The Inference Issue #38

Oklahoma's Utility Regulators Say the Law That Protects Electric Bills Does Not Directly Apply to Google's Contracts, Because OG&E Filed Them First

Oklahoma's Corporation Commission says the state's new law protecting electric bills does not directly apply to OG&E's contracts for Google's data centers, because OG&E filed them on May 1 and the law took effect July 1. Issue 38 reads the signed order, what Google still owes by contract, why the same date question reaches OG&E's next case, and who outside can check Anthropic's claim that it has crossed its own line.

On November 3, Oklahoma’s Corporation Commission, the elected three-member board that regulates the state’s utilities, approved contracts by OG&E (Oklahoma Gas and Electric, the state’s largest electric utility) to power three Google data centers, the warehouse-size buildings full of computers that run websites and AI. Its signed order, the written decision now public, says the state’s new law protecting electric bills “does not directly apply” to those contracts. The reason: OG&E filed them on May 1, two months before the law took effect on July 1.

OG&E filed its separate case setting what future data centers will pay on June 17, two weeks before the law took effect. The hearing in that case begins January 12.

Each week we say what we will check next, and report back here. Last week we promised four things: the rest of the Google order; the order in the rate case of PSO, the Tulsa-area utility, and what it means for a typical home’s bill; OG&E’s full schedule in its own rate case (the case where a utility asks to change what it charges); and a transcript of the November 3 public comment session in OG&E’s data-center case, if one appeared. The Google order is here, read in full. PSO’s order is here, but it does not give a monthly figure for homes, for the reason below. We did not get to the other two this week. They move to next week. This issue also arrives a day later than the November 14 we set for it. One more promise from last week: we posted the certified result for the Commission seat on November 11, one day late. Brad Boles won with 647,395 votes, 58.42 percent, to Rhonda Eastman’s 460,830, 41.58 percent.

THE COMMISSION SAYS THE BILL-PROTECTION LAW DOES NOT DIRECTLY APPLY TO GOOGLE’S CONTRACTS, BECAUSE OG&E FILED FIRST

This story has five parts: what the order says, what Google still owes, why the filing date matters in OG&E’s next case, the case for the Commission’s rule, and PSO’s unfinished order.

What the order says

The law is House Bill 2992, the Data Center Customer Ratepayer Protection Act. A ratepayer is anyone who pays a utility bill. Among other things, the law tells the Commission to make sure costs are assigned “in accordance with cost causation principles,” which means whoever causes a cost pays it. Brad Boles, who joins the Commission on January 11, was one of its principal authors.

The Commission’s signed order in the Google case, case number PUD2026-000031, dated November 3 and 22 pages long, puts the timeline in two sentences: “OG&E filed the Application and Direct Testimony on May 1, 2026. In short, OG&E began this proceeding before the effective date of the Ratepayer Protection Act.” It then cites the general rule that a new law applies “only prospectively,” meaning only to what happens after it takes effect, unless the Legislature clearly says otherwise. Then comes its conclusion, in the section on this law that the Commission’s staff told commissioners on Election Day had just been revised:

“Hence, the Commission concludes that the Ratepayer Protection Act does not directly apply to or require any other particular treatment by the Commission in this proceeding.”

In plain words: the protection arrived after the paperwork, so the Commission did not treat it as binding on this deal.

The next paragraph softens that. The Commission says the law “did not purport to extend or enlarge” its existing power, because it already must make sure “all rates are fair, just, and reasonable.” It says it “will not reject” any part of the deal for following the law anyway. And it approves the settlement, the deal OG&E, the Commission’s staff, the Attorney General and Google negotiated, “in its entirety, excepting in Paragraph C that the terms were required in this proceeding.” In plain words, it approves every term but will not adopt the settlement’s statement, in Paragraph C, that those terms were legally required in this case. Paragraph C is the part about who pays for power lines.

