cycle data has been disclosed. The primary GTM risk is that Decart's revenue concentration in a small number of hyperscaler DOS licensing deals and a large, mostly non-revenue-generating developer base creates a barbell revenue structure — a handful of large contracts offset by high-volume but low-monetization developer API usage — making aggregate revenue quality difficult to assess without a customer-by-revenue breakdown. [CI005, CI006, CI007, CI015, CI016, CI022]

4.3 Cost structure, gross margin drivers, and capital intensity

Decart's cost structure is dominated by compute and R&D payroll, with no owned hardware capital expenditure given its asset-light model using leased Amazon Trainium3, NVIDIA GPU, and Google TPU capacity. The gross margin profile of a managed inference API business at scale would typically run 50–80% for software-defined services, but Decart's actual margins depend critically on the terms of its hyperscaler compute agreements, the cost of Trainium3 versus standard NVIDIA instance pricing, and whether DOS licensing revenue carries substantially higher margins than API usage. Publicly, the DOS optimization layer is described as delivering over 80% Model FLOPS Utilization on Trainium3 — a hardware efficiency metric cited by AWS VP Nafea Bshara — and reducing per-video production cost from hundreds or thousands of dollars to less than $0.25 per video. If accurate, these efficiency gains translate directly into gross margin improvement relative to peers running on standard inference infrastructure. However, these are company-claimed benchmarks with no independent third-party audit. DOS 2.0, announced alongside the $300M raise, claims 1,600+ tokens per second for agentic inference versus an industry average of approximately 200 tokens per second, and up to 100 frames per second for world models. If the cost advantage is as described, Decart's gross margin on API revenue could materially exceed commodity inference API margins. R&D payroll for 60+ employees (many with deep low-level systems and HPC expertise) adds a significant fixed cost base. The company has not disclosed COGS, gross margin, headcount cost, or any unit-economics metric publicly. [CI009, CI010, CI011, CI021, CI024, CI034]

4.4 Capital adequacy and financing dependency

Decart's total capitalization exceeds $450 million after its $300M Series C in May 2026. The funding chronology comprises approximately $53 million in seed and early-stage rounds, a $100 million Series B at a $3.1 billion valuation (August 2025, led by existing investors Sequoia, Benchmark, and Zeev, plus new investor Aleph VC), and a $300 million Series C at approximately $4 billion (May 2026, led by Radical Ventures with strategic participants NVIDIA, Toyota Ventures, Adobe Ventures, eBay Ventures, and Amazon as strategic customer). The SEC Form D filed September 2025 by JSL Decart.AI Coinvest, L.P. (CIK 0002084011) confirms a Delaware LP co-investment vehicle was formed for the Series B, though the document provides minimal additional detail on investor composition or valuation mechanics. CEO Leitersdorf stated in August 2025 that Decart had spent less than $10 million of its $153 million in capital, with operating costs covered by DOS licensing revenue. In June 2026, he revised this to "drastically less than $100 million" in lifetime spend — consistent with, but less precise than, the August 2025 claim. Both figures are CEO-stated with no independent audit or board attestation provided. If the lifetime spend claim is taken at face value, Decart enters mid-2026 with roughly $350–$440 million in cash, implying multi-year runway even at materially higher monthly burn than the CEO's implied sub-$1M/month figure. Strategic investors participate commercially: NVIDIA as hardware partner and DOS distribution vehicle, Amazon as customer and Bedrock distribution channel, and Toyota, Adobe, and eBay as potential customers and validators of enterprise demand. This creates a financing-commercial co-dependency where investment terms may be partially predicated on commercial commitments, reducing the independence of either signal. [CI001, CI002, CI012, CI013, CI017, CI020]

