| |
Since last week: a bankruptcy auction produced two bids for one company's internal records, a union objection delayed the sale, and a third company then made a higher offer after the auction closed. A judge let the core privacy claims against an AI notetaker survive. OpenAI cut its flagship price on its own rate card. And Klarna's Q2 put AI in a written filing, with no number attached. |
The week in three numbers: $10,000,000, Google's winning bid for a failed airline's internal corporate data · 30,865,471, the customer call recordings on that same schedule marked Not Included · 20% and 33%, the cuts to OpenAI's flagship input and output prices, promotional through at least November 21.
|
In this issue
01 · The One Thing · 02 · Do This Week · 03 · The Signal · Skim ends here.
04 · The Margin-Proof Tracker · 05 · The Synthesis · 06 · Where the Minds Disagree
Then: What We're Watching · Worth Your Time
|
|
01 · The One Thing
A bankruptcy auction just put a price on the corporate data few companies govern like an asset: the internal operating record. Primary source.
On August 14 Spirit Airlines auctioned its data and Google won at exactly $10,000,000. We read the filing. Major customer datasets are Not Included, down to 30,865,471 call recordings. Included is the exhaust: 100 million emails across 80,000 staff accounts, 500 million Teams items, code. Nothing has been sold. A judge adjourned the approval hearing to September 9 after a union objected. (The auction filing · the union's objection)
|
| |
The executive shift: you already know your customer data has a market value. Here the process drew two auction bids for the internal operating record, and a later offer. Mercor.io was the court-recognized backup at $7.5 million. Then, on August 19, after the auction had closed, the AI data startup Micro1 sent Spirit's lawyers an unsolicited $12.5 million offer. Reported. Business Insider first reported the offer. Scope it carefully: Micro1 missed the auction, the offer is not on the docket, and experts disagree about whether the court will entertain it. Analysis: the winning number is not the point. Three buyers pursued this material even though the major customer datasets were excluded. (Business Insider · Inc.) |
1 · Ask your general counsel: if this company were sold or wound up tomorrow, what would be on the data asset schedule? Not the customer database, which everyone thinks of. The staff mailboxes, the chat archive, the ticket system, the code repositories. Stakes: a court filing this week put a real price on exactly that list. If nobody has looked at your version, you do not know what you hold or what a future owner could do with it.
2 · Ask your IT or procurement lead for a list of every AI meeting-notetaker running in the company, and what each does by default with the recording. Stakes: a judge has just let the core privacy claims against one of these vendors survive dismissal. Our read: companies using these tools may still carry separate notice or consent duties, particularly where outside participants are recorded. Have counsel map them. Section 03 has the ruling.
3 · Ask your AI vendors, in writing, which rates are standard and which are promotional, and for each promotion, the guaranteed-through date and any stated rate after it. Stakes: both just moved, in opposite directions. On August 10 Anthropic made an introductory rate permanent and canceled a scheduled 50% rise. OpenAI cut its flagship pricing by 20% on input and 33% on output, and calls that one promotional. Anthropic's is a standard rate you can bank. OpenAI's is available at least through November 21, with no later rate announced.
This week's question: if the model keeps getting cheaper and the material keeps getting bought, which one are you actually paying for?
