Clark is an intelligence infrastructure company that builds the world's first Hierarchically Detached Federated Mixture-of-Experts reasoning system — a 70-billion-parameter backbone that routes intelligence to 10,000 specialist expert models (each 100M to 500M parameters), producing structured, verifiable, traceable outputs at a fraction of the cost of equivalent monolithic systems. The backbone is the circuit. The experts are the lightbulbs. The circuit does not produce light. It is the infrastructure through which every specialist illuminates exactly what they were trained to illuminate, precisely when required.

Together, a 70B backbone and 10,000 experts provide the knowledge coverage of a system with over 5 trillion parameters, but at the inference cost of activating only the handful of experts relevant to each specific query. This is the architecture the entire AI industry will eventually converge toward. Clark is building it first, from Chennai, and deploying it to the world.

The AI industry has convinced itself that intelligence scales by making one model larger. This is correct but incomplete. Scaling a dense model improves average performance across all domains simultaneously, which sounds appealing until you consider the economics. To improve a dense model's IFRS accounting performance by 10%, you must increase the entire model's capacity — all the chemistry knowledge, all the history, all the code — by a proportional amount. You are paying the computational cost of a trillion-parameter system to improve one domain.

Clark improves IFRS performance by training a better IFRS expert. The cost is proportional only to the IFRS expert's size — 200M parameters, trained exclusively on accounting literature. The leverage is different by orders of magnitude. And when a user asks a question that spans IFRS accounting, Indian company law, and international transfer pricing simultaneously, Clark activates three experts in parallel. A dense model activates no experts — it activates everything and hopes the correlation patterns in its training data surface something coherent.

The circuit produces no light by itself. Every lightbulb illuminates exactly what it was trained to illuminate. Together they light up a room that no single bulb, however powerful, could illuminate alone.

Clark launches in India because India provides the optimal calibration environment: 1.4 billion potential users, 17 million developers, cost-sensitive economics that demand genuine architectural efficiency rather than compute brute-force, and 22 scheduled languages that require multi-lingual expert coverage from day one. India is where the circuit gets wired correctly under real conditions.

But the architecture is language-agnostic and jurisdiction-agnostic by fundamental design. Training a Tamil legal expert or an English common law expert is the same protocol applied to different corpora — both are 100M–500M parameter models registered to the same backbone. Training a German tax code expert or a Japanese medical terminology expert follows the same path. Global expansion is not a rebuild. It is a registration event. New language, new jurisdiction, new domain: train the expert, certify it, register it. The circuit already exists. The socket is already there. You are adding a bulb.

The global addressable market — ₹39.6 trillion annually across individual users, enterprise, and the API-developer ecosystem — is the total worldwide demand for reliable, structured, verifiable intelligence across every domain, language, and jurisdiction. Clark is building the infrastructure to serve that demand from a base in Chennai that the entire world connects to.

The prevailing consensus assumes that artificial intelligence will advance primarily through scale — bigger models, more parameters, more training data. This assumption is deeply embedded in how every major AI lab allocates capital and sets research priorities. It is also the source of a fundamental exploitable error.

Scaling a dense model cannot produce the same depth per domain as a specialised model trained exclusively on that domain. A 70B dense model must distribute its 70 billion parameters across every domain of human knowledge simultaneously. Clark's 200M IFRS expert concentrates all its parameters on one domain. The depth comparison is not even close. And when the user's question spans five domains, Clark activates five specialists simultaneously. The dense model guesses from blurred memory. Clark illuminates from focused expertise.

The entire Clark model rests on one testable assumption: that a 70B backbone trained for orchestration, routing queries to 100M–500M parameter specialist experts, produces outputs that are more reliable, more verifiable, and more economically efficient than a monolithic model of comparable or greater total parameter count. This is not assumed to be true — it is designed to be empirically tested at Month 4.

Across students, developers, analysts, founders, and enterprise decision-makers, complaints about current AI systems converge with striking consistency. They are not complaining about speed, access, or intelligence. They are complaining about structure and reliability — which are architectural properties, not parametric ones. Verbatim: 'It gives me an answer but I don't trust it.' 'I still have to check everything manually.' 'It breaks on real problems.' 'I spend more time fixing its output than doing the work myself.' 'It sounds completely confident and is completely wrong.' 'I can't use it for anything that actually matters.' 'I need five tools to finish one task and none of them talk to each other.'

The pattern is unambiguous. Users are not asking for a more capable model. They are asking for a more reliable system. You cannot make a monolithic model more reliable by making it larger — you can only make it more capable on average. Clark's circuit-and-lightbulb architecture is designed specifically for reliability, because reliability requires decomposition, specialisation, routing, and verification. None of these are achievable in a single forward pass through a dense model.

