# SoTO Labs — Pitch Analysis > "Unlocking the world's most valuable private data for frontier AI" > Infrastructure connecting AI labs to proprietary institutional datasets across MENA, ASEAN and Africa. Raising $10M seed at $66.7M post-money. 15% to investors. --- ## What They're Building SoTO is a two-sided marketplace and infrastructure layer sitting between **frontier AI labs** (who need differentiated training and evaluation data) and **large regulated institutions** in the Global South (who hold it but cannot safely commercialize it). The three-step model: 1. **Source** — identify institutions with valuable private data through sovereign relationships (MENA, ASEAN, Africa) 2. **Prepare** — deploy anonymization software inside the client environment. Data never leaves the building. 3. **Connect** — license certified, lab-ready datasets to OpenAI, Anthropic, xAI via pre-qualified relationships Business model: **15–50% commission per transaction**. Self-described as Palantir-style infrastructure + investment-bank licensing + AI compliance tooling. The more interesting move: they don't just sell raw data. They convert it into **RL environments** — the full stack of input data → golden references → rubrics → tasks — which takes a $100K dataset to $750K+ in value. That's where the real margin lives. --- ## Why Now — The Timing Signal Three converging forces make this window real: 1. **Meta/Scale AI deal** weakened Scale's neutral supplier position and opened procurement gap. Labs are actively reallocating spend. 2. **Frontier shifted to post-training** — from bulk web pretraining to reinforcement learning, agent training, and bespoke evaluation environments. The most valuable inputs now sit inside institutions, not on the public web. (Anthropic alone reportedly spending $1B+ on RL environments.) 3. **Supply infrastructure barely exists** for regulated, private, institutional data from emerging markets — especially MENA/ASEAN/Africa. The category is genuinely nascent. This is the exact market shift described in [[model layer and above]]: *"the shift from static training data to agentic learning environments is a major directional arrow of AI progress."* SoTO is betting on that arrow. --- ## Traction | Signal | Detail | |---|---| | Institutional partners | 3 secured — medical group, TV/radio broadcaster, large dam construction company | | Lab relationships | OpenAI, Anthropic, xAI (active procurement) | | Validated data | 130K real medical scans run through anonymization platform | | RL environments | In progress — being packaged into frontier-lab-ready evaluation suites | --- ## The Team - **Elyas Felfoul** (President) — sovereign networks, MENA institutional access, AI policy. WISE, MILA AI Fellow, XPRIZE, LKY Singapore - **Younes Mourri** (CEO) — Stanford AI lecturer, 2M+ learners, Founder LiveTech.AI, ML for governments - **Hassan Hayat** (CTO) — Ex-Two Six Capital, built $160B data platform. AI@Stanford, Math@PennState *"Anchored in Qatar. Serving the world."* --- ## Vault Principles That Apply Here ### 1. The Moat Is Geographic, Not Technical From [[Technical Moat Assessment Framework]]: > **Durable (rare):** Sovereign/political positioning in specific geographies. Customer trust built through successful production deployments. Proprietary training data from customer operations. > The pattern: individual AI components commoditize fast. Integration and orchestration commoditize slowly. Customer relationships and political positioning **don't commoditize at all.** SoTO's sovereign access in MENA/Africa is explicitly the category that doesn't commoditize. An American competitor cannot replicate Elyas's sovereign networks or Younes's government ML relationships in 12 months. This is the real defensibility claim — and it's sound. The anonymization technology and the licensing infrastructure are table stakes. The access is the moat. ### 2. Consultancy-to-Platform Risk Is the Core Execution Risk From [[Consultancy-to-Platform Transition]]: > The trap is staying in bespoke mode and calling it a platform anyway. Jumping to "platform" language before doing the ugly, vertical-specific work is how teams get stuck. SoTO's Source → Prepare → Connect pipeline currently looks like a transaction service: they identify, they deploy software, they broker. Each deal probably requires significant bespoke work. The question is whether the anonymization deployment and the data certification process can become increasingly standardized — turning each new institution into faster, more margin-rich supply. Watch: **how many hours does it take to onboard a new institution from first contact to lab-ready dataset?