4.5 Revenue quality, disclosure gaps, and financial verdict

Decart presents a compelling capital-efficiency narrative and a multi-stream revenue architecture, but nearly all financial claims originate from the CEO and cannot be independently verified. As of mid-2026, no ARR, gross margin, EBITDA, customer count by tier, net revenue retention, burn rate, or audited accounts have been disclosed publicly. The company is a private Israeli entity with no statutory obligation to file public accounts in its operating jurisdiction. Decart's only public financial instrument is the SEC Form D for the co-investment vehicle JSL Decart.AI Coinvest, L.P., which is a procedural filing confirming a Delaware LP and providing no financial performance data. The TechCrunch review of Oasis 3 (June 2026) identifies materially adverse technical limitations — physics-consistency failures, environmental degradation over long sessions, and context-window constraints — that the CEO himself acknowledges as active research problems. These limitations are directly relevant to the financial outlook for the Oasis AV simulation revenue stream: if physics fidelity cannot be guaranteed at scale, enterprise AV customers will defer or limit adoption. The $300 million raise in May 2026 — approximately nine months after the CEO claimed the company barely needed capital — implies a strategic pivot toward aggressive growth investment (hiring, infrastructure, model training, US expansion) that is in tension with the prior capital-efficiency narrative. Whether the company is spending toward a product-market-fit inflection or toward a growth investment cycle is unknowable without quarterly management accounts. The five blocking diligence questions are: (1) confirmed ARR and revenue mix by stream, (2) gross margin by product line, (3) monthly burn rate and cash position as of Q2 2026, (4) terms and concentration of DOS licensing contracts, and (5) a reconciliation of "millions in revenue" against the $3.1–4.0 billion valuation that implies $300M+ ARR at standard SaaS multiples. [CI018, CI019, CI022, CI023, CI025, CI031]

4.6 Exhibits

5.1 Stack definition and module map

Decart positions the company as a vertically integrated research lab rather than a single-model vendor. The public stack separates into three layers: DOS, the optimization and inference layer; Lucy, a family of real-time visual transformation and try-on products for commerce, streaming, gaming, and live media; and Oasis 3, a promptable world model for closed-loop physical-AI simulation. The product surfaces are accessible through a cloud API with official SDKs and published pricing, which is materially more concrete than many frontier-model launches. At the same time, Decart is still moving quickly across narratives: the 2024 Oasis launch centered on a Minecraft-like interactive demo, Lucy then became the visible production product for live video use cases, and the 2026 Oasis 3 launch reframed the company around physical AI and autonomous-vehicle simulation. That breadth is strategically interesting, but it also means buyers still need diligence on how much of the stack is productized versus still research-led.[CE001, CE002, CE003, CE004, CE005, CE006]

5.2 Oasis 3 API mechanics and workflow

The Oasis 3 documentation exposes a real operating workflow rather than a vague demo narrative. Developers open a session against a hosted gRPC endpoint, initialize the scene with a text prompt, and then repeatedly send fixed action chunks made of four throttle-and-steering pairs. The service returns four RGB frames for each advertised camera stream, which the docs describe as left-forward, front, and right-forward. That interface supports a specific claim: the product is built for closed-loop inference rather than one-shot video rendering. The docs also include an RL example and collision-risk demo, which makes the product meaningfully more developer-ready than research previews that lack reproducible mechanics. Still, public evidence does not yet show the richer simulation modalities many AV stacks want, such as lidar, segmentation, or explicit physics-state export, so the current public product appears strongest for camera-first experimentation and early workflow integration rather than a full simulator replacement.[CE011, CE012, CE013, CE014, CE015, CE016]

5.3 Performance claims, hardware coupling, and differentiation

Decart’s strongest differentiation claim is that it can run world and video models in real time because the company co-designs model architecture with its DOS infrastructure layer. Company materials claim DOS 2.0 can exceed 1,600 tokens per second for agentic inference, run across NVIDIA GPUs, Google TPUs, and Amazon Trainium, and push full-HD video or world-model workloads toward 100 FPS. Independent coverage partly supports the hardware-optimization story: AI News and partner coverage describe Decart optimizing Lucy on AWS Trainium and targeting lower latency on Trainium3, while TechCrunch explains why Oasis 3 frame generation quickly becomes a token-throughput problem. However, much of the cost-efficiency advantage, hardware portability, and multiview coherence story still comes from company or partner framing rather than third-party benchmarking. Against alternative world-model approaches such as Genie 3, Decart is differentiated today more by immediate API access, explicit control loops, and production intent than by independently verified long-horizon realism.[CE021, CE022, CE023, CE024, CE025, CE026]

