The Intel spinout building domain-specific generative AI for the industries general models keep getting wrong.
In a market where nearly every startup claims its AI "simply works," Articul8 AI made a narrower, harder wager: build generative AI that holds up when the answer has to be exactly right - inside a power grid, a semiconductor fab, or a bank's compliance desk. Spun out of Intel in January 2024 and headquartered in Santa Clara, California, the company builds domain-specific GenAI systems that run on a customer's own proprietary data rather than on shared, general-purpose models.
The premise is simple to state and difficult to execute. General chatbots are trained to be broadly useful; regulated industries need AI that is narrowly, verifiably accurate and that leaves a trail auditors can follow. Articul8 sells to energy, manufacturing, aerospace and defense, semiconductors, and financial services - sectors where a wrong answer is expensive and an unexplained one is unacceptable.
Founder and CEO Arun Subramaniyan, a former vice president and general manager in Intel's Data Center and AI Group, began his career designing jet engines and rocket systems. That background shows in the company's posture: less interested in demos, more interested in measurable accuracy on tasks that matter to engineers. The approach has attracted 29 paying customers and, by early 2026, a valuation north of $500 million.
Most enterprises sit on decades of proprietary data - engineering documents, sensor logs, maintenance records, regulatory filings - that general models were never trained on and cannot safely reason about. Articul8's answer is a full-stack platform that turns that data into what the company calls expert-level generative AI, deployed inside the customer's environment so sensitive information never leaves it.
The differentiator is specificity plus control. Where competitors offer one large model for everything, Articul8 builds families of smaller, domain-tuned models and wraps them in an orchestration layer that checks its own work. The result, the company argues, is higher accuracy, lower cost, and outputs that a compliance officer can actually trace.
"ModelMesh acts as an Agent-of-Agents, enabling our customers to optimize for multiple KPIs."
Articul8's architecture pairs a reasoning engine with a growing library of domain-specific models. ModelMesh decides which model to trust and what to do next; the A8 model families supply the expertise.
An autonomous "agent of agents" that combines Bayesian systems with specialized language models to verify outputs, choose next actions, and hold complex industrial workflows together.
A model evaluation and dynamic routing layer that scores tasks and sends each request to the best-fit model instead of forcing one model to do everything.
Domain-specific models for chip engineering, reported to write Verilog roughly twice as accurately as leading models while running 50-100x smaller.
Built with EPRI and NVIDIA for grid optimization, predictive maintenance, and equipment reliability across the power sector.
Described as the first domain-specific GenAI family for manufacturing supply chains, reporting ~92% accuracy on industrial reasoning tasks.
A product that lets enterprises rapidly explore complex proprietary data to surface insights and unlock value without heavy setup.
Articul8 competes in the space between general-purpose foundation-model providers - OpenAI, Anthropic, Google - and enterprise AI platforms such as Databricks, Palantir, Cohere and Scale AI, as well as the in-house builds many large companies attempt themselves. Its argument for existing is that the biggest models are optimized for breadth, while regulated industries need depth, control, and auditability.
That positioning has drawn notable validation. Articul8 was named the foundational GenAI provider for EPRI's Open Power AI Consortium alongside NVIDIA, selected as a launch partner for Google's Agent-to-Agent interoperability protocol, and earned AWS's Generative AI Competency and Agentic AI Specialization. It was also chosen for the exclusive Meta and AWS "Startups: Building with Llama" program.
Subramaniyan builds domain-specific AI models within Intel's Data Center and AI Group - including a pre-ChatGPT model code-named "LLaMa," roughly nine months before Meta shipped its own.
Launched as an independent company in January 2024 with Series A backing from Intel Capital and DigitalBridge at roughly a $100M valuation.
Ships A8-SupplyChain, A8-Energy and A8 Essential, and announces ModelMesh Dock & InterLock alongside a Google Agent-to-Agent partnership.
Raises ~$70M led by Adara Ventures and completes the round via a strategic partnership with a global industrial software leader, crossing a $500M valuation - about 5x its Series A.
It builds domain-specific generative AI platforms and models for regulated, industrial enterprises, deployed on the customer's own data and infrastructure with full auditability.
It was founded by Arun Subramaniyan and spun out of Intel as an independent company in January 2024.
ModelMesh is Articul8's autonomous "agent of agents" engine that orchestrates specialized models, verifies outputs, and decides next actions across complex workflows.
It raised a Series A in 2024 and a ~$70M Series B in 2026 led by Adara Ventures, surpassing a $500 million valuation.
Reported customers include Hitachi Energy, Franklin Templeton, AWS and Intel - around 29 paying customers and over $90M in contract value.
Looking for video? Search "Articul8 ModelMesh demo" and "Arun Subramaniyan Articul8 interview" on YouTube for product walk-throughs and founder conversations.