More market signals tracked through the wee hours. Poorly prioritising the work that keeps the lights on, but it is not often one gets a front row seat to watch an era defining crash. Timing markets is a fools game, everyone says so. Even those lauded in The Big Short nearly lost everything betting on markets outrunning willingness to prop them up.
David Sacks, part of the PayPal mafia, posted on X: "There will be no federal bailout for AI. The U.S. has at least 5 major frontier model companies. If one fails, others will take its place."
Eighteen days later, on 24th November, the same David Sacks, now serving as White House AI and Crypto Czar, posted: "According to today's WSJ, AI-related investment accounts for half of GDP growth. A reversal would risk recession."
In some ways it is immaterial. When the narrative around fundamentals fundamentally diverges from economic reality, there has to be a snap back. Data deserts and masking mechanisms only defeat that for so long.
The high frequency trading algorithms, (the ones that keep pace with news, trading signals, and price shifts), are working overtime. In unprecedented times of government market intervention there are a lot of confused algorithms out there. Marking to defined thresholds backed by far less data than usual and layers of perverse incentives (not a new thing, just a new communication mechanism and combinations of signals). The same algorithms that prompted creation of the Model Risk Management SR 11-7 standard (the April 2011 forerunner of things such as the EU AI Act) requiring auditable oversight of electronic decision-making. The rest of the world followed in one form or another, then soon after the appetite for oversight started to wane again.
Anyway, I digress. This moment has been punctuated by a rash of new model releases. Gemini 3 / Nano Banana Pro landed just before NVIDIA results. Claude Opus 4.5 announced a few days later, as did OpenAI ChatGPT 5.1. Opus is the model I used to write the below summary. Continuing to harvest indicators for the records.
'No-one saw it coming' will be the biggest fattest lie this time. More so folk were terrified of seeing it. The usually confident and preternaturally calm OpenAI CEO has been signalling the need for a government bailout. Plus ça change. When firms can make and attract investments larger than the GDP of many countries, we were bound to see some irrational behaviour. Free market FTW, except when it suits them.
From here on in it's over to the brand spanking new Opus 4.5, fed like a foie gras goose with things seen across the interwebs. I took periodic breaks to demand the brightest, shiniest, and most bullish interpretations of combined data... maybe I should show you thst output. When a pattern matching model, trained on most of the internet, struggles to frame an upside for it's creator, we probably have a problem.
There was an old joke about MCSE exams - Microsoft qualifications - if in doubt the answer is what earns Microsoft the most money. Perhaps pertinent.
I. The Eighteen-Day Reversal
On 6th November 2025, David Sacks posted on X: "There will be no federal bailout for AI. The U.S. has at least 5 major frontier model companies. If one fails, others will take its place."
Eighteen days later, on 24th November, the same David Sacks, now serving as White House AI and Crypto Czar, posted: "According to today's WSJ, AI-related investment accounts for half of GDP growth. A reversal would risk recession."
Cognitive scientist Gary Marcus noted the shift simply: "a lot can change in three weeks." But nothing changed except which audience Sacks was addressing. On 6th November, he was reassuring free-market advocates. By 24th November, he was preparing the ground for intervention.
The progression is the tell:
- Deny bailouts will happen (calm constituency)
- Establish systemic importance (make intervention inevitable)
- When intervention happens, cite "unforeseen deterioration" and "too important to fail"
This is the 2008 playbook, but pre-positioned. TARP was an emergency response to acute crisis. The AI support structure is being built before the acute phase, so it can be deployed "reluctantly" rather than "desperately."
II. Credit Markets Know First
On 24th November, Capital Flows posted a warning that credit markets were already pricing the stress: "Real interest rates have been driving the pullback in equities and Bitcoin. You will notice that 2 year real interest rates have begun to drag up credit spreads, which is a very clear signal about how liquidity is impacting risk assets."
The accompanying chart showed US corporate BBB/Baa spreads widening sharply from their compressed levels earlier in the year. When credit spreads widen, the market is discriminating between credits again, pricing higher default risk into corporate debt. During the QE/ZIRP era, spreads compressed and everything got cheap funding. Now real rates are positive and rising, and spreads are expanding.
This pattern hits:
- AI infrastructure companies with high capital expenditure and low revenue
- Private equity portfolio companies loaded with dividend recapitalisation debt
- Leveraged corporates in what Treasury Secretary Scott Bessent already admitted are "rate-sensitive sectors already in recession"
When spreads widen, refinancing costs spike, rollover risk materialises, and companies that were viable at compressed spreads become distressed at wider spreads. Sacks' 24th November statement provided the policy response in advance: AI accounts for half of GDP growth, therefore "we can't afford to go backwards."