5.4 Governance, privacy, safety, and operational surfaces

Unlike many early-stage frontier-model products, Decart exposes a visible operating surface for production users: published pricing, FAQ guidance on client tokens, terms allocating ownership between input and output, an acceptable-use policy, and a public status page. These materials are useful for diligence because they clarify practical boundaries. Browser and mobile integrations are expected to use short-lived client tokens rather than permanent API keys, active sessions survive token expiry, and the company publishes uptime history through a dedicated status site. The legal surface is more cautionary. Terms assign output rights to users but also reserve broad rights for Decart to use content to improve the platform and for marketing. The privacy policy says the company may process inputs, outputs, generated media, and live audio or video recordings for product improvement and research. The AUP forbids deceptive deepfakes, military use, and safety-critical workflows where failures could cause severe harm, which is notable given Decart’s marketing toward physical-AI and autonomous-driving use cases.[CE033, CE034, CE035, CE036, CE037, CE038]

5.5 Independent caveats and unresolved technical risk

The public evidence base is strongest on interfaces and weakest on long-horizon simulation quality. TechCrunch’s hands-on reporting is the clearest independent caution: the model can produce compelling photorealistic driving scenes, but scene identity can drift over time and other vehicles can pass through one another. That lines up with Decart’s own caveats. The 2024 Oasis research post openly described issues around domain generalization, memory, distant-detail fuzziness, precise inventory or object control, and the need for scaling to address them. The 2026 Oasis page itself says the system is not a physics engine. These are not fatal flaws for an early world-model platform, but they matter for procurement because they narrow where current deployments are credible. Decart looks most credible today as an API-accessible real-time visual or simulation layer with unusually complete developer tooling; it looks less proven as a drop-in replacement for mature physical simulators or as a fully validated training substrate for safety-critical autonomy.[CE020, CE024, CE026, CE027, CE028, CE029]

6.1 Customer segments and buyer map

The fetched sources point to a multi-layer customer base rather than a single wedge. At the top end, Decart sells infrastructure and model capacity to cloud providers, AI labs, and hyperscale buyers through DOS and partnership-driven distribution. The $300 million company publication and CTech coverage both frame Amazon as a strategic customer, while also saying Decart is already generating revenue from contracts with large cloud providers and AI laboratories. A second layer is enterprise application buyers in commerce, media, streaming, and advertising, where Lucy supports virtual try-on, livestreaming effects, and dynamic content workflows. A third layer is physical-AI developers and AV teams using Oasis 3 for camera-first simulation. Finally, a self-serve developer community sits underneath the enterprise motion, supported by public docs, pricing, SDKs, examples, and GitHub repositories. That mix is promising because it diversifies product surfaces, but the public record still leaves open which layer contributes most of current revenue.[CU001, CU002, CU003, CU004, CU011, CU012]

6.2 Adoption signals and self-serve onboarding

Decart’s customer acquisition strategy appears to rely heavily on a developer-first front door. The pricing page exposes self-serve rates, the try-on walkthrough shows how a merchant would integrate live try-on into a product page, the GitHub organization exposes SDKs and reference implementations, and partner coverage repeatedly says Bedrock distribution lowers integration friction for AWS-native teams. This matters because the company is trying to turn realtime AI from a bespoke enterprise sale into an accessible API motion. TechCrunch reports a community of more than 100,000 developers, many building on Lucy in e-commerce and livestreaming, and the public try-on repo count provides a smaller but more concrete artifact-level proof of onboarding effort. However, the public surface still stops short of true funnel transparency. There is no disclosed conversion rate from registered developers to paying accounts, no named list of production merchants using try-on, and no public churn or renewal metric that would demonstrate durable monetization from the developer base.[CU005, CU006, CU007, CU008, CU009, CU010]