Trust A judge let the core privacy claims against an AI meeting-notetaker survive, and narrowed the case. Primary source, and we read the order, not the coverage. On August 13, in In re Otter.ai Privacy Litigation (N.D. Cal.), Judge Eumi K. Lee granted Otter's motion to dismiss in part and denied it in part. The surviving claims include federal wiretap, California privacy, Illinois biometric privacy, unfair competition and unjust enrichment. Several others were dismissed with leave to amend, and six were withdrawn by the plaintiffs. This is a pleading-stage ruling on whether the complaint states a claim, not a finding that Otter broke the law. Two other AI-notetaker cases, neither this week's news: Granola (filed July 30) pleads the same wiretap claims, and Fireflies.AI (December 2025) was voluntarily dismissed in March and is closed. Our read: the defendants are the vendors, not their customers, and what this means for everyone else is a separate and unsettled question. Section 02 has the move. ( The order · coverage · Granola complaint · Fireflies docket)
Work A publisher built a copy-editing agent out of 30,000 of its own editor's past edits. Vendor claim, because the founder is describing his own company. Dan Shipper of Every told Platformer on August 20 he collected "a dataset of 30,000 of her historical edits" from editor in chief Kate Lee, then built and tested a prompt against them. Two precisions the coverage invites you to skip. It is a prompt refined against past work, not a trained model. The headcount figure, 15 to 30, is Every's own. Analysis: any rival can rent the same model. Nobody else has those 30,000 edits. ( Platformer)
Deploy Salesforce's own implementation partners report Agentforce is not driving bookings. Reported, and you cannot go read the source. It is a TD Cowen note to clients, relayed by The Register on August 21. Of the partners surveyed, "none were seeing Agentforce become a driver of bookings activity." No sample size is published, so treat it as an analyst survey, not a study. A KeyBanc note in July pointed the same way. Analysis: the partners are paid to make it work. When the channel is not earning, the product is not yet compounding. ( The Register)
Horizon Frontier pricing moved too, but under a promotion rather than a new standard rate. Primary source. On August 21 OpenAI cut GPT-5.6 Sol to $4 per million input tokens and $20 per million output, down 20% and 33%. It is on the rate card, and the first cut to Sol since it launched. OpenAI calls it promotional, and says it is "available at least through November 21, 2026." Read that precisely: a guaranteed-through date, not an expiry, and no rate after it has been announced. Anthropic went the other way. Its Sonnet 5 rate of $2/$10 became the permanent standard on August 10, and the September 1 rise to $3/$15 "will not occur." Analysis: the pressure is now visible at the top tier, not only at the cheap end. Anthropic's is now a standard rate. OpenAI's remains promotional, with no announced successor rate. Only the former belongs in a two-year baseline without a contingency. ( OpenAI's Sol model page · Anthropic pricing)
▼ Below the Cut
You may read that KE Holdings credited its best margins in three years to AI. Management never said it. The company reported on August 21: adjusted net income up 74.9%, net margin 13%, a three-year high. Real numbers, in its own release. But asked what drove profit past revenue, the CFO named a lower cost baseline, operating efficiency, and business mix. AI was not among them. The AI attribution sits in a third-party transcript page's standfirst, whose footer says parts of the article "were created using Large Language Models." Analysis: an AI-written line appears to have invented an AI story the company did not tell. We were about to add this row to our Tracker. We opened the filing instead. ( The company's release · the transcript page)
One question worth putting to your team this week: what would be the most valuable line on our asset schedule?
| End of skim · deep read begins |
| 04The Margin-Proof Tracker |
| |
One row was tested this week and it did not move. Issue 009 printed a dated test: Klarna's Q2 on August 18, and whether its roughly $60M AI saving would appear in writing. It reported on the day. A written mention now exists and the number does not. N = 12. |
What companies claim AI is worth, against what shows up in their financial statements. None of the twelve companies in this table has reached Stage 4. The evidence ladder: 0 · Narrative (a story, no numbers) · 1 · Operational (activity counted) · 2 · Financially linked (a number tied to AI, mixed with other causes) · 3 · P&L-attributed (a reported profit or margin change the company credits to AI) · 4 · Sustained (Stage 3 held four quarters).
Filing types in plain English. An 8-K is a US company reporting an event as it happens. A 10-Q is its quarterly report. A 6-K is what Klarna files instead, as a foreign company listed in the US. Opex means operating expenses.
