Products do labs (and labs do products) (10 minute read)
Product companies like Ramp are publishing foundational AI research as they become agent-first platforms, while AI labs like OpenAI and Anthropic are building consumer products, creating a convergence from both directions.
What: Ramp, an AI-native financial services platform valued at $32 billion, has launched Ramp Labs to publish research on topics like KV cache optimization and activation steering—work traditionally done by AI research labs, not fintech companies. Meanwhile, OpenAI and Anthropic announced consumer products like ChatGPT for Clinicians and Claude Design.
Why it matters: As companies build products that are fundamentally agent-based systems consuming billions of tokens, they need to do foundational research on model behavior, memory systems, and inference optimization to serve both their own efficiency and their customers (Ramp reports 13x increase in customer AI token spend since January 2025). This represents a structural shift where the product IS the agent infrastructure, not just a UI over APIs.
Takeaway: If you're building agent-heavy products, consider whether foundational research on model behavior, token efficiency, and agent architecture belongs in your roadmap rather than treating the LLM as a black box.
Deep dive
- Ramp processes $100B+ in transactions across 50,000 companies and has grown from $13B to $32B valuation in 2025, building all internal operations and customer products on agent loops from the ground up
- Ramp Labs publishes applied AI research directly on social media (not peer-reviewed), including work on multi-agent KV cache compaction for memory sharing and activation steering for concept control
- These research topics seem unusual for a fintech company but make sense when tokens become a major operational expense—if your product burns billions of tokens, agent architecture and inference optimization become core business concerns
- Historical precedent exists for companies doing domain-specific research: Google on ranking/retrieval, Netflix on recommendations, Uber/Airbnb on marketplace dynamics, Stripe on payment fraud—but those were adjacent to product features
- Ramp's research is different because the agent infrastructure IS the product layer, not supporting infrastructure, blurring lines between foundational model research and product development
- Ramp customers show 13x increase in monthly AI token spend since January 2025, meaning Ramp's research insights directly apply to their customer base's operational challenges
- The inverse trend: OpenAI launched ChatGPT for Clinicians (specialized consumer product), Anthropic launched Claude Design for prototyping/slides—both moving up the stack into curated applications
- This convergence means product companies move down into foundational research while labs move up into consumer products, meeting somewhere in the middle
- The shift reflects a world where "your code is the LLM plus instructions and an infinite loop" (Ramp CTO) rather than traditional software with AI features bolted on
- Companies doing this research publicly benefit from reputation building while advancing industry knowledge that helps their customers and ecosystem
Decoder
- Agent labs: Companies whose core product is composed of autonomous AI agents rather than traditional software, requiring foundational AI research as part of product development
- KV cache: Key-value cache used in transformer models to store previous token computations, avoiding redundant processing; optimizing this reduces memory usage and improves multi-agent system efficiency
- Activation steering: Technique for controlling model behavior by directly manipulating internal neural network activations at inference time, rather than only using prompts
- Down the stack: Moving from high-level product features toward lower-level infrastructure and foundational systems (in this case, from fintech products to model internals)
- Harness-engineering: Designing the orchestration layer that coordinates multiple agents, manages their interactions, and controls execution flow
- Token: Unit of text processed by language models; companies are "burning" billions of tokens as they run agent systems continuously at scale
Original article
Product companies can do faster, less formal, more product-driven experimentation and research, while labs build products that push and inspire companies to build better, more curated services for consumers.