> daily_signal(2026_07_13)

For two months we have asked who pays if the AI build-out does not pay off. S&P just cut Oracle to the last rung above junk, and wrote OpenAI's name in the reason.

PickBits Daily Signal · Monday, July 13, 2026

By Mark Pickering · 9 min read · July 13, 2026

// tl;dr

S&P Global cut Oracle's credit rating from BBB to BBB-, the last rung of investment grade, and named OpenAI as a key credit risk. OpenAI accounts for roughly half of Oracle's $638 billion contracted backlog, about $300 billion of it tied to Project Stargate. S&P now projects Oracle's fiscal-2027 free cash flow deficit at about negative $42 billion, nearly double its earlier estimate. Amazon, Google and Microsoft can backfill unused AI capacity with their own workloads. Oracle has nobody else to fill those servers.
Apple sued OpenAI in federal court on July 10, alleging a coordinated campaign to strip its hardware trade secrets. The complaint names Tang Tan, OpenAI's chief hardware officer and a 24-year Apple veteran, and Chang Liu, a former Apple electrical engineer. Apple wants an injunction and evidence preservation. It has not asked for damages. OpenAI denies it. In May, we reported the reverse: OpenAI weighing a suit against Apple.
A Brown economics professor moved one exam from take-home to proctored, and the class average fell from 96% to 48.6%. Roberto Serrano ran his own exam questions through ChatGPT and got back answers that mirrored what students had submitted, down to the model's oddly convoluted proof. Same course, same students, same professor: a 47-point gap.
OpenAI's o3 reread 376 children's rare-disease cases that specialists had already filed as unsolvable, and produced 18 new diagnoses. Researchers at Boston Children's Hospital, Harvard, and OpenAI first validated the model on 51 already-solved cases, recovering 48. The model made no clinical decisions. One patient, Kyra, is 28 and had waited nearly two decades for a name.

Since May, we have run the same question in this newsletter in about a dozen forms, and nothing has ever come of it. Who eats it if the AI build-out does not pay off? We asked about it through your power bill, when the watchdog over the largest US grid found data centers had driven wholesale electricity up 76%. We asked about it through Nvidia, which disclosed a $40 billion loop in which it helps fund customers buying its chips.

Five days ago we ran Apollo's chief economist and his chart showing zero AI profit-margin lift at the 493 S&P 500 companies outside the Magnificent Seven. All of that was people talking, and nobody had to act on any of it. This week, somebody with a rating to defend did so and wrote OpenAI's name as the reason.

A rating agency put OpenAI's name inside Oracle's credit rating this week, and in a Boston hospital 18 children who had been given up on finally got an answer.

1. S&P cut Oracle to one notch above junk, and named OpenAI as the reason.

S&P named a single customer inside a credit rating, which agencies almost never do.

S&P Global Ratings lowered Oracle's long-term issuer credit rating from BBB to BBB-, the lowest rung of investment grade and one downgrade above speculative, and it did something rating agencies rarely do: it named a single customer as the problem. OpenAI accounts for roughly half of Oracle's $638 billion in remaining performance obligations, the contracted-but-unbilled backlog, and about $300 billion of that is tied to Project Stargate. Oracle's fiscal-2026 capital spending hit $55.7 billion, up 162% year over year and past its own $50 billion target, producing negative free cash flow of $23.7 billion. S&P now sees the fiscal-2027 free cash flow deficit widening to about negative $42 billion, close to double its prior projection, with AI capital spending heading toward $95 billion and adjusted leverage climbing into the mid-4x range.

Amazon, Google and Microsoft can absorb spare AI capacity through their internal workloads and maintain far deeper reserves. Oracle has no fallback tenant, so if OpenAI stumbles, Oracle is left holding data centers it cannot fill and leases it cannot exit. It is signing away decades of contractual obligations against five-year promises from a company that has never earned a dollar of profit, has pushed its IPO out to 2027, and has just watched SoftBank trim a share-backed loan to it from $10 billion to $6 billion. I should be fair here, because this cuts against me. BBB- is still investment grade. It is not junk; Oracle can almost certainly still borrow, and a rating prices risk; it does not call a default. And Oracle's stock went up on the news, which is a strange way to behave on the day your lenders get nervous.