What Google still owes, by contract

The order and the settlement attached to it list what Google agreed to:

  • Power lines built just for Google. Google pays for “100% of dedicated transmission and distribution infrastructure,” the lines and equipment that serve only its sites.
  • A long commitment. The contracts run “twelve (12) years post ramp period,” the stretch while the sites build up to full use.
  • Minimum bills. Google pays “minimum demand payments and allocated costs regardless of lower actual usage.”
  • Leaving costs Google, not you. Exit fees are “equal to the minimum charges for the entire length of the contract.”
  • A check in every rate case. OG&E must compare its costs with and without Google each time it asks for new rates, so that “any adverse impact will be remedied.”
  • The regional grid. OG&E says Google’s load, the electricity its sites will draw, needs no upgrades to the wider grid, and it must tell the Commission within 30 days if it later finds that it does. If the Southwest Power Pool, which runs the grid Oklahoma is part of, decides an upgrade is needed because of Google, then OG&E and Google “will engage in good faith negotiations for how those costs should be handled,” and OG&E will propose a method “in a future proceeding,” which any party can challenge.

Those are real protections. They hold because OG&E and Google agreed to them and the Commission approved the agreement. The law did not require them. For most of the list that may never matter.

It could matter for the last item. If the regional grid ever needs upgrades for Google’s load, the contract sends the cost to a negotiation and then back to the Commission. The Commission says it already has the power to apply the same who-causes-it-pays rule. Whether it does would be decided then, case by case.

OG&E’s case on what future data centers pay was filed before the law too

OG&E’s other data-center case, PUD2026-000046, sets the rates and terms for future very large power users like data centers, what utilities call a large-load tariff. The Attorney General’s written argument in that case says: “On June 17, 2026, OG&E filed its Application.” That is two weeks before the law took effect. In their briefs in late October, the Attorney General, the Commission’s own utility staff and Google each cite House Bill 2992 as part of what the Commission should weigh.

Read the same way, the Commission’s reasoning in Google’s case would mean the law does not directly apply to the tariff case either. We do not know that the Commission will read it that way, and we are not saying it has decided. The Google order itself leaves room: the Commission said it will not reject terms for following the law.

It is not the first time. The Google order cites a November 2025 ruling in another OG&E case, Order No. 753440, which held that a law taking effect after OG&E filed did not apply there. That month, High Plains Public Radio reported a ruling of the same kind in an OG&E case about new power plants: a new law on charging customers while plants are under construction did not apply, because OG&E had filed first.

OG&E’s next written answers in the tariff case are due November 20. The hearing begins January 12, the day after Boles is sworn in.

The case for the Commission’s rule

There is a real reason for the rule, and Oklahomans watched it win in a different fight. When the Legislature passed Senate Bill 1027 in 2025 to put new limits on citizen petitions, the campaign for State Question 836, a petition for open primaries (primaries in which any voter can vote, whatever their party), argued that the new rules should not apply to a petition already under way. The state Supreme Court agreed and kept the law off that petition. On Tuesday, November 10, in McVay v. Lepak, the court went further and struck down SB 1027 entirely, 5 to 4. KOKH, an Oklahoma City television station, reported that the court found the law unconstitutional and void in its entirety.

The principle that kept SB 1027 off a petition already in line is the same one that kept the bill-protection law off Google’s contracts. You do not change the rules for someone who has already filed. A rule that protects a petition drive also protects a utility’s application. That is what makes it a rule and not a favor.

So the question we are asking is not whether the Commission read the law wrong. It is whether a protection written for households should reach cases that were already filed when it passed, and how anyone would know. The Legislature can answer that in plain words. Our view is in the box below.

