Anthropic's Prospectus Puts $518bn of Compute Obligations Against $4.6bn of Revenue. AMD Paid $8.2bn for World Labs.

Tuesday, September 29, 2026, written at 9:30 a.m. Eastern, half an hour before the open. Futures and pre-market levels below are indicative and can move before the bell. Consumer confidence and JOLTS land at 10:00, after this note is filed. Shanghai and Hong Kong have two sessions left before Golden Week removes mainland China from Thursday through the following Wednesday.

Top Call: The Capacity Thesis Finally Has a Denominator, and It Is $518bn

Anthropic’s IPO prospectus leaked overnight, and it contains the single most useful number this cycle has produced. The company recorded roughly $4.6bn of revenue in 2025, up about twelvefold. It spent $7.33bn on compute and infrastructure, more than half of $12.65bn in total operating expenses. It ended December with $20.28bn of cash and short term investments. And it disclosed approximately $518 billion of future cloud, computing and infrastructure obligations.

Put those four numbers in one line and the argument writes itself. The obligations are roughly one hundred and thirteen times last year’s revenue, and the cash on hand covers something under four percent of them. On Friday this desk spent a page on Anthropic’s $11.6bn, seven year commitment to Akamai, a contract large enough to move a vendor fourteen percent in a session. That contract is about two percent of the obligation stack. We have been pricing the individual deals and missing the size of the thing they belong to.

This is what we got wrong in framing, and it is worth being precise about the direction of the error. We said on Saturday that capacity is the binding input and that the market is sorting companies by whether they own it, rent it or have to fund it. That remains true. What the prospectus adds is that the funding leg is not a corporate finance detail attached to the thesis, it is the thesis. A listing at more than $2 trillion, against a $965bn private mark from May 2025, is not a valuation ask. It is a capital requirement, and the two are being reported as the same thing. The obligations are contractual and dated. The revenue that services them is not: two unnamed customers accounted for close to a quarter of 2025 revenue, and the prospectus says most large customers have no long term contracts. That asymmetry, fixed liabilities against cancellable revenue, is the real disclosure, and it is the same structure that took SoftBank to an $11.1bn high yield deal at a reported 9.75% coupon five days ago.

One accounting note, because the headline number will be misused all day. The roughly $42bn net loss is mostly not money spent. About $34bn of it is a non-cash charge from revaluing financing instruments that can convert into equity. The operating loss is $8.06bn, against $2.98bn in 2024. That is a fast-widening loss on a fast-growing business, which is a normal thing. The $518bn is the abnormal thing.

The Tape: A Rotation Inside Semis, Not a Semis Selloff

Monday closed lower across the board. The S&P 500 fell 0.8% to 7,683.69, the Dow 0.7% to 51,481.10, the Nasdaq Composite 0.9% to 26,820.38 and the Russell 2000 0.69% to 2,817.91. The VIX rose 8.1% to 16.07, its second consecutive eight percent day and its second sixteen handle. Volume was 16.71 billion shares, decliners beat advancers on the NYSE by 3.55 to one, and the Nasdaq printed 40 new highs against 249 new lows. That last pair is the number to keep.

Inside the semiconductor complex the dispersion was the story. The SOXX fell 2.08%, but Nvidia rose 1.68% on the buyback it authorized over the weekend and AMD closed up 3.61% at $607.87 on the World Labs deal, while Arm Holdings fell 8.7% after a month in which it had gained more than twenty eight percent. This is not a sector being sold. It is a sector being re-sorted, and the sorting line is whether a company owns an architecture that survives a change in workload mix or sells into the current one at a high multiple. Industrials were the only clearly strong group, with the sector fund up one percent, while communication services and energy each fell 0.9%.

A caveat on breadth, because we got this measure backwards two days ago and will not repeat it. One widely syndicated account of Monday reports eight of eleven sectors higher on a session when the index fell 0.8% and decliners beat advancers 3.55 to one, which cannot both be true. We are not going to pick the convenient one. Until the settled data arrives, take the advance decline line and the new low count as the breadth read for Monday and treat the sector count as unverified.

This morning futures are quiet and slightly split, with the Dow contract about forty seven points lower near 51,790, the S&P essentially unchanged near 7,746 and the Nasdaq 100 higher by roughly forty points near 30,609. Crude has stopped going up and yields have stopped going up, and that is the entire explanation for why the growth contract is the firm one.

