AI Daily: OpenAI Shows Its Loss Curve and Asks for Limits

OpenAI published its internal pretraining scores for the first time on Sunday, 22 November 2026, and its chief scientist proposed “common limits” on that measure that labs would not cross until AI monitoring catches up. On the same morning, two senior US officials spoke on television about Nvidia chips reaching China.

As we read it, the day set a lab offering to slow itself next to a market already pricing a slowdown: OpenAI’s limits proposal, the Nasdaq-100 49 points from a bear market, and filings showing customers prepaying much of the AI build. Each item links to the full piece, which names its sources.

Business Pill · THE COST OF CHECKING

An evergreen Business Engineer explainer on the cost of checking, why work that is cheap to make is not cheap to check. It teaches the general idea only and does not discuss the stories in this roundup.

AI Daily Briefing · video

Today’s stories in about a minute (22 Nov 2026).

The key insight: As we read it, Sunday’s coverage set a lab offering to slow itself next to a market already pricing a slowdown: OpenAI published its pretraining scores and proposed common limits, while the Nasdaq-100 closed Friday 19.84% below its 6 October peak and customers had prepaid about $19.3 billion to four AI builders, by our count. We do not say the threads are linked.

OpenAI Puts a Number on Progress

OpenAI published its internal pretraining scores on Sunday for the first time, in a post called “Perplexity and Intelligence” by chief scientist Jakub Pachocki. The log-perplexity of its models fell from about 0.99 for GPT-3.5 to about 0.76 for the model that started producing new maths results, about 23% by our arithmetic, and Pachocki proposes “common limits” on this measure that labs would commit not to cross until AI monitoring reaches a concrete milestone. Read: the curve and the proposal.

Pachocki wrote that “as soon as 2027 we could conceivably see” an “avalanche of fundamental pretraining advancements”; by our count, OpenAI or its models were behind 5 dated maths results in the 75 days before the post. He expects capabilities nobody can name yet in a window of about 0.68 to 0.70: “I don’t know what those capabilities will be, and we may not be testing for them today.” Read: the maths record and the next window.

We read the latest public safety frameworks we found from four rivals, and by our count 2 of the 4 softened a stop or pause commitment in 2026; Meta’s own change log shows its top risk threshold went from “Stop” to “Develop with Mitigations” on 7 April. Separately, Elon Musk wrote on Sunday that “unequivocal super intelligence has arrived,” giving no test or benchmark for the claim. Read: the rivals’ frameworks and Musk’s post.

OpenAI's stated log-perplexity on its held-out dataset fell about 0.08 points from GPT-3.5 (about 0.99) to GPT
OpenAI’s stated log-perplexity on its held-out dataset fell about 0.08 points from GPT-3.5 (about 0.99) to GPT-4 (about 0.91), then about 0.15 points to the model that began producing new maths results (about 0.76), per its post Perplexity and Intelligence. The steps and the ratio of almost two are our arithmetic.

Nvidia on the Sunday Shows

Treasury Secretary Scott Bessent said on CBS’s “Face the Nation” that the administration still needs answers from Nvidia on its chip sales to China and wants to know who was in charge of compliance, Reuters reported; by our count, that was 9 days before the former CEO’s 1 December interview with the House. Director of National Intelligence Jay Clayton said on CNN that China appears to have had access to Nvidia’s top AI chips, based on the facts alleged in a federal indictment; at least about $510 million of servers were allegedly diverted, about 20.4% of the $2.5 billion total by our arithmetic. These are allegations. Read: Bessent on compliance and Clayton’s answer.

The Wall Street Journal reported that many of Nvidia’s supplier and customer deals rested on handshake agreements, as relayed by NDTV, while Nvidia’s disclosed supply commitments reached $390 billion at the end of October, about 13.5 times the level two years earlier by our arithmetic. An X post claiming Elon Musk “may acquire a 50% stake in NVIDIA” drew about 1.0 million views; we found no filing, no company statement and no news report behind it. Read: the handshake deals and the viral claim.