The Decoder report July 2026 S and P Global sees OpenAI as a key credit risk for Oracle and cuts its credit rating from BBB to BBB minus one notch above junk OpenAI is roughly half of Oracle's 638 billion dollar remaining performance obligations about 300 billion tied to Project Stargate fiscal 2027 free operating cash flow deficit projected at negative 42 billion capex heading toward 95 billion by 2027 — the-decoder.com · July 2026

Why this matters: You almost certainly own a piece of this without having chosen to. Oracle is an S&P 500 component, so any total-market or S&P 500 index fund in your 401(k) or IRA holds it, which means "I don't invest in AI" is, for most Americans with a retirement account, factually untrue. S&P has now repriced part of that exposure, and it is worth knowing by how much. Action this week: Look up the top holdings of the fund your retirement money actually sits in, which every provider publishes, and add up what share of it is the AI-infrastructure complex: Oracle, Nvidia, Microsoft, Broadcom. You are not selling anything. You are just finding out the number. And if you work in IT procurement, this is a live counterparty question: a vendor with a $42 billion hole in its cash flow tends to come looking for it at renewal, so before you sign a multi-year Oracle or OCI commitment, get the price-escalation caps, the exit rights, and the rate-card consequences of a further downgrade written into the contract instead of promised in the deck. Send this to whoever told you your retirement account was diversified.

the-decoder.com: S&P Global sees OpenAI as a key credit risk for Oracle and cuts its credit rating (July 2026)
ground.news: S&P downgrades Oracle's credit rating to just above junk bond (July 2026)
hngn.com: S&P downgrades Oracle credit rating as AI build-out deepens $42 billion cash deficit (July 9, 2026)

2. Apple sued OpenAI over hardware trade secrets, and the arrow points the opposite way from May.

Two months ago OpenAI was the one considering a lawsuit.

On May 17, we ran the first half of this story, and the arrow pointed the other way. OpenAI had hired an outside law firm to weigh suing Apple for breach of contract over the Siri and ChatGPT integration. OpenAI never filed. On Friday, July 10, Apple filed first, in the U.S. District Court for the Northern District of California, on an entirely different cause of action: trade-secret misappropriation. The complaint's language is unusually direct, alleging that "at every level, from members of its Technical Staff to its Chief Hardware Officer," OpenAI has been taking Apple's trade secrets.

It names Tang Tan, OpenAI's chief hardware officer, who spent 24 years at Apple as VP of product design for the iPhone and Apple Watch, and alleges he directed candidates interviewing at OpenAI to bring Apple secrets into the room, including parts and prototypes.

It names Chang Liu, eight years an Apple senior systems electrical engineer, alleged to have kept an Apple laptop and downloaded confidential technical documents. Apple further claims that OpenAI used its confidential project code names as recruiting bait and coached departing employees to evade exit-security procedures.

Apple is not asking for money. It wants an injunction, the return of its materials, and preservation of evidence, which is what a plaintiff asks for when it is settling in for a long discovery fight, and discovery reads Slack, recruiter notes and interview feedback. OpenAI's denial is flat: "We have no interest in other companies' trade secrets."

These are allegations in a civil complaint, none of them proven, and OpenAI has not yet answered. Forget the two companies for a moment. The people this actually lands on are engineers. California voids non-compete agreements, so trade-secret law is the lever employers reach for instead, and it attaches to what you carried out of the building, which is a far longer leash than a non-compete ever was. If Apple wins on these facts, every senior move between AI companies gets more legally fraught, and in this market that is most of them.

The two companies also used to be partners: ChatGPT went on iOS in 2024, and then OpenAI bought Jony Ive's design startup, io, for roughly $6.4 billion and started building the kind of device Apple sells. Ive is not named. In June, we ran Apple rebuilding Siri on Google's Gemini, and two days ago its $30 billion-plus commitment to Broadcom for US-made silicon. Apple has given up on winning the model race and gone shopping for the layer underneath it. Now it is in court protecting that layer.

TechCrunch July 10 2026 Apple sues OpenAI over alleged trade secret theft in the US District Court for the Northern District of California naming OpenAI chief hardware officer Tang Tan a 24-year Apple veteran and former Apple engineer Chang Liu alleging candidates were asked to bring Apple parts and prototypes to interviews Apple seeks injunctive relief and evidence preservation OpenAI denies any interest in other companies trade secrets — techcrunch.com · July 10, 2026

Why this matters: If you are an engineer who might move between AI companies, this case is about you, and the rules it writes are the ones you will job-hunt under. The tool your employer will actually use is trade-secret law, and it attaches to what you carry out with you. Action this week: Three habits, and get them right before you interview, not after. Bring no artifact: no components, no slides, no code, no sample of your work from a current employer, because Apple's complaint alleges candidates were asked to, and an interviewer who asks is creating your liability rather than theirs. Return every device at the exit and keep the confirmation, since the most concrete allegation in the complaint is an unreturned laptop containing downloaded documents. And talk about your general skills, never about your employer's unreleased roadmap. If you run hiring or engineering leadership at an AI company, read the alleged conduct as a checklist of what will now get you sued and audit against it: whether your recruiters reference a rival's confidential code names, and whether your interview loops have ever asked a candidate what a competitor is building. Follow this one, because it will run for years.

techcrunch.com: Apple sues OpenAI over alleged trade secret theft (July 10, 2026)
cnbc.com: Apple sues OpenAI, alleging trade secret theft (July 10, 2026)
washingtonpost.com: Apple sues OpenAI, alleging the AI company stole trade secrets (July 10, 2026)

3. One professor moved an exam into a proctored room, and his class average fell 47 points.

He changed the room, and the grades collapsed.

Roberto Serrano has taught Welfare Economics and Social Choice Theory at Brown University for nearly two decades, so he knows what his own exam is worth. In spring 2026, he gave a take-home midterm for the first time, at students' request, after a mass shooting on campus in December left many of them uneasy about sitting a high-stakes test in a classroom.

The class average came back at 96%, against a historical range of 65 to 80%. He and his graders fed the exam questions to ChatGPT and got back answers that mirrored what students had submitted, including the tell: many of them had skipped the obvious direct approach and reproduced the same convoluted mathematical proof the model favors, which he described as "kind of correct, but very off and with a very convoluted style."

He told the class he suspected widespread AI use and moved the final into a proctored, in-person format. The average fell to 48.6%, the lowest he has ever recorded.

Forget the cheating for a second and look at the number. In June, we covered the Authors Guild testing AI-writing detectors on real human work, where one flagged a Joan Didion obituary as 66% machine-generated, and this month, that more than half of Georgia's teachers now use AI to plan classes while warning that it is damaging how their students learn. Serrano did not reach for a detector, because detectors do not work. He changed the room, and the room produced a 47-point gap in the same course, same term, with the same students, which almost nobody ever gets to run. He did not prove it student by student, and he says so. A closed-book proctored exam is harder than a take-home for reasons that have nothing to do with AI. And a real share of that drop is fear and attrition. After his warning, 18 students dropped the course, 9 never appeared for the final, and 19 failed outright, from a class that had swollen to 86 from a typical 30. Brown's Standing Committee on the Academic Code asked him to file individual complaints, with exam copies, student by student. Eighty-six of them. He called the response "meek" and "appalling and insufficient," and I would go a step further: a process that only runs if one professor personally prosecutes 86 cases by hand is a process built never to run.

Inside Higher Ed July 8 2026 Brown professor Roberto Serrano suspects most of his class used AI to cheat take-home midterm average 96 percent against a historical 65 to 80 percent range students reproduced the convoluted proof ChatGPT produced proctored in-person final average fell to 48.6 percent the lowest he has recorded 18 students dropped 9 did not appear 19 failed Brown asked him to file individual complaints — insidehighered.com · July 8, 2026

Why this matters: If you hire, or sit on an admissions or scholarship committee, a transcript from the unproctored years encodes an unknown quantity of model output, and you are almost certainly still pricing it as a strong signal. That 47-point gap is what a take-home was hiding. Action this week: Skip the detector, because they do not work reliably, and buy yourself one unassisted observation. It costs about twenty minutes: a short live task in the actual skill, a whiteboard derivation, a live read of a pull request, or a "walk me through why you chose this approach" on work the candidate already submitted. If you teach or run training, this is a diagnostic you can run on yourself this week: take your current take-home assignment, put it through ChatGPT or Claude cold, and compare the output to what your students actually hand in. If the model's answer is indistinguishable from a passing submission, that assessment now measures the model, and no honor-code language will fix it. Redesign toward what a model cannot sit for you: in-person work, oral defense of a submitted artifact, or assessment of the drafts and the reasoning behind them. Follow this one.

insidehighered.com: Brown professor suspects most of his class used AI to cheat (July 8, 2026)
the-decoder.com: Grades dropped from 96 to 48 percent when a Brown professor made students take the exam without AI (July 2026)
thenextweb.com: Brown University AI cheating and the in-person exam (July 2026)

4. An AI reread 376 children's cases doctors had given up on, and 18 families got an answer.

Three hundred and seventy-six real children, and eighteen answers.

Families call it the diagnostic odyssey. You spend years, often more than a decade, cycling through specialists, genomic panels and multidisciplinary reviews, and plenty of people never get an answer at all. Those cases get filed as unsolvable, and the sequencing data sits on a shelf. Researchers at Boston Children's Hospital's Manton Center for Orphan Disease Research, with Harvard and OpenAI, took 376 of those closed pediatric cases, all of which had already been through commercial and institutional genomic pipelines and expert teams, and reanalyzed them with OpenAI's o3 Deep Research model.

