Nvidia proved the AI trade. PCE proved the problem.

The AI trade needed a print, and it got one.

Nvidia beat every number that mattered. Data center revenue at $89 billion, up 117% from a year ago. Q3 guide at $108 billion against a Street looking for $104 billion. Stock up 4% after hours. But the sentence that will echo through every strategy note this morning came from CFO Colette Kress: top-five hyperscaler capex jumping from $800 billion this year to $1.3 trillion in 2027.

A $500 billion year-over-year increase in one spending category, forecast by the company that sells the picks and shovels for it. If she is right, the AI infrastructure cycle is not late. It is barely halfway.

The problem is the bond market has other reading material. July PCE came in at 3.7% headline, hotter than expected. Core stayed at 3.3%. Real consumer spending was flat. Kevin Warsh takes the Jackson Hole podium in under 48 hours.

Two arguments, both loud, both important, arriving in the same 72-hour window.

Nvidia gave the AI bulls a $500 billion year-over-year gift

Numbers first. Nvidia (NVDA) reported Q2 revenue of $96.2 billion, up 106% from a year ago, beating the $91.9 billion consensus by $4.3 billion. Adjusted EPS of $2.22 versus $2.08 expected. Data Center revenue hit a record $89 billion, up 117% year-over-year. Guidance for the current quarter came in at $108 billion plus or minus 2%, well ahead of the $104.2 billion Street estimate. Gross margin guide of 74%.

Shares closed the regular session at $209.66, down 1.59% in the run-up. In after-hours, they climbed to roughly $218, giving Nasdaq futures a 1% pop overnight.

The headline story is that the beat was clean. The deeper story is what Kress said next. Alongside the print, Amazon (AMZN) and Nvidia announced Amazon Web Services will buy 2 million Nvidia GPUs and use the company's new Vera CPU. That is one hyperscaler locking in silicon at industrial scale on the same day Kress told analysts capex among the top-five buyers is set to move from $800 billion in 2026 to $1.3 trillion in 2027.

Sit with the second number. $1.3 trillion is nearly 5% of US annual GDP spent by five companies on one category of infrastructure. Kress framed it as a base case, not a stretch. If she is right by even 70%, the AI capex cycle runs deep into 2027 without needing a fresh catalyst. Data Center Hopper shipments to China were less than 1% of revenue and the guide assumes zero China again. That is a call option nobody is pricing.

The market signal?

Nvidia's Q3 guide resets AI capex modelling for every desk on the Street. As long as hyperscaler capex is compounding faster than Nvidia's own revenue, the bear case needs a customer to blink first. Nothing on this call suggests one is about to.

Get the #1 Stock on Elon Musk's $2.1 Trillion Hit List

The SpaceX IPO made Elon Musk the wealthiest man who has ever lived.

Now he's spending it... buying up everything he needs to complete his dynasty.

A handful of companies stand between him and total control of the space economy.

And #1 on his list is a rival he's been launching satellites for... while they quietly eat into his most profitable business.

Click here to get the name and ticker before Elon makes his move.

PCE ran hot. Spending stalled. The Fed just got a harder job.

Wednesday's cash session went nowhere. The S&P 500 closed at 7,675, down 0.02%. The Nasdaq slipped 0.08% to 26,130. The Dow lost 0.21%. Everyone was waiting for two things: PCE at 8:30 and Nvidia at 4:00.

PCE landed first, and the print was not clean. Headline came in at 0.2% month-over-month and 3.7% year-over-year, both hotter than the 0.1% and 3.6% consensus. Core, the number the Fed actually looks at, held steady at 3.3% annually and 0.2% monthly, in line. The mix underneath is what matters: personal income rose 0.4%, but inflation-adjusted consumer spending was flat. Households are earning more and buying less. That is a defensive consumer, not one running out of money.

Financial services and insurance prices jumped 1.2%. Gasoline fell 2.7%. Services inflation is still hot and goods deflation is doing the polite work of holding the average down. That is not a mix the Fed can cut into. Traders trimmed September hike odds to the mid-30s%. The fact that any live probability of a hike exists, when it was near zero six weeks ago, is itself the story.

The market signal?

Watch the two-day sequence, not the day. If Warsh sounds patient Friday and August payrolls surprise soft next week, September hike odds collapse and long-end Treasuries rally. If either goes the other way, the 30-year retests 5.30% and the AI trade's discount rate goes up regardless of what Kress just promised.

Colette Kress just underwrote the entire 2027 hyperscaler bill

The call sentence to circle: capex at the top-five hyperscalers moving from about $800 billion in 2026 to $1.3 trillion in 2027. Kress, speaking to analysts Wednesday evening, delivered it as a matter of course. It landed like a policy announcement.

Consider what that number requires. Microsoft, Alphabet, Meta, Amazon, and Oracle combined spent $166 billion on capex in the June quarter alone, up 87% from a year earlier. Extrapolate flat and you get roughly $700 billion for the year, in the neighborhood of Kress's $800 billion figure. Getting to $1.3 trillion means every one of those companies raises capex again in 2027 by an average of 60%, on top of a base that is already the highest in corporate history.

The market has been debating whether hyperscaler capex is about to plateau. Kress's answer: not close. And the AWS-Nvidia deal announced alongside the print puts one of those five customers on the record for 2 million GPUs plus Vera CPUs. That is not a slide-deck commitment. Silicon in shipment.

There is a self-interest problem with the source. Nvidia's CFO forecasting more capex from Nvidia's customers is not neutral. But the disclosure is specific enough to be checkable against 10-Q filings over the next four quarters. That is a hostage.

The market signal?

If Q4 hyperscaler guides in early 2027 come within 15% of the $1.3 trillion path, Nvidia is a hold at any multiple the market wants to put on it. Below $1.1 trillion, Kress just handed the bears a receipt.

Why are companies flying spy planes over Elon's closely-guarded AI lab?

Elon did the seemingly impossible – far faster than anyone expected...

And it's sent the tech industry into PANIC MODE.

ChatGPT, Claude, Google Gemini, and DeepSeek could soon become obsolete.

And three little-known firms could soar 10X or higher as a result.

Get the details here.

The AI-kills-software thesis died in a single after-hours session

For most of 2026, a specific fear haunted enterprise-software valuations: if AI can do what software charges for, the software companies get eaten. Zscaler's May guidance miss was Exhibit A. Goldman's Intuit downgrade in June was Exhibit B. The trade was to short the SaaS incumbents into their AI-native replacements.

Wednesday after 4:00 pm, that thesis had a bad evening.

Salesforce (CRM) beat and jumped double digits in extended trading. CrowdStrike (CRWD) posted Q2 revenue of $1.47 billion against a $1.44 billion consensus, EPS of $0.31 versus $0.29, and net new annual recurring revenue of $333 million, up 51% year-over-year and a company record. Management raised the full-year net new ARR growth outlook by 630 basis points. Shares ran 11% after hours. Okta (OKTA) reported and pushed higher into extended too. Add Nvidia's beat and it was a clean sweep for the AI-adjacent complex.

CrowdStrike's net new ARR is the tell. A 51% year-over-year gain, after a Q1 that already hit a record, does not sound like the top of a cycle. It sounds like AI-native competitors are still marketing while enterprises keep writing checks to the incumbent.

The lesson is not that software is safe. Companies owning the trusted data layer (Salesforce), the security perimeter (CrowdStrike), the identity graph (Okta), or the compute floor (Nvidia) are being paid to sit on top of the AI stack rather than get replaced by it. The companies at real risk are the ones between: point solutions with a workflow a foundation model can replicate in a weekend.

The market signal?

When four AI-adjacent large caps beat and pop the same night, the SaaSpocalypse trade is done for this cycle. The next test is Q3 net new ARR at CrowdStrike and Agentforce contribution disclosure at Salesforce. Fail there and the thesis returns.

Meta paid $16.7 billion to change what teenagers see at 11 pm

Meta (META) settled with 29 state attorneys general Wednesday for up to $17.1 billion, base payment $16.68 billion, ending a federal trial in Oakland in its second week. The case alleged Facebook and Instagram were engineered to be addictive to minors, safety features were misrepresented, and data was collected in violation of the Children's Online Privacy Protection Act.

The hit is bounded. Meta will accrue roughly $10 billion in Q3 legal expense. Against 2025 net income of nearly $70 billion, that is a bruise, not a break. Shares gained 1% on the day. The number the market cares about is not the payment. It is what Meta agreed to change.

Under the settlement, Meta imposes daily use caps for teen accounts. Nighttime blocks restrict access during school hours and overnight windows. Age-verification gets stricter. Parental controls expand. The consent judgment runs, effectively, for a decade.

That matters commercially because the teen cohort is upstream of Meta's engagement metrics, not incidental to them. Restricting when 14- to 17-year-olds can scroll compresses one of the highest-frequency user segments on the platform. Advertisers pay for attention, and attention among teenagers just got a court-ordered curfew. Meta's own trial filings claimed potential penalties could theoretically reach $1.4 trillion. Settling at $17 billion is the price of not facing a jury. TikTok and YouTube are next in the states' queue.

The market signal?

This is a template for how state AGs will approach every large social platform for the next three years. If the same terms hit ByteDance and Alphabet, the entire teen-engagement economy resets. Watch Snap and Pinterest for the read-through.

America's oil savings account is running into the laws of physics

Not every finance story is about basis points. Some are about salt.

The US Strategic Petroleum Reserve fell to 298.7 million barrels earlier this month, below the 300 million mark for the first time since January 1983. The reserve sits in 60 salt caverns thousands of feet under the Gulf Coast, Texas and Louisiana. Design capacity is 714 million barrels. Current fill: roughly 42%.

The reason it is here is the war. Before the US-Israel strikes on Iran in late February, the reserve held about 415 million barrels. Trump authorized a 172-million-barrel release in March to soften the Hormuz shock. That drawdown, plus the follow-on IEA release, pulled the number under 300.

Here is where a headline becomes an engineering problem. As oil is drawn from a salt cavern, brine floods in to maintain pressure. That process is largely one-way. Each drawdown-and-refill cycle can cause salt-wall dissolution, thermal shock, and eventually cavern deformation. A May GAO report found more than a quarter of SPR inventory was already unavailable for drawdown due to construction and cavern outages. Rapidan Energy estimates 103 million barrels of the current inventory are effectively inaccessible.

The theoretical max extraction rate is 4.4 million barrels per day. At current fill, real-world extraction runs closer to 1 to 1.4 million, because low reservoir pressure slows the flow. The reserve can be drained, but not as fast as headline capacity suggests, and not without accelerating damage to the caverns holding whatever remains. Amos Hochstein, a former Biden energy advisor, told CNBC: "I don't know anyone who believes we can go below 300."

The market signal?

The SPR is now a hedge that is losing usable capacity while inflation runs at 3.7%. If Hormuz stress returns and Washington cannot draw at scale, the entire premium Brent has priced out over the last two months returns in a session. That is a reason to keep energy exposure in the portfolio even after oil's summer round-trip.

Landmark Executive Order 14241 Unleashes

TRUMP’S NEW DOLLAR

Republican or Democrat – whether you support or oppose Trump’s New Dollar – every American could soon be forced to use it

Discover three critical moves to help you prepare, before it’s too late

• Warsh delivers his first Jackson Hole keynote Friday at 10:00 am ET, and the market has already priced a neutral speech. 69% of BofA-surveyed fund managers expect no forward guidance. That is exactly the setup where silence reads hawkish, because the July minutes already documented three dissents in favour of hiking. If Warsh does not push back, September hike odds move higher by Monday's open.
• The BLS preliminary payrolls benchmark revision also lands Friday. Last year's revision erased 911,000 jobs after the fact. Another meaningful downward mark would give the doves their first hard evidence in months that the labour market has been softer than the monthly prints suggested, and it arrives the same morning as Warsh.
• Weekly jobless claims and Q2 GDP second estimate drop Thursday at 8:30 am. GDP consensus is 3.3% annualised. A downward revision would compound Wednesday's PCE weakness and give the more dovish committee members a talking point going into the weekend.
• The market signal across all three? Every catalyst this week bends toward one question: does the bond market get its rate cut evidence, or not. Friday morning is when the answer starts to land.

That's it for today's Slate. Nvidia validated the trade, PCE complicated it, Warsh decides it Friday. Stay close to the feed.

Today's reply prompt: If Warsh says nothing Friday, is that dovish, hawkish, or a tell that he's already lost the room?