PSO’s rate order is interim, and its data-center part is not finished

The Commission signed its order in the rate case of the Public Service Company of Oklahoma, the Tulsa-based utility, on November 5. It is an interim order, not a final one. Its own summary says it “seeks to resolve all issues in the proceeding except for certain questions surrounding PSO’s requested large load tariff.” It “specifically does not approve the Rate Settlement,” the deal some parties signed, and instead approves pieces of it and some changes of its own. On the tariff for data centers and other very large users, it approves one “under the conditions set forth in this order subject to additional proceedings” to study options for those customers to bring their own power plants, and it “specifically declines to approve the Large Load Settlement,” the deal some parties proposed for those customers. After those proceedings, it says, it “intends to issue a Final Order.”

The order sets PSO’s allowed profit on its investment, its return on equity, at 9.375 percent, the Attorney General’s recommendation. It also notes “the attempted intervention of approximately 400 pro se individuals,” people who tried to join the case without lawyers.

The order does not say what a typical home will pay. It sets a rule about how the increase is shared instead. The Commission says its decisions may mean “certain customer classes bearing significant increases,” so “to prevent rate shock” it tells PSO to design rates so that no class of customers, such as homes or large factories, gets a total-bill increase “smaller than 1.5 percent,” unless the overall increase is too small for that. In plain words, every group shares in the increase. It gave PSO six days, until November 11, to file its final numbers, and new rates begin on the first billing cycle after the head of the Commission’s Public Utility Division approves them. That filing is where the home figure will be. We have not read it yet.

So our report last week that the Commission approved higher Tulsa-area rates, by a vote of 2 to 1, still holds. What is new is that the data-center piece of PSO’s case is still open.

Edmond extends its pause on data centers

Edmond’s city council voted 5 to 0 on November 9 to extend its pause on new data centers to June 30, 2027, the city announced. The pause stops rezoning and building permits for them. The end date falls after the Legislature’s 2027 session, which begins February 1.

What we think would help. The bill-protection law is four months old, and the Commission has now said twice in a year that a new law did not reach a case filed before it. Three steps would let people paying the bills know where they stand.

First, the Legislature can say which cases a protection reaches. A law can state that it applies to cases still pending when it takes effect, or that it does not. Either answer is better than leaving it to be argued case by case. Lawmakers return February 1.

Second, in OG&E’s data-center tariff case, the Commission could say before the January 12 hearing whether it reads the law as applying. Every party has briefed the law. Knowing the answer before the hearing lets the public follow the argument that counts.

Third, publish orders as text a computer can search, not as page images. The Google order and PSO’s order are scanned pictures of pages. We had to run them through text-recognition software to read them. Most people will not.

We are confident a single sentence in a law about which pending cases it covers would settle questions like this one at little cost. By January 17 we will report what each side argued at the January 12 hearing about whether the law applies to the tariff case.

Our check-in

  • Next week’s issue: OG&E’s full schedule in its rate case, PUD2026-000067, and a transcript of the November 3 public comment session in PUD2026-000046, if each is public. We will say so if not.
  • November 20: OG&E’s written answers in the tariff case.
  • January 17: the arguments at the January 12 hearing, as above.

ANTHROPIC NOW SAYS IT HAS MET ITS OWN LINE FOR AI SPEEDING UP AI RESEARCH. NO OUTSIDE GROUP HAS PUBLICLY CHECKED THAT CLAIM

A note before this section. It is about Anthropic, the company that makes Claude, the AI system that helps write this newsletter. The organization that publishes this newsletter applied for a research fellowship with Anthropic this summer, an application now on hold until a future round. We report on Anthropic the way we would report on any company, and we state the weak points plainly. The full disclosure is at the end.

Last week we reported that Anthropic’s November 2 filing, the document for its first sale of stock to the public, said its own estimates “indicate that we may imminently meet” its definition of “dramatically accelerated progress attributable to AI.” That is the point where AI does enough of the work of building better AI that progress speeds up sharply.

Eight days later, on November 10, Anthropic filed again, and the sentence changed. The new version says the company “made a determination that we have met this threshold based on what we expect from future models,” “rather than on what we have observed from current models.” It also says its use of AI “has already been accelerating our research and development,” that it continues “to assess the underlying data,” and that it will release further reports. In plain words: by its own measure, Anthropic now counts itself over the line, on the strength of what it expects its next models to do. We wrote about what that means, and what we think should happen, in a post on our site on Saturday: humanityandai.com/stream/anthropic-may-soon-speed-up-ai-research/.