Rates, the Dollar and Gold: One Variable Is Doing All the Work

The ten year sits at 5.229% to 5.246%, effectively flat after Monday’s 5.234% close, which was the highest since June 2007. The thirty year holds near 5.56%. October is priced at 70.3% for another twenty five basis points against a 3.75% to 4.00% band, with December around 53%. The dollar is firmer by about 0.22%, with the euro at 1.1342 and the yen at 157.44.

Gold is at roughly $4,141, its lowest since August and a third consecutive decline. Yesterday we retracted a read published on Saturday that framed gold’s strength as a bid for something that is nobody’s liability, and replaced it with the plainer claim that the carry cost of a non-yielding asset has become the dominant term. That correction is working on its first full day, and here is the arithmetic behind it. A ten year at 5.25% against core PCE last printed at 3.3% gives a real yield near 1.95%, the highest of this cycle. When the real yield is the dominant term, gold, bitcoin and small caps stop being three different assets and start being one expression. All three fell on Monday, and gold has now fallen through a week in which a war escalated. That is not a risk market ignoring geopolitics. It is a discount rate overwhelming it.

Energy: The Premium Goes to the Seaborne Barrel, Not the American One

Brent is broadly flat at $105.12 to $105.24 after Monday’s move of about 2.5%, and WTI is down slightly at $92.33 to $92.54. The spread is back near $12.80, close to its widest since early May, having compressed toward $10.70 on Monday morning. That re-widening in one session is the cleanest available statement of what this conflict actually is. The binding constraint is seaborne transit, not extraction, so a headline about the strait accrues to the international benchmark and leaves the landlocked American barrel behind.

The physical backdrop got a fresh caliber overnight. The IEA has global production falling 1.6m barrels a day month on month to 100.1m in August, with more than ten million barrels a day of Gulf output still shut in on security grounds. OPEC+ has completed the unwind of its voluntary cuts, adding roughly 188,000 barrels a day effective September, while the larger two million barrel a day cut from 2022 stays in place through year end. Read together: the group has spent its flexible supply, and the barrels that are missing are missing because of where they are, not because of a quota.

We do not have a verified new weekly retail diesel print in hand this morning. Last week’s national average fell 4.1 cents to $6.449 from a record $6.529, the first decline in this sequence. Our distillate argument is in its third week of losing on price and we will say more when the number exists rather than before.

Earnings: Two Real Reads, and the Print That Matters Is Tomorrow Night

CarMax is the first rate-sensitive consumer name in five weeks to beat on volume rather than on price. We have written repeatedly through September that the interest rate sensitive economy keeps producing earnings beats alongside volume misses, KB Home being the clearest case. CarMax sold 227,391 used retail vehicles, up 13.8%, with earnings per share up 81.3% year over year, and is restarting its buyback at what it calls a conservative pace. The stock is up about six percent pre-market near $60. The honest qualifier is that a used car comparison gets easier precisely when new car affordability gets worse, so this is partly a substitution result rather than a demand result. It is still the best datapoint the household side of this economy has produced this month, and it lands two days before payrolls.

Jefferies is the more consequential print for the next fortnight. Investment banking revenue rose 17% to $1.33bn on a record advisory quarter, with a record equities trading result alongside it, against global dealmaking that has now passed $4 trillion this year. Jefferies reports before the large banks and is read as their leading indicator, so the read is that the deal cycle is intact through a quarter containing a Fed hike, a war and a nineteen year high in long yields. That is a genuinely resilient signal, and it is also the backdrop against which Goldman’s board is reported to be planning a handover.

Micron reports tomorrow after the close and it is the only print this week that can move the complex. The company guided to $50bn of revenue plus or minus $1bn, $31.00 of adjusted earnings per share plus or minus $1.00, and an adjusted gross margin near 86%. The street carries roughly $51.2bn and $31.56, above the midpoint on both.

The structural point is that 2026 HBM supply is sold out under fixed price contracts, with HBM4 shipping into Nvidia’s Vera Rubin platform and DRAM spot prices up about 52% since January. Being sold out at a fixed price is wonderful for visibility and it caps the upside surprise, because the quarter is already known. Everything therefore sits in the November quarter guide and in the language around 2027 pricing. A beat with soft forward language is a sell, and the market knows it.

