I spent the day taking raw notes at the Cypher AI Summit. The room was packed, the discussions were grounded, and instead of the usual high-level AI hype, speakers actually focused on operating realities, market mechanics, and business models.

Here are the key takeaways, field notes, and open questions I jotted down directly from the some of the keynotes, panels, and debates.

1. Autodesk Keynote: Judgment, Models, and Cost

James Bradley (Autodesk)

• Intelligence is outpacing execution: Access to raw model intelligence is scaling much faster than an organization’s ability to actually act on it.
• The necessity of a decision layer: Adding a distinct decision layer after a model’s prediction is becoming essential. You shouldn’t just take a frontier model’s output as an automatic decision.
• The world model mismatch: Frontier models don’t necessarily evaluate context the way humans do. When frontier models were asked to label comments as toxic, they flagged ~50% as toxic, whereas human ground-truth labellers evaluated only 10% as toxic. Models tend to over-index on risk without local human context.
• The economics of frontier models: Relying purely on a massive frontier model incurred higher operating costs than using a simpler Bag-of-Words pipeline paired with an explicit decision layer.
• The core takeaway: Everyone has access to answers now; raw answers aren’t the scarce resource. The real challenge is making judgment explicit so organizations can learn what actually scales.

2. Panel: The GCC to GDC Shift

Panel discussion featuring Karan Thapar

The shift from Global Capability Centres (GCCs) to Global Delivery Centres (GDCs) is forcing companies to re-evaluate how they operate across four distinct segments:

• The execution layer is solved: India no longer holds a pure “IT arbitrage” or numbers-game advantage. Companies won’t invest purely based on headcount or seat capacity anymore, because routine execution is increasingly handled by AI.
• Leveraging institutional experience: The flip side is that decades of deep IT and software service experience give Indian talent an edge in leveraging AI tools to multiply baseline efficiency.
• The GCC real estate anomaly: Why does GCC real estate vanish as quickly as it appears? Because enterprise headcounts are uncertain. Companies don’t know how many physical seats they will need in 12–18 months. Side thought: Is there an opportunity to build dynamic real-estate modeling tools specifically around AI-driven headcount fluctuations?
• Forward Deployed Engineers (FDEs) and talent: Frontier labs and global enterprises are hunting for FDEs—engineers who can go straight into the field, sit with users, understand domain-specific problems, and deploy custom pipelines on the ground.
• The courage to bet early (The Infosys/OpenAI lesson): A fascinating historical reference came up regarding Infosys missing out on a major early stake in OpenAI back in 2015. A leadership clash between founder vision and traditional governance ended up blocking a $1B bet. It makes you think deeply about the sheer organizational courage required to back emerging paradigms before they are proven.

3. Panel: AI as the Enterprise Operating System

What happens when AI moves from being a simple developer tool to the core operating system of an entire enterprise or GCC?

• Domain understanding is the true moat: The differentiator isn’t raw model access; it’s how deeply you understand a specific business domain and translate it into system prompts and workflow logic.
• Context over data: Agents don’t just need raw data lakes; they require business context, implicit rules, and environmental state to deliver meaningful outcomes.
• Unit economics & decision logic: We need to keep a tight focus on unit economics—the fundamental revenue and cost generated per individual unit of production (e.g., cost per task completed or cost per API query vs. human labor saved). If the decision logic is bloated, the unit economics fall apart instantly.
• The senior talent gap: An insightful question was raised from the floor: If AI completely automates entry-level engineering tasks, how will companies train the next generation of senior domain experts? Without the trial-by-fire of entry-level execution, the traditional talent pipeline breaks.

4. Debates & Quick Definitions

• Horizontal technology: AI is fundamentally horizontal, meaning it acts as a foundational, cross-cutting layer (like electricity or the database) that powers every vertical industry rather than standing alone as a single niche product.
• Corroborate: To confirm, validate, or give support to a statement or finding with additional evidence.
• Cross-operating efficiency numbers: Highlighted during discussions on NITI Aayog’s economic growth reports, referring to metrics that evaluate how efficiently capital, labor, and technology interact across multiple interconnected sectors.

5. Physical AI & Infrastructure Reality

• Systems across layers: The frontier of physical AI involves systems that operate seamlessly across multiple layers—from low-level IoT sensors up to high-level decision engines.
• The need for focus: It’s easy to get distracted by shiny physical AI demos. You have to take a step back and selectively choose the specific domain constraints you want to master.
• Local advantages (e.g., Pharma): Physical AI deployment depends heavily on local market conditions and domain assets. For instance, India’s massive position in global pharmaceutical manufacturing provides a unique, real-world data and testing advantage for physical AI automation.
• Infrastructure is king: Hardware constraints, local latency, specialized compute, and environment design ultimately dictate what kind of physical AI can actually survive deployment.