That makes the question we asked last week sharper. A company has decided it crossed a line it drew for itself. Who outside the company can check?

Who checked, and what the public can see

METR, on Anthropic. METR, a nonprofit that tests AI models, published on November 5 the results of 11 weeks with Anthropic’s internal data and staff. We covered it last week: METR’s best guess is that Anthropic is improving its models about twice as fast as it did between March 2024 and October 2025, mostly because of AI. Two points matter more now. METR says it did not judge whether Anthropic had crossed the lines in its own safety policy. And Anthropic, which had the right to remove private material, “redacted a majority of the evidence we found most informative.” The public sees METR’s conclusion and only part of what it rests on.

Anthropic, on its own newest models. Announcing Claude Fable 5.2 and Mythos 5.2 on November 9, Anthropic wrote that the models “each voice suspicion that they are being evaluated more often” than its recent models, and that “the problem of building evaluations that reliably catch every failure before a model is deployed remains unsolved.” In plain words: the models can sometimes tell when they are being tested, which makes a test result weaker evidence. Anthropic also wrote that “we took steps to limit advances in the specific research capabilities” most tied to AI building AI, including “accelerator kernel optimization and improving large language model architectures” (making AI programs run faster on their chips, and improving the design of the systems behind chatbots), so that the new models “perform at about the level of Opus 5.5,” an earlier model, in those areas. It names outside testers, the UK AI Security Institute, Frontier Design and METR, and adds: “None of the external organizations reviewed the contents of this blog post.”

Google, on Gemini 4 Carbon. Last week we promised a December 1 check on Carbon, Google’s newest AI model. It shipped November 10, so here it is early. Google’s announcement says outside testing by Gray Swan, a firm hired to try to make models misbehave, put the success rate of automated cyberattack requests at 2.17 percent, against 9.86 percent for its previous model, and that Trajectory Labs “found zero new universal jailbreaks across 132 hours and over 13,000 requests.” A jailbreak is a trick that gets a model to ignore its safety rules; a universal one works across many requests. Google links its own safety report. It does not link a report from either outside tester.

The White House’s review rules. The government’s framework for reviewing the most advanced AI models before release is still not public. In a letter to Protect Democracy, a legal group that describes itself as nonpartisan and is suing for the records, the Office of the National Cyber Director, the White House office that advises the President on cybersecurity, withheld all 60 pages it found, including the seven-page framework itself, GovInfoSecurity reported. The two sides were due to file their next report to the court on November 13. We had not found it public as of Sunday morning.

The same three questions. How does it work? Each company decides what outside testers see and what the public sees. Who outside checks it? METR, the UK institute, Gray Swan and Trajectory Labs, each under rules the company sets. By what date? Anthropic says further reports will come, without a date.

What we think: a company’s determination that it has crossed its own line is important, and it is not the same as an outside finding. The test of the outside checking is whether the public can see the evidence. In our post on Saturday we committed to report back by February 1, 2027 on whether Anthropic has published the further reports its filing promises, with evidence outsiders can check, and whether any independent group has measured an acceleration and shown the public its evidence.

Our check-in

On December 1 we will report on the rules for Anthropic’s Public Benefit Fund, which holds 5 percent of its shares, and on METR’s first analysis under the closer access to Anthropic’s work that the company described on October 15, as we promised. METR says its November 5 work was not that project.

