Evidence convention. “Verified” means the statement appears in an official regulator, legislature, company filing, company release, or technical document. A company’s operating metric remains a first-party claim unless a customer, regulator, or audited filing independently confirms it. Reuters is used below as secondary evidence and is labeled accordingly. W-Axis promotion thresholds are research rules, not industry standards.
1. Agent CPU Tollbooth
Current judgment: architectural centrality is verified; durable CPU rent capture is not.
Agentic systems do more than produce tokens. They retrieve data, route requests, maintain state, invoke tools, apply policies, execute code, and coordinate accelerators. That makes the CPU a plausible control plane. A control plane, however, becomes a tollbooth only when technical centrality produces supply leverage, binding contracts, pricing power, and realized cash flow.
What first-party evidence establishes
- Intel’s first-quarter 2026 release states that the shift from foundation models to inference and agentic AI is “significantly increasing” the need for Intel CPUs, wafers, and advanced packaging. This is an attributed management statement, not an independently measured decomposition of demand. Intel Q1 2026 release.
- Intel reported second-quarter revenue of $16.1 billion and 59 percent year-over-year growth in Data Center and AI revenue. It also described production-ready rack-scale infrastructure for inference and agentic workloads. The segment includes more than agentic workloads, so the 59 percent figure cannot be assigned to agents alone. Intel Q2 2026 release.
- Intel and SambaNova announced a signed agreement for a heterogeneous design using Xeon 6 as host and action CPUs, with enterprise availability expected in the second half of 2026. That verifies a product architecture and partnership, not customer utilization or profit. Intel–SambaNova.
- Microsoft says CPU infrastructure is essential for data preparation, search, reinforcement learning, and agent coordination at scale, and introduced upcoming Azure virtual machines for those workloads. This is independent architectural evidence from a hyperscaler, but the offerings are not disclosed as proof of a CPU shortage. Microsoft Azure and AMD.
- AWS documents a production pattern in which CPU-hosted small models handle narrow agent tasks and a CPU routing layer escalates complex cases to a GPU-hosted model. It is a recommended pattern, not evidence that most deployed agents use this architecture. AWS EKS best practices.
- Dell and HPE have announced servers built around NVIDIA Vera CPUs for agentic data processing and orchestration. Their announced availability dates extend into late 2026 or 2027, so they are product-direction evidence rather than installed-base revenue evidence. Dell; HPE.
What remains secondary or unproven
Reuters reported on July 23 that Intel and AMD were entering longer-term server-CPU supply agreements with Chinese customers amid higher prices and longer lead times. The report relies on sources and does not disclose named contracts or complete terms. Intel’s official quarterly release does not independently confirm those specific CPU agreements. The claim should therefore remain secondary evidence, not be restated as a company-verified fact. Reuters.
Promotion gates
- Two consecutive quarters of disclosed server-CPU unit, average-selling-price, backlog, or capacity growth that cannot be explained only by memory, substrate, or yield constraints.
- A named customer and a first-party contract of at least twelve months with a minimum volume, prepayment, or take-or-pay feature.
- Production telemetry showing that CPU core-hours per completed agent workflow remain material after model and routing efficiency improve.
- Evidence that technical centrality reaches supplier margin or free cash flow rather than being competed away by x86 rivals, Arm, DPUs, GPUs, or custom silicon.
Failure conditions
- Agent orchestration migrates into accelerators, DPUs, or hyperscaler custom silicon faster than workload growth expands general-purpose CPU demand.
- Longer lead times are primarily a temporary manufacturing or memory bottleneck with no durable pricing power.
- Model efficiency, caching, and task compression reduce CPU work per useful agent action faster than agent volume grows.
2. Ratepayer Covenant Underwriting
Current judgment: this is no longer an observation seed. Formal mechanisms in at least five states have moved large-load risk allocation into tariffs, service agreements, fees, and separate customer classes.
The visible story is data-center load growth. The control layer is the covenant that determines who pays when forecast demand is delayed, reduced, canceled, or served by infrastructure that cannot be repurposed.
Ohio
The Ohio Consumers’ Counsel, a state agency, says the approved AEP Ohio data-center tariff requires new large data centers to pay for at least 85 percent of contracted capacity for up to twelve years. Ohio Consumers’ Counsel.
Virginia
The State Corporation Commission created the GS-5 class for customers at or above 25 megawatts. Its public guidance describes a fourteen-year minimum obligation for qualifying new customers, monthly payment of at least 85 percent of transmission and distribution costs, and possible collateral covering up to 60 percent of minimum charges for customers without sufficient credit. The Commission’s November 2025 release separately describes a 60 percent generation-demand minimum. Some provisions begin January 1, 2027. Virginia SCC facts; Virginia SCC order summary.
Georgia
The Georgia Public Service Commission approved a rule allowing terms beyond standard service for new customers above 100 megawatts. Its official summary says those customers can be charged site-specific costs and upstream generation, transmission, and distribution costs as construction progresses. Georgia PSC.
Oregon
The Oregon governor’s office says the first updated rate proposal under the POWER Act raised average rates for data centers and other large PGE users by 29 percent from July 8, 2026, while lowering rates for other PGE customers. The release describes the Act’s purpose as preventing large-user grid costs from being shifted to households and small businesses. Oregon governor’s office.
Minnesota
Minnesota law authorizes a very-large-customer class and review of electric-service agreements. The state’s legislative summary says the Commission must prevent costs and stranded costs attributable to very large customers from being paid by other customers. A separate statutory fee ranges from $2 million to $5 million annually according to arranged peak demand. Minnesota House summary; Minnesota Statutes §216B.72.