Clark's total addressable market cannot be read from any existing market research report because no existing report models the convergence of individual intelligence tools, enterprise reasoning platforms, and API infrastructure into a single architecture. Clark is building the infrastructure layer beneath all three simultaneously.

When Clark opens its expert registration protocol to third parties — universities, enterprises, research institutions, practitioners — the knowledge base grows without Clark paying for that growth. Consider: IIT Bombay's materials science department trains a 200M metallurgy expert and registers it to Clark's backbone. IIT Bombay earns 70% of every API call that routes to their expert. Clark earns 30% with zero development effort. The backbone gets smarter. The marketplace deepens. Users gain access to genuine materials science expertise no monolithic model can match. This is market creation.

Clark's competitive situation is architecturally unusual. The difference between Clark and its competitors is not a matter of degree — it is a matter of kind. OpenAI, Google, Anthropic, and Mistral are all building increasingly capable versions of the same architecture: a single dense transformer model, pre-trained on broad data, fine-tuned for specific behaviours. Clark is building a fundamentally different architecture: a routing backbone plus a registry of detached specialist experts. These two approaches are not on the same evolutionary path. You cannot evolve a dense model into Clark's architecture without rebuilding from scratch.

Strength: brand dominance, developer ecosystem, fastest model iteration. GPT-5.4 pricing: $2.50/1M input tokens — verified at OpenAI Pricing ↗. Structural blindspot: OpenAI is architecturally committed to scaling dense models. Every research team, every infrastructure investment, every product decision assumes that making one model larger and better-trained is the path to better AI. Pivoting to detached MoE would require dismantling the research programme that defines their identity. It is not a pivot they can make without becoming a different company.

Strength: distribution across search, productivity tools, cloud, and advertising. Deepest financial resources. Gemini 2.5 Pro pricing: $1.25–$2.50/1M — verified at Gemini Pricing ↗. Structural blindspot: Google's business model depends on search query volume. A system that replaces search with direct expert-level answers reduces advertising inventory. Google must develop AI within the constraint of not fully cannibalising search revenue. This is a permanent structural conflict, not a temporary tension. Clark has no such conflict.

Clark's primary user is defined by mindset rather than demographics: precision-seeking, efficiency-obsessed, with zero tolerance for confident-sounding incorrect answers. This user has been burned by AI-generated outputs that looked right and were wrong. They have paid the cost — in time, in credibility, in bad decisions — of trusting a system that was not actually reliable. Three defining fears drive their behaviour: the fear of invisible mistakes in plausible-looking outputs; the fear of being outpaced by peers with more capable tools; and the fear of losing ownership of their own thinking process. Clark's architecture addresses all three: traceable expert outputs eliminate invisible mistakes; genuine expert-level depth provides competitive advantage; the system amplifies rather than replaces judgment by showing its reasoning chain.

Enterprise budget holders — CTOs, Heads of Operations, Chief Product Officers — are not buying AI capability. They are buying three things: operational leverage (more output with the same headcount), risk reduction (AI-assisted decisions that can be audited and defended), and strategic competitive advantage (the ability to deploy genuine domain expertise across the organisation without hiring 100 additional specialists). Clark's expert marketplace addresses the third point in a way no competitor can. An enterprise can access 10,000 domain experts simultaneously at a fraction of the cost of maintaining even 10 in-house specialists.

The current market charges 5–10× the actual cost of delivering expert-level intelligence. A chartered accountant charges ₹3,000–15,000/hour for IFRS compliance analysis. Clark's IFRS expert delivers comparable structured analysis at ₹150–500 per query. A patent attorney charges ₹8,000–25,000/hour for prior art analysis. Clark's patent law expert and chemistry expert, co-activated by the backbone for a pharmaceutical patent query, deliver traceable, citable analysis at ₹500–2,000 per query. The gap exists not because technology is expensive — Clark's inference cost for a 300M expert is fractions of a rupee — but because specialist knowledge was previously locked behind years of human specialisation.

The backbone does not answer questions. It is trained to understand the structure of questions — to recognise what kind of problem has arrived, what sub-problems it can be decomposed into, which specialist domains each sub-problem requires, and how the outputs of those specialists should be synthesised. Think of the backbone as a highly experienced project manager who has worked across every domain of human knowledge. They do not need to know how to perform brain surgery — they need to know enough about brain surgery to recognise when a brain surgeon is needed, to understand the surgeon's output, and to integrate it with the radiologist's and pharmacologist's outputs. The backbone's training data is problem-decomposition patterns, expert-output evaluation criteria, and multi-domain synthesis — not domain content itself.

Clark's routing infrastructure is stateless — backbone model weights are read-only after training. Every routing request can be handled by any available backbone inference node without session-specific state. Horizontal scaling = adding nodes with no architectural changes. Expert models are served through a similar stateless layer. Only a small subset (3–12 experts per query) activates per request, keeping per-query compute cost low regardless of total registry size.