6.3 Public customer proof and ecosystem evidence

The strongest named proof is Amazon. Multiple fetched sources—Decart’s own publication, CTech, and JNS—say Amazon joined as a strategic customer as part of the 2026 financing and partnership ecosystem. That signal is stronger than a generic integration because it is repeated across company and independent reporting and linked to concrete Bedrock or Trainium distribution language. Beyond Amazon, the evidence base becomes more cohort-level than logo-level. The company and third-party sources describe revenue from large cloud providers, AI labs, hyperscalers, and deployments across commerce, streaming, social, gaming, and physical-AI scenarios, but they do not publish a comparable list of named production customers. Developer-facing proof is much richer: the GitHub organization, Python SDK, Quest XR project, try-on guides, and AWS partner articles all show how users can actually adopt the product. That split matters for diligence because it suggests strong top-of-funnel and ecosystem credibility, but thinner public proof on enterprise breadth than the headline funding round might imply.[CU001, CU002, CU003, CU012, CU013, CU019]

6.4 Retention, concentration, and adverse disclosure gaps

The main customer risk is not a lack of surface-level activity; it is the narrowness of public disclosure. Amazon is the only clearly named strategic customer across the fetched sources, while the rest of the commercial story depends on broad but unnamed categories such as cloud providers, AI labs, hyperscalers, commerce deployments, and developers. That raises a concentration question: if a meaningful share of near-term revenue depends on a small number of infrastructure customers or on Amazon-linked distribution, downside could be sharper than the company’s diversified marketing story suggests. The other gap is durability. None of the fetched sources publish NRR, GRR, churn, contract length, or even an exact active customer count. Public pricing and example repos make onboarding legible, but they do not prove repeat spend or enterprise standardization. Startup Fortune is the clearest skeptical source because it frames Oasis 3 as a bet on renting simulation before larger incumbents internalize the capability and notes current modality limitations versus more mature internal stacks.[CU025, CU026, CU028, CU029, CU030, CU031]

7.1 Risk Overview and Severity Framework

Decart faces a multi-layered risk stack that is elevated relative to its $4 billion valuation and early revenue stage. The company operates at the intersection of three high-risk regulatory zones: generative AI copyright law (training-data liability), the EU AI Act (real-time AI systems), and US state-level AI governance frameworks that are accelerating in 2026. On top of the regulatory overhang, the company's core technology — autoregressive world models — has documented reliability limitations that are fundamental to the architecture and not cosmetic: TechCrunch's independent review of Oasis 3 (June 2026) found physics-consistency failures and context-window degradation within minutes of generation, issues the CEO acknowledged as ongoing research problems. Operationally, Decart's concentration in AWS Trainium3 infrastructure and its go-to-market dependence on Amazon Bedrock creates a single-throat-to-choke failure mode at the infrastructure layer. Competitively, Google DeepMind (Genie 3), Waymo (internal world model on Genie 3), Runway ($5.3B, $315M Series E, Feb 2026), and Luma AI ($900M Series C) are all advancing with significantly more resources or defensible strategic positions. The risk heatmap (FR001) plots these themes by likelihood and residual impact after reported mitigations. The most concentrated risks — copyright liability and competitive displacement — are in the high-likelihood, high-impact quadrant. [CR001, CR002, CR003, CR004, CR005, CR006]

