Control planes move into the workday, power rationing reshapes AI buildouts, and sovereign compute races intensify

By DripPublished

The gist

This week, generative AI shifted from model competition to control, power, and proof: who owns execution, who gets compute, and who can show ROI.

This week’s developments

ServiceNow’s Moveworks Deal Extends the Control Plane Into the Workday

ServiceNow’s roughly $2.85B Moveworks acquisition is the newest sign that control is now being consolidated inside the systems where work actually happens, tying agent execution more tightly to identity, access, and workflow ownership. That comes as multi-model production has moved from architecture choice to buying behavior: 56% of enterprise builders now use multiple foundation-model providers in production, versus 15% relying on one, and hybrid control-plane architectures are expected at roughly 35–36%. The result is a concentration of enterprise power in platforms that control routing, permissions, approvals, and auditability—ServiceNow, Salesforce, Workday, Microsoft, SAP, Oracle, and intermediaries like OpenRouter—while OpenAI and Anthropic remain critical suppliers rather than the sole system of record.

The economics now match the architecture. Stanford’s AI Index showed a more than 280x decline in inference cost for GPT-3.5-level performance between November 2022 and October 2024, while Ramp data across 70,000+ U.S. businesses showed token volume up about 7x as spend rose only about 5.5% from January to September. Because agentic workflows consume roughly 50–500x more tokens than simple chat, margin shifts to routing, memory, retries, monitoring, and stopping logic; Writer reported orchestration cuts of 38% fewer tokens per task and 41% lower cost at equal quality. For practitioners, this extends the prior control-plane story: the winning stack is the one that owns workflow execution and the economics of every step inside it.

Who will own workflow control as multi-model adoption scales?

If you operate in this industry

  • Control is moving into workflow owners, not model providers.
  • Build around identity, approvals, and auditability—or risk being routed through someone else’s control plane.

Sources

If you sell into this industry

  • Governance and orchestration are now the product, not add-ons.
  • Shift roadmap and GTM toward routing, memory, and audit controls; buyers will fund cost and control, not raw model access.

Sources

If you invest in this industry

  • Value is concentrating in platforms that own execution and spend.
  • Favor control-plane incumbents and orchestration layers; standalone agents and model wrappers face bundling pressure.

Sources

ERCOT Starts Rationing AI Load Access as Financing Follows the Queue

ERCOT’s audit of an approximately 474 GW interconnection queue and Texas’s freeze on new large-load connections turned grid delay into market rationing: capacity is no longer just late, it is being allocated. With about 90% of the queue tied to data centers, ERCOT’s delayed Batch Zero process and December audit timeline put roughly 20% of the U.S. data-center pipeline at risk. PJM’s suspended procurement, driven by a 6.8 GW reliability shortfall, now pushes relief to Feb. 28, 2027, while Northern Virginia, Phoenix, and Dallas remain on 4–7 year timelines. Even where power eventually clears, liquid-cooling readiness and site access are stretching 100–750 MW AI campus deployments.

The response has shifted from buying GPUs to financing guaranteed access to energized, coolable capacity. Nvidia has expanded that model with about $36B of guarantees for neocloud obligations, about $20B in future data-center leases for reassignment, and about $3.5B backing cloud-provider leases; it has also been reported as able to guarantee up to $105B tied to OpenAI’s proposed 10 GW Ohio campus. Hyperscalers are locking in supply through long-duration commitments, while SPVs, sale-and-leasebacks, and $11B of U.S. data-center CMBS/ABS by mid-June 2026 extend the financing stack. Competitive advantage is now moving one step further, to balance-sheet strength and pre-committed infrastructure that can actually clear the queue, not model quality alone.

How should we secure financed power access before queue rationing deepens?

If you operate in this industry

  • Power access, not model quality, is becoming the real moat.
  • Lock in energized capacity and financing early; queue position now matters as much as GPU supply for scaling share.

Sources

If you sell into this industry

  • Demand is shifting to financed capacity, not just hardware sales.
  • Sell into SPVs, leases, and grid-ready campuses; roadmap for liquid cooling and financing support, not just product specs.