| Company |
Evidence |
Grade |
Next test |
| Klarna |
EvidenceChanged this week. Opex +16% against +27% revenue, "supported by AI-enabled productivity gains and continued cost discipline." Unquantified, credit shared, and the ~$60M figure appears nowhere. 6-K, Aug 18 |
Grade2 |
Next testQ3, does a figure ever attach? |
| Duolingo |
EvidenceGross-margin rise "reflecting continued reductions in per-unit third-party AI costs." 10-Q |
Grade3 |
Next testQ3, two quarters to Stage 4 |
| IBM |
EvidenceAI signings; company says signings are not revenue. Q2 |
Grade2 |
Next testQ3, bookings or revenue? |
| Latch / DOOR |
Evidence~65 roles, $10 to 12M expected, not booked. 8-K |
Grade2 |
Next testQ4, does it get booked? |
| Visa |
Evidence$563M severance; AI's share never stated. 8-K |
GradeProvisional |
Next testQ4, capex and hiring mix |
| Infosys |
Evidence8.2% of revenue labeled "AI," alongside cut guidance. Q1 FY27 |
Grade2 |
Next testQ2 FY27, share up and guidance up? |
| Equifax |
Evidence$150M AI cost-reduction goal. Q2 |
Grade2, a target |
Next testQ3, booked or restated |
| ServiceNow |
EvidenceAI contract value past $1B. Committed, not earned. Q2 |
Grade2 |
Next testQ3, recognized in results |
| Alphabet |
EvidenceCloud +82% to $24.8B; AI credited, not separated. Q2 |
Grade2 |
Next testQ3, is AI revenue separated? |
| Bank of America |
EvidenceEfficiency ratio 59%; no stated link to AI. Q2 |
Grade1 |
Next testQ3, linked in writing? |
| JPMorgan |
EvidenceAI-linked headcount cut. No primary document found. |
Grade1 |
Next testQ3, any written attribution |
| Etsy |
Evidence220 roles, ~$35M. The AI denial is in the staff memo, not the 8-K. 8-K |
GradeNot scoreable |
Next testQ3, does product-dev spend rebuild? |
A row we declined to add. KE Holdings looked like a clean Stage 2 on the headline. We opened the release and the attribution is not there. Not added. Rows other than Klarna are unchanged and carry the evidence verified when each last moved.
The model is rented. The corpus may be owned. The system makes it useful. And the rights decide whether any of it can travel.
For four issues we have asked one question in different clothes. If everyone can buy the same capability, who keeps the excess return once competition catches up? That excess return is what economists call a rent. This week sharpened the answer into four parts, and the fourth is the one nobody budgets for.
The model is rented, and its price keeps falling. OpenAI cut its cheapest tier about 80% in late July, its mid tier about 20%, and this week its flagship by 20% on input and 33% on output. Anthropic canceled a scheduled 50% rise. The pressure is now visible at the top tier, not only at the cheap end. If your advantage is access to a good model, it is getting cheaper for your competitors at roughly your own rate.
The corpus may be owned. The sale process produced two auction bids for one, $10 million from Google and $7.5 million from a vetted backup bidder, and a later reported offer of $12.5 million from outside the auction. Not the major customer datasets, which were excluded from the package. The internal record. In the same week 18 advocacy groups asked the Federal Trade Commission to investigate AI companies buying and destroying physical books, and framed it deliberately as a competition problem. The letter says it "does not ask the FTC to regulate AI." It asks whether locking up scarce inputs is unfair competition.
The system makes it useful, and it is harder to copy than we said last week. A corpus with no working system around it is a dormant archive. And a system is not just a project a rival can run next year. It accumulates integrations, evaluation history, operating habits and the trust of the people who use it daily. Those compound in calendar time too. Every's editor's agent is a good example: the 30,000 edits are the scarce part, but the reason it works is that staff call it by name inside their normal workflow.
And the rights decide whether any of it can travel. This is the part Spirit adds that we would not otherwise have seen. The contract requires stripping personal data "while preserving referential integrity." The union's objection is that those two requirements fight each other. Analysis: a corpus you cannot lawfully move, or cannot anonymize without destroying its usefulness, is worth less than its size suggests. That question goes to a judge on September 9. The ruling binds only this case. Every other estate holding a data asset will read it anyway.
The honest counter-case. One week is not a trend, and the only priced instance is a distressed sale, so the price may not represent a normal-market valuation. Our own Tracker is the second piece of counter-evidence: we set a dated test for Klarna, it reported on the day, and the AI saving still has no number. And a Carnegie Mellon study released in August, run with the analytics firm Larridin, looked at 564 companies. Firms with the most specific AI disclosures achieved "an 8% advantage in revenue growth." Its own conclusion: AI is "a top-line story, not yet a cost story." Revenue moved. Margin did not. A rent you cannot see in a margin is a hypothesis. Ours remains one. (the study)
The reusable test. Stop asking whether your AI works, which is table stakes, and ask the question the Spirit schedule forces. If this company were sold tomorrow, which line on the asset schedule would a buyer actually bid for? If the honest answer is the model subscription, that is an operating expense and any competitor can match it by Friday. If it is a body of work only you have, that is the asset. Then ask the two follow-ons: can we actually use it, and are we allowed to move it?
| 06Where the Minds Disagree |
| |
If you can fund only one this year, do you build the system or preserve the corpus?