7.2 Regulatory, Legal, and IP Risks

Training-data copyright is Decart's most structurally ambiguous legal risk. The EU Parliament's Legal Affairs Committee published its report "Copyright and Generative AI — Opportunities and Challenges" (February 2026, 17-3 vote) calling for a rebuttable presumption that any GenAI model placed on the EU market has used copyrighted works for training unless full transparency obligations are met. This presumption — if enacted into EU law — would reverse the burden of proof onto Decart, and a 5–7% flat-rate turnover fee was proposed for retroactive licensing pending a formal framework. Separately, the US Copyright Office's ongoing AI rulemaking (Part 3 of its AI Report) is developing guidance on licensing frameworks for AI training; the Office has not yet issued binding rules but its guidance shapes litigation posture and future statutory risk. Osborne Clarke's February 2026 legal analysis describes the EU situation as a "watershed moment" where existing opt-out systems are insufficient and new frameworks are likely binding on non-EU AI providers serving EU markets (market-of- destination principle). The 2025 Trump Executive Order on AI (December 2025) signals federal intent to preempt state AI regulation in the US but does not eliminate risk — Gunder LLP's 2026 AI laws update notes that states like Colorado and California have enacted comprehensive AI governance statutes that continue to operate and are already shaping vendor contracting. Decart's AUP (updated February 2026) explicitly acknowledges the EU AI Act, DSA, and DMCA as governing frameworks, and prohibits use in military/weapons, real-time biometric surveillance, and social scoring contexts. No public IP litigation against Decart has been identified as of June 2026, but the absence of lawsuits does not confirm training-data licensing compliance. The diligence ask is a full training-data provenance audit with licensing agreements for any copyright-protected sources. [CR007, CR008, CR009, CR010, CR011, CR012]

7.3 Operational, Model Quality, and Security Risks

Decart's operational risk is dominated by the fundamental architectural limitations of its autoregressive world models and its deep dependency on AWS infrastructure. On model quality: TechCrunch's independent June 2026 test of Oasis 3 documented three distinct failure modes — (1) thematic coherence degradation within minutes as the context window fills; (2) physics-consistency failures where vehicles pass through other vehicles because the model lacks proper collision simulation; and (3) control responsiveness issues preventing precise navigation. CEO Leitersdorf acknowledged all three as ongoing research problems, attributing the context degradation to the fact that generating video at 30fps consumes ~8,000 tokens per frame and fills context windows very rapidly at hundreds of thousands of tokens per second. The company is researching longer-context and memory-compression techniques. For AV customers, physics failures are not cosmetic — they mean synthetic training data may contain physically invalid scenarios that could degrade downstream model behavior. This creates a direct training-data quality risk for Decart's target AV market. On infrastructure: Decart's Lucy model runs on AWS Trainium3 and is distributed via Amazon Bedrock, creating a concentrated single-cloud dependency. A sustained AWS outage would halt all inference operations. The Trainium3 hardware is newly announced and limited in broader availability, creating hardware-sourcing risk during scale-up. On security: as a platform generating real-time video from customer-supplied prompts, Decart faces prompt injection risks, jailbreak risk for AUP-prohibited content (deepfakes, CSAM, weapons synthesis), and API key compromise risk. The AUP (February 2026) addresses prohibited uses but independent audit of content moderation effectiveness has not been identified. [CR015, CR016, CR017, CR018, CR019, CR020]

7.4 Partner, Dependency, and Customer Concentration Risks

Decart's go-to-market and infrastructure strategy creates layered concentration risks. The most material single dependency is AWS/Amazon: Decart's Lucy model runs on AWS Trainium3, models are distributed via Amazon Bedrock, and Amazon itself is listed as a strategic customer — creating a three-way concentration where Amazon is simultaneously cloud provider, distribution channel, and revenue source. If Amazon develops competing internal world-model capabilities (plausible given their AWS AI infrastructure investment) or reduces Bedrock shelf-space for third-party models, Decart's distribution channel collapses at the same time as its infrastructure. Similarly, Nvidia is both Decart's primary alternative compute supplier and an equity investor (Nvidia participated in both the $100M Series B and the $300M Series C). A material Nvidia relationship dispute or competitive pivot would simultaneously affect Decart's hardware roadmap and its cap table optics. Customer concentration is amplified by the investor/customer overlap: Toyota Ventures, Adobe Ventures, and eBay Ventures are equity investors whose parent companies are also listed as strategic customers. This arrangement is positive for signal quality but creates adverse incentive dynamics — a customer churning from the platform while holding equity sends a particularly damaging market signal. Startupfortune (June 2026) noted that Decart's real market is the "dozens of AV programs, robotics labs, and drone startups" below the Waymo/Tesla/Nvidia tier — meaning its monetizable customer base is smaller, less well-capitalized, and less strategically sticky than the strategic investor/customers. The developer community (100,000+ registered) is large but API revenue at $0.02/second is highly volume-dependent and the revenue per developer at typical usage patterns is uncertain. [CR022, CR023, CR024, CR025, CR026, CR027]