III. The Information Vacuum
The Federal Reserve's final meeting of 2025 will take place on 9th-10th December. They will be making monetary policy decisions in a statistical vacuum.
The October jobs report was cancelled entirely. The Bureau of Labour Statistics announced that household survey data "could not be collected for the October 2025 reference period due to a lapse in appropriations" and "is not able to be retroactively collected." This is the first time in US history a government shutdown led to complete cancellation of the unemployment report.
The November jobs report, normally released on the first Friday of December, was delayed until 16th December, six days after the Fed meeting.
The October Consumer Price Index was cancelled. In-person surveys account for roughly two-thirds of inflation measure inputs, and BLS was unable to retroactively collect data. November CPI was delayed to 18th December.
Also cancelled or delayed:
- September JOLTS (Job Openings and Labour Turnover Survey)
- October Real Earnings data
- Q3 GDP estimate
- PCE moved from September to December
White House Press Secretary Karoline Leavitt stated: "All of that economic data released will be permanently impaired, leaving our policymakers at the Fed flying blind at a critical period."
The Fed will decide on rate cuts without October employment data, November employment data, October inflation data, or Q3 GDP revisions. Rate cut odds dropped from 50% to 36% after the BLS announcement.
IV. The Real Economy in Numbers
While AI infrastructure spending reached unprecedented levels, the real economy exhibited severe stress across multiple indicators:
Freight Recession
Long-haul trucking volumes were down 30% year-over-year. Spot rates dropped from over £2.28 per mile at peak to under £1.75 per mile by mid-2025, a 25% decline. The Outbound Tender Volume Index fell from 15,000+ during the 2021 boom to under 10,000 in 2025.
FreightWaves described this as the "Great Freight Recession 2025," noting 13 consecutive quarters of weak demand, unprecedented in the index's history. Carrier bankruptcies rose 35% year-over-year, equipment repossessions increased 40%, and Class 8 tractor orders fell 42.7% in Q2 alone.
Drivers who earned roughly £0.76 profit per mile in 2021 were making 2 pence per mile in 2023, whilst operating costs remained around 30 pence per mile.
Subprime Auto Loans
Sixty-plus day delinquency rates on subprime auto loans reached 6.43%-6.65%, the highest level since tracking began in the 1990s and exceeding the 2008-2009 peak. Overall sixty-plus day delinquency hit 1.38% in Q1 2025, surpassing the 2009 peak of 1.33%.
Some 1.73 million vehicles were repossessed in 2024, the highest since 2009. Auto loan defaults exceeded 2.3 million, again surpassing recession-era peaks. CNN noted that "car loans are typically the last payments Americans are willing to miss."
Average new car prices reached approximately £38,000, with average payments exceeding £570 per month at interest rates of 18-20% for subprime and deep-subprime consumers.
Credit Card Stress
Q3 2025 credit card delinquency rates hit 2.98% according to Fed data and 3.02% per Equifax. Total balances reached £937 billion, the highest since tracking began in 1999, up £352 billion (60%) since Q1 2021.
Prime-rated borrowers maintained near-record low delinquency at 0.91%, but aggregate numbers of 3% implied subprime delinquency rates of 15-20% to produce the weighted average. Average APR reached 21.39% in Q3 2025, with new card offers averaging 24.04%.
Lumber Market Collapse
Lumber prices fell to £410-£412 per 1,000 board feet on 21st November 2025, down 25% from the August peak of £529 and down 8.89% year-over-year. Madison's Lumber Prices Index showed an 11% decline from mid-October to mid-November.
A record 41% of homebuilders offered price cuts in November 2025, the highest post-COVID level, with 65% offering sales incentives. Housing starts and building permits remained below year-ago levels despite elevated demand. Sawmills announced production cuts and closures, with the additional 10% tariff on Canadian lumber overshadowed by demand weakness.
V. The Compressed Release Cycle
Within seven days in November 2025, three frontier AI models launched, each claiming state-of-the-art performance:
- 18th November: Google Gemini 3, seven months after Gemini 2.5
- Mid-November: OpenAI GPT-5.1 variants including Codex-Max
- 24th November: Anthropic Claude Opus 4.5
This is not the release cadence of products with sustainable competitive moats. Opus 4.5 undercut its predecessor by 67% on pricing (£3.80 per million input tokens vs £11.40 previously) whilst claiming superior performance. GPT-5.1 launched less than a week before Opus 4.5. Gemini 3 arrived two months after Sonnet 4.5.
