Editor’s Note
This week’s market narrative was not that AI demand is fading. It was that the AI boom is becoming more expensive, more physical, and more financially demanding. The investment question is shifting from “who has the best model?” to “who controls scarce infrastructure, who can fund the buildout, and who can earn acceptable returns at today’s prices?”
AI Moves From Software Story to Capital Cycle
The dominant theme was that AI’s next bottleneck may be capital, not imagination. Across chips, memory, clouds, and data centers, the boom is pushing deeper into the physical economy.
At the semiconductor level, the market is rediscovering the importance of memory bandwidth. In the Qualcomm memory discussion, Dylan Patel of SemiAnalysis framed the opportunity clearly: “Instead of, you know, stacking the HBM separately from the chip, you stack the memory directly on the chip, and that makes your bandwidth explode.” That is why Samsung, SK Hynix, Micron, and Taiwan Semiconductor remain central to the AI supply chain. The surprise is Qualcomm, whose potential edge is using cheap low power memory refined through years of smartphone design, with Meta and Microsoft reportedly already placing orders for chips when available.
The memory market also looks tighter than feared. Concerns that China’s Zhaoxin, also known as CXMT, could flood the market have not changed the current evidence: memory prices have roughly doubled in the past year, contributing to price increases from Apple and Microsoft on the Xbox. The big three producers have shifted capacity toward high bandwidth memory, tightening consumer-device memory supply. Apple reportedly seeking U.S. government permission to buy from a Chinese company blacklisted by the Pentagon supports the point: “Apple isn’t going to China because it’s cheap.” It is going because qualified supply is scarce.
That scarcity is showing up in hyperscaler plans. Meta reportedly plans to double its AI computing capacity to 14GW in 2027 and begin manufacturing its in-house Iris AI chip in September. Iris is part of Meta’s MTIA program, designed with Broadcom and manufactured by TSMC. Even when hyperscalers move toward custom silicon, they still drive demand for foundry, memory, storage, and networking, including Meta’s long-term supply agreements with Samsung, SanDisk, and Sumitomo Electric.
The Model Layer Faces Price Pressure
While hardware scarcity remains powerful, the model layer looks increasingly price-sensitive. OpenAI reportedly raised $122 billion in March 2026 at an $852 billion valuation, with $35 billion contingent on an IPO and $20 billion contingent on SoftBank financing. It has reportedly committed to $600 billion of data center spending over the next 5 years, or about $120 billion per year, against current annualized revenue of roughly $30 billion.
Demand is real. Anthropic’s annualized revenue reportedly rose from about $10 billion at the end of 2025 to $45 billion by May 2026, while OpenAI reached about $30 billion. But customers are pushing back as billing shifts from subscriptions like $20 or $200 per employee per month to token-based usage. Some customers reportedly saw costs triple. GPU rental prices for Nvidia H100s rose from around $1.70 per hour in February 2026 to as high as $3.20 in May, then fell 30% in June.
Sam Altman’s June 2nd, 2026 line captured the pain: “My company spent my entire 2026 budget in Q1.” He also said, “Every enterprise now is thinking about spend and the value they’re getting exchanged for AI.” Palo Alto Networks’ CEO added that token efficiency may need to improve to roughly 1/5 to 1/8 of current levels in the next 12 months, and eventually closer to 1/10.
That pressure strengthens the case for open-source AI and cloud toll collectors. A Databricks engineer called the current moment the “open-source moment for AI,” with demand described as “astonishing.” Coinbase, Shopify, Airbnb, Uber, Siemens, Palantir, and Microsoft were all cited as exploring or discussing open-source AI adoption. If enterprises increasingly download, fine-tune, and run models themselves, Google, Amazon, and Microsoft may benefit regardless of whether workloads use OpenAI, Anthropic, or open-source models.
Valuation Discipline Returns to the AI Trade
Howard Marks’ warning gave the week its investing spine: transformative technology can still become a money-losing bubble. His Anthropic example was blunt:
“If somebody will tell me what they think Anthropic net earnings will be in 2036, I’ll bet them that they’re not within 50% of the truth.”
The SpaceX example made the valuation issue concrete. SpaceX reportedly had $6.8 billion in operating cash flow, raised $85.7 billion, and was valued around $2.5 trillion at the time of recording. Even assuming 20% annual growth for 10 years, a terminal valuation of 20 times free cash flow, a required 10% annual return, and treating operating cash flow as free cash flow by ignoring current CapEx, the estimated intrinsic value was $440 billion, with no margin of safety. That implies a price roughly six times the generous valuation estimate.
Marks compared today’s AI enthusiasm with railroads in the 1860s, radio in the 1920s, automobiles, computers in the 1950s and 60s, and the internet in 2000. His warning: “If this technological innovation with its exuberance doesn’t produce a money-losing bubble, it’ll be the first.”
The practical takeaway is not to avoid AI. It is to size risk intelligently. One portfolio framework this week suggested 20% in pure AI infrastructure companies, 20% in companies related to AI but not entirely dependent on it, and 60% in businesses with almost no AI exposure. The reminder from 2000 is that real technologies can still produce 80% to 90% drawdowns in the wrong securities.
Strong Markets, Fragile Financing
Markets are not acting broken. The S&P 500, NASDAQ 100, and Russell 2000 all hit record highs in June. The S&P 500 has posted 24 all-time highs in 2026, after 39 in 2025 and 57 in 2024. From the March lows, the S&P 500 gained 19% over 9 weeks, the 16th largest 9-week gain since 1950.
Leadership has broadened. Emerging markets are up 24%, small caps up 22%, mid caps up 17%, large-cap value up 16%, and international stocks up 14%, versus about 10% for U.S. stocks, 5% for large-cap growth, and the Magnificent Seven down a few percent. Earnings also support the tape: S&P 500 earnings are expected to rise 19% year-over-year in Q2, full-year 2026 growth is projected at 24%, and profit margins hit a record 14.8% in Q1.
But AI financing is flashing yellow. JP Morgan estimated that combined free cash flow of Amazon, Meta, and Google could move from $125 billion in 2025 to negative $80 billion in 2027. Google reportedly raised over $80 billion by selling its own stock after years of buybacks, while Amazon said it would issue $25 billion of new debt and would not raise more money for the rest of 2026.
Macro adds pressure. CPI rose 1.9% from January to May, moving from roughly 2.3% to 4.2%. The June jobs report showed 57,000 jobs versus expectations of 113,000 or 115,000, with 74,000 in downward revisions. The Fed minutes showed no clean pivot, and odds of rates being higher than today by year-end remained around 85%.
Looking Ahead
Next period, watch whether AI infrastructure spending keeps broadening into memory, power, networking, and financing, or whether hyperscaler balance sheets start to slow the buildout. Also watch token pricing, enterprise AI cost sensitivity, and GPU rental rates for signs that model-layer margins are compressing faster than demand is growing.
The broