Nvidia Guides to $108B but Concedes Margin to Memory Costs. Salesforce Gains 12% on Claudeforce.

Thursday, August 27, 2026, written at 9:20 a.m. Eastern, ten minutes before the cash open. The night that this book has been pointing at for a week has resolved, and it resolved with more nuance than either camp wanted.

Top Call: The Tape Pays for Proof, Not for Harvest

Nvidia printed $96.2 billion of revenue against roughly $92.2 billion expected, more than doubling year on year, and guided the October quarter to $108 billion against $104.2 billion with no China data centre revenue assumed at all. Then Colette Kress told the call that fiscal 2028 grows about seventy percent, that the figure is what supply allows rather than what demand wants, and that gross margin falls from 75% today to 74% next quarter and 71% to 72% by the January quarter because memory costs have exceeded the company’s own prior expectations and are heading higher still.

Both halves are true and the second one is the new information. The stock initially traded down about one percent on the release, then recovered through the call and is up roughly six percent pre-market near $220. Salesforce, Okta and CrowdStrike all beat, all raised and all rose. That combination forces a correction to yesterday’s edition, which is below, and it points the argument somewhere more useful than a beat-or-miss scorecard.

Owning Yesterday’s Frame

This book wrote yesterday that the tape had stopped funding investment stories and would pay only for a harvest, using Intuit down twelve percent on a beat, Alibaba down ten on an equity raise and Samsung down 8.7% on a capital return as the evidence. That frame is half wrong and the correction belongs at the top rather than buried.

Salesforce is the purest attach and investment story in enterprise software. It grew revenue 10.8% to $11.345 billion, raised the fiscal 2027 range to $46.1 billion to $46.4 billion, put Agentforce annualised revenue above $1.5 billion on 240% growth, and announced a deepened Anthropic partnership called Claudeforce. It rose twelve percent. Okta raised full-year sales guidance and rose seventeen. CrowdStrike posted record net new annual recurring revenue of $333 million, up 51%, lifted its full-year net new ARR growth outlook by 630 basis points, and rose ten. None of those are harvests.

The distinction that actually survives the evidence is not harvest against investment. It is proof against promise. Intuit asked the market to fund a spending year on a narrative and was refused. Salesforce showed $1.5 billion of annualised agent revenue, current remaining performance obligations of $33.5 billion growing 14% in constant currency, and the strongest quarter of net new annual order value in four years, then raised. The market paid for the receipt, not for the plan. Synopsys is the control case: it beat by 6.5%, raised full-year guidance, and fell six percent, because after the run it had already been paid.

Nvidia in Full

The revenue lines are not the story because they were always going to be fine. The three items worth carrying are the margin path, the ramp and the financing.

On margin, this is the AI buildout taxing its own supplier. Nvidia sells the accelerators whose deployment has bid high bandwidth memory into scarcity, and it now buys that memory at prices set by the demand it created. Management expects fiscal 2028 to recover only to 72% or 73%, and only because price increases take effect. Reuters has already reported that customers were told some AI server systems shipping in early 2027 rise more than fifteen percent on memory. So the sequence is legible: the cost of a unit of compute is going up, the vendor is passing it through, and the buyer is a hyperscaler whose own capital expenditure is increasingly debt and equity funded. Somebody absorbs that, and this quarter it is Nvidia’s gross margin.

On the ramp, Vera Rubin production shipments began in August, purchase orders are in from every major hyperscaler and OEM, and management expects it to be roughly twenty percent of data centre revenue already in the October quarter. The number that matters is the revenue opportunity per gigawatt of deployed capacity: about $40 billion for Vera Rubin against $25 billion for Blackwell and $18 billion for Hopper. Nvidia has more than doubled what it extracts from a fixed unit of power in two generations, which is the correct response to a world where the binding constraint has moved from silicon to electricity.