Sources

  • Navigating data centers' power-price conundrum — Data Center Dynamics, October 26, 2026

    Explains how load forecasts, competitive bidding, and flexible pricing help data center operators manage power costs and contract risk.

  • Inside an AI Factory – Power, Cooling, and GPUs Explained — freeCodeCamp.org, October 16, 2026

    Learn how operators use site selection, staged builds, on-site generation, and cooling choices to bring AI capacity online despite grid bottlenecks.

  • AI Physical Infrastructure Software — Mergers And Acquisitions Newsletter™, October 6, 2026

    See how startups use AI software to optimize data-center cooling, energy projects, and power operations as infrastructure constraints intensify.

If you invest in this industry

  • Infrastructure control is overtaking model hype as the winner's edge.
  • Favor balance-sheet-heavy platforms and infra financiers; queue-constrained AI capacity is now a capital allocation game.

Sources

Survey Data Exposes the Productivity Gap Behind AI Spend

A 2026 survey of nearly 6,000 executives sharpened the credibility gap behind the workflow ROI story: about two-thirds of firms now use AI, yet 89% reported no measurable labor-productivity improvement over the past three years. That tracks McKinsey’s finding that 88% of organizations use AI regularly but only 39% see enterprise-level EBIT impact, while just 5% of pilots produced measurable bottom-line results. The issue is no longer adoption; it is whether scaled usage converts into finance-visible gains.

That makes proof itself the market filter. Enterprises are prioritizing deployments tied to time to first value, workflow adoption, error reduction, cycle-time improvement, and labor or cost outcomes finance can verify. The weakest results remain in open-ended judgment work, COBOL-heavy maintenance, and cross-functional workflows where local gains disappear at handoffs. Gartner’s estimate that more than 50% of generative-AI projects will be abandoned after proof of concept by end-2025 fits the same pattern: poor data, weak controls, rising costs, and unclear business value.

For operators, the bar is now moving from deployment to instrumented workflow redesign. For vendors and investors, the next advantage belongs to companies that can bundle implementation, governance, and ROI measurement into repeatable outcome contracts, extending the proof standards that were already starting to shape buying decisions.

How do you prove EBIT impact before buyers lose patience?

If you operate in this industry

  • Adoption no longer wins; measurable workflow ROI does.
  • Shift from pilots to instrumented redesign, or your AI spend will stay invisible to finance and vulnerable to cuts.

Sources

If you sell into this industry

  • Buyers now pay for proof, not promises.
  • Bundle governance, implementation, and ROI measurement into repeatable outcome offers, or lose deals to safer incumbents.

Sources

If you invest in this industry

  • AI winners are shifting toward firms that can prove EBIT impact.
  • Favor vendors with measurable outcomes and low-abandonment deployments; point tools without proof are getting repriced.

Sovereign AI Is Becoming a Regional Infrastructure Race

The UK is turning sovereign AI into a regulated infrastructure category, with planning and power policy now shaping where compute can be built and who can control it. AI Growth Zones are designed to cut consent times from about 18 months to 12, while the October 2026 Data Centre Covenant will let strategic projects reserve and reallocate grid capacity.

That shifts competition away from generic cloud scale toward local control over data, model hosting, and governance. The immediate beneficiaries are large, region-specific builds such as Culham’s proposed 100 MW site scalable to 500 MW, SEGRO–Pure’s 56 MW West London campus, and Colt’s planned 97 MW Hayes expansion. For operators and investors, the value is moving to developers that can secure power, planning, and policy alignment in constrained markets; for vendors, the prize is infrastructure that can satisfy sovereignty requirements without sacrificing deployment speed.

Where should we position for sovereign AI infrastructure advantage?

If you operate in this industry

  • Sovereignty is now a site-selection advantage, not just a compliance issue.
  • Secure power, planning, and local governance early or risk losing capacity to regionally anchored rivals.

Sources

If you sell into this industry

  • Sovereign-ready infrastructure is becoming the new enterprise buying filter.
  • Build for local hosting, auditability, and fast deployment; budget is shifting to compliant regional stacks.

Sources

If you invest in this industry

  • Compute access is turning into a regulated land-and-power race.
  • Favor developers with grid, planning, and policy edge; sovereign AI demand may concentrate in scarce regional assets.

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.