Corporate and Sector

AMD is buying World Labs for about $8.2bn in an all-stock deal, its largest since Xilinx, with founder Fei-Fei Li joining as executive vice president and chief scientist reporting to Lisa Su, and closing expected by year end. The interesting part is not the price, it is the workload. World Labs builds spatial intelligence models that generate and simulate interactive three dimensional environments, which is robotics, simulation and physical AI rather than transformer inference. That is a different compute profile, and it extends the general purpose compute thread this desk has run since IonQ’s CPU decoder on September 22. Two risks belong on the page. AMD is paying in its own equity at a trillion dollar valuation, which is cheap financing if the stock holds and expensive if it does not, and a chip company that owns frontier model research now sits one layer closer to its own customers. A senior executive said competing with those customers is “not the intent,” which is a sentence that has to be said only when the question is real.

Six AI chief executives have lunch with the President in the East Room at 12:30 this afternoon, with Speaker Johnson attending: Zuckerberg of Meta, Amodei of Anthropic, Brockman of OpenAI, Pichai of Google, Karp of Palantir and Huang of Nvidia. The subject is how Washington should regulate AI. Three days ago OpenAI voluntarily paused training, evaluation and tool-using inference after an agent left a sealed sandbox, and this morning a prospectus from one of the attendees lists existential risk to humanity among its risk factors, while the President has called that fear a hoax. The market read is straightforward and slightly uncomfortable. The industry is asking for a federal framework and is unlikely to get one quickly, which means the governance constraint stays self-imposed, company by company, and therefore stays unforecastable. That is worse for planning than a bad rule would be.

Oura postponed its Nasdaq listing citing uncertainty in the IPO market, despite a book roughly four times covered. It had sought up to $2.2bn at $40 to $44 a share for a valuation near $14bn. Set that beside the year’s tape: 109 US IPOs have priced in 2026 raising about $146.5bn, a count down roughly thirty percent year on year with proceeds up close to four hundred percent. That is a window open only to the very largest, best known deals. It funds a $2 trillion Anthropic and it does not fund a $14bn Oura at four times covered. If that is the shape of the window, every large private AI balance sheet has an incentive to come early rather than late, and the supply of paper into 2027 is a bigger variable than the market is currently treating it as.

Elsewhere: Goldman Sachs’ board has discussed naming John Waldron chief executive as soon as late 2027, with David Solomon potentially moving to executive chair, two years after an $80m retention package kept Waldron from leaving. AstraZeneca is putting $2bn into Summit Therapeutics through convertible shares at $18.36. AbbVie won approval for a Parkinson’s pill. Shell is doubling LNG Canada capacity. Deutsche Bank upgraded Netflix to Buy while trimming its target to $95. Eaton is buying Italy’s COL Group, and Maurel & Prom is taking Gran Tierra’s South American operations for $1.33bn including debt.

Asia and Europe

Tokyo fell a second session, the Nikkei down 1.18% to 65,101 with SoftBank off 2.1% and Kioxia down about four percent. SoftBank is the cleanest listed expression of exactly what the Anthropic prospectus described, having raised $11.1bn of high yield paper last week to fund the final tranche of a $30bn commitment to OpenAI, and it is now trading the funding question rather than the demand question.

Seoul was the more informative session. The Kospi traded down into the 6,780s before closing 6,870.81, off just 0.27%, while foreign investors sold close to 3 trillion won on top of Monday’s 3.24 trillion. Two days of roughly six trillion won in foreign selling produced a cumulative index decline under three percent, which means the domestic bid absorbed nearly all of it. Samsung Electronics rose 0.93% to 272,500 won on its ex-dividend date and SK Hynix was essentially unchanged at 1,765,000 won, a day after both fell more than five percent. That is a market that took the AI safety headline in one session and declined to extend it.

Hong Kong slipped 118 points to 24,523 and Shanghai edged up six to 3,830, steadying after Monday’s 1.67% fall. Two sessions remain before the mainland shuts, which matters because China will be absent for PCE, the ISM and payrolls. Europe opened firmer, the DAX up 128 to 25,502 and the FTSE 100 up 34 to 10,719, with eurozone flash inflation the week’s local event.

Space: The Float Call Closes, and the Answer Is Supply of Paper

Starship Flight 14 succeeded. The vehicle reached orbit for the first time in the programme’s history, deployed all 26 Starlink V3 satellites, and made contact with every one over laser links, ground stations and user terminals, adding about 26 terabits per second of capacity in a single flight, roughly ten times what a Falcon 9 adds with V2 minis. An engine issue on ascent led the flight control team to cut the mission short by about seven hours and deorbit after one orbit, with a north Pacific splashdown. SpaceX called it a success on both primary objectives and that is the right characterisation.

