Power Becomes the AI Moat, Inference Costs Shift Value to Serving, and Provenance Moves Inward

By DripPublished

The gist

This week, AI competition shifted from model quality to infrastructure control, workflow governance, and trust layers that determine who captures durable enterprise value.

This week’s developments

Power Access Becomes the AI Moat

Power constraints are now the binding AI bottleneck: the limiting factor is shifting from model efficiency to power delivery, permitting, and long-lead electrical equipment. PJM and ERCOT remain repeatedly cited as constrained, Dublin has effectively paused large new grid connections, and since January 2026, 86 projects were canceled or extensively delayed, 76 because of local opposition. Even where capital is available, only about one-third of roughly 12 GW of U.S. capacity scheduled for 2026 was under active construction in one cited assessment, with equipment shortages a major cause.

That changes competitive positioning across the stack. The winners will be operators that can secure land, utility relationships, private generation, and permitting execution alongside GPUs. Value shifts toward hyperscalers, power-system vendors, and infrastructure owners, while capital-light AI entrants face higher execution risk, slower deployment, and a rising effective cost of capacity.

How do we secure power access before competitors do?

If you operate in this industry

  • Power access, not model quality, is now the AI capacity moat.
  • Secure grid, land, permits, and backup generation early or your roadmap slips behind better-connected rivals.

Sources

If you sell into this industry

  • AI demand is moving from compute features to power-delivery execution.
  • Shift GTM toward utilities, EPCs, and infra owners; budget is following power, permitting, and equipment lead times.

Sources

  • Sponsored: All aboard the AI factory powertrain — Data Center Dynamics, November 4, 2026

    Learn how standardized, adaptable powertrain designs and early stakeholder collaboration can reduce time to compute for high-density AI facilities.

If you invest in this industry

  • Capacity winners will be the ones who control power, not just GPUs.
  • Favor hyperscalers and infrastructure owners; capital-light AI names face slower growth and higher execution risk.

Sources

Inference Cost Compression Shifts Value to Serving Infrastructure

Revolut’s reported November move of roughly 2,000 engineers from Claude Code and OpenAI Codex to self-hosted Z.ai GLM models on 56 NVIDIA B200 GPUs is the clearest sign that serving-stack gains are now forcing workload replatforming, not just benchmark wins. The company said blended inference fell to about $0.30 per million tokens, roughly 90% below prior proprietary-API costs, using GLM-5.1 to GLM-5.3 with NVFP4 quantization, context and hyperparameter tuning, and separate prefill and decode paths, while routing vision to Gemini and GPT and keeping coding on GLM.

That extends the cost race beyond one deployment. IBM Research’s llm-d 1.0 reported 5–10x lower per-token costs on open-weight GLM 5.2, including an 11.5x reduction on a quantum-error-correction workload with about 98% prefix-cache reuse and blended cost near $0.11 per million tokens. New tooling is attacking the same bottlenecks: Buoyant’s Kinference added Kubernetes-native routing and KV-cache-aware load balancing on Nov. 8; XorbitsAI’s Xavier added cross-replica prefix-cache sharing on Nov. 7; and Fireworks Forge introduced half-price background mode and multi-day prompt caching on Nov. 3. The strategic shift is clear: value is moving from premium token pricing to routing, caching, orchestration, and deployment flexibility, while API-centric margins compress.

Where should we invest as inference shifts to serving infrastructure?

If you operate in this industry

  • Inference is becoming an infrastructure game, not a model-brand game.
  • Replatform high-volume workloads to self-hosted, cache-heavy serving; keep premium APIs only where multimodal or frontier quality still matters.

Sources

If you sell into this industry

  • Serving-stack efficiency is where budget is moving now.
  • Shift roadmap to routing, KV/prefix caching, quantization, and Kubernetes-native orchestration; sell TCO, not just model access.

Sources

If you invest in this industry

  • API margins are getting squeezed as serving infrastructure captures value.
  • Favor infra and orchestration winners; assume proprietary-token pricing power erodes as open-weight deployment economics keep improving.

