OpenAI's Revenue Is $20 Billion Smaller Than Investors Believed. Lumentum Is Sold Out Through 2029.

Friday, October 9, 2026, written at 9:15 a.m. Eastern, fifteen minutes before the open. Every future, yield, currency and commodity level in this note carries that stamp and can move before the bell. Seoul and Taipei were closed for holidays. University of Michigan preliminary sentiment lands at 10:00 and is previewed below.

Top Call: The Numerator Broke Before the Cost of Capital Did

For three sessions this desk has argued that the artificial intelligence buildout’s weak point had moved to its funding. That was the wrong place to look first. What actually gave way this week was the revenue line the funding is underwritten against.

The Financial Times reported Thursday that OpenAI told investors its annualised revenue was approaching $50bn at the end of September. Investors had been carrying roughly $70bn. The chip index fell 3.4%, Micron fell 4.8%, and the Nasdaq Composite fell 1.25% to 27,193.34 two days after a record close.

The mechanism behind the gap is the part that matters, because it is not fraud and it is not a shortfall. Investors took an estimated $40bn August figure, applied the 70% growth OpenAI had disclosed, and arrived at $70bn. Nobody at OpenAI published $70bn. Separately, Anthropic counts revenue booked through cloud partners such as AWS and Google Cloud in its own run rate, and OpenAI does not. So the two largest private borrowers in this cycle are reporting on different bases, and the sell side has been comparing them as though they were not.

Put that beside what launched this week. Roughly $150bn of financing is in syndication across the Broadcom and Anthropic lease package, the SpaceX chip raise and the Broadcom and OpenAI Nexus programme. Every one of those structures is sized against a run rate that is unaudited, undefined across issuers, and in at least one case was never published by the company at all. A lender can diligence a 210 megawatt lease with an investment grade counterparty. A lender cannot diligence an annualised number whose denominator changes depending on who is counting.

The equity market reached the same conclusion from the other direction this morning, and it did so with a price. Firmus, the Nvidia backed data centre operator, withdrew its Australian listing, citing market volatility. The offer was priced at A$11 for a valuation near $30.6bn, roughly three times the $10.5bn mark from its own August round, with analysts for the deal’s managers estimating about $30bn of debt against it. Fund managers said the price did not reflect assets that are largely unbuilt. The company will now raise privately and look at a Nasdaq listing later. That is the second AI infrastructure book to close in two days.

So the hierarchy of capital has reordered itself inside one week. Public equity said no at $30.6bn. Private credit is still saying yes at $150bn. And the rescue bid is now sovereign, with SoftBank reported to be seeking up to $100bn from Gulf investors for an AI acquisition vehicle, its shares down more than 3% in Tokyo on the news. When the marginal buyer of a sector’s risk migrates from listed funds to private credit to sovereign wealth inside five sessions, the sector has not run out of money. It has run out of investors willing to price it in public.

Against all of that, the single most useful datapoint of the morning points the opposite way. Lumentum’s chief executive said optical component capacity is sold out through early 2029. Lumentum rose about 7% pre-market, Coherent and Applied Optoelectronics around 5%, and Nvidia put $2bn into each of Lumentum and Coherent earlier this year precisely to lock that supply. Memory tells the same story: Jefferies has servers at 60% to 70% of memory demand against roughly 30% before this cycle.

That is the whole note in two facts. Physical capacity is contracted out more than two years. The financial representation of the demand behind it just shrank by $20bn on a definitional dispute. We would rather own the constraint than the claim, which means optics, memory and foundry over the application layer and over anything whose equity value rests on a private mark. We are not bearish on the buildout. We are bearish on the arithmetic that has been used to size it, and this morning’s rally does not resolve that.

Rates: The Thirty Year Cleared, and We Set the Test on the Wrong Instrument

Yesterday we wrote that the 1:00 p.m. auction was the direct test of the $150bn stack, and that a tail would widen every spread product behind it. The auction was fine. Treasury’s $22bn thirty year reopening printed a 5.62% high yield with a 2.54 bid to cover against a 2.41 average and indirect bidders at 72.3% against 68%. Taken with Wednesday’s ten year, which stopped through with indirects at 80.3%, the conclusion is that duration demand at these yields is real at both tenors.

The curve agreed. The ten year is 5.26% this morning against Thursday’s 5.35% intraday twenty four year high, the thirty year is 5.63% from 5.73%, and the two year is 4.80%. Nine basis points of relief at the long end is not a trend, but it does settle the question we posed. The sovereign market is not the binding constraint. Pimco, per the FT, still sees the thirty year reaching 6%, which would be the first time since 2000, and that view deserves respect rather than dismissal given where this curve has gone since late September.

The Fed has not changed. Waller said further hikes will likely be needed with flexibility on pace. Williams said there is no need for urgency and Bowman said she sees no urgent need for further action, which is the dovish wing restating the same calendar from the other side. October is priced near 16% to 20% and December sits around 76% to 81%, with the target range at 3.75% to 4.00%. A committee that intends to hike once more but not this month is exactly what both wings are describing.

University of Michigan preliminary sentiment is at 10:00 with consensus near 47.6 against 48.1, which would be a third consecutive decline and would leave the index under three points above the 44.8 record low set in May when fuel prices peaked. The year ahead inflation expectation was 4.6% in September and the five to ten year gauge 3.4%. Our view is that the expectations series matters more than the headline today. The New York Fed’s one year measure already rose to 3.9% from 3.6% this week, and a confirming move in the Michigan series would be the second independent survey showing households repricing inflation while the committee debates one more hike.

Energy and Geopolitics: The Principal Denied the Report

Brent is $102.94 and down 1.1%, WTI $90.79 and down 0.8%, both heading for a weekly loss of roughly 1%. Brent settled at $104.28 on Thursday after a 4% gain and traded close to $106 intraday.

The reversal came from Trump himself. He posted that the United States is having productive discussions with Iran and will not attack at any time prior to the November 3 midterms. That is a direct denial of the Axios and Atlantic reporting we led with yesterday, issued by the only person whose view on it is dispositive, and it removed the dated strike window that had been worth about five dollars of Brent. The blockade stays in force and an Israeli official told Axios that attacks before the midterms are not ruled out, adding that after the midterms the chances increase significantly. So the premium has been deferred by twenty five days rather than cancelled.

Three other things are pulling crude lower and none of them is a new barrel. China is resuming refined fuel exports after Golden Week, the IEA’s accelerated releases are coming, and the dollar eased. Three things are holding the premium in: UKMTO has logged nine Hormuz attacks this month against half of September’s total for the strait and Gulf combined, the Houthis are striking Saudi Arabia, and Hurricane Isaias has roughly 500,000 barrels a day shut in across the Gulf of Mexico. CENTCOM struck five IRGC linked tankers after attempted Iranian attacks on an American warship, and Tehran is reported to be preparing a maritime exclusion zone outside the strait where it would stop vessels transiting without permission.

Putin offered to help end the war when he met President Pezeshkian ahead of a regional summit in Turkmenistan, and the EU is adding ships to its Red Sea mission. We would not trade either headline. The exclusion zone is the one to watch, because it is the first Iranian proposal that would make the legal status of transit, rather than the physical security of it, the binding constraint on flows.

Equities: Thursday Was a Chip Event, Friday Is a Relief Rally

Thursday closed split. The S&P 500 fell 0.47% to 7,765.36 and the Nasdaq Composite 1.25% to 27,193.34, while the Dow rose 51.77 points or 0.10% to 51,231.64. Energy led all eleven sectors and technology was the worst. The breadth detail is the one worth keeping: the Nasdaq posted 33 new highs against 319 new lows. A 1.25% index decline with a ten to one new low ratio is not a tech selloff, it is a tech selloff inside a market that was already thin underneath.

This morning all four contracts are higher, with S&P futures at 7,842.50 and up 0.34%, Nasdaq 100 futures 31,191.75 and up 0.72%, Dow futures 51,519 and up 0.05% and Russell futures 2,815.00 and up 0.19%. The VIX is 15.27 and down 0.9%. Nvidia and Micron are up 1.5% to 1.6% pre-market and Oracle 1.8%, after Bloomberg reported OpenAI expects to reach or exceed $70bn of annualised revenue by year end on enterprise growth.

We would be careful with that rebound. The Bloomberg follow up restores the $70bn figure as a forecast, not as a current run rate, which is a different statement from the one investors were working with on Wednesday. The gap between what a company expects to earn in December and what it is earning in September is exactly the gap that a lease or a credit agreement has to survive. Note also the leadership: the Dow was green on Thursday and the Nasdaq was down 1.25%, so for one session the rotation out of the buildout actually went somewhere rather than into cash. LSEG IBES now has third quarter S&P 500 earnings growing more than 30%, which is the strongest fundamental argument the bulls have and does not depend on any private mark.

