The walkthrough, annotated

The counter-signal I have to carry into the room

The regulator is already running its own tool. On June 25, 2026, NRC's chief data officer said AI had shortened some licensing reviews that once took four years to as little as nine months, using an internal tool called SimplifAI built on Azure OpenAI. NRC has also issued a strategic plan for reviewing applications that themselves used AI, and expects such applications within a few years.

This cuts two ways. It weakens "drafting isn't review" as a blanket objection: review time is falling, and the regulator says AI is why. It strengthens the walk-away condition on slide 3: the regulator having its own copy is no longer hypothetical, it's the state of play, so a vendor whose roadmap depends on the regulator not having one is already behind the facts.

What's behind a website

The method for any company that appears on this desk. The marketing page is a claim; these are the record.

What roles are being hired for, and what each one tells

Read job postings as the company saying, in its own words, what it can't yet do.

• Evaluation. The scarce skill across labs in 2026 is designing evals that catch real regressions; it separates candidates more than model familiarity does. A lab hiring evaluators is a lab that doesn't trust its own benchmarks. That's good.
• Forward-deployed engineers. Every posting for one is an admission that the product doesn't work without a person inside the customer. Growing, constrained by pool size. For the desk: forward-deployed headcount is a lagging measure of how far the technology is from the marketing.

Who is talking about it, and how

The same fact set is discussed in different rooms with different vocabularies, and the rooms don't read each other.

• Power and permits is discussed by utility planners, state commission staff, LBNL and Belfer, data-center REITs in earnings calls, and grid-policy shops. The vocabulary is megawatts, queue position, cost allocation, show-cause. Almost nobody from AI governance is in those dockets.
• AI governance is discussed by labs, think tanks, and state legislators. The vocabulary is frontier models, transparency, safety frameworks, preemption. Almost nobody from a state commission is in those rooms.
• Implants are discussed by neurosurgeons, FDA, and patient communities, in the vocabulary of feasibility studies, adverse events, and PMA. Consumer-tech coverage borrows the words and drops the standing.

The role is to be the person who reads all three and can say the same fact in each dialect. That's slide 3.

Reviewing comments not from bots

Where human commentary with standing actually lives, and how to tell it from generated volume.

• FERC dockets. RM26-4, the large-load rulemaking, drew hundreds of public comments before FERC chose show-cause orders over a final rule. Comments from RTOs, utilities, and data-center developers are signed, costly to write, and reveal position.
• NRC public meeting transcripts and ADAMS. Slow, dense, and human. The reviewer who said "Revision 0" is in there somewhere.
• Federal Register comment periods. Same property: a signature and a cost.
• GitHub issues on open-weight repos. Maintainers arguing with users about a release is the closest thing to a live safety debate with names attached.
• Researcher Discords and seminar Q&A. The person who asked the unfriendly question.
• The test. A comment counts if it names a specific document, costs its author something, and comes from someone with standing to be wrong. Volume, sentiment, and reach are not evidence of anything except distribution.

Writing for models, speaking for people

Two registers, and the desk should keep them separate on purpose.

• The written record is now read by models as well as people. Anything I want retrieved later gets structure: dated, primary-linked, status-marked, one claim per line. That's why the desk looks the way it does.
• The room is human. Slides carry claims, not structure. The sentence to argue with is a sentence, not a table.
• The counter I should hold: writing for retrieval is writing for extraction. Sources who talk to me should know which register their words will land in, and the podcast protocol's last question, "on the record: observed, inferred, contested, speculative", is the mechanism for that.

Stated framework against the record

Not a character judgment. A comparison of two documents: what a leader said the company would do, and what the company's own record shows it did. Inference throughout; the record is public.

The honest version: I don't have the inside view on any of these. What I have is whether the outside documents agree with each other.

Reading a system card

Threat model, eval methodology, what was excluded, and whether mitigations are pre- or post-deployment.

Sources. DOE Office of Nuclear Energy (March 25, 2026); NRC reporting and public AI-review materials (June 2026); NRC strategic plan for reviewing AI-assisted applications; FERC RM26-4; EDGAR/Form D; company documentation, system cards, and GitHub records.

Five companies, and what each one is evidence of

Who is acquiring, and what they're buying

• AMD acquired Taalas, August 6 2026: inference silicon and dataflow optimization, into the Instinct line.
• OpenAI bought Promptfoo and Astral. Anthropic bought Vercept and Bun. Google completed its $32B acquisition of Wiz on March 11, 2026.
• Between March 2024 and January 2026, Google, Microsoft, Amazon, and Meta spent over $20B hiring away founding teams without acquiring a company. Partial-acquisition structures dominate frontier consolidation through the first half of 2026; Nvidia, OpenAI, Salesforce, ServiceNow, Datadog, and Snowflake are the most active outright acquirers.

The read: the acquirers are buying evaluation tooling, developer infrastructure, inference silicon, and security. They are buying the means of judging and running models, not models. The acqui-hire structure can consolidate a field in a form antitrust review reaches unevenly. The FTC examination of Microsoft and Inflection is the counterexample worth keeping in view. That belongs on the portfolio map because the structure changes who holds the people and the work even when it is not a conventional acquisition.

Fifteen companies to watch, by domain

Three per domain. Each with the one thing I'd watch, not a profile.

Nuclear and advanced energy

1. Everstar (Gordian). Whether the secure pilot with approved safety analyses names NRC staff as participants.
2. Endurance Energy. Which agency claims jurisdiction for a seafloor plant, and when the first permit application is filed.
3. Commonwealth Fusion Systems. Google has invested and agreed to buy power from the first commercial plant. Watch the interconnection request, not the physics.

Brain-computer interfaces 4. Synchron. Likely first premarket approval application for a permanently implanted communication BCI; COMMAND study had six patients and zero serious adverse events at twelve months. Watch the post-market surveillance plan in the PMA. 5. Precision Neuroscience. FDA 510(k) clearance April 2025 for up to 30 days of intraoperative recording; first peer-reviewed clinical data published. Watch whether the temporary clearance becomes the wedge for a permanent one. 6. Neuralink. About 21 participants implanted by early 2026 across four countries; no PMA expected before 2027 or 2028. Watch what the informed-consent language says about aftercare after the study ends.

Frontier AI 7. Goodfire. Whether the life-sciences partnerships produce a claim a regulator accepts. 8. Adaption Labs. Who joins from which lab. The hiring list is the thesis. 9. Axiom Math. Whether a machine-checked proof gets cited by someone who isn't the company.

Adjacent and emerging 10. Cerebras. First quarterly reports; what the concentration of revenue looks like, and whether the G42 relationship reappears in the filings. 11. Taalas, inside AMD. Whether inference-dataflow optimization ships in Instinct or disappears. Acquisitions of this shape often do. 12. A neuromorphic company with a customer. I don't have one I can name with a paying deployment. That absence is the finding.

Outside AI 13. Promptfoo, inside OpenAI. Evaluation as an acquired product. Whether it stays open-source. 14. Wiz, inside Google. $32B says security is the constraint on enterprise deployment. Watch the cloud-security posture become the de facto AI deployment standard. 15. Digital Realty or Equinix. The rate-case counterparty. Their filings in Virginia and Texas are where the public cost lands.

The next frontier, and what we're not looking at

What the five companies above have in common: none of them is a bigger model. Continuous learning, formal proof, interpretability as product, inference cost, and firm power from a new source. If the next frontier is anywhere, it's in the layer that decides whether a model can be trusted, afforded, and powered, not in the model.

What the rooms aren't looking at, as a list I'd keep open:

• Insurance as the regulator. Who underwrites an AI-drafted safety case, a BCI implant, or a data center's power contract. Underwriters write the real standard.
• Procurement language. How "open" and "safe" get defined in government and enterprise purchasing terms, which is where the definition fight is actually decided.
• Rate cases. State commission decisions on who pays for large-load interconnection. The public cost of AI is a utility docket.
• CFIUS as AI governance. The Cerebras and G42 episode. Foreign-investment review is doing governance work nobody calls governance.
• Acqui-hire structure. Consolidation in a form antitrust review reaches unevenly.
• Aftercare. What happens to implanted people when a company exits. No business model funds it.
• Seafloor permitting. Endurance. A slow clock with no agency yet.

Sources. Company funding announcements and SEC/EDGAR filings; AMD on Taalas; Google on completion of the Wiz acquisition (March 11, 2026); FTC public records on Microsoft/Inflection; FDA device records; company hiring and product documentation.