** If that number isn't falling with each engagement, the unit economics don't compound. The data flywheel only kicks in when institutional onboarding becomes configuration, not consulting. See [[Azraq Data Flywheel]] for what a well-designed flywheel looks like: Ingest → Learn → Codify → Accelerate. ### 3. The "3 Hard Truths" Are Being Applied Correctly From [[3 Hard Truths of Deep Tech Commercialization]]: > You need to do things that are irrational and not scalable initially. You need custom, application-specific products first before generalisation. SoTO is doing this right. Starting with medical imaging (130K scans validated), media archives, and industrial datasets — three distinct verticals, each requiring different anonymization approaches. This is correct sequencing: find the patterns that repeat, then build the platform that automates them. The risk is not the approach; it's the discipline to recognise when to abstract upward. ### 4. Key-Person Risk — Distributed Enough From [[Key-Person Risk in Deep Tech]]: > Bus factor = 1 means the company dies if one person leaves. In deep tech, this is common and dangerous. The three-founder structure here is deliberately anti-fragile: sovereign access (Elyas), lab access (Younes), infrastructure (Hassan). No single person holds all the keys. However: if Younes's Stanford/AI lab relationships are the primary channel into OpenAI/Anthropic/xAI, and those relationships are personal rather than institutional, there is latent key-person concentration. Worth exploring whether lab relationships are documented and distributed. ### 5. Post-Training Demand Signal Is Confirmed From [[model layer and above]]: > Multi-agent reinforcement learning verifiers are emerging to score subjective outputs—from legal writing to manufacturing workflows—turning them into feedback that models can learn from. This shift from static training data to **agentic learning environments** is a major directional arrow. SoTO's move up the value chain from raw data → RL environments is exactly this. The $1B+ Anthropic spend on RL environments is the real demand signal. Expert tasks, rubrics, and golden outputs from regulated industries (medical, engineering, infrastructure) are precisely the high-context, non-synthetic data that cannot be generated by another model. This is defensible supply. ### 6. The Global South Infrastructure Gap Is Real From [[2025 Year-End Reflection - What Landed, What Didn't]]: > The future is here, but it's very unevenly distributed, especially across the global south. SoTO is explicitly building in the gap. No one else is building the trust, anonymization, and licensing layer for MENA/ASEAN/Africa institutions. The category leadership window is real — and speed is the variable they correctly identify in their use-of-funds framing. --- ## Opportunity Framing > [!tip] Core Bet > SoTO is betting that the supply side of post-training data infrastructure will consolidate around trusted intermediaries with geographic/sovereign positioning — the same way investment banking consolidated around trusted intermediaries with regulatory and relationship positioning. If that's right, the moat is: - Trust (data stays in-building, anonymization certified) - Access (sovereign relationships not replicable by outsiders) - First-mover in markets with high regulatory friction (MENA, ASEAN, Africa) - RL environment know-how accumulates with each domain The comparison to Mercor and Turing in their pitch ($10M → $300M Y1→Y3) is cheeky but the directional logic holds: once inside frontier lab procurement cycles, trusted suppliers scale fast. --- ## Open Questions / Watch Points > [!question] Key diligence questions > 1. **Onboarding velocity** — how many hours/weeks to onboard a new institution? Is it falling? > 2. **Lab contract depth** — are OpenAI/Anthropic/xAI relationships formal procurement or informal conversations? > 3. **Anonymization IP** — is the on-premise anonymization software genuinely proprietary, or built on open-source (Presidio, ARX, etc.)? > 4. **Exclusivity** — do institutional data agreements grant SoTO exclusivity, or can the institution sell to others? > 5. **Younes's lab relationships** — personal or institutional? Key-person concentration risk here. > 6. **Data rights** — how are contracts structured when an institution's domain experts produce golden references and rubrics? Who owns the RL environment? > 7. **Regulatory arbitrage or friction?** — being in MENA/Africa could mean looser data protection (easier to move) or stricter sovereignty rules (harder to export). What's the actual legal framework per market? --- ## Related Notes - [[Technical Moat Assessment Framework]] - [[Consultancy-to-Platform Transition]] - [[3 Hard Truths of Deep Tech Commercialization]] - [[Key-Person Risk in Deep Tech]] - [[Azraq Data Flywheel]] - [[model layer and above]] - [[Defensibility Principles MOC]] - [[2025 Year-End Reflection - What Landed, What Didn't]] --- Tags: #investing #ai #deeptech #data-infrastructure #mena #firstprinciple