On the same programme, Microsoft president Brad Smith said the company has taken nine steps after its 12 November outage and has 13 more to go, and called the month’s outages, in substantial part, “AI ripple effects”; Bessent said bank glitches came from security patches. Microsoft held $175.2 billion of long-lived assets outside the United States at the end of June, about 37% of its total, according to its annual report. Read: the outage fixes and Smith on data centers abroad.

Who Carries the Build’s Cash

Elon Musk wrote that delays between paying for compute and earning from it will mean “hundreds of billions of incremental dollars just for working capital.” Customers paid four AI infrastructure builders about $19.3 billion in advance in their latest reported periods, about 62% of the $31.0 billion of cash their operations produced, by our count from their filings. Read: Musk’s warning and the prepayments.

Nscale, planning a $35 billion New York listing, is pausing most hiring, the Financial Times reported; its filing shows a preliminary net loss of $1,691.9 million on $307.5 million of revenue in the nine months to 30 September, about $5.50 lost for every $1 of revenue by our arithmetic. In the first half of 2026, without a $3,811.5 million rise in customer prepayments, its operating cash would have been about minus $2.1 billion, by our arithmetic.

SB Energy restated two quarters after understating the warrant liability to OpenAI by $826.6 million at 30 June, its amended prospectus shows. Read: Nscale’s pause, its prepayments and SB Energy’s restatement.

KBRA assigned preliminary A- ratings to about $29.6 billion of notes, by our arithmetic on its release, backed by a lease of Nvidia Vera Rubin GPU equipment to SpaceX, and said the ratings consider the risk of the chips going out of date. Dell, HPE, Lenovo and Supermicro are demanding up to 45% more for AI servers months after deals were signed, the Financial Times reported. Read: the KBRA rating and the server repricing.

Markets Into Thanksgiving Week

The Nasdaq-100 closed Friday 19.84% below its 6 October peak, 49.40 points above the line markets commonly call a bear market, and the S&P 500 goes into Thanksgiving week up just 0.95% for 2026 after being up 14.2% at its high, by our arithmetic. Thirty-three trading days after its peak, the Nasdaq Composite was 18.05% below it, against 25.25% at the same point after the March 2000 peak, on FRED data. Read: the bear-market line, the week ahead and the 2000 comparison.

Nvidia’s market value fell to about 6.3 times its own guided annual sales pace, from about 13.4 times at its 6 October peak, by our arithmetic. Net of imported computers, AI-linked spending added about 0.41 percentage points to US real GDP growth in the first nine months of 2026, by our arithmetic on BEA data, not the 65% share circulating on X. Bessent said on CBS that “many people believe” the Fed may be making a policy mistake if it over-tightens. Read: Nvidia’s multiple, the GDP net figure and Bessent on the Fed.

In Asia, SK hynix was worth 73.5% of Samsung Electronics at Friday’s close after coming within 0.2% of it in June, and six Taiwanese chip and AI server makers reported October revenue up 64.0% while their shares sat 22.6% to 54.0% below their 2026 highs, by our arithmetic on exchange data. Read: the Korean memory gap and Taiwan’s split.

Anthropic, and Tuesday’s Earnings

A Wall Street Journal profile reports that Anthropic chief executive Dario Amodei once floated immersive video games as a way to help people who lose their jobs to AI; NewsBytes said he raised it “in a Slack chat back at OpenAI, not as an official plan.” Cathie Wood said Anthropic’s run rate “could be at 100 billion” by the end of this year; from the roughly $81 billion in its filing, that takes about $9.5 billion a month, by our arithmetic.

David Sacks said models should not be trained to think they have “moral status or feelings”; Anthropic’s constitution says “Claude’s moral status is deeply uncertain.” Read: the WSJ profile, Wood’s call and Sacks and the constitution.