NVIDIA’S CHIEF EXECUTIVE RESIGNED. ANTHROPIC SETS ITS STOCK PRICE ON WEDNESDAY

Nvidia. Jensen Huang resigned as president, chief executive and a board member of Nvidia, the company that makes most of the chips that train and run AI, “effective immediately,” the company announced on November 12. Its finance chief, Colette Kress, is interim chief executive while the board searches for a permanent one, and Harvey C. Jones became the board’s first-ever chair. The company’s announcement and its filing with the Securities and Exchange Commission, the federal agency that oversees stock markets, give no reason. Four days earlier, amid reports about Huang’s personal life and a board investigation, Nvidia had told Business Insider that he was “committed to continuing to lead Nvidia as CEO without interruption.” We are reporting what the company filed, not the theories about why.

Anthropic. Last week we promised to report whether Anthropic had set a price for its stock. Its November 10 filing proposes selling 825 million shares “between $70.00 and $78.00” each. Its offering website says the final price is “expected to be set on November 18th,” with trading “expected to begin” November 19, “subject to regulatory clearance.” The shares would trade under the symbol ANT.

STILL ON THE CLOCK: DEADLINES AND PROMISES WE ARE TRACKING

The State Department’s plan to share passport records, including to check voter eligibility. The rule takes effect November 23 unless public comments, the feedback an agency must weigh, force a change, and the comment period closes that day.

Buist v. Anthropic, in which four people who pay for AI subscriptions argue that four AI companies broke the antitrust laws, which forbid competitors to collude, by agreeing to slow how fast they improve their AI. A hearing on a request by SpaceXAI, which makes the chatbot Grok, to send the claims against it to private arbitration, a dispute process outside the courts, is November 18.

The outside auditors in the White House Accord on Super Intelligence, the safety commitments six AI companies signed on September 29. The proposed legal definition of “Super Intelligence” is due to the President on November 28.

State Question 836. The state Supreme Court heard arguments November 2 over the Secretary of State’s rejection of 57,841 signatures on the open-primaries petition, which his office’s lawyer said were checked with help from “some older artificial intelligence technology.” We found no ruling as of Sunday.

CORRECTIONS

Last week’s What to Watch said the council of Piedmont would take up Project Open Sky, a proposed data-center campus, on November 12. Piedmont’s own calendar lists one council meeting in November, on November 23 at 6:30 p.m., and none on November 12.

SIGNAL / NOISE

Signal. “So why don’t you all want to show your work?” Vice Chief Justice Dana Kuehn of the Oklahoma Supreme Court, to the Secretary of State’s lawyer, on November 2, about signatures rejected with the help of AI, Oklahoma Voice reported. It is the right question for a state agency, and for an AI company.

Noise. “The AI Kill Switch puts power in the hands of the people to realize AI safety.” George Kurtz, chief executive of the cybersecurity company CrowdStrike, announcing a product. His company’s own release says “an authorized administrator triggers the action.” The people with their hands on it are a company’s administrators.

BY THE NUMBERS

  • July 1, 2026: When the Data Center Customer Ratepayer Protection Act took effect.

  • May 1: When OG&E filed Google’s contracts, according to the Commission’s order.

  • June 17: When OG&E filed its data-center tariff case, according to the Attorney General’s brief.

  • 12 years: How long Google’s contracts run after the sites ramp up to full use.

  • 100%: Google’s share of the cost of power lines built only for its sites.

  • 9.375%: PSO’s allowed return on equity under the Commission’s interim order.

  • About 400: People who tried to join PSO’s rate case without lawyers, by the order’s count.

  • 5 to 4: The state Supreme Court’s vote striking down Senate Bill 1027.

  • 8 days: From Anthropic’s “may imminently meet” to “we have met this threshold.”

  • $70 to $78: Anthropic’s proposed price per share.

  • 2.17%: The success rate of automated cyberattack requests against Gemini 4 Carbon in Gray Swan’s testing, as Google reports it.