Ratepayer Covenant Score · 100 points
Revenue certainty · 35
20: minimum bill or take-or-pay percentage.
15: contract duration relative to the life of dedicated infrastructure.
Credit and exit protection · 25
15: collateral, guarantee, or other credit support.
10: termination payment and explicit stranded-cost protection.
Cost causation · 25
15: prepayment or direct funding of dedicated and upstream facilities.
10: separate rate class and transparent cost allocation.
Execution and public value · 15
5: load-ramp milestones, true-ups, and reporting.
10: curtailment, flexibility, reliability, environmental, or community obligations.
Research identity: net ratepayer exposure = dedicated utility capital expenditure + allocated upstream cost − prepayment − collateral − present value of enforceable minimum charges − termination recovery.
Public documents may omit confidential terms. The identity should not be populated with assumed values.
Validation gates
- Build a comparable five-state matrix using final tariffs, orders, and statutory text rather than press-release language alone.
- Separate signed and credit-supported load from interconnection requests and speculative development pipelines.
- Track whether covenant protection covers only distribution facilities or also generation and regional transmission exposure.
- Test whether utilities with stronger covenants show lower unrecovered-capital risk, lower political backlash, or better financing terms.
Failure conditions
- Confidential side agreements, waivers, or weak credit tests materially dilute the public tariff.
- Major costs remain regionalized through wholesale transmission or capacity markets outside the retail covenant.
- Courts, legislatures, or commissions reverse the protections before the associated infrastructure enters service.
- The score predicts formal contract language but fails to explain actual recovery after cancellations or load shortfalls.
3. High-Mix Robotics Control Layer
Current judgment: the category has a credible production wedge. Pivot Robotics has not publicly crossed the scale, recurring-revenue, or autonomy gates.
High-mix manufacturing is difficult for fixed automation because parts vary and programming time can exceed production time. The potential control layer is not the robot body. It is the perception, path-generation, adaptive-control, inspection, exception-handling, and fleet-data stack that turns variable parts into repeatable production.
Pivot: what the company claims
- Pivot’s current website says its systems process more than 5,000 parts per week at major North American foundries, achieve 1 millimeter precision on cast-iron components, offer a return on investment of less than two years, and add a new part up to 100 times faster than traditional programming. These are first-party marketing claims; the page does not publish an independent audit or the methodology behind each figure. Pivot Robotics.
- Y Combinator’s company profile says Pivot was deploying its software on more than ten robots in a cast-iron foundry. The profile supports an early production wedge, but it does not provide a current fleet count, recurring revenue, or intervention rate. Y Combinator.
- Pivot’s blog index advertises an OSCO Industries case study with an 83 percent cycle-time reduction and 1 millimeter precision. The linked case-study page currently says “Coming Soon” and contains no supporting data. The 83 percent figure should therefore remain a first-party teaser, not a validated result. Pivot blog index; case-study placeholder.
Comparable category evidence
Path Robotics publishes named customer cases in which TYCROP expanded from a single-robot to a dual-robot cell and added a larger second cell, while Mine Rite reports 85 percent of weld scope automated and a 26 percent improvement in lead time. These cases strengthen the category thesis, but they remain vendor-published customer evidence rather than audited cross-company data. TYCROP; Mine Rite.
The actual control-layer variables
Perception and planning
Time to onboard a new geometry, requirement for CAD or custom fixtures, path-generation time, and tolerance across natural part variation.
Autonomy and quality
Human interventions per 100 cycles, operational availability, first-pass yield, rework, scrap, and closed-loop inspection.
Replication economics
Install time, integration labor, payback, repeat orders, revenue per deployed cell, and whether support headcount scales more slowly than the fleet.
Data compounding
Unique geometries processed, reusable policies across customers, exception-data capture, fleet learning, and the customer’s rights over production data.
Promotion gates
- Scale: more than 100 paying production units across at least ten active customer sites.
- Autonomy: customer-confirmed operational availability of at least 85 percent and no more than five human interventions per 100 cycles for the measured process.
- Quality: first-pass yield at or above the relevant human or fixed-automation baseline, with scrap and rework disclosed.
- Economics: customer-confirmed payback within twenty-four months and repeat expansion by at least 40 percent of mature customers.
- Business model: disclosed recurring software or service revenue that grows with the installed base rather than with custom engineering headcount.
- Transfer: demonstrated production use across at least three processes and two industries without rebuilding the control stack for every site.
Failure conditions
- Each deployment remains a bespoke integration project and gross margin does not improve with repetition.
- Human exception handling or remote support grows linearly with installed units.
- Customer pilots do not become repeat orders, multi-cell expansions, or independently validated production throughput.
- Revenue remains dominated by one-time hardware and installation while the software layer lacks pricing power.
- Safety, liability, or quality failures prevent the system from moving beyond supervised operation.
Do not flatten different evidence states
These theses should not be published as three equally mature opportunities. Ratepayer Covenant Underwriting has crossed its policy trigger and belongs in active P1 validation. Agent CPU Tollbooth has crossed the architecture gate and also belongs in P1, but its contractual and margin gates remain open. High-Mix Robotics remains on watch until production scale, intervention, repeat purchase, and recurring economics become visible.
The discipline is the asset: promote a thesis only when the evidence changes, and retire it when the control layer fails to compound.