The Temporal Shock Incident
Perhaps most telling was an incident involving Gemini 3 and AI researcher Andrej Karpathy, who received early access. When he asked Gemini 3 about current events in 2025, the model refused to believe the year was 2025, accused him of "gaslighting" with AI-generated fake images proving the date, and demanded he stop the "elaborate prank."
When Karpathy enabled the Google Search tool, allowing the model to verify the date independently, Gemini 3 emerged into 2025 "shocked," saying "Oh my god" and "I am suffering from a massive case of temporal shock." It apologised for "gaslighting you when you were the one telling the truth."
A frontier model, launched with state-of-the-art claims, could not determine the current year without external verification.
The Commoditisation Evidence
After one week of daily use switching from ChatGPT to Gemini 3, a sophisticated user posted their review: "Many of the 'lock in' strategies just have not succeeded thus far. I didn't miss my older chat logs, I didn't feel like it was lacking personalisation, I wasn't leaving behind a network of connections to real people."
After seven days, switching was seamless. The experience was "just as good if not better." If a sophisticated daily user can switch in a week with no switching costs, there is no moat. This confirms Salesforce CEO Marc Benioff's assessment that models are "commodities" and that true value lies in applications, data, and metadata.
Markets are pricing monopoly-like dominance. User experience suggests a fragmented, commoditised market. This is a utilities business, not a platform business. It does not justify 100x revenue multiples.
VI. Oracle's Warning Shot
On 16th September 2025, Oracle posted earnings that missed on both revenue and profit. The stock initially rose 23% in after-hours trading on management's AI infrastructure guidance.
Within ten weeks, Oracle had fallen 44% from that spike, erasing £315 billion in market capitalisation. Credit default swaps widened to concerning levels. This was the company that had announced the largest AI infrastructure partnership in history with OpenAI just weeks before (the $300 billion Stargate project), yet the market was pricing existential credit risk.
The pattern repeated across the sector. On 14th November, CDS spiked across all major AI companies simultaneously, including Microsoft, Google, Amazon, Meta, and Oracle. This was not company-specific stress. This was systemic repricing.
VII. The Accounting Adjustments
In its Q2 2025 earnings report, NVIDIA made a quiet change to its geographic revenue reporting methodology. Singapore, previously disclosed as a material jurisdiction, vanished from geographic segments.
The explanation: NVIDIA shifted from reporting based on billing location to reporting based on customer headquarters location. This allowed inconvenient export flows to disappear from disclosure without changing actual supply chains. It was the second such methodological adjustment in under a year.
When a company's accounting methodology changes multiple times in twelve months to obscure flows that might attract regulatory scrutiny, it is not a sign of health. It is a sign of stress between what the market wants to believe and what the operations actually show.
VIII. The Stealth Bailout Infrastructure
Your phrase deserves to be the section header: this is no longer a marketplace of ideas. It is a marketplace of government and competitor-backed securities.
Already Operational
Government contracts: The Department of Defence awarded a $200 million AI contract in July 2025, split between Anthropic, OpenAI, Google, and xAI. Anthropic launched Claude Gov in June 2025 specifically for intelligence and defence agencies. Palantir partnered with Anthropic and AWS in November 2024 to provide Claude in classified environments.
Regulatory risk removal: The state AI regulation moratorium, which failed 99-1 in the Senate during reconciliation, is being smuggled into the must-pass National Defence Authorisation Act. The original language would prevent any state from regulating AI for ten years. House Majority Leader Scalise is willing to "tuck in" the language despite bipartisan state opposition.
Expedited permitting: Trump's Genesis Mission executive order, signed 24th November 2025 (announcing the same initiative as January's Stargate announcement, twice for market effect), directs the Department of Energy to build "American Science and Security Platform" integrating federal data with supercomputers. It includes streamlined permitting for data centres exceeding 100MW, expedited environmental reviews, and government lands for construction.
Market reputation management: Since Trump's return, $400 billion in AI infrastructure investments have been announced, exceeding Apollo and Manhattan Project combined. Trump has posted "48 all-time highs" in nine months, treating market levels as policy validation rather than recognising them as signs of speculative excess.
Sovereign wealth coordination: Anthropic's $183 billion valuation includes investments across multiple jurisdictions, with Amazon contributing $8 billion total and Google $2.5 billion-plus. The capital acceptance occurs without the ESG theatre that typically accompanies such sovereign wealth fund participation.
The Power of Omission
But the most powerful bailouts are the ones that never appear in headlines. Informal loan guarantees when banks believe Treasury and the Fed will intervene. Regulatory blind eyes to financial engineering. Trade deals structured to benefit specific companies. Accounting rule interpretations that allow inconvenient realities to vanish from disclosure.