Asia: Hong Kong Internet Bounces, the Mainland AI Chain Does Not

Hong Kong was the session. The Hang Seng rose 425 points or 1.8% to 24,211.35 on HK$209.2bn of turnover, up 1% on the week, led by internet names. The Hang Seng Tech index rose 3.1% to 4,197 and is still close to a two year low, which tells you how far that index has fallen rather than how good the day was.

The mainland went the other way under the surface. The Shanghai Composite added one point to 3,813 on ¥888.56bn of turnover and the CSI 300 rose 0.2% to about 4,317, but the index fell nearly 1% on the week for a ninth consecutive weekly decline, and the selloff in AI supply chain names deepened. The 5G Communication Index fell as much as 6% to a two month low, Zhongji Innolight fell 1%, coal and consumer staples gained. So on the same morning that Lumentum said western optical capacity is sold out to 2029, China’s largest optical module maker fell and its communications complex hit a two month low. Either the mainland supply chain is being repriced on domestic demand rather than on the global constraint, or one of these two tapes is wrong. We think it is mostly the former, and it is the cleanest illustration this week of how bifurcated the AI trade has become by geography.

Tokyo closed effectively flat at 69,030.92 with SoftBank down more than 3% on the Gulf fundraising report. Australia rose 0.6% to 8,716.60 and India 1.2%. Seoul and Taipei were shut, which means the two markets that would have had the strongest opinion on the memory and foundry story did not get to express it, and Monday’s Korean open is worth watching for a catch up to both the OpenAI headline and the Lumentum one.

Europe: France Backs Off the Edge

The Stoxx 600 is 629.96 and up 0.71%, recovering from Thursday’s close at the lowest in nearly four months. The DAX is 25,025.14 and up 0.88%, the FTSE 100 10,514.88 and up 0.70%, the CAC 40 7,782.51 and up 0.68%, the FTSE MIB 49,584.49 and up 0.58%.

The French trade has turned. The ten year OAT is 4.83% and down six basis points, the Bund 3.47% and down two, and the spread is near 135 basis points against 159 a week ago and against the 142 session high we flagged yesterday as the widest since 2012. The proximate cause is political rather than fiscal: Le Pen, the presidential frontrunner, offered reassurances on restraining spending. Lagarde’s position remains that any backstop depends on Paris staying inside EU fiscal rules, which is both correct and the reason the spread cannot compress much further on words alone. The budget debate opens October 13 and the deficit is still above 5% of GDP, so we would treat twenty basis points of compression on a campaign statement as a trade rather than a repair. The euro is back near 1.12 after Monday’s seventeen month low.

Corporate: SpaceX Buys the Spectrum, Apple Cuts the Orders

SpaceX agreed to pay Grain Management about $8bn in cash for a nationwide portfolio of up to 14 megahertz of paired 800 MHz spectrum, and the FCC separately authorised its next generation 15,000 satellite direct to device constellation. The company said the low band licences close the key remaining technical gap for Starlink Mobile to operate as a major American carrier. AT&T fell 7.3%, T-Mobile and Verizon 6.6% on Thursday, and they are down another 6% to 7% pre-market, with Deutsche Telekom off about 7% in Frankfurt on its 54% stake in T-Mobile US. SpaceX is up 3.8% to $166.70.

This is the most important competitive event of the week and it is being read correctly. When SpaceX said in the spring it would launch a mobile offering, the carriers fell 2% to 2.5%, because a satellite overlay is a coverage product. Buying cleared low band spectrum is not an overlay, it is entry. The three incumbents formed a satellite joint venture last week without Starlink, which now looks like a response written before the problem was fully understood. We would not buy the dip in the carriers on valuation alone until there is clarity on FCC approval conditions and on what Starlink Mobile actually prices at.

Apple has asked suppliers to cut October iPhone 18 Pro and Pro Max component orders by at least 15%, with some supply chain sources putting the range at 15% to 20%, per Nikkei Asia. The Pro models start at $1,199 and $1,299, each $100 above the prior generation, and Nikkei links the increase directly to the memory squeeze driven by AI data centres. The base iPhone 18 was held back to early 2027. The stock is down about 1.6% to 2% pre-market.

