Marvell Beats and Falls 8% on Lower Custom-Silicon Margins. Advent and Stripe Walk Away From PayPal.

Friday, August 28, 2026, written at 9:15 a.m. Eastern, forty-five minutes before the cash open and forty-five minutes before Kevin Warsh speaks. Yesterday’s tape was one of the narrowest good days of the year. Today’s has a policy event in the middle of it.

Top Call: The Argument Has Moved to Margin

Marvell printed a record quarter, beat on revenue and earnings, raised both fiscal 2027 and fiscal 2028, and fell as much as eight percent. The reason given was gross margin, and specifically that a rising share of custom AI silicon carries a lower margin than the company’s standard products.

That is the second time in three nights that an AI supplier has been marked down on margin rather than on demand. Nvidia guided its own gross margin from 75% to 71% or 72% by January on memory costs. Put the two together and the trade has changed shape. Nobody is arguing about whether the compute gets bought. The argument now is about who keeps the money.

Marvell: A Good Quarter With an Expensive Mix

Every line above is good except the last one, and the last one moved the stock. The mechanism deserves to be stated plainly, because it is counterintuitive and it generalises.

Marvell’s margin falls when it wins. Custom silicon is designed to a customer’s specification, is priced closer to cost plus a fee, and does not carry the pricing power of a merchant part sold into a broad market. So the same contract wins that produce the revenue beat and the raised outlook are the ones that dilute the margin. Nine days ago Marvell handed Google a warrant on up to 58.97 million shares, worth as much as $12.2 billion at an exercise price of $206.58, vesting in 240 tranches of roughly 240,000 shares, one for every $500 million of custom product revenue, toward a notional $120 billion of purchases running to January 2033. The stock rose more than eleven percent that day. Last night the market priced what that revenue actually earns, and it liked it less.

Our take is that this is a re-rating of quality rather than of quantity, and it is not a reason to abandon the name. Bank of America, UBS, Barclays, Wells Fargo and Citi all stayed constructive this morning. The move also looks larger in the headline than it does on the tape: the stock fell 7.6% after hours and was down about 7.9% earlier in pre-market, and by 9:15 it had recovered to roughly $241.45, off about 1.5%. A stock up around 187% on the year giving back one and a half percent on a raised guide is not a verdict, it is a mix adjustment.

Why the Mix Keeps Shifting

The reason custom silicon share keeps rising is visible in a chip that was detailed this week. OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first inference ASIC, taped out on TSMC N3P in a development cycle measured in months rather than years. The B0 stepping delivers 13.4 petaflops of MXFP4 at 700 watts against roughly 900 to 1,150 watts for Nvidia’s Rubin, pairs HBM4 at 15.4 terabytes per second, and posts more than 700 tokens per second per user on DeepSeek R1 and about 1,400 on GPT-OSS. OpenAI’s own benchmarks put it at 1.5 to 1.9 times the work per watt of Nvidia across three model families, and SemiAnalysis, which visited the labs, reported it beat Blackwell on performance per watt in nearly every scenario tested and was not narrowly tuned to OpenAI’s own models.

Take that at a discount, since it is a vendor benchmark. The direction still holds. When the binding constraint is power rather than silicon, a purpose-built inference part wins on the metric that actually rations capacity, and every hyperscaler now has a reason to route more of its inference budget through a custom program. Marvell and Broadcom collect that revenue. They collect it at a lower margin than Nvidia collects on a merchant accelerator, and Nvidia collects its own at a falling margin because the memory that goes into all of these parts has been bid into scarcity.

So the industry’s economics are compressing at both ends at once. Demand is not in question. The share of the value chain that ends up as operating profit, and where it sits, very much is. That is the frame we would carry into September, and it is a different frame from the one that has driven the trade since June.

PayPal: A $53 Billion Buyout Dies and the Stock Lands on the Bid

Advent and Stripe have abandoned their pursuit of PayPal. The consortium had offered $60.50 per share, valuing the company at more than $53 billion, which would have ranked among the largest leveraged buyouts ever attempted. PayPal’s board considered the bid inadequate and pointed to regulatory and financing hurdles. The stock fell as much as sixteen percent in pre-market and was quoted near $61.47.

That last number is the interesting one. The board rejected $60.50 as too low, the buyers walked, and the market has repriced the company to roughly the level of the rejected bid with no bid attached. Either the board was right about value and is now owed a re-rating on execution, or the bid was the value and the equity has simply returned to it.

The broader read matters more than the single name. This was the largest financing test in the market and it failed on financing and regulatory grounds, not on price discipline alone. With the thirty-year at 5.20% and its highest since 2007, the cost of very large leveraged capital is doing exactly what that yield implies. Anyone underwriting a take-private of scale in the second half of this year should treat this as the marker.

Warsh at Ten

The thirty-year Treasury sits at 5.20%, the highest since 2007. The two-year is at 4.19% and the ten-year at 4.66%. That is a steepening curve, and it is the shape a market takes when it is willing to trust the front end and is not yet willing to trust the back.

Kevin Warsh delivers his first Jackson Hole keynote as chair at ten o’clock, into a symposium themed on financial innovation and payments that runs through tomorrow. He has been criticised repeatedly for declining to spell out a reaction function, and this is his clearest opportunity to supply one. Investors positioned in long bonds are explicitly asking him for a hard commitment to price stability, which would also serve Treasury Secretary Bessent’s effort to arrest the selloff at the long end and contain the interest bill.