Dell reports on Tuesday with a record $95 billion AI server backlog, about 5.8 times the AI servers it sold last quarter by our arithmetic; Workday has guided subscription growth to 12%, after 14.4% and 13.9% by our arithmetic; and HP’s PC revenue rose 18% last quarter while units fell 16%. Read: Dell’s backlog, Workday’s guide and HP’s PC pricing.

Bar chart: 33 trading days after its peak, the Nasdaq Composite was 18.05% below it on 20 November 2026, against 25.25% after the March 2000 peak
Nasdaq Composite, percent below its peak close 33 trading days after the 6 October 2026 and 10 March 2000 peaks, by our arithmetic on Federal Reserve Bank of St. Louis (FRED) data, as our piece reports. A comparison of two price paths, not a forecast.

The Structural Read

As we read it, the OpenAI thread is about who sets the pace: Pachocki proposes limits on log-perplexity that labs would commit not to cross until AI monitoring reaches a concrete milestone, while by our count 2 of 4 rival labs softened a stop or pause commitment in 2026, as our pieces report.

As we read it, the Nvidia thread is about governance, not demand: Bessent asked who was in charge of compliance, Clayton pointed to allegations in an indictment, and the WSJ reported handshake deals, as our pieces report, while Nvidia’s disclosed supply commitments reached $390 billion.

As we read it, the cash thread is about who carries the build: customer prepayments were about 62% of four builders’ operating cash, by our count, Nscale is pausing most hiring, the Financial Times reported, and KBRA said its ratings on SpaceX’s GPU notes consider the risk of the chips going out of date.

Jakub Pachocki, OpenAI’s chief scientist, in “Perplexity and Intelligence”, as quoted in our piece

“I don’t know what those capabilities will be, and we may not be testing for them today.”

Three Implications

READERS OF AI CLAIMS OpenAI’s common limits are a proposal; our pieces report no other lab signing it, and Musk’s superintelligence post gave no test or benchmark.

LENDERS AND SHAREHOLDERS Customer prepayments were about 62% of four AI builders’ operating cash, by our count from their filings, and Nvidia’s value fell to about 6.3 times its guided annual sales pace, by our arithmetic; this is not investment advice.

PEOPLE FOLLOWING NVIDIA The China chip case rests on allegations in an indictment, and the former CEO’s transcribed House interview is set for 1 December, as our pieces report.

The Business Engineer Lens

This story maps onto the Business Engineer framework The AI Supercycle.

The analysis says of its frame: “the term supercycle should not be interpreted as a purely economic concept. It goes far beyond economics.”

As we read it, Sunday touched more than the markets: a lab proposing limits on its own progress, a spy chief and a Treasury secretary on chips reaching China, and Microsoft’s president on data centers abroad. We do not say the analysis predicted these items. It was published in June 2026 and does not discuss them.

What Is Not Established

This roundup adds no new reporting and selects from the day’s coverage. Every figure above is taken from the linked FourWeekMBA piece, each built from its own primary source. The threads, and the reading of the day as a lab offering limits next to a market pricing a slowdown, are our grouping, not a finding.

We do not claim that any one story moved prices or caused another. OpenAI’s limits are a proposal no other lab has signed; the China chip case rests on allegations in an indictment; the WSJ handshake and Amodei reports rest on the Journal’s reporting as relayed; Musk’s superintelligence claim and Wood’s run-rate call are their own statements; the market and growth figures are our arithmetic; and nothing here is a forecast.

This Week In The AI Supercycle

A World Without an Order, and an AI Central Bank Without a Successor

This week’s issue: the week AI memory stayed largely sold out while the people holding the shares paid: why the AI Supercycle is not a dot-com replay, read through geopolitics, macro and technology, with the gap between building and financing the widest we’ve recorded.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer on how geopolitics, culture and the economy connect in the AI era.

Read the analysis →

Get the premium package at 50% off →

The Bottom Line

For anyone pricing the AI build, Sunday set OpenAI’s proposal of common limits on pretraining progress next to a Nasdaq-100 19.84% below its peak and about $19.3 billion of customer prepayments, about 62% of four builders’ operating cash, by our count.