On the financing, two disclosures deserve more attention than they will get. Days sales outstanding rose to sixty days on extended payment terms granted to major customers. And Nvidia has now put nearly $50 billion into frontier AI labs while standing up financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR that are expected to raise more than $500 billion of third-party capital. Cloud industry backlog is above $2 trillion. Read together, Nvidia is lengthening its own receivables, seeding its customers’ equity and arranging their debt. That is not fraud and it is not even unusual for a supplier at the front of a capital cycle, but it means the quality of the revenue is now a function of the capital markets rather than of the customers’ cash flow. That is the line to watch over the next four quarters, and it is more consequential than any single guide.

The Other Prints

Salesforce deserves a second paragraph for Claudeforce specifically. The arrangement makes Claude the default model across Slack AI, Slackbot, Agentforce Coworker and Salesforce’s own engineering organisation, and it ships a plugin with thirty-seven prebuilt sales skills that lets a seller work live customer relationship data from inside Claude rather than from inside Salesforce. Pilot customers now, open beta in September. Marc Benioff has spent a year answering questions about whether agents eat seat-based software, and the answer he has chosen is to put the system of record inside the assistant instead of defending the application layer. That is a real strategic decision with a real risk attached, and the market gave him twelve percent for making it rather than deferring it. Note the symmetry with the paragraph above: the same investor base that refused to fund Intuit’s spending year paid for this one, because this one arrived with $1.5 billion of annualised agent revenue behind it.

Dollar General is the day’s non-technology signal and it is a good one. Sales up 5.2%, earnings 23.5% above consensus, same-store sales guidance lifted to 2.5% to 2.9%, and full-year earnings taken up roughly eight percent at the midpoint. Set against Abercrombie’s record quarter yesterday and Dick’s nineteen percent guidance cut on Monday, the consumer aggregate is not deteriorating. Execution dispersion inside it is extreme, and the trade-down channel is still gaining share.

Marvell reports after the close tonight, guided to $2.70 billion plus or minus five percent with data centre at roughly 76% of revenue, and the stock is up about 180% this year. It is the cleanest read on whether custom silicon and optical interconnect participate in what Nvidia just described, or whether the merchant accelerator absorbs the whole budget.

Macro: No Excuse Left for Patience

Initial jobless claims for the week ending August 22 came in at 203,000 against 208,000 expected, down 4,000, with the four-week average at 205,500 and continuing claims falling to 1.778 million from 1.796 million. That is a labour market near cycle-low distress readings.

Put it beside yesterday’s inflation report, where core personal consumption expenditures held at 3.3% for a sixty-fifth consecutive month above target and the headline hit a three-year high at 3.7%, and the Federal Reserve’s dovish case has lost its best supporting argument. Kansas City Fed President Jeff Schmid said yesterday that current rates are very accommodating. Kevin Warsh gives his first Jackson Hole keynote tomorrow at ten o’clock, into a symposium themed on financial innovation and payments that opens today and runs through Saturday.

The bond market is not positioning aggressively for it. Two-year, ten-year and thirty-year yields are each about a basis point higher at 4.23%, 4.66% and 5.19%, which is a market that has decided the speech matters more than any datapoint preceding it and is unwilling to be short into it. September pricing sits close to a coin flip. Our reading is that the risk into tomorrow is skewed hawkish rather than balanced, because a chair delivering a first keynote with inflation above target for five years and claims at 203,000 has very little incentive to open by softening.

Overseas, the Bank of Korea raised its base rate 25 basis points to 3.00%, a second consecutive hike and the first back-to-back pair since early 2023, while simultaneously lifting its 2026 growth forecast to 3.3% from 2.6% on semiconductor exports. A central bank tightening into an upgrade is a rare and unambiguously constructive combination, and it is worth noting that the AI cycle is now visible in a sovereign growth forecast rather than only in earnings decks.

Asia and Europe Did Not All Buy It

Korea took the print and ran with it, adding 1.53% with Samsung Electronics up 1.34% and SK hynix up 1.90%, helped by the growth upgrade that arrived with the rate hike. That is now a third consecutive advance and a full recovery of Monday’s Samsung shock, which settles the question this book raised at the start of the week.