Our view is that the risk into ten o’clock is two-sided in a way yesterday’s was not. A credible inflation commitment is bullish the long bond and bearish the front, because it raises the near-term path while lowering the term premium. A speech that avoids the reaction function again leaves the thirty-year where it is and probably higher. Note the VIX at 14.44 and falling: the options market is charging almost nothing for an event that can move the curve in either direction by a meaningful amount.

Two data points land inside the same half hour. The Chicago PMI is due at 9:45 with 57.6 expected, and the final University of Michigan sentiment reading at 10:00 with 55.2 expected, one-year inflation expectations at 4.2% and five-year at 3.3%. Sentiment fell about eight percent this month, ending two consecutive months of improvement. If the five-year expectation prints above 3.3% in the same minute Warsh begins speaking, that is the combination to watch.

The Other Prints

Elastic is the one worth a second look for anyone tracking AI attach in software. A twelve cent beat on a fifty-eight cent number with only a one hundred million dollar full-year revenue nudge is a margin story rather than a growth story, and it still cleared seventeen percent, which tells you what the bar looks like for mid-cap infrastructure software right now. Affirm and Gap are both fine and neither changes a sector thesis. Affirm confirms that the buy-now-pay-later funding market is open and that credit performance has not cracked. Gap is a management change and a beat on soft comparisons.

Yesterday’s edition said Dell reports after the close alongside Marvell. That was wrong. Dell reports Tuesday, September 1, with roughly $44.9 billion of revenue and $4.89 of earnings expected against a record $51.3 billion AI backlog carried out of the April quarter. It is the single best read available on whether the custom silicon margin story extends into systems.

Asia and Europe

Korea and Japan went opposite ways for related reasons. The Kospi fell 1.16% back below 6,900 on profit taking in AI chips, with Samsung Electronics and SK hynix both down and foreign outflows against a won at 1,380. Selling the memory complex the morning after Nvidia’s best session since April 2025 looks perverse until you read Nvidia’s margin guide properly: the cost inflation Nvidia is absorbing is Korean revenue, the market already owns that view at size, and the natural next step is a customer pushing back on price. Japan rose 0.81% to 66,665.64 with SoftBank leading on reports it is in final talks to buy all of SP.LINKS from Blackstone and Sony Bank, which is a domestic payments deal rather than an AI one.

Tokyo core inflation excluding fresh food rose 1.8% year on year in August from 1.7%, a third consecutive acceleration achieved even as the government pushed energy costs down. That is a live Bank of Japan meeting in September, and it belongs beside the thirty-year Treasury at 5.20% rather than in a separate box. Two of the three largest bond markets in the world are being repriced toward higher terminal inflation at the same time.

Europe is flat and near records, which is the least informative tape available and is fine.

Oil Falls Through a Week of Refinery Strikes

Brent is $88.22, down 0.34%, and WTI is $82.53, down 1.20%, both lower on the week. Overnight, Ukrainian drones struck the Yaroslavl refinery, one of Russia’s largest, along with the Kuibyshev refinery in Samara and the Afipsky refinery in Krasnodar, while Russia sent 164 Shahed-type drones at Ukraine, roughly half of them jet-powered, of which 137 were intercepted.

Crude fell anyway, and that is correct rather than complacent. Refinery strikes destroy refining capacity, not crude supply. In the near term they can raise the volume of Russian crude looking for an export home as domestic runs fall, and they show up in product cracks rather than in Brent. The variable the crude market is actually trading is Hormuz, where Iran and Oman have agreed shares of the strait’s waters and revenues, the White House says ten million barrels transited on Tuesday, and this month’s sanctions package spared Iran’s trading partners and landed softer than feared. Tehran continues to insist the blockade is lifted first.

We would separate the two channels rather than net them. Hormuz is a crude supply story trending toward more barrels. Ukraine is a products story trending toward fewer refined ones. If you are watching this for the inflation channel, and that is the reason to watch it at all, diesel cracks are where the risk is accumulating, not Brent.

Separately, the administration is preparing a 7.5% tariff on Chinese goods over industrial overcapacity, layering onto the 10% to 12.5% announced last month and restoring the second-term rate to roughly 20%, a level Beijing has previously said is consistent with the truce. Officials want the overcapacity inquiry published before the September 24 meeting in Washington. The market is pricing this at approximately zero, which is probably right on the level and probably wrong on the tail.

Key Events Today and Ahead

The Read

Three things into the open.

First, the AI trade is now a margin argument and not a demand argument, and the two sides of that argument are not symmetric. Nvidia’s margin falls because the input it needs got expensive. Marvell’s falls because the business it is winning is structurally lower margin. The first is cyclical and reverses when memory supply catches up. The second does not reverse, because custom silicon is not a phase, it is the shape hyperscaler procurement takes when power is the constraint and every large buyer wants a second source to Nvidia. Position for the first to mean-revert and the second to persist.

Second, PayPal is a credit signal wearing a fintech costume. A $53 billion buyout collapsing on financing and regulatory grounds, on the same morning the thirty-year prints its highest yield since 2007, is one fact observed twice. Large leveraged equity is expensive right now, and it will stay expensive until the long end stops repricing. That is a constraint on the entire take-private pipeline into year end, not a PayPal problem.

Third, everything today runs through ten o’clock. The curve is steepening, the front end is bid, the long end is at a nineteen-year high, and the VIX is at 14.44, which means nobody is paying for protection into a speech that can move the term premium in either direction. Yesterday resolved the earnings risk. Today is the policy risk, and it is the half of the page that has been unresolved all week.

Written pre-market. Every level above is a pre-open mark, and the ten o’clock keynote can rewrite the macro half of this note inside a single paragraph of delivered text.