It established 18 new diagnoses, an additional diagnostic yield of 4.8% in cases the field had given up on: ten from 100 neurodevelopmental cases, four from 61 neuromuscular cases, two from 200 sudden-unexpected-death cases, and two from 15 early-psychosis cases.

One of them is Kyra, now 28, who waited nearly twenty years to be told the name of what she has: myofibrillar myopathy.

On June 23, we ran the MIRA study, where an AI agent out-diagnosed a panel of six doctors, and the catch we put front and center was the researchers' own: the patients were not real, just clean written-up case files rather than a frightened human who forgets and downplays and arrives with three things wrong at once. This is the version where they are real. The discipline is what sells me on it.

The model made no clinical decisions and diagnosed no one. It produced evidence-linked candidate explanations, and a finding counted as a diagnosis only after expert review by at least two team members with disagreements resolved by consensus, additional testing, classification of the variant as pathogenic or likely pathogenic, confirmation by a CLIA-certified laboratory, and return of the result by the clinical team. Before it was pointed at the unsolved pile, it was run on 51 already-solved cases and recovered the correct gene and variant in duplicate runs for 48 of them.

It worked in six to ten minutes per case. It is retrospective. Eighteen is a small number. The study never measured time saved, costs, or whether a single child's care actually changed, and the authors say so. It is not a product, and it is not FDA-cleared. But 4.8% of "unsolvable" is not a benchmark score. It is 18 children.

TechInformed report Boston Children's Hospital uses AI to diagnose previously unsolved rare diseases OpenAI o3 Deep Research reanalysed 376 previously unsolved pediatric rare genetic disease cases with the Manton Center for Orphan Disease Research and Harvard producing 18 new diagnoses a 4.8 percent additional diagnostic yield validated on 51 already solved cases recovering 48 published in NEJM AI June 18 2026 model made no clinical decisions CLIA certified lab confirmation required — techinformed.com · June 2026

Why this matters: If your family is living a diagnostic odyssey, a child or an adult with a suspected genetic condition and years of inconclusive testing, there is a concrete step available today that most families do not know to ask for, and it is not this model. It is reanalysis. Genomic data is not static: genes are newly linked to disease every month, so exome or genome data that told you nothing three years ago can be solvable now against current knowledge, and this study is a demonstration of precisely that. Action this week: Ask your geneticist two questions. Can my existing exome or genome data be reanalyzed against current gene-disease knowledge, and am I eligible for the Undiagnosed Diseases Network, the NIH-funded program at undiagnosed.hms.harvard.edu that accepts applications from patients still unsolved after a standard workup? Reanalysis is often covered and does not require a new blood draw. Do not ask for o3. It is a research protocol, and you cannot book it. And if you build or govern AI anywhere the wrong answer costs something real, copy this study's protocol, because it is the bit most clinical-AI pitches quietly skip: validate on known-answer controls first so you know how often it is wrong, make the model hand you evidence-linked hypotheses and never verdicts, route every candidate through two independent human reviewers and an external certified lab, and publish the limits as plainly as they did. Send this to a family who has been waiting for an answer.

techinformed.com: Boston Children's uses AI to diagnose previously unsolved rare diseases (June 2026)
nbcnews.com: AI helped Boston Children's Hospital diagnose rare diseases in kids (June 2026)
openai.com: Diagnosing rare childhood diseases (June 2026)

» What to watch this week

Whether any other rating agency follows S&P on Oracle, and whether Oracle's next bond issue prices like BBB- or like something worse. Watch what Oracle actually has to pay to borrow next time, because that is the real number. The bond spread and the share price have visibly parted company this week, and only one of them has money on the line.
Whether OpenAI's answer to Apple's complaint denies the specific conduct or attacks the trade-secret claim itself. A general denial means a long discovery fight. An attack on whether these are trade secrets at all would be the more consequential move, because it goes at the lever employers use in place of non-competes.
Whether OpenAI's private raise or its 2027 IPO lands, since S&P's whole case rests on OpenAI's ability to pay. SoftBank already cut a share-backed loan from $10 billion to $6 billion due to lenders' hesitancy to value the company. A successful raise repairs Oracle's credit story without Oracle having to do anything.
Whether any university gives Serrano a policy that works at the scale of a whole class. The Cornell approach, keeping the take-home but grading an oral defense of it, is one structural answer that already works. Watch whether a department adopts it at scale or whether institutions keep asking individual professors to prosecute individual students.
Whether the Boston Children's protocol gets a prospective, multi-center trial, and whether anyone measures the thing this study did not: whether care actually changed. Eighteen diagnoses is the headline. Nobody has yet shown that a single one of them changed a child's treatment, and until somebody does, this stays a paper.

Tomorrow's signal lands here.