Clark's pricing is anchored to the economic value of expert-level, verifiable output — not tokens consumed. If Clark's IFRS expert replaces ₹15,000/hour chartered accountant time for structured compliance analysis, pricing is anchored at ₹1,000–5,000 per complex query — a fraction of the displaced value, while maintaining 85%+ gross margin. This reframes the conversation entirely: users are not comparing Clark's price to OpenAI's price per million tokens. They are comparing Clark's price to the professional services cost it replaces. Clark wins that conversation decisively.

Clark's most powerful distribution mechanism requires no paid marketing spend: the expert contributor programme. When a domain expert trains and registers an expert model, they receive 70% of every API call that routes to their model. This creates an immediate economic incentive to promote Clark's platform within their professional network. A tax attorney who registers a GST compliance expert and earns ₹50,000/month in passive API revenue becomes an advocate for Clark in every professional conversation. More expert contributors → more professional network exposure → more users discovered → more API revenue → more expert contributors attracted. This flywheel is self-reinforcing and cannot be purchased by a competitor — it must be earned through having the circuit first.

There is a philosophical resonance between Clark's product architecture and its hiring philosophy: Clark builds a system where specialist experts are orchestrated by a generalising backbone. Clark hires specialists orchestrated by a leadership backbone. The hiring principle is the same: hire people who are genuinely expert in a narrow domain rather than generally capable across many. A candidate who 'knows a bit about' ML infrastructure, security, and frontend is less valuable than a candidate who knows ML infrastructure at a depth where they have made original contributions.

Clark's deepest moat is the combination of routing intelligence and expert registry depth. Both are earned through deployment. Both compound with time. Together they create a competitive position that widens every day the system is used.

Every major AI regulation enacted or proposed globally — India's DPDP Act, the EU AI Act, the US AI Executive Orders, and emerging ISO and NIST standards — converges on three requirements: traceability (outputs must be explainable and auditable), data sovereignty (personal data must be processed in compliant jurisdictions), and human oversight (high-stakes decisions must have audit trails enabling human review). Clark's architecture satisfies all three structurally, because these properties were designed as core architectural features rather than compliance afterthoughts.

The circuit-and-lightbulb architecture is uniquely suited to regulatory compliance. Every conclusion is tagged to the contributing expert model. Every expert's domain scope is certified and documented. Every reasoning chain is traceable step by step. When a regulator asks 'how did this AI system reach this conclusion?' Clark's answer is not a probabilistic explanation — it is a specific, ordered sequence of expert contributions with documented provenance. No dense monolithic model can provide this.

India is not simply 'a large market.' It is the optimal calibration environment for Clark's architecture. Cost sensitivity forces genuine efficiency — you cannot hide behind compute brute-force when your customer base has a maximum individual willingness to pay of ₹500–2,000/month. This constraint forces Clark to build the most efficient possible routing architecture, and that efficiency becomes a structural competitive advantage when the platform expands to higher-paying markets where competitors are still paying for inefficiency.

India also provides the linguistic and jurisdictional diversity that makes the detached MoE architecture's breadth claim credible from day one. A system that can handle Tamil legal questions, Gujarati business analysis, Hindi technical writing, and English financial modelling simultaneously — with certified expert depth in each language — is demonstrably more capable than any English-optimised system. India is where the expert registry must prove its multi-lingual depth.

Clark does not try to sound like the future. It tries to become indistinguishable from how serious work gets done. The brand is built around four properties: precision (every word used is the correct word, no superlatives), reliability (the brand voice is as consistent as the system's outputs), depth (never shallow, never trend-chasing), and clarity (complex architecture explained in terms a tenth-grader could follow and a domain expert would not find imprecise).

Clark's ambition is not to be the best AI company. It is to define what the category of Intelligence Infrastructure Systems means — to set the standards by which all future systems are measured, to establish the expert model certification protocols that the entire industry adopts, and to create the open ecosystem through which any qualified entity can contribute expertise to the global intelligence layer.

The expert certification standard that Clark develops internally will, within five years, be proposed as an industry standard through NASSCOM, ISO, and IEEE. Just as TCP/IP is the standard through which all internet traffic flows, Clark's expert model interface specification will be the standard through which all deployed intelligence is routed. This is category ownership at its deepest.

In fifty years, Clark is not a company. It is the infrastructure. The circuit is wired into every institution, every profession, every domain of human knowledge. No one asks which intelligence infrastructure they use — the way no one asks which electrical infrastructure powers their building. Clark is the circuit through which the entire accumulated expertise of human civilisation is made accessible to every person on Earth, in their language, in their context, at the cost of electricity.