Deep Cogito raises $43M for enterprise-owned specialized AI models

Deep Cogito raised $43 million in a Series A round led by TQ Ventures, with participation from Benchmark, Nexus Venture Partners, Atreides Management, South Park Commons, and Zscaler, to build specialized AI models that enterprises can train on proprietary data and own. The funding, announced Aug. 26, brings Deep Cogito's total funding to more than $56 million. Co-founders Drishan Arora and Dhruv Malrana, former Google employees, aim to shift enterprise AI value toward post-training and model ownership.

Deep Cogito raises $43M for enterprise-owned specialized AI models TQ Ventures led the Series A, with Benchmark, Nexus, Atreides, South Park Commons and Zscaler backing former Google employees Drishan Arora and Dhruv Malrana. By RuntimeWire Staff /author/runtimewire-staff · Published Primary source: The Wall Street Journal https://www.wsj.com/cio-journal/deep-cogito-aims-to-put-companies-in-control-of-their-own-ai-50ddf44b Why it matters Deep Cogito is betting that enterprise AI value will shift toward post-training and model ownership, giving customers control over data, deployment and recurring inference costs. Drishan Arora and Dhruv Malrana raised $43 million for Deep Cogito https://www.deepcogito.com/?ref=runtimewire to build specialized AI models that enterprises can train on proprietary data and own, according to The Wall Street Journal https://www.wsj.com/cio-journal/deep-cogito-aims-to-put-companies-in-control-of-their-own-ai-50ddf44b?ref=runtimewire . The Series A, announced Aug. 26, brings Deep Cogito's total funding to more than $56 million. Arora, Deep Cogito's co-founder and chief executive, spent nearly eight years at Google working on the machinery behind question answering in Search and Assistant, followed by Google's large-language-model and generative-search efforts, according to his 3AI biography https://3ai.glueup.com/en/event/natural-language-processing-achieving-human-level-ai-performance-40569/?ref=runtimewire . He studied electrical engineering at IIT Delhi before earning a master's in computer science at Columbia, where his work focused on machine learning and natural-language processing. Malrana came at the problem from the product side. His work spanned Google DeepMind, Google Search and Microsoft Office, according to South Park Commons https://www.southparkcommons.com/companies/deep-cogito/?ref=runtimewire , where he was a Founder Fellow. The pair's shared experience inside mass-market search products helps explain Deep Cogito's enterprise pitch: general models are useful, but companies eventually want intelligence shaped around their own data, costs and operational constraints. TQ Ventures led the financing. Existing backers Benchmark and South Park Commons participated alongside Nexus Venture Partners, Atreides Management and cybersecurity company Zscaler. Deep Cogito previously raised a reported $13 million seed round https://finsmes.com/2025/08/deep-cogito-raises-13m-in-seed-funding.html?ref=runtimewire led by Benchmark. The Series A valuation was not disclosed. From Google's answer box to owned models Deep Cogito started with an ambitious research thesis rather than an enterprise software package. The lab says it is pursuing general superintelligence through advanced reasoning and iterative self-improvement, while releasing model weights that developers can download and modify. Its practical foundation is post-training. Deep Cogito starts from open pretrained checkpoints and applies its own training methods to improve reasoning, tool use and task-specific performance. The first Cogito preview https://www.deepcogito.com/research/cogito-v1-preview?ref=runtimewire , released in April 2025, included models from 3 billion to 70 billion parameters built from Llama and Qwen bases. A July 2025 release https://www.deepcogito.com/research/cogito-v2-preview?ref=runtimewire expanded the family to a 671 billion-parameter mixture-of-experts model. The lab calls its method Iterated Distillation and Amplification, or IDA. A model spends additional computation searching for a stronger solution, then Deep Cogito distills that reasoning process back into the model's parameters. The intended result is a model that reaches better answers with less inference-time searching. Deep Cogito's models can answer directly or switch into a reasoning mode for harder tasks. Its November 2025 Cogito v2.1 release https://www.deepcogito.com/research/cogito-v2-1?ref=runtimewire used a fork of an open-licensed DeepSeek base model, followed by post-training conducted by Deep Cogito. The lab publishes benchmark comparisons while cautioning that those tests do not fully measure real-world performance. That distinction matters for the business Arora and Malrana are building. Deep Cogito is selling the work that happens after a base model exists: tuning intelligence around a company's data, workflows and desired outcomes, then giving the customer control of the resulting weights. Enterprises can deploy the model in a chosen environment instead of routing every sensitive request through a general-purpose API. Ownership still leaves infrastructure dependencies. Deep Cogito has used Together AI https://www.together.ai/customers/deep-cogito?ref=runtimewire for custom GPU clusters, inference and quantization, including training work on H100 and H200 hardware. Together says Cogito models have passed 1 million downloads across Hugging Face and Ollama, a provider-reported adoption figure that combines researchers, developers and other users rather than identified paying enterprises. Zscaler alliance identifies a security use case Zscaler's participation gives the round a more concrete enterprise angle than another set of model benchmarks. In June 2026, Zscaler named Deep Cogito as a technology alliance partner in Project AI-Guardian, an effort to connect security data, identity context and enforcement systems across enterprise AI tools. Arora argued in the Zscaler announcement https://ir.zscaler.com/node/16146/pdf?ref=runtimewire that defending against advanced threats requires specialized models post-trained on a security team's own data and outcomes. Cybersecurity is a favorable test case for Deep Cogito's thesis because threat patterns change, enterprise data is sensitive and defenders need models optimized for their own environments. Zscaler's announcement described a technology alliance and planned integrations rather than a customer contract or deployment. The relationship could make Zscaler a potential enterprise design partner as Deep Cogito develops its approach to specialized post-training. The same model could apply in other domains where generic intelligence runs into proprietary processes, regulation or strict deployment controls. Deep Cogito will face established competitors. Mistral AI https://mistral.ai/news/mistral-3/?ref=runtimewire offers open models and enterprise customization, while Cohere https://cohere.com/private-deployments?ref=runtimewire sells private-cloud, on-premises and air-gapped deployments. Infrastructure providers also let companies fine-tune and serve open models without adopting a separate model lab's research stack. What the $43 million has to prove Deep Cogito said in July 2025 that it spent less than $3.5 million training its first eight models, including data generation and more than 1,000 experiments. That company-reported figure covers its early model-development work rather than the broader cost of building an enterprise business. The larger check gives Arora and Malrana room to move beyond model releases and make specialized post-training repeatable across enterprises. That requires research talent, long-running GPU clusters, evaluation systems tied to customer tasks and software that turns customization into a manageable deployment process. The financing also reflects a widening split in the AI market. Frontier labs are concentrating capital around giant general-purpose systems. Deep Cogito is betting that enterprises will want smaller circles of control: downloadable weights, proprietary training data, specialized behavior and deployment choices that remain theirs. Arora and Malrana have already shown that Deep Cogito can ship models across a wide range of sizes. The harder commercial test begins with this round. Deep Cogito has to turn model ownership from an appealing procurement argument into deployments that outperform general models on the narrow work companies actually pay to improve.