7.5 People, Execution, and Financing Risks

Decart's execution risks center on key-person dependency, geopolitical exposure, and the financing gap between its $4 billion valuation and publicly disclosed revenue metrics. Dean Leitersdorf (CEO, 27 years old) has been the primary public face, fundraiser, and technical narrative driver for every round from seed to Series C. His departure would simultaneously create a fundraising, technical credibility, and media-narrative gap at a critical scale-up phase. The company's leadership team outside the two co-founders is not publicly documented in depth, with Dr. Kfir Aberman (formerly Snap and Google, co-creator of DreamBooth) being the most senior publicly named hire as head of the SF R&D center. This creates execution risk in the event of any co-founder or Aberman departure. The Israeli operational HQ creates geopolitical risk: a sustained escalation in the Israel-Gaza conflict or broader regional instability would affect the Tel Aviv engineering team's ability to operate, access visas for global travel, and maintain relationships with risk-sensitive enterprise customers. Several large enterprises (particularly in defense-adjacent sectors) have procurement policies against vendors in active conflict zones. Financing risk: Decart has raised $450M+ but claims lifetime burn of "less than $100M" (CEO, June 2026). Even if this is accurate, the burn rate will accelerate significantly as the company scales AV/physical-AI simulation infrastructure, expands headcount from 60+ employees, and builds out enterprise sales capacity. At the $4 billion valuation, the next round would need to be at $5–6 billion or higher to avoid a flat/down round, which requires demonstrable revenue traction that is not yet publicly evidenced. [CR028, CR029, CR030, CR031, CR032]

7.6 Mitigations, Kill Criteria, and Diligence Asks

Decart has taken meaningful but incomplete steps to mitigate its top risks. Regulatory: the AUP (February 2026) explicitly references EU AI Act, DSA, and DMCA; this demonstrates regulatory awareness but is not a substitute for training-data transparency obligations or a licensing framework. Technical: the CEO has publicly acknowledged physics-consistency and context-window limitations and stated active research programs; the next model version is expected to address video-seeded world generation, which may improve consistency. Infrastructure: the AWS Trainium3 deal provides performance efficiency but increases AWS lock-in; no public multi-cloud fallback has been identified. People: the SF R&D center hire of Dr. Kfir Aberman represents geographic and talent diversification but does not resolve co-founder concentration. The mitigation and kill-criteria table (TR005) maps each major risk to monitorable triggers and investment-thesis implications. Diligence asks prioritized by urgency: (1) training-data licensing agreements — the single most material unresolved item for EU market access; (2) independent security and content-moderation audit; (3) AV customer physics-accuracy acceptance criteria — whether Decart's physics simulation meets safety-case requirements for any AV company; (4) revenue composition by product line and customer — DOS licensing vs. Lucy API vs. Oasis 3 enterprise; (5) retention metrics for e-commerce and gaming customers using Lucy. [CR033, CR034, CR035, CR036, CR037]

7.7 Exhibits

8.1 Funding History and Capital Structure

Decart has raised over $450 million across multiple rounds since its founding in 2022, with a trajectory that places it among the fastest-valued Israeli AI startups on record. The most recent round — a $300 million Series C closed in May 2026 — brought the post-money valuation to approximately $4 billion. The round was notable for its strategic composition: lead investors include technology companies that are simultaneously paying customers (Toyota, Adobe, eBay Ventures), creating an unusual alignment structure where capital and commercial adoption are bundled. Radical Ventures led, with participation from Nvidia, Sequoia Capital, and Benchmark. Prior to the Series C, Decart raised $100 million at a $3.1 billion valuation in August 2025 (Series B), implying a 29% step-up in approximately nine months. The cumulative dilution and pre-money ownership structure have not been publicly disclosed. The $450 million+ war chest is substantial for a company at Decart's commercial stage, and implies a multi-year runway even at elevated compute spending, but burn rate figures are not available for validation. The strategic investor base introduces governance complexity: customers who are also investors may have conflicting incentives around product roadmap, enterprise pricing, and potential exit routes. [CV001, CV002, CV003, CV004, CV005]

