Editor’s Note
AI moved from a growth narrative toward an economic test. Demand remains visible in contracts, shortages, revenue, and capital spending, but the market is increasingly asking who converts that demand into cash flow and who merely finances the capacity. Beneath a record S&P 500 close, software dispersion, semiconductor losses, and changing market leadership showed that the index can remain calm while individual investment theses break violently.
The Multiple Is Not the Thesis
Forward valuations across AI beneficiaries appear unusually low. Marcos Milla placed NVIDIA ($NVDA) at 16.2x forward PE 2027, Micron ($MU) at 5.2x, Samsung at 3.2x, Oracle ($ORCL) at 12.2x, and Salesforce ($CRM) at 11.9x.
The difficulty is not calculating those ratios. It is deciding whether the earnings denominator is durable.
Samsung is expected to generate roughly $1.25 trillion in operating profit over the next three years. According to one author, reaching that figure would require at least $40 trillion of global AI spending, or approximately $13 trillion annually, compared with stated current worldwide IT spending of $7 trillion. That does not disprove the forecast, but it suggests Samsung’s 3.2x multiple may represent an extraordinary assumption rather than an ordinarily cheap business.
Physical scarcity is real. DigiTimes reported that the three major memory suppliers have pre-sold their available DRAM and high-bandwidth memory capacity for all of 2027, while suppliers can satisfy only 60 to 70% of requested capacity. Apple ($AAPL) is also facing MacBook Air shortages, with many new orders not arriving until September.
Scarcity improves near-term pricing power, but it does not establish the duration of peak economics. Average analyst estimates shared by @JonErlichman call for five-year annualized sales growth of +54% at Micron, +48% at SK Hynix, +44% at AMD, and +31% at Nvidia. TSMC is expected to grow +24%, while Applied Mat, Lam, ASML, and KLA range from +18% to +20%. The market is already underwriting a broad and persistent infrastructure cycle.
Mohnish Pabrai’s response is to place memory and AI in the too hard pile despite patents, scale, engineering expertise, and supply allocation. His preferred contrast is Kaspi, which produces $2 billion in annual cash flow, trades at roughly five to seven times cash flow, and has a dividend yield approaching 10%. Turkey is uncertain optionality, while the Kazakhstan operation supplies the core cash flow.
That distinction became more important with a 10-year Treasury yield of 4.6% and a 30-year Treasury yield around 5.20% to 5.21%. When bonds offer material yields, distant or cyclical equity earnings face a higher opportunity cost.
AI’s Returns Are Separating by Layer
The strongest application-layer evidence came from Palantir. Q2 2026 revenue reached $1.94B, up 93% YoY. U.S. commercial revenue rose 149% to $764M, U.S. government revenue increased 90% to $809M, and GAAP net income reached $1.06B, a 55% margin. Adjusted free cash flow was $1.22B, a 63% margin.
Those are realized economics, not a 2030 capacity forecast.
Infrastructure demand is also real, but its returns are harder to observe. Amazon, Google, Microsoft, and Meta spent a combined $165 billion on capital expenditures during the quarter, 87% more than a year earlier and 393% more than three years earlier. Google reported negative free cash flow for the first time in its history, while Meta’s free cash flow fell 91% to $784 million.
Aswath Damodaran said marginal returns on invested capital have fallen sharply at Meta, Alphabet, and Microsoft. His concern is conditional: unless earnings become commensurate with the tens of billions invested, these companies may become more capital intensive businesses with lower returns on invested capital.
Oracle captures the financing problem. It reported 93% cloud infrastructure growth, 119% infrastructure CPU and GPU growth, and $648 billion in remaining performance obligations. Yet projected cash outlays are rising from $50 billion to $70 billion as debt and preferred-share issuance increase. Capacity can be scarce and revenue can grow rapidly while shareholders still earn inadequate returns.
SpaceX offers more operating evidence, alongside even larger assumptions. Q2 revenue was $7.8B, up 92% Y/Y, while AI revenue reached $2.6B, up 247% YoY. AI EBITDA moved from a $600M loss in Q1 to a $1.1B profit. SpaceX also reported $18.4B in CapEx, including $15.8B attributed to AI.
SemiAnalysis projected a $305 billion annualized revenue run rate by Q4 2027, including $235 billion from AI compute. That forecast assumes 6 to 8 gigawatts of new 2027 capacity, possibly more than 10 gigawatts, at an estimated $50 billion of capital expenditure per gigawatt. Reported construction speed supports execution capability, but turning essentially no AI compute revenue into 77% of a projected $305 billion run rate within eight quarters requires financing, customers, and premium pricing to align.
Expectations Are Now the Binding Constraint
Western Digital beat the cited revenue, adjusted EPS, gross-margin guidance, and expense guidance expectations, yet its shares fell about 12% after rising more than 200% year to date. Barclays reported that 85% of S&P 500 companies beat Q2 EPS estimates while reactions to both beats and misses were unusually negative.
Software displayed the same separation. Airtable agreed to sell for less than $1.3 billion after a nearly $12 billion peak valuation, while HubSpot and Datadog each fell 19%. Atlassian rose 35% after its most profitable quarter since 2021, and Twilio gained more than 20%.
Cloudflare reported Q2 revenue of $696.1 million, up 36% year over year, but non-GAAP gross margin declined from 76.3% to 73.1%. Its agentic commerce products may become valuable, yet management declined to quantify Workers or Workers AI revenue. Distribution is evidence of opportunity, not monetization.
The week’s cleanest capital-structure lesson came from Leopold Aschenbrenner’s Situational Awareness fund. The bottleneck thesis generated extraordinary gains, but the fund reportedly borrowed as much as $4 for every $1 invested. A synchronized decline turned volatility into forced selling. The thesis did not fail first. The financing did.
Counter-Thesis and Risk Watch
Market skeptics focused on valuation, leverage, fiscal pressure, and geopolitical inflation. GMO forecast an average 8.1% annual real loss for US stocks over seven years, equivalent to a 47% decline, while another speaker said mean reversion to a P/E ratio of 16 would imply a 60% market crash. The same material acknowledged that following Jeremy Grantham’s bearishness could have meant missing the bull market since 2022.
@wallstreetmillennial reported that US-listed leveraged ETFs hold $177 billion in assets and may represent around $400 billion of exposure. South Korea supplied the warning case: SK Hynix fell 54% from its June 2026 peak, while its single-stock leveraged ETF declined more than 80% over the previous month.
@clearvaluetax9382 reported expected Treasury borrowing of $739 billion from July through September 2026 and $628 billion from October through December, roughly $1.36 trillion in the second half. The source placed federal debt at $39.7 trillion, with 33% maturing within 12 months. Its prediction of trillions in additional borrowing and money creation during a crisis is a forecast, not an observed outcome.
Geopolitical sources supplied conflicting interpretations of Iran, Yemen, and US policy. Ben Norton argued that Washington is losing its war with Iran, while @geopoliticaleconomyreport disputed the characterization of Ansar Allah as simply an Iranian proxy and separately claimed disruptions at the Strait of Hormuz and Bab al-Mandab affect roughly 30% of global oil supply. The geopolitical conclusions are the sources’ views. The investable risk is narrower: an energy disruption could challenge the cooler inflation case while long-term financing costs are already elevated.