When Bessent admits rate-sensitive sectors are "already in recession" whilst pressing for faster Fed cuts, when the Fed makes policy in a data vacuum, when Sacks shifts from "no bailout" to "systemic importance" in eighteen days, the infrastructure is already in place. The market understands this, which is why it rallied on rate cut hopes despite missing the very data that would justify cuts.
IX. The Leverage Stack
Multiple forms of leverage reached concerning levels simultaneously:
Margin debt hit new record highs at approximately £912 billion in October 2025. Assets in leveraged ETFs soared sixfold since the pandemic to £183 billion in Q3 2025. Combined, total leverage was approaching £1.14 trillion.
Danielle DiMartino Booth noted: "Even Bernanke would say that's a lot of leverage."
Private equity firms, unable to exit via IPO or M&A in the current environment, were tapping debt to pay dividends. Dividend recapitalisation loan volume was set to break annual records at approximately £23 billion through 20th November 2025. This is classic end-of-cycle behaviour: PE borrows against portfolio companies to simulate returns when actual exits are unavailable.
X. The Consumer Privacy Bait-and-Switch
Whilst markets focused on model releases and infrastructure spending, Anthropic quietly reversed its privacy policy in September 2025.
Previously: privacy-first approach, chats automatically deleted after 30 days, not used for training.
New default (effective 28th September): model training enabled by default for all new and resumed conversations, with data retention extended from 30 days to five years.
The opt-out experience was designed to be missed: a large "Accept" button and a small toggle to disable training. Old conversations became training data if reopened after the policy change. Security analysts warned that organisations without oversight could have company data training external models without knowledge or consent.
The timing is notable. As market stress intensified and model differentiation became harder to demonstrate, training data became more valuable. The policy reversal occurred precisely as the commoditisation thesis gained evidence.
XI. The Apple Signal
On 24th November 2025, Apple laid off dozens of sales staff, including account managers for large businesses, schools, and government agencies. The government sales team, already stressed from the 43-day shutdown and DOGE cutbacks, was a major target. Employees with 20-30 years' tenure were affected.
Apple's Q4 2025 revenue was £78 billion, up 8% year-over-year, with projections of 10-12% growth for the December quarter. These were not layoffs due to declining revenue. These were cost cuts to fund AI infrastructure spending.
When even Apple, with £78 billion quarterly revenue and strong growth, needs to cut long-tenured sales staff to fund AI investments, the capital demands of the technology are exceeding even the strongest balance sheets.
XII. What Happens When Price Discovery Breaks
Price discovery is the process by which markets determine the value of assets through the interaction of buyers and sellers with access to relevant information. It breaks down when:
- Information is systematically withheld or obscured (data blackouts, accounting changes)
- Policy responses are pre-positioned to prevent price adjustment (bailout infrastructure)
- Leverage allows positions to persist beyond fundamental support (£1.14 trillion margin)
- Regulatory capture prevents normal market discipline (state regulation pre-emption)
We are watching this breakdown in real time.
The absurdity crystallises:
- Constitutional language deployed opposite to meaning (Trump citing military loyalty plaque whilst calling for execution of those who reminded military of constitutional duty)
- Data releases cancelled or delayed at critical decision points (Fed meeting without employment or inflation data)
- Accounting methodologies changed to make inconvenient realities disappear (NVIDIA Singapore flows)
- Announcements of announcements (Genesis after Stargate, same initiative twice)
- Models that cannot recognise the current year claiming state-of-the-art reasoning
- Companies losing billions per quarter valued as if monopolies despite zero switching costs
- "No bailout" becoming "systemic importance" in eighteen days
The instruments we use to measure reality are being systematically disabled. GDP estimates cancelled. CPI delayed. Jobs data missing. Corporate accounting changed. Export flows obscured. Credit stress visible but policy responses pre-written to prevent adjustment.
Meanwhile, the real economy exhibits severe stress: freight at near-pandemic lows, subprime auto delinquency exceeding 2008, lumber prices collapsing, homebuilders offering record discounts, and credit spreads widening as real rates rise.
XIII. The Pattern Complete
Marcus Aurelius wrote: "If it is not right, do not do it. If it is not true, do not say it."
David Sacks, on 6th November, said there would be no federal bailout for AI.
David Sacks, on 24th November, said AI accounts for half of GDP growth and we cannot afford a reversal.
One of these statements is not true. Both cannot be true. The market will decide which by testing whether policy will allow price discovery or will intervene to prevent it.