8.2 Comparable Company Analysis

The nearest public comparables for Decart are private-market peers in AI video, world modeling, and AI gaming infrastructure. Runway AI — the closest functional comparable — raised $315 million at a $5.3 billion valuation in February 2026, and has publicly indicated it is expanding from creative video editing toward world models, making it a direct future competitor with a $1.3 billion valuation premium over Decart. World Labs raised $200 million in February 2026 with an anchor investment from Autodesk, targeting professional 3D workflows at a similar ~$5 billion implied valuation; it is narrower in scope than Decart but has a clearer commercial vertical in the $25 billion 3D design software market. Luma AI raised $900 million in Series C funding from HUMAIN (Saudi Arabia's state AI fund) in November 2025 at a ~$900 million valuation — a case where total capital raised ($450 million at time of round) approximately equals post-money valuation, suggesting significant investor caution on the video generation sub-segment. Inworld AI, valued at over $500 million with $100 million raised, is the primary AI-native games infrastructure peer; its focus on NPC character engines is narrower but complementary. The peer group average post-money valuation is approximately $2.9 billion, and Decart's $4 billion represents a 38% premium over the peer group average. The premium is defensible if real-time world simulation proves to be a higher-value architecture than post-production video generation, but is currently assumption-led given Decart's opaque commercial metrics. [CV006, CV007, CV008, CV009, CV010, CV011]

8.3 Total Addressable Market

Decart addresses three overlapping market segments with different maturity profiles. The global gaming market, sized at approximately $210 billion in 2025 by BCG's Global Gaming Survey, is the near-term wedge: simulation infrastructure embedded in game engines could capture a fee-per- session or API-call model targeting 3-5% of developer tooling budgets — implying a serviceable addressable market of $1-3 billion in gaming simulation alone if Decart achieves toolset integration. BCG specifically identifies generative AI as a primary structural change in the gaming industry, projecting that ~20% of new Steam games disclosed AI use in mid-2025 (double the prior year), and that the gaming industry is exiting its three-year post-pandemic slump. The Business Research Company estimates the generative AI in gaming market at $2.1 billion in 2025, growing to $10.5 billion by 2030 at a 30% CAGR — Decart's real-time world simulation is a premium offering within this segment. Research and Markets similarly sizes the generative AI gaming market at $2.1 billion in 2025 growing at a 24% CAGR through 2030. The second segment — autonomous vehicle simulation — is potentially the largest value pool but carries the most technology risk given Decart's documented physics-consistency limitations. The third segment — e-commerce and retail try-on — is already generating API revenue via the Lucy model and is the clearest near-term ARR path, though the TAM is smaller ($500M-1B for AI virtual try-on). McKinsey's State of AI report confirms enterprise AI adoption rates have crossed 72% in large organizations, validating the enterprise demand backdrop, though Decart's share of that spend is unquantified. Statista sizes the global mobile games market at approximately $92 billion in 2025, a subset of the total gaming TAM directly accessible through Decart's API. [CV013, CV014, CV015, CV016, CV017, CV018]

8.4 Revenue Scenarios and Financial Modeling

In the absence of any publicly disclosed revenue data, valuation modeling requires assumption-led scenario construction. The primary revenue drivers are: (A) Lucy model API fees for e-commerce and livestreaming (current traction); (B) Oasis 3 gaming simulation API fees at $0.02/second (launched June 2026); and (C) potential AV enterprise contracts (unannounced, no public evidence yet). CBInsights confirms no ARR or financial metrics have been publicly disclosed by Decart. In a bull scenario, Decart converts its 100,000+ developer community into paying customers at even modest retention, achieves 3-5 enterprise gaming or AV contracts, and reaches $30 million ARR by end of 2026 and $200 million by 2028 — yielding a 20x forward multiple that sustains current valuation. In a base scenario, adoption is slower and Decart reaches $15 million ARR by end of 2026 and $80-100 million by 2028 at a 20x multiple, implying a significantly discounted valuation of $1.6-2.0 billion versus the $4 billion current pricing. In a bear scenario, physics failures close the AV market and competition from Runway and Google limits gaming adoption; Decart reaches $25 million ARR by 2028, implying a $500 million valuation at 20x — an 87.5% discount to current pricing. Roblox Corporation's SEC 10-K filings provide a comparable public company anchor: Roblox generated approximately $3.9 billion in revenue in fiscal 2025 at a market capitalization near $26 billion, implying a price/revenue multiple of approximately 6-7x for a mature gaming platform — suggesting that even bull-case Decart revenues would need to be exceptional to sustain a $4 billion market value at maturity-stage multiples. [CV020, CV021, CV022, CV023, CV024, CV025]

8.5 Valuation Framework, Risk Factors, and Thesis Breaks

The current $4 billion valuation can be decomposed into three components: (1) demonstrated technical optionality — real-time world model architecture that is unique at current latency and fidelity specs; (2) infrastructure partnership premium — AWS Trainium3 production deal and Amazon Bedrock distribution create scale that most startups at this stage do not have; and (3) strategic investor premium — the customer-investor structure implies high early adoption probability. Against this, the adverse thesis is material: TechCrunch's June 2026 independent review documented physics-consistency failures in Oasis 3 that the CEO acknowledged as an active research problem, indicating core product limitations are not cosmetic. CBInsights rates Decart as high-risk due to lack of financial disclosure. The physics limitation is the key thesis-break trigger for the AV vertical, which arguably accounts for 30-50% of the bull-case valuation. Three additional thesis-break triggers are: (A) failure to demonstrate enterprise-grade ARR by Q4 2026; (B) a material copyright enforcement action targeting Decart's training data in the EU; or (C) Runway or Google shipping a real-time world model at competitive price-quality levels within 12-18 months. At current comparable valuations, the AWS and Nvidia strategic partnerships do not justify the premium to peers: Runway has similar strategic relationships (Adobe acquisition candidate) and disclosed commercial traction. A diligence session must include audited revenue, gross margin, burn rate, and customer concentration data before any recommendation can be upgraded beyond "track." [CV027, CV028, CV029, CV030, CV031, CV032]

8.6 Recommendation, Diligence Asks, and Disclosure Limitations

Recommended stance: TRACK. Confidence: Medium. Risk rating: High. Valuation stance: Stretched. The investment case for Decart is asymmetric — the technical differentiation, AWS partnership, and strategic investor quality are genuinely superior, but the $4 billion price point embeds execution assumptions that are not yet evidenced by any public financial disclosure. The recommendation is to track rather than pass because the TAM is large enough and the technical moat is real enough that a premature "pass" risks missing a generational platform opportunity; but a "buy" recommendation requires financial validation that is currently unavailable. The upgrade trigger to "buy" is: (1) disclosure of ARR at or above $25 million with demonstrated YoY growth; (2) at least one public AV enterprise contract with disclosed TCO; (3) publication of a physics-consistency roadmap with measurable milestones. The downgrade trigger to "pass" is: any material copyright enforcement action, confirmed physics failures blocking AV adoption, or a Runway/Google world model launch at price parity. Disclosure limitation: all financial projections in this chapter are assumption-led; Decart has not publicly disclosed revenue, ARR, bookings, burn rate, gross margin, customer count, or retention metrics as of June 2026. This chapter presents scenario ranges only and explicitly cannot support a specific investment price recommendation. Investors should treat the $4 billion valuation as a hypothesis to be tested, not a validated market price. [CV034, CV035, CV036, CV037, CV038, CV039]

8.7 Exhibits