Sources

Bounded Autonomy Becomes the Enterprise Workflow Control Layer

Snowflake’s Nov. 3 launch of CoCo/CoWork and CoCo Automations pushed agents into the operating path: background AI coworkers can act inside business tools, route recurring work, switch models dynamically, and pause for human sign-off with audit trails. Box followed on Nov. 5 with Box Architect, extending the same logic into content workflows, including routing high-risk contract clauses to attorneys and orchestrating banking onboarding across forms, extraction, signing, and account setup. Mistral’s Nov. 5 managed agents and workflow engine added coordinated execution and approvals, while Smartsheet’s Nov. 4 agentic workflows turned plain-language requests into recurring autonomous tasks delivered to designated recipients.

The pattern is clear: vendors are moving from AI in the interface to AI in workflow control, but only within bounded autonomy. OECD reporting this week showed agents entering compliance-sensitive processes such as regulatory monitoring and policy conflict detection, while vendors emphasized checkpoints, role-based permissions, observability, and human review. Adoption is broad enough to matter — KPMG said 62% of IT leaders are building or deploying agents — but TCS found only 10% allow cross-functional decisions without human approval and only 9% can securely orchestrate workflows across enterprise applications. The buying criterion is shifting from model quality to safe ownership of routing, approvals, and escalation, where trusted automation and pricing power will concentrate.

Where will workflow control create the next defensible enterprise moat?

If you operate in this industry

  • Workflow control is becoming the moat, not model quality.
  • Build or buy bounded-autonomy orchestration fast; routing, approvals, and auditability now decide enterprise share.

Sources

If you sell into this industry

  • Governed workflow automation is now the enterprise buying trigger.
  • Shift roadmap and messaging to permissions, checkpoints, and observability; model performance alone won’t close deals.

Sources

If you invest in this industry

  • Value is moving to platforms that own safe enterprise execution.
  • Favor vendors with workflow control and compliance depth; point agents without governance look increasingly commoditized.

Sources

Provenance Controls Move Into the Model and File Layer

On November 3, 2026, Anthropic expanded invisible text watermarking to four older Claude models—Mythos 5, Opus 4.7, Opus 4.6, and Sonnet 4.6—and broadened signed C2PA provenance metadata for supported files. The watermark is a statistical token-selection pattern with no user or conversation data, designed to remain detectable after copying and limited editing. Anthropic is framing it as provider-origin detection tied to EU AI Act Article 50, not a user-facing disclosure label.

That distinction matters: detection can suggest Claude processing, but it does not prove authorship, identify the user, or establish an unbroken chain of custody. The control point is shifting from workflow governance to machine-readable traceability embedded in outputs themselves. India’s DEPA launch points the same way through secure processing, immutably logged activity, cryptographic authentication and authorization in audit trails, and purpose-limited consent via consent managers. The U.S., by contrast, is easing disclosure obligations without a single federal architecture for auditability, consent, or end-to-end provenance. For operators, provenance is becoming a deployment requirement in regulated sectors; for vendors and investors, value is moving into model, file, and logging layers that determine market access and enterprise pricing power.

Where does provenance control create the next defensible moat?

If you operate in this industry

  • Provenance is becoming a gate to regulated deployment, not a nice-to-have.
  • If you sell into regulated buyers, build traceability into outputs and logs now or risk losing deals to compliant rivals.

Sources

If you sell into this industry

  • Auditability is moving into the model and file layer buyers will pay for.
  • Shift roadmap and GTM toward provenance, signed outputs, and logging; that's where enterprise budget and switching costs are rising.

Sources

If you invest in this industry

  • Value is migrating to provenance infrastructure, not just model quality.
  • Favor vendors owning model, file, and audit layers; disclosure-light markets still reward traceability where regulation bites.

Sources

Stay ahead in Generative AI

Get the weekly Generative AI brief in your inbox — the developments, what they mean by vantage, and what to do next.