Katie Martin, the Financial Times's markets columnist, has a name for it: “The world of one trade – AI.” The indices bear her out. Since 14 August the S&P 500 has gone nowhere. It closed on Friday at 7,811.54, up 0.33 per cent. Beneath that flat line the market has split. The NASDAQ is up 2.10 per cent. The Dow, whose thirty members include Nvidia, Microsoft, Apple and Amazon, is down 4.06 per cent. The Russell 2000, the index of America's smaller companies, is down 8.52 per cent. Even an index that holds the AI leaders has lost ground, so the rally is narrower than the S&P 500's level suggests, and the smaller companies are falling fastest. Issue 39's reading of the long bond stands. This issue asks a larger question: can the industry carrying the index pay for what it has promised to build?

The evidenced layer. Anthropic's IPO prospectus, as reported by Reuters, shows 2025 revenue of about $4.6 billion against $7.33 billion of computing costs. It also shows at least $518 billion of computing commitments over a decade with six partners, around 80 per cent of them non-cancellable. Google ($111.1 billion) and Amazon ($110 billion) hold the largest shares. One clause reads: "If our actual spend falls short, we must pay Google the difference." Two customers account for nearly a quarter of revenue. The costs are fixed for a decade, while much of the revenue, pay-as-you-go usage and monthly subscriptions, can leave at a month's notice. On Thursday the Financial Times reported that OpenAI's annualised revenue was about $50 billion at the end of September, some $20 billion below the figure investors had extrapolated. On Friday Bloomberg reported that OpenAI expects $70 billion by year-end. Anthropic's annualised run-rate is about $65 billion, but the two companies do not count alike. Anthropic books sales made through its cloud partners gross, while OpenAI books some partner sales net. Axios estimates that on a net basis Anthropic's figure would be 6 to 10 per cent lower and still about $20 billion ahead.

The capability question. Apple researchers (GSM-Symbolic, 2024) found that leading models' mathematics scores moved by up to ten per cent when only the names and numbers in a problem were changed. In this year's AAAI proceedings, Gary Marcus and Vaishak Belle argue that scaling has not cured the problem. This matters because the industry's fastest-growing product is the "agent": software that carries out tasks on a customer's behalf, often for hours at a time, with no one checking each step. A business will only pay for that if it can trust the work to be right without a person looking over its shoulder. The industry asks investors for the same act of faith, on a larger scale: revenue figures that each company calculates for itself, with no one checking each step.

The bull case, at its strongest. Anthropic's revenue grew roughly twelvefold in a single year, which is rare for a company of any size. Businesses are already paying, every day, for AI tools that write and check software. Much of the critical research tested earlier models, and each new generation performs better than the last. The cost of running these models, measured per question answered, also keeps falling. On this view, the $518 billion of commitments is not a reckless gamble. It is a reasonable bet on demand that customers are already showing.

This publication’s reading. On the first reading, revenue compounds, the IPO settles the financing and the index's concentration is earned. On the second, the industry has fixed its costs, left its revenue floating and has no moat, because a customer can switch to any capable model. The OpenAI episode is less a demand collapse than a measurement problem. The figure behind a trillion-dollar valuation moved by $20 billion depending on who did the sum. "Annualised revenue" is an unaudited snapshot, and each issuer defines it for itself, gross or net. The IPO, rumoured for before Thanksgiving, is where such numbers first meet audited standards. Leadership passing from OpenAI to Anthropic within a year is the moat argument in a single statistic. For now the bulls control the price, and the IPO is the gate. Set against a disorderly unwinding of half a trillion dollars of commitments, the Gilt market's quarrel with the Burnham government is local.

The evidenced layer. On Friday President Trump announced that Russia would supply 300,000 tonnes of diesel immediately, 500,000 in November and more thereafter. The Treasury licensed Russian diesel until 7 April 2027, a month after the President signed a law tightening sanctions on Moscow. The Iran conflict is in its eighth month, and about a tenth of global refining capacity has been lost since late spring. US diesel averages $6.23 a gallon (AAA). Diesel futures fell about 4 per cent on the news but remain more than 110 per cent above January. President Zelenskyy: "Gifts to Putin will not bring peace."

This publication’s reading. Issue 39 described two thermostats, one for equities and one for crude. Friday added a third, for diesel, four weeks before the 3 November midterms. Its effect on supply is doubtful. Russia has banned its own diesel exports since July, and 300,000 tonnes is roughly half a day of American demand. It moves the dial, not the temperature. Cheaper diesel helps the Fed's cutting cycle. But a government that waives its own sanctions to move a pump price before an election has shown how it will meet the next shock, and intervention is paid for at the long end. Europe, a diesel importer, gains on price and carries the sanctions rift.

The evidenced layer. Kemi Badenoch closed conference season with a series of pledges. She would abolish inheritance tax on family homes, add a £1 million allowance for couples, scrap the mansion tax and halve employer National Insurance for workers aged 21 to 24. The party says these are funded largely through welfare savings, but has published no costing. The IFS's Stuart Adam: “The biggest beneficiaries would be those with the highest wealth and the most valuable homes.” Andy Burnham's Liverpool speech set out reform of the triple lock from 2030, a National Care Service and a six-month cut of VAT on electricity. The PollCheck average (7 October) stands at Labour 28.4, Reform 23.1, Conservative 19.3, Liberal Democrat 9.7 and Green 9.6.

This publication’s reading. The Daily T's Camilla Tominey and Tim Stanley see pro-wealth Conservatives against a state-led Labour that is rousing but thin on fiscal substance. The steelman is that triple-lock reform is the first structural cut to age-related spending that any government has attempted. Both readings agree that nobody offered consolidation. With the right split 23 to 19 and a pact ruled out, Burnham faces little discipline from Parliament. The discipline comes from his creditors, and the OBR's verdict on 28 October is their test.

The market has already spoken. The ten-year Gilt closed at 5.44 per cent, up 7 basis points and 20 over Treasuries, double the gap in Issue 39. At 196 basis points over Bunds, the Gilt spread is the widest in the G7: 59 basis points more than France and 87 more than Italy. Different policy rates explain part of that. They do not explain why the gap widened in a week when Treasuries rallied. Paris shows how a sovereign without margin is repriced, and London is already priced above it.

The verdict, as ever, is delivered below the surface.