Fifteen policies, now and upcoming

Dated, with the one thing each one decides. The ones that move the slide-3 argument are marked.

Who is adopting policies

Adoption is a different question from enactment. Who has actually changed a document because of a rule.

• Grid operators. All six FERC-jurisdictional RTOs have to file by the show-cause deadline. PJM was separately directed to write new rules. The tariff filings are the adoption.
• Frontier developers. California SB 53 makes publication of qualifying frontier-developer safety frameworks a legal duty. The adoption is the framework; embodiment is the comparison on slide 1.
• States. Colorado adopted, then unadopted most of it. New York adopted. Texas adopted a narrower model. The EU adopted and deferred. The direction of travel is toward transparency and away from duty-of-care, everywhere at once.
• The regulator itself. NRC adopted an internal AI tool before any applicant's tool was reviewed. That's the most important adoption on this list.

Think tanks and research shops, sorted by what they can prove

Not a list of who's smart. A list of who holds primary data or a docket seat.

• Hold the data on the slow clocks: LBNL Electricity Markets and Policy (Queued Up), Belfer Center (Texas and Virginia load), Grid Strategies, RMI, Clean Air Task Force.
• Nuclear licensing and policy: Breakthrough Institute, Nuclear Innovation Alliance, Third Way's clean energy program, and the ANS Nuclear Newswire as the trade record.
• AI policy with a docket seat or a testimony record: CSET, RAND, IAPS, GovAI, Brookings, CNAS.
• Civil society with standing on rights: AI Now, Ada Lovelace Institute, CDT, EPIC, Data & Society.
• Labor and work: EPI, Roosevelt Institute, and union research departments, which are think tanks nobody counts.
• The test I'd apply: whether the shop has ever filed a comment in a docket above, and whether it publishes its data.

Which of these ON already funds is my first internal question.

Sources. FERC RM26-4; NRC Part 53, 91 FR 15696; NRC AI strategic plan; DOE Gordian announcement; European Commission AI Act and GPAI guidance; DOJ intervention in xAI v. Colorado (April 24, 2026); California, New York, Colorado, and Texas statutes; Texas SB 6; Virginia SCC Case PUR-2025-00058; FDA.

Five episodes I'd want to do first

Each one is one case, two accounts that disagree, one affected or operating person, seven fixed questions, and the last question on the record: observed, inferred, contested, speculative. None of these is an interview with a founder about their vision.

1. The Revision 0 draft. An Everstar or national-lab engineer who built the Gordian run, and an NRC or national-lab reviewer who has read chapter 5, on the same 208 pages. Question that decides it: what a reviewer checks that a drafter doesn't. The output is the public rubric, or the reason there isn't one.

2. Who pays for the wire. An RTO interconnection planner and a state ratepayer advocate, on one large-load request, with the show-cause filing on the table. Question: what a speculative filing looks like from inside, and who carries the cost when it withdraws.

3. After the company exits. A BCI trial participant and a rehab clinician, with an FDA device reviewer in the third chair, on what the consent form says happens when the study ends. Question: who holds the neural data on wind-down, and who pays the clinic. This is the room slide 4 says I wouldn't convene until a participant agrees. The episode is how I ask.

4. Against the scaling race. Sara Hooker on continuous learning, opposite an open-weight maintainer or lab researcher who thinks scale wins, on what "open" and "adaptive" mean in a procurement contract. Question: which eval standard the next release cites, and who wrote it.

5. What was acquired. A founder whose team was hired away in one of the $20B of non-acquisitions, opposite an antitrust lawyer or FTC alum, on the deal document. Question: what the acquirer bought, what it left behind, and whether any process reviewed it.

Where Founder's Corner and Smartphone Nation already exist, none of these fits inside either. That's the gap the slide names.

Sources. DOE and NRC licensing records; FERC RM26-4 and RTO filings; FDA device records and trial materials; company and funding announcements; public antitrust records on acquisitions and acqui-hires.

Leading indicators, with where each one is read

Sources. LBNL Queued Up; FERC RM26-4; DOE/NRC announcements; Texas SB 6 and PUCT large-load proceedings; Virginia SCC Case PUR-2025-00058; FDA device database; model and system cards; SEC Form D; federal and state procurement notices.