Oct 8, 2025, 8:58 AM ETPanel, Conference Presentation

WEKA, Intel, Graphite, InstaDeep & AWS: Scaling Without Crashing. AI Framework for Enterprise Growth

2023 was characterized by creating experiments and proofs of concept to derive initial value from AI.
Last year focused on moving AI from experiments into production, a process that remained largely manual.
Current focus (2024-2025) has shifted to scaling AI through the deployment of autonomous agents.
Agents are defined not just as API calls, but as systems requiring monitoring, observability, hallucination handling, and context engineering.
Adoption status: AI is no longer optional for businesses; companies must integrate it to remain competitive, regardless of industry maturity.

Cost pressure: High expectations and operational costs are significant barriers, with unit economics becoming a primary focus for AI companies.
Pricing wars: A consensus is emerging to reduce cached input pricing by approximately 75% relative to standard input pricing.
Market shift: Approximately 80% (4 out of 5) of AI dollars are projected to move toward inference rather than training.
Model economics: Companies like Cursor faced backlash and had to revise terms due to unsustainable token economics, signaling a broader industry "Uber moment" for AI pricing models.
Future compute cost: The speaker predicts that compute cost competitiveness will become the deciding factor between winners and losers in the AGI era.

Transformation vs. Initiative: AI should be integrated as a central enabler within broader transformation programs (driven by regulation, transactions, or digitalization) rather than treated as a standalone technology project.
Trust and Governance: Success requires a top-level (ex-co) AI framework to ensure data transparency, regulatory compliance, and ethical decision-making.
The "Move 37" Concept: Companies must establish a virtuous loop of data, domain expertise, and simulation to allow AI to discover novel insights faster than human capability alone (analogous to AlphaZero's rapid evolution).
SPARC Framework: A recommended approach for agent swarms involving Specification, initial pseudocode, Architectural planning, Reflection/Looping, and Completeness (security/compliance).
Context Engineering: A major challenge is aggregating disparate data sources (Slack, Jira, Figma) into a unified context for agents, creating new security risks regarding data visibility.

Mid-sized Enterprises (€250M – €1B revenue): These companies face unique challenges regarding skills acquisition and must define AI's role in transformation to maintain trust among employees.
Biotech and Health:
- Progress is described as being on the "verge" of capabilities far exceeding current LLMs, moving toward agentic multi-modal analysis.
- Microsoft agents recently demonstrated diagnostic capabilities that significantly outperformed human doctors.
- Applications include personalized cancer vaccines and next-generation genomics, utilizing "zero-copy" data strategies to handle massive datasets.
Software Engineering:
- The developer role is shifting from line-by-line coding to prompting and reviewing; the Pull Request is becoming the central interface of the developer's universe.
- Graphite addresses second-order effects where code volume increases 3-10x, requiring AI-driven code review agents to ensure quality and security at scale.
- Tools like Cursor and Windsurf are viewed as complementary to code review platforms, forming an end-to-end workflow.

Open Interoperability: Intel advocates for open standards (e.g., Unified Acceleration Foundation) to prevent hardware vendor locking, allowing developers to focus on software regardless of the underlying engine (CPU, GPU, FPGA).
Small Language Models (SLMs): A strategic pivot toward SLMs for edge deployment and specific tasks to improve cost-effectiveness and reduce latency.
Storage and Memory:
- Performance bottlenecks are shifting from compute to memory and storage latency.
- Weka's "Neural Mesh Axon": A new memory product designed to accelerate the inference market by reducing latency for model loading and caching.
- Zero-copy approach: Essential for managing exabytes of data efficiently without excessive data center real estate or power consumption.

AGI Reality: Speakers argue that AGI is economically happening now, with AI models nearing 140 IQ points, necessitating immediate strategic adaptation.
Talent Evolution: The volume of software production will explode; developers will need deeper domain expertise or broader product/design skills rather than just coding proficiency.
Risk Mitigation: While hallucinations are a concern, the primary risk in enterprise scaling is security by obscurity breaking down as agents aggregate previously siloed data.
Regulatory Drivers: New regulations (e.g., mandatory electronic invoicing in France) are cited as catalysts for rethinking operating models and accelerating AI adoption.