The Real Reason AI Agents Are Wasting 10x More Power | Avinash Lakshman, Weilliptic

Summary of the video’s main points

1) AI infrastructure will be commoditized below the “operating system/management” layer

• Avinash Lakshman argues that open-weight LLMs will win and that LLMs will become commodity.
• He compares this shift to specialized hardware becoming standard “dumb” servers, while intelligence moved into the software stack.
• Rather than paying escalating prices for base models, he suggests value will shift to layers that manage outcomes and optimize ROI, including:
  • Running and supervising agent workloads
  • Monitoring, observability, and governance for agent actions
  • Providing an “OS” layer for autonomous systems

2) The “AI bubble” question is unclear, but AI itself won’t disappear

• When asked whether AI infrastructure is in a bubble, the guest avoids a definitive yes/no, saying he’s not sure and that the economics are “nebulous.”
• Even if parts of the market contract (e.g., some model providers), he argues AI will continue as a powerful tool.
• The market hype, in his view, migrates toward higher layers that deliver real operational value.

3) Power and efficiency constraints will push innovation toward agent infrastructure

• A key concern is that the industry is building agentic systems and often implementing them in Python.
• Lakshman claims Python is highly inefficient for large-scale agent deployment compared to systems languages (e.g., Rust).
• He uses back-of-the-envelope reasoning, citing a comparison like:
  • Python “~15W” vs Rust “~0.7W” per agent
• If scaled to millions of agents, power budgets could become untenable—forcing data center expansion and making the approach impractical.
• He links power efficiency to broader constraints like water usage and carbon footprint.
• To make agent scaling feasible, he argues the focus must shift to efficient runtime/sandboxing infrastructure—and points to WebAssembly as an example for sandboxing/guardrails.

4) Circular financing for inference providers may not doom the sector

• The guest addresses concerns about “circular financing,” such as NVIDIA financing GPU purchases that are then used for inference services.
• He argues this isn’t automatically dangerous, comparing it to financing dynamics in other industries (e.g., car purchasing) where evolution still occurs.

5) Inference providers will likely consolidate

• The video describes a fragmented landscape with “hundreds or thousands” of inference providers, including specialized ones.
• Lakshman predicts consolidation, drawing parallels to earlier domains (e.g., search or consumer social media) consolidating into fewer major players.
• He notes inference is B2B in practice, which may shape how consolidation plays out.

6) Why cloud/managed services won’t fully commoditize—yet overlays can add resilience

• The discussion uses an analogy: storage hardware became commodity, and intelligence moved into software and cloud.
• The guest claims cloud is closer to commodity than specialized hardware, but not fully cheap because:
  • Businesses pay for reliability, regional redundancy, and availability
  • Outages can directly translate into lost revenue
• He argues value will appear in overlay solutions that sit on top of multiple cloud providers to handle outages and multi-cloud realities.

7) Elliptic: an “AI governance / agent OS” approach

• The guest explains Elliptic (Weilliptic / “elliptic” in subtitles) as focused on governance and infrastructure for autonomous AI:
  • A “ledger” for AI activity: he claims enterprise AI is often “shadow AI,” where requests, invoked tools, and transmitted data aren’t auditable.
  • He proposes tracking agent/tool actions as immutable audit trails, with transparency similar in spirit to blockchain (even if not implemented identically).
  • Monetization framing: pricing could be tied to audit events, treating each action/tool/agent call as an auditable unit rather than charging only for compute or model calls.
  • Data/compute sovereignty: a design to run cryptographically secure compute in region-specific “pods,” keeping data/compute within jurisdictions and mitigating regional outages via overlays across clouds and data centers.

8) Personal/professional background and prior companies

• The guest recounts system-building work, including Dynamo, Cassandra, and Hedwig, emphasizing that infrastructure building takes long build cycles and requires reliability before revenue scales.
• For Hedwig specifically:
  • He built a hybrid/multi-cloud storage fabric designed to integrate without forcing application changes.
  • He references acquisition around 2019 (Comvault mentioned) as a “good deal,” with benefits for both sides.

---

Presenters / contributors

• Siddhhat Alia (host; founder of Neon Fund / “Neon show”)
• Avinash Lakshman (guest; founder/associated with Weilliptic / elliptic; storage and AI governance background)