The session is led by Kate Niedermeyer (AWS), joined by Davis Stone (Arcee) and Arbaaz Khan (Graphon AI), focusing on enterprise AI deployment challenges around cost, control, data access and hallucination. 

Key takeaways

  • Shift from experimentation to production 
    • Generative AI is moving beyond proof-of-concept towards measurable outcomes; solutions must either generate revenue or reduce costs. 
    • A disciplined, workload-level approach is required, starting small and scaling based on demonstrated economic value. 
  • Control as the core enterprise value proposition 
    • Enterprises are increasingly prioritising ownership of models, data and infrastructure to address privacy, compliance and cost concerns. 
    • Deploying models internally converts variable API costs into fixed infrastructure costs and mitigates data leakage risks. 
  • Geopolitics and model sovereignty emerging as investment themes 
    • A perceived dominance of Chinese open models has created demand for domestically developed alternatives, particularly for government and regulated industries. 
    • This creates an opportunity for investors in sovereign or regionally aligned AI infrastructure. 
  • The ‘data problem’ outweighs the ‘model problem’ 
    • Enterprise data volumes far exceed current model context windows, creating a bottleneck in practical AI deployment. 
    • Solutions such as Graphon’s data representation layer aim to unlock value from vast, multimodal datasets. 
  • Hallucination mitigation through context engineering 
    • Hallucinations are linked to excessive or poorly structured input; reducing context to only relevant data materially improves outputs. 
    • Traceability and linking outputs back to source data are becoming key enterprise requirements.
       
  • CVCs as critical go-to-market enablers 
    • Corporate venture investors play a key role in providing early customers and facilitating top-down adoption via CTOs and heads of AI. 
    • Startups benefit from access to business units and real use cases rather than purely financial backing. 
  • Emerging moats: speed, customisation and proprietary data layers 
    • Competitive advantage is shifting towards rapid model iteration, domain-specific tuning and continuous learning loops. 
    • Novel data architectures (e.g. graph-based representations) are positioned as defensible IP in an otherwise commoditising model landscape. 
  • Clear ask for corporate investors 
    • Both companies prioritise strategic partnerships that enable deployment across portfolios and co-development, alongside capital. 

This is an AI-generated summary, which has been lightly edited by GCV staff.