The Synthesis says you need both. Budgets say you pick. This is the sequencing argument.
Fund the system first, and Deloitte is the case for it. Its survey, fielded April to June and published August 12, found just one in five leaders said their organization was prepared to redesign processes to run autonomously with AI agents. Scope it precisely: every organization surveyed was already piloting agentic AI, and the answer is a leader's self-assessment, not an audit. So it tells you most surveyed leaders do not yet feel prepared, which is a weaker and more honest claim than "most companies have not built it." Conflict label: Deloitte sells agentic transformation consulting, and the population it surveyed was self-selected toward buyers.
Preserve the corpus first, and Every is the case for it. Dan Shipper says he built his editorial agent from 30,000 edits by one specific editor. Conflict label: he is describing his own company's product in an interview, and no part of it, including whether it works, was independently checked.
Our read: fund the system, and start preserving the corpus now. A system is expensive and slow, and Deloitte suggests most surveyed leaders did not feel prepared. But preserving governed, legally usable work product is relatively inexpensive, and missed history cannot be recovered later. A proprietary corpus accumulates in calendar time and may be slower to replicate than access to a model. Note the honest limit, which this issue's own lead supplies: Spirit shows a rival can sometimes just buy a corpus.
What would change our mind: if systems built without a proprietary corpus perform just as well, or if firms with identifiable, legally usable corpus advantages fail to turn them into better quality, growth or margin over time.
|
Monday, August 31 · last day for the California legislature to pass bills this year, with a number of AI bills still in play. A correction to how this date is usually described: the legislature's own calendar shows August 31 as the last day to pass bills and the start of Final Recess. The session does not formally end, which the legislature calls adjournment sine die, until November 30. The governor's deadline to sign or veto is September 30. ( Official 2026 legislative calendar)
Wednesday, September 9 · the adjourned hearing on the Spirit data sale. The decision: whether a court approves selling a company's deidentified internal records as an AI training asset, and on what conditions. What would change our read: conditions attached to deidentification, or any ruling on the late Micro1 offer. ( The union's objection)
Thursday, October 15, 11:59pm · comments close on NIST draft SP 1353, a guide to using generative AI for Cybersecurity Framework analysis and reporting. If your security team will be asked to use AI on framework work, this is the moment to say what does and does not work. ( NIST)
Tuesday, October 20 · comments close on the CFTC's request for comment on listing compute derivatives, contracts whose underlying commodity is "access to rented compute capacity." This is a request for comment, not a proposed rule. It is still the clearest sign yet that compute is being treated as a scarce traded commodity. ( Federal Register, RIN 3038-AF77)
Zapier's survey of 835 people actually running AI pilots. Managers and above, companies over 100 staff, fielded in May and published July 14, so it is background rather than news. The plainest available answer to how many pilots reach production, and unlike most vendor surveys it publishes its sample and its margin of error. ( Zapier)
The FTC letter on books, for the argument rather than the outrage. Eighteen organizations, and they deliberately declined to make it an AI-safety complaint. It is an input-foreclosure argument, and worth reading if you want the competition-law shape of the training-data fight. ( Letter, Aug 21)
Anthropic on protein design. Claude designed protein binders against 15 targets and succeeded on 14, with the physical lab work done by Adaptyv Bio and Twist Bioscience. A vendor's own research, not peer reviewed, and Anthropic says it intends further work to confirm its hit rates. ( Anthropic, Aug 18)
Cloudflare on driving an open-source issue queue toward zero. Open issues on the Astro project went from over 200 to about 30, using agents with a human confirming each fix. A vendor claim about its own tooling, but an unusually honest one: it states plainly that it did not get there by auto-closing old tickets. ( Cloudflare, Aug 4)
How we label evidence: Primary source · Corroborated · Reported · Vendor claim · Analysis. Written and edited by Mario Suarez · Independent analysis · Every link in this issue was opened and confirmed before send; Tracker rows other than Klarna carry evidence verified when each last moved.
|