Executive Summary
All three sources point to the same shift. The scarce inputs in enterprise AI are no longer the ones money or LLMs can produce. A hyperscaler cannot buy a state permit. A buyer cannot get accumulated agent working context out of a vendor through a data export. Polished documentation no longer proves that software works. For the next 24 months, the questions that matter are who controls power permits, who holds the agent control plane, and who can prove their product is correct. In each case the advantage goes to whoever controls the hard-to-reproduce layer. Clients who plan around capital, model choice or vendor artifacts will misread where they are exposed.
What Changed
The 2028 GPU supply now depends on state elections. SemiAnalysis treats 2027 as locked: about 75% of the 33.8 GW due that year is already under construction. The exposure is 2028, when 62.2 GW of US additions are mostly unstarted. The 13 states with the largest 2028 additions hold 70% of that capacity, and 10 of them elect a governor on Nov. 3. Nine of the 20 nominees want a construction pause. If every pause candidate wins, about 30 GW is exposed. If the less restrictive candidates win, it falls to 5.7 GW. Texas carries about 30% of 2028 US capacity, and its permits have been frozen since Sept. 21. If the freeze runs through the 2027 legislative session, the delay grows by 5–9 months. Relocating to another state does not fix this, because power, permits and construction still have to meet the same delivery date.
Rival vendors are now contesting one position. Salesforce–Microsoft agent interoperability (Oct 27) and the OpenAI–Oracle workspace are moves to become the agentic control plane: the layer that decides what data an agent sees, which instructions are current, who approves actions and where outputs go. Owning the data no longer protects an incumbent, because Fabric-style in-place access removes the migration barrier.
Vendor artifacts are weaker evidence. Kingsbury's point is that LLMs make polished docs, active commit history and full test suites cheap to produce whether or not the system underneath works. Expect engineers to treat any new software project with default suspicion.
Cross-Expert Synthesis
None of the sources address each other, but they connect in two ways.
Capital loses leverage. Jones cites hyperscaler capex of about $957B this year and more than $1.5T next year. SemiAnalysis shows the limit on that spending: money absorbs conditions-only regimes (paying grid costs, losing tax breaks, water reporting), but it cannot make a state issue a permit it has stopped issuing. The 2028 constraint is capacity slip, not demand or budget.
Lock-in and trust both come down to what is hard to copy. Jones puts durable lock-in in accumulated corrections, instructions, semantic definitions and tool wiring. Kingsbury puts trust in verification that cannot be faked. They are describing the same thing from two sides. The valuable layer is the one that took real work and real failures to build. That also means Kingsbury's discount applies to Jones's own evidence. Salesforce's Agent Context Engine is in preview, the OpenAI–Oracle workspace is unreleased, and Jones cites no deployment metrics. Interoperability announcements are themselves cheap signals.
Enterprise Implications
- 2028 capacity plans have a political risk factor. Clients relying on hyperscaler GPU expansion in Texas, Ohio, Arizona or Georgia (Tier 1) should expect 2028 capacity to arrive late and cost more. This is an inference, not SemiAnalysis's claim: owned or colocated on-prem capacity provisioned now becomes a hedge against that slip, beyond the usual sovereignty argument for private AI.
- The control-plane decision is made by default unless someone makes it. If no one chooses, it lands wherever employees already work. That gives the incumbent ownership of the context layer, which is where switching costs build up.
- Agent context is a corporate asset with no export format. Jones's recruiter example shows that switching providers means rebuilding the working knowledge, not moving the records.
- Vendor due diligence for young AI infrastructure is weaker than it looks. Agent frameworks, vector stores and inference platforms are exactly the categories where documentation now proves little.
- Internally generated docs and tests carry the same discount. When one model writes both the code and its tests, the tests share the code's mistakes.
What To Do About It
- Track the Nov. 3 results by state, not by party. Party is a poor guide: pause and conditions positions split about evenly across both parties. The next dates to watch are ERCOT's Dec. 10 report, the Texas legislative session (Jan. 12–May 31, 2027), and Oregon's permit freeze, which runs through July 1, 2027. For clients with 2028 hyperscaler GPU commitments, ask which region each reservation is in and what the contract says about late delivery.
- Put context portability into procurement contracts. Require that instructions, corrections, semantic definitions and tool wiring can be exported in a usable form, not just the raw records. Treat agent swappability as a hard requirement.
- Assign control-plane placement as an architecture decision with a named owner. Segment by function, since finance and marketing need different controls. Build in-house only if the client can staff it permanently.
- Weight evaluations toward evidence that is expensive to fake. That means adversarial testing (Jepsen-style), benchmarks reproduced in the client's own environment, production references and documented failure modes. Keep verification separate from the tooling that generated the work.
- Customer question: "If you switched agent vendors next year, what would you have to rebuild by hand, and who would know how?" Most clients cannot answer it, and that gap shows where their lock-in is building.
Risks and Blind Spots
SemiAnalysis's 30 GW figure assumes every pause candidate wins, so treat it as an upper bound. Governors' permit actions can be reversed, and utility commissions in Georgia, Arizona and Oklahoma have staggered terms that slow any change of policy. Jones's control-plane thesis is directional and unproven, and a single dominant winner is possible but not likely before 2028. The Kingsbury excerpt states the problem but offers no fix, and adversarial verification costs more than most mid-market clients will pay. That leaves a gap advisors can fill, and one vendors will try to fill with certification claims that are themselves cheap to fake.