Editor's Note

August made one thing clear: AI demand is not the weak part of the story.

The weak point is the bill attached to it.

Memory, power, debt, leases, customer financing, and long construction timelines became just as important as chips and models.

AI Spending Moved From Orders to Financing

At the start of the month, the biggest warning was simple: the largest technology companies were spending much more.

Amazon, Google, Microsoft, and Meta spent a combined $165 billion on capital expenditures during the quarter. That was 87% more than a year earlier and 393% more than three years earlier.

That was the opening signal.

By the end of the month, the story had moved beyond capex. It had become financing.

Nvidia was working with Apollo Global, KKR, Brookfield, BlackRock, and Goldman Sachs on a $500 billion funding package. Nvidia was also reported to provide up to $105 billion in financing for an OpenAI data center in Ohio.

Broadcom was reportedly exploring an AI financing package that could approach $100 billion.

Nebius priced an upsized $5.0 billion convertible notes offering.

IREN said cash and committed financing totaled $14 billion, with new GPU financing of $2.8 billion funding 90% of capex.

This is the bigger change.

AI infrastructure is no longer just a supplier order book. It is becoming a credit cycle.

That does not mean demand is fake. It means the return question has moved to the balance sheet.

Who funds the buildout?

Who guarantees the loan?

Who keeps the margin after interest, dilution, depreciation, and customer prepayments?

Memory Became the Clearest Bottleneck

Memory was the strongest repeated signal.

Micron said data center customers are asking for roughly 50% more supply than it can commit. CEO Sanjay Mehrotra told CNBC, “We see no end when supply catches up with demand.”

Micron has signed more than 16 five-year Strategic Customer Agreements. Customers commit to specified volumes under take-or-pay terms, and they can extend the agreements.

That is different from a normal spot-price memory cycle.

Other memory numbers were even more aggressive, but less supported. One post claimed $100 billion of guaranteed minimum revenue and $22 billion in customer deposits. The digest said those figures were not independently supported.

The safer point is enough:

Customers are committing years ahead to secure memory.

Nvidia also said memory pricing has become “extreme,” with costs rising faster than expected and supply likely constrained through FY28. AI server prices were reportedly set to rise more than 15% in many cases because of higher memory costs.

That creates the next test.

If Nvidia, server suppliers, and neoclouds can pass those costs through, memory scarcity supports margins.

If customers slow orders because each dollar buys less compute, the same scarcity becomes a headwind.

Nvidia Proved Demand, But Custom Silicon Got Louder

Nvidia delivered the month’s cleanest proof of AI demand.

Revenue was $96.2bn. Data centre revenue was $89bn, up 117 per cent year over year. Net income was $59.7bn, and gross margin was 75 per cent.

The company guided current-quarter sales to $108bn, plus or minus 2 per cent.

The market reaction was huge. Nvidia added roughly $453B in market value, the largest one-day gain recorded in the U.S.

But the competition story also got louder.

OpenAI’s Jalapeño inference chip, built with Broadcom, was reported to deliver up to 1.9x more throughput per watt and up to 3.6x lower latency than Nvidia’s Blackwell.

That does not hurt Nvidia’s current demand. It does challenge the assumption that all future AI spending flows through GPUs at today’s economics.

Google’s Marvell warrant made the custom silicon theme more visible. Google received the right to buy up to 58,970,907 Marvell shares at $206.58 per share. Full vesting corresponds to a $120 billion revenue ladder, but the source cautioned that this is not a revenue commitment.

The pattern is clear.

AI demand is expanding.

The fight is over who captures it.

Neoclouds Have Contracts. Now They Need Proof.

CoreWeave, Nebius, and IREN became the month’s highest-intensity infrastructure stories.

CoreWeave reported Q2 revenue of $2.6B, adjusted EBITDA of $1.5B, and a 59% margin. Backlog was roughly $104B. Active power more than tripled year over year to 1.5GW, and contracted power reached 4.2GW.

Nebius reported Q2 revenue of $582.3 million, up 454% year over year. AI Cloud revenue was $575 million, up 514%. Annualized recurring revenue ended June at $3.0 billion, up 58% quarter over quarter.

IREN’s transition was more expensive. Q4 AI Cloud revenue more than doubled from $33.6M to $70.5M and passed Bitcoin mining revenue for the first time. But Q4 net loss was $684.0M, including $450.4M in impairments.

That is the month in one section.

The business is scaling.

The accounting still shows the cost of getting there.

ARR is not GAAP revenue. Backlog is not profit. Contracted power is not shareholder return.

The neocloud winners need to prove they can turn scarce power and GPUs into cash flow after financing costs.

Software Was Not Dead

The early fear was that AI would crush software. The month did not support a simple version of that story.

Palantir reported Q2 2026 revenue of $1.94B, up 93% year over year. U.S. commercial revenue rose 149% to $764M, and adjusted free cash flow was $1.22B, a 63% margin.