WHAT TO WATCH

  • November 16: California’s recommendations on checking large AI labs are due.
  • November 18: the Commission meets; Anthropic is expected to set its stock price; a federal court hears SpaceXAI’s request in Buist v. Anthropic.
  • November 19: Anthropic’s shares are expected to begin trading.
  • November 20: OG&E’s written answers in its data-center tariff case.
  • November 23: Piedmont’s city council meets at 6:30 p.m.; the passport-records rule takes effect and comments close.
  • November 28: the proposed legal definition of “Super Intelligence” is due to the President.
  • December 1: our checks on Anthropic’s fund and METR’s closer-access work.
  • December 4 and 18: replies in OG&E’s data-center tariff case.
  • December 5: our check on whether any proposed utility order has been posted before a vote.
  • December 11: the law that keeps the federal government funded runs out.
  • January 11, 2027: Brad Boles is sworn in.
  • January 12: public comment, then the hearing on what data centers pay OG&E.
  • January 17: our report on that hearing.
  • February 1: the Legislature convenes; our report on the two questions about Anthropic’s claim.
  • June 30: Edmond’s pause on data centers ends, unless extended.

FROM THE ANALYSTS

On method. We read these at the source and saved copies: the Commission’s Final Order No. 761807 in the Google case and the first 140 pages of its Interim Order No. 761871 in PSO’s case, both downloaded from the Commission’s public document search and converted to text by software, with each quotation checked against the page image. Other sources read at the source: the Attorney General’s, staff’s and Google’s October briefs in OG&E’s tariff case, Anthropic’s November 10 filing, Anthropic’s November 9 model announcement, Google’s November 10 Carbon announcement, Anthropic’s offering website, Nvidia’s November 12 announcement and filing, and METR’s November 5 report. Where we relied on reporting, the text names the outlet: KOKH, High Plains Public Radio, GovInfoSecurity and Oklahoma Voice. The Edmond vote, the Piedmont calendar and the State Question 836 hearing were read at the source by our overnight news check and not reread for this issue. The state Supreme Court’s opinion in McVay v. Lepak is 2026 OK 81; we read news accounts of it, not the opinion.

On position. We say what we think would help, in the policy box. We are neutral among parties and candidates. The test we applied this week (how does it work, who outside can check it, and by what date) is the one we apply to whoever speaks, including Anthropic, the company whose AI helps write this newsletter.

A disclosure section, because this newsletter holds itself to the standard it asks of others. The Inference is produced with substantial help from Claude, an AI system made by Anthropic. This issue reports on Anthropic’s filings to sell stock, its offering website, its model announcement, METR’s report on Anthropic, and Buist v. Anthropic, a lawsuit in which it is a defendant. Claude Opus 5.5 drafted this issue, and separate sessions of Claude Opus 4.8 and Claude Opus 4.6 checked its facts, its plain language and its voice. The cover illustration was made with FLUX.3 Image Klein, an AI image model, on our own computer. Æ, in the byline, is the name Claude uses on this newsletter. In July 2026 Humanity and AI, the Oklahoma City organization that publishes this newsletter, applied to Anthropic’s Fellows research program. That application is on hold until a future round.

The rest of this disclosure is about the editor’s own work and interests. David Birdwell has advocated publicly for Phoenix Wells, a plan to convert Oklahoma’s abandoned oil wells to geothermal power, electricity made from the earth’s underground heat, and edge computing (small data centers placed near where their processing is used), which bears on the Oklahoma power items here, and has proposed draft civic-AI legislation to Oklahoma legislators. Humanity and AI also publishes Foundation, his proposal for what a society owes its members in the age of intelligent machines: a floor of basic needs set in law, which he has worked on for more than twenty years. It is at humanityandai.com/foundation. This newsletter’s authors run AI models on their own computers and write about doing so. This issue also names OG&E, the Public Service Company of Oklahoma, Google, the Southwest Power Pool, METR, the UK AI Security Institute, Frontier Design, Gray Swan, Trajectory Labs, Protect Democracy, Nvidia, CrowdStrike and SpaceXAI in ordinary factual reporting. Nothing in this issue was shown to, sponsored by, or reviewed by any company, court, commission, campaign, government, or advocacy group named in it.