Japan is the more interesting tape. The Nikkei opened higher on the same headlines and closed down 130.18 points at 66,131.98 as buying in artificial intelligence and semiconductor names faded into profit-taking, with Advantest, a direct Nvidia supplier, among the drags. Europe did something similar in a different shape: semiconductors were the strongest sector, ASML up one percent and STMicroelectronics up four, yet the Stoxx 600 is down 0.40% because consumer goods, chemicals and food fell one to one and a half percent, with Pernod Ricard off seven percent on weak sales.

That pattern is repeating in American futures right now. The Nasdaq 100 contract is up 0.86% while the Dow contract is down 0.21% and the Russell is down 0.15%. A blowout quarter from the largest company in the world is lifting exactly the names that sell into it and nothing else. The VIX is down to 15.07 as event premium comes out. If you want a single sentence on the internals, it is that this is the narrowest possible version of a good day.

Hormuz: A Functioning Strait With a Burning Tanker In It

Crude has stopped falling. Brent is $88.52, up 0.77%, and WTI is $82.37, ending a four-session decline that took roughly eighteen dollars out of the price from the late-July high.

The reason is that the two tracks running through the strait diverged again. On the diplomatic side, Iran and Oman have outlined a phased framework for a temporary joint shipping corridor plus mine clearance, and Oman’s foreign minister says he is hopeful of formalising it soon. President Trump said Wednesday that the waterway has been fully cleared of mines and called it a very functioning strait, allowing that every once in a while there will be a drone or a rocket. On the other side, a tanker was struck by an unidentified projectile off the Omani coast and caught fire, with the blaze extinguished and the crew safe, and American allies have publicly challenged the claim that the mines are cleared. Iran’s deputy foreign minister says the strait does not fully reopen until Washington meets its obligations under the June interim framework.

Traffic through Hormuz has trended lower this week and remains far below pre-war levels even with the United States escorting vessels along the Omani side. That is the number that adjudicates the dispute, and it says the strait is not functioning in the sense that matters to a charterer. Our view is unchanged from yesterday on direction and firmer on pace: the reopening is real and it is slower than a two percent day in Brent implies, because insurance, escort and salvage economics do not re-rate on a communiqué. The reason to care is inflation. Every dollar off crude removes upside risk from the path that has driven the entire 2026 tightening narrative, and it is precisely that channel that has kept the front end calm through a three-year high in headline inflation.

Key Events Today and Ahead

The Read

Three things to carry into the open.

The first is the correction. This book’s harvest-against-investment frame from yesterday does not survive a night in which the three biggest gainers were an agent platform, an identity vendor and a security company, all of which are spending. Replace it with proof against promise. The tape will fund an expensive strategy if you show it revenue that the strategy has already produced, and it will not fund the same strategy described in the future tense. Intuit and Salesforce made nearly the same argument nine days apart and were treated in opposite ways for exactly that reason.

The second is that Nvidia’s best quarter contained the first honest accounting of what the buildout costs the people building it. Seventy percent growth guided a year in advance is extraordinary, and it is a supply figure, which means demand is higher and the constraint is memory and packaging. That same constraint drives the gross margin from 75% to 71% or 72% within two quarters. You cannot celebrate the growth number and dismiss the margin number, because they are the same fact observed from two ends. The AI trade in the second half of 2026 is going to be an argument about who eats the component bill, and last night the answer was the vendor with the most pricing power in the industry.

The third is that today’s tape is narrower than the headline suggests and tomorrow’s is not about earnings at all. Dow and Russell futures are red on a day Nvidia is up six percent. Japan sold the news and Europe’s index fell while its chipmakers rose. Meanwhile claims are at 203,000, core inflation is 3.3%, oil has stopped falling, and a new Fed chair takes the podium at ten tomorrow morning with a clean sheet and no reason to be gentle. The earnings risk resolved overnight. The policy risk has not.

Written pre-market. Every level above is a pre-open mark and tomorrow morning can rewrite the macro half of this page.