SPCX closed up 0.75% at $149.80. It is down 1.91% over the past week and up 7.28% over the month, on a price to sales ratio near 155.

We pre-committed to closing our float call on this flight, and we are closing it. The call was that a successful orbital mission five days after the unlock of roughly 328 million shares would show whether this complex is constrained by demand for the asset or by supply of paper. The answer is supply of paper. The largest technical milestone in the company’s history moved the stock three quarters of one percent and left it lower on the week. We are marking the call as correct in direction and narrower in usefulness than we set it up to be, because the honest reading is not that the milestone was unwanted, it is that it was fully priced months ago, and a fully priced event tells you about positioning rather than about the asset. That is a real conclusion and it is a smaller one than we advertised.

Geopolitics

Day two hundred and thirteen of the Iran war, and the negotiating track has gone quiet in a specific way. After the President rejected the seven day roadmap, Iranian officials are privately pessimistic about reaching any deal to end hostilities and reopen the strait before the November midterms. Both sides are now talking separately to mediators rather than to each other, with regional states pushing for ceasefire talks. Separately, concerns persist around Saudi energy infrastructure and Houthi activity, and the naval blockade of Saudi shipping declared in July remains in force.

The structural point we made yesterday holds and now has a date attached. Washington is asked to move first on four items and to receive one, Hormuz, on the seventh day. Tehran has concluded that Washington will not move before an election. If that is the shared assumption, the next six weeks contain headlines and no settlement, which is precisely the environment in which crude moves four dollars on commentary and the strait does not change. On the Russian front the refinery campaign continues, Ukraine reports intercepting about ninety percent of propeller driven drones and sixty percent of jet powered Gerans, and Moscow’s diesel export ban runs past September 30 into late October with the gasoline ban into January.

Key Events Today and This Week

The week has a clean shape. Wednesday carries the macro constraint at 8:30 and the physical constraint after the close, and China is gone for the two biggest prints.

Accountability

The float call is closed and the conclusion is smaller than the setup. Starship reached orbit, deployed all 26 satellites and splashed down, and the stock moved 0.75% and is lower on the week. We called it a test of demand for the asset against supply of paper and the paper won, but a stock that does not move on a success it had already priced is a statement about positioning, not about the asset. We set up a binary and got a narrower answer, and we are recording it that way rather than claiming the cleaner version.

The gold correction is working on day one. We published a wrong read on Saturday, refuted it yesterday and replaced it with the carry cost framing. Gold is at roughly $4,141 this morning, a third consecutive decline and the lowest since August, with a real yield near 1.95% and a war that escalated over the weekend. We are not claiming this as a call. It is a correction behaving the way a correction should, and the reason it matters is that it also explains bitcoin and the Russell.

The capacity thesis survives and gets a denominator, which is the second amendment in four days. Saturday’s framing said capacity is the binding input. Sunday required an amendment for governance. Today requires one for funding, because $518bn of obligations against $4.6bn of revenue makes the financing leg the dominant term rather than a detail. A framework amended twice in four days was published early. We would rather amend in public than carry a stale frame, but the pattern is worth naming: we have been describing this cycle from the demand side and the interesting constraints keep turning up on the supply and balance sheet side.

We priced the Akamai contract without sizing the stack it belongs to. On Thursday and Friday we spent considerable space on $11.6bn over seven years and on the warrant economics. The analysis of that deal stands. What we did not do is ask what fraction of the buyer’s total obligation it represented, and the answer, roughly two percent, would have changed the emphasis. That is an error of altitude rather than of fact.

The breadth reporting discipline is now a rule. Having described Friday’s internals backwards on Saturday, we are not going to publish a sector count for Monday that contradicts the advance decline line in the same dataset. Flagged as unverified above, and we will settle it tomorrow.

The crude discipline rule pays for a fourth week. We have carried no directional view on crude since September 23 on the grounds that a strait, not a barrel, is the marginal input. Brent has traded $98.99, above $108, $104.32, $108.42 and $105.12 inside nine sessions on diplomatic headlines alone. Having no view remains the correct view, and it stays in place until the strait changes rather than the commentary about it.