This deserves more attention than a single order cut normally gets, because it is the first clean evidence of the AI buildout destroying demand somewhere else in the economy rather than merely crowding out capital. Server memory went from roughly 30% of demand to 60% or 70%, DRAM and NAND repriced, Apple passed $100 of it to the consumer, and the consumer declined. That chain is now empirically complete, and it says the memory upcycle is not a free lunch for the handset and PC complex. Digitimes is already forecasting lower PC shipments on the same mechanism.

Elsewhere, American Express was fined $350m over anti money laundering controls, FICO is cutting about 15% of staff as competition in mortgage scoring intensifies, and Porsche is cutting a further 4,100 jobs under the Volkswagen overhaul. In storage, SanDisk is up more than 2% pre-market with Western Digital and Seagate up about 1%.

Earnings: Delta Prices the Oil Shock

Delta Air Lines opened the season and gave the market the first full corporate quantification of what a $100 barrel costs.

Our take: this is a cost print, not a demand print. Revenue grew 15.7% and Bastian still expects about $4.5bn of pre-tax profit while absorbing a $6bn increase in fuel. The guidance cut is almost entirely the jet fuel line. The detail that matters is the assumption: fourth quarter guidance of $1.15 to $1.65 against a $1.47 consensus assumes the forward curve as of October 2 and an all in price near $4.25 a gallon including a $0.40 refinery benefit. Brent traded above $105 after that date. So the Q4 range was built on a curve that is already stale, and unless crude keeps falling from here the risk sits at the low end. The stock is down about 5% near $78 and we think that is the right first reaction and probably not the last one.

The read across is to every transport and consumer name reporting over the next three weeks. This is now the fourth consecutive print in which the operating business is healthy and the forward number falls on an input cost, after Lamb Weston, Constellation Brands and PepsiCo. That pattern, not any single company, is the reason we are more worried about margins in the fourth quarter than about volumes.

Key Events Today and Ahead

The Michigan expectations series is the only release today that can move the December pricing. Everything else this morning is positioning into a weekend with a deferred strike window and an open Hormuz question.

Accountability

We led yesterday with a strike window that the president closed the same afternoon. Our top call framed a dated window before November 3 on the Axios and Atlantic reporting, and Trump explicitly ruled out an attack before the midterms hours later. Brent went from $105 to $103 and gave back most of the move. We have spent two weeks writing that supply side policy responses in this conflict are worth about one session. The symmetrical statement, which we did not make, is that reported intentions sourced below the principal are worth even less. We bought a report and the principal denied it. That is the clearest miss of the week and the lesson is specific rather than general.

The auction test resolved against us and we named the wrong instrument. We wrote that the thirty year reopening was the direct test of the $150bn stack and that a tail would widen every spread product behind it. It cleared at 5.62% with a 2.54 cover and indirects at 72.3%, the ten year is nine basis points lower, and by our own stated test the funding thesis is early rather than right in the sovereign market. What broke instead was the equity window, with Firmus pulling a $5bn book at a $30.6bn valuation. The thesis survives and the test did not. We were looking at Treasury when we should have been looking at the IPO calendar, and we flagged the Firmus book closing on Wednesday without drawing the conclusion from it.

The capacity call keeps getting stronger and we have no claim on it. We said after TSMC’s September revenue that the constraint is where Lisa Su and Sanjay Mehrotra keep saying it is. Lumentum selling out optical components through early 2029 is a third independent confirmation. We set that frame rather than forecast it, and we claim nothing more.

We have been describing the memory upcycle as a one sided benefit and Apple just priced the other side. For six weeks we have written about server memory demand as a straightforward positive for the memory complex. The order cut of at least 15% on the iPhone 18 Pro, traced by Nikkei straight to a $100 price increase caused by the AI memory squeeze, is the cost of that upcycle landing on a different industry’s volumes. We should have been modelling the handset and PC demand destruction as part of the same trade and we were not.

The bifurcation claim gets its first real test Monday. We are arguing that physical capacity is the asset and the financial representation of AI demand is the liability. Korea and Taiwan were closed today, so the two markets with the most direct exposure to that split have not voted. If Seoul opens Monday following Hong Kong’s internet rebound rather than the mainland’s AI supply chain selloff, the geographic read in our Asia section is wrong and we will say so.