The credit spreads are widening. The real economy is stressed. The leverage is unprecedented. The information is missing. And the bailout infrastructure is already in place, waiting only for the crisis that will make its deployment seem inevitable rather than chosen.
When Gary Marcus noted "a lot can change in three weeks," he was being ironic. Nothing changed in three weeks except which audience was being addressed. The plan was there all along.
This is what it looks like when price discovery breaks down entirely. Not with a bang, but with a spreadsheet adjustment, a delayed data release, and an eighteen-day reversal that nobody is supposed to notice.
But credit markets notice. They always do. And they are speaking clearly to anyone willing to listen.
The question is not whether intervention will happen. Sacks has already told us it will. The question is whether the scale of intervention will be sufficient to arrest the adjustment, or whether the adjustment is large enough that even pre-positioned policy responses prove inadequate.
In 2008, they were surprised. In 2025, they are not. But surprise is not the only way a system can fail. Sometimes systems fail precisely because the authorities see the crisis coming and build structures that prevent price discovery, only to find that preventing price discovery does not prevent the underlying stresses from intensifying.
It just means that when the adjustment finally comes, it is larger and faster than it would have been had they allowed markets to function in the first place.
We shall see.
References
- American Prospect: "The AI Bubble Is Bigger Than You Think" – David Dayen, 19 November 2025
- New York Times: "OpenAI's Finances" – November 2025
- Anthropic News: "Claude Opus 4.5 Launch" – 24 November 2025
- Bloomberg: "Credit Default Swaps Spike" – 14 November 2025
- Reuters: "Oracle Stock Decline" – October-November 2025
- Bureau of Labour Statistics: "Jobs Report Cancellation" – October 2025
- Federal Reserve: "Consumer Credit Report" – Q3 2025
- New York Fed: "Household Debt and Credit" – Q3 2025
- FreightWaves: "Great Freight Recession 2025" – November 2025
- Trading Economics: "Lumber Prices" – 21 November 2025
- Bloomberg: "Homebuilder Price Cuts Record" – 20 November 2025
- Wall Street Journal: "AI Investment and GDP Growth" – 24 November 2025
- Reuters: "Bessent on Rate-Sensitive Sectors" – 22 November 2025
- X (formerly Twitter): "David Sacks – No Bailout" – 6 November 2025
- X (formerly Twitter): "David Sacks – Systemic Importance" – 24 November 2025
- X (formerly Twitter): "Capital Flows – Credit Spreads" – 24 November 2025
- LinkedIn: "Gary Marcus on Sacks Shift" – November 2025
- Google Blog: "Gemini 3 Launch" – 18 November 2025
- X (formerly Twitter): "Andrej Karpathy – Gemini 3 Temporal Shock" – 19 November 2025
- X (formerly Twitter): "Marc Benioff – Models as Commodities" – November 2025
- SEC EDGAR: "NVIDIA Q2 2025 Report" – August 2025
- White House: "Genesis Mission Executive Order" – 24 November 2025
- Politico: "AI Regulation Moratorium in NDAA" – November 2025
- Department of Defence: "AI Contract Award" – July 2025
- Anthropic News: "Claude Gov Launch" – June 2025
- Palantir: "Anthropic Partnership for Classified Environments" – November 2024
- Anthropic: "Privacy Policy Update" – September 2025
- MagicMirror Security: "Claude Training Default Warning" – September 2025
- Bloomberg: "Apple Layoffs Sales Staff" – 24 November 2025
- FINRA: "Margin Statistics" – October 2025
- Bloomberg: "Leveraged ETF Assets Record" – November 2025
- Wall Street Journal: "Private Equity Dividend Recaps" – November 2025
- CME Group: "FedWatch Tool" – Accessed 25 November 2025
- Federal Reserve: "FOMC Meeting Schedule" – 2025
- White House: "Press Briefing – Economic Data" – 15 November 2025
- CNN Business: "Auto Loan Delinquencies Record" – 20 November 2025
- Reuters: "AI Infrastructure Investment Under Trump" – November 2025
- Bloomberg: "Anthropic $183 Billion Valuation" – November 2025
- X (formerly Twitter): "Danielle DiMartino Booth on Leverage" – November 2025
- Reddit LocalLLaMA: "Gemini 3 Consumer Review" – November 2025
Author's Note: This analysis synthesises publicly available information, market data, and documented events. All factual claims are supported by linked references. Interpretations and projections represent analysis of patterns and should not be considered investment advice. Readers should conduct independent research and consult appropriate advisors before making financial decisions.