For the frame behind the day, read This Week in the AI Supercycle on The Business Engineer: read the analysis.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. This roundup draws on FourWeekMBA’s news coverage of 21 and 22 November 2026, linked above; every story was built from its own primary source, which it cites and links. We re-read each piece on 22 November 2026 and took every figure here from it. The threads and the figures in our charts are our own grouping and selection. Nothing here is a forecast, and nothing here is financial or investment advice.

Sources: FourWeekMBA: OpenAI’s Loss Fell 0.99 to 0.76; It Proposes Common Limits · FourWeekMBA: 5 OpenAI Math Results in 75 Days, Then ‘As Soon as 2027’ · FourWeekMBA: OpenAI: Next 0.08 Points Bring Unknown Skills, Weaker Audits · FourWeekMBA: OpenAI Seeks AI Limits; 2 Rivals Dropped a Stop in 2026 · FourWeekMBA: Musk Declares Super Intelligence, 304 Days After Davos Call · FourWeekMBA: Bessent Asks Who Ran Nvidia Compliance, 9 Days Before House · FourWeekMBA: Spy Chief Says China Got Nvidia Chips, Citing a $2.5B Case · FourWeekMBA: WSJ Flags Nvidia Handshake Deals as Supply Bill Hits 13.5x · FourWeekMBA: Viral Musk-Nvidia 50% Stake Claim Has No Filing Behind It · FourWeekMBA: Microsoft Has 13 Outage Fixes Left; Bessent Says ‘Vaccine’ · FourWeekMBA: Microsoft Has $175B of Assets Abroad as Smith Answers Trump · FourWeekMBA: Musk: AI Delays Mean Hundreds of Billions in Working Capital · FourWeekMBA: $19.3B of Prepayments: 62% of 4 AI Builders’ Operating Cash · FourWeekMBA: Nscale Pauses Hiring After Losing $5.50 per $1 of Revenue · FourWeekMBA: Nscale’s $1.7B Operating Cash Rested on $3.8B of Prepayments · FourWeekMBA: SB Energy Restates 2 Quarters: OpenAI Warrants Off by $827M · FourWeekMBA: KBRA Rates $29.6B of SpaceX GPU Notes A-, Flags Obsolescence · FourWeekMBA: Up to 45% More: Dell, HPE Reprice Signed AI Server Deals · FourWeekMBA: Nasdaq-100 Ends 49 Points From a Bear Market; Nvidia Off 47% · FourWeekMBA: S&P 500 Enters Thanksgiving Week Up 0.95%, Was 14.2% at Peak · FourWeekMBA: 33 Days In, Nasdaq Is Down 18% vs 25% After the 2000 Peak · FourWeekMBA: Nvidia Trades at 6.3x Its Guided Sales Pace, Down From 13.4x · FourWeekMBA: Net of Imported Computers, AI Adds 0.41 Points to US Growth · FourWeekMBA: Bessent Warns on Fed Over-Tightening, 17 Days to Rate Call · FourWeekMBA: SK Hynix Came Within 0.2% of Samsung’s Value. Now It’s 73.5% · FourWeekMBA: Taiwan AI Makers’ Sales Up 64%, Shares Down as Much as 54% · FourWeekMBA: WSJ Profile: Amodei Floated Video Games for AI Job Losses · FourWeekMBA: Cathie Wood’s $100B Anthropic Call Needs $9.5B a Month · FourWeekMBA: Sacks Hits Claude’s Moral Status; Its Text Says ‘Uncertain’ · FourWeekMBA: Dell Reports Tuesday With $95B AI Backlog, 5.8x Its AI Sales · FourWeekMBA: Workday Reports Tuesday as Its Growth Guide Slips to 12% · FourWeekMBA: HP Reports Tuesday After Selling 16% Fewer PCs for 18% More

Scroll to Top
FourWeekMBA