CrowdStrike rose more than 20% after beating estimates and raising guidance. ARR reached $5.84B, up 25%, and record net new ARR was $333M, up 51%.

Salesforce gained more than 22% after a strong quarter, a higher outlook, and its Anthropic “Claudeforce” partnership.

But software was not treated as one bucket.

Airtable agreed to sell for less than $1.3 billion after a nearly $12 billion peak valuation. HubSpot and Datadog each fell 19%. Atlassian rose 35%, and Twilio gained more than 20%.

The lesson is not “software is fine” or “software is dead.”

It is that AI is separating software companies by workflow ownership, distribution, and ability to turn AI into customer value.

The Market Looked Strong, But Not Calm

The S&P 500 reached record highs during the period, and reported earnings were strong.

With 88% of companies reported, second-quarter S&P 500 revenue rose 15% and earnings rose 50% year over year.

But the earnings surprise needed an asterisk. The 29% earnings surprise fell to 11% excluding Google and Amazon, because both reported large unrealized gains tied to investments including SpaceX and Anthropic.

That matters because AI exposure is now showing up in several places at once:

Operating earnings.

Private-market marks.

Corporate bond issuance.

Customer financing.

Equity stakes.

ETF flows showed the same caution. QQQ attracted $10.9 billion in August through the reported date, while SMH lost $2.8 billion and IGV lost $610 million.

Investors still wanted technology exposure. They seemed less willing to make a concentrated semiconductor or software bet.

Rates and Policy Raised the Hurdle

The cost of capital was the other big story.

Total federal government debt exceeded $40 trillion. The 30-year Treasury yield reached 5.34% last week, up from 4.82% in late June.

July PCE inflation was 3.7% year over year, slightly above the 3.6% forecast. Core inflation remained at 3.3%.

Fed Chair Kevin Warsh said high inflation is “concerning” and called price stability the Fed’s “predominant focus.” Markets raised the implied chance of a September rate increase to roughly 58%, from about 35% the prior day.

Tariffs added another cost risk. Canada announced retaliatory tariffs on $19.94 billion of U.S. goods, with rates from 15% to 50%, taking effect September 8.

This is the uncomfortable setup.

AI spending is becoming more capital intensive at the same time long-term money is not cheap.

Shein Was the Warning From Outside AI

Shein gave the cleanest non-AI lesson.

It is seeking a valuation near $27 billion in a Hong Kong IPO, far below its $98.2 billion valuation from a 2022 private funding round.

The business also slowed. Shein reported a $99 million loss in the first quarter of 2026, compared with a $395 million profit in the same period a year earlier. Revenue increased 1.1%.

Winston Ma of New York University School of Law said: “Public investors are no longer paying for hyper-growth.”

That line applies beyond Shein.

Growth stories are still getting funded. But public investors are asking harder questions when growth slows, margins fall, or financing becomes the thesis.

Counter-Thesis and Risk Watch

The strongest risk watch is circular financing and capital intensity.

Wall Street Millennial argued that circular financing has played a major role in the AI boom, pointing to financial relationships among Nvidia, OpenAI, and cloud providers. George Gammon compared Nvidia with Cisco during the internet boom and argued that Nvidia is taking equity and balance-sheet exposure across the AI ecosystem.

Another source estimated Big Tech capital expenditure at 2.4% of US GDP in 2026 and 3.1% in 2027, compared with a Dot-Com broadcasting and telecommunications peak of 1.2% in 2000.

The second risk is competition.

The Jalapeño risk-watch source said analysts expect Nvidia’s inference share to fall from more than 90% toward 20 to 30% by 2028. That is a source claim, not an outcome. But custom silicon is now a real part of the investment question.

The third risk is macro and geopolitics.

ClearValue Tax, George Gammon, Bull Theory, and Geopolitical Economy Report all framed debt, dollar pressure, Treasury buybacks, Iran, Canada tariffs, and foreign Treasury demand as risks. Their most severe forecasts were opinions, not established outcomes. The reported facts were enough: debt exceeded $40 trillion, long yields rose, Canada retaliated, and Iran sanctions expanded.

The fourth risk is demos without business models.

Wall Street Millennial argued that robot success depends heavily on the environment. Amazon ended Scout in October 2022, while more than 700,000 Kiva robots worked better in controlled warehouse spaces. Serve Robotics said lower-than-expected utilization hurt Uber delivery volume in Q2 2026 and did not currently expect renewing the Uber agreement to make sense when it expires in early 2027.

A demo is not a business.

Looking Ahead

The open questions are now specific.

Can memory suppliers keep high margins as new capacity arrives?

Can neoclouds turn ARR, backlog, and contracted power into reported free cash flow?

Can Nvidia pass through memory costs without pushing customers faster toward custom silicon?

Can hyperscalers earn enough from AI to justify capex, leases, and financing commitments?

And if long-term rates stay near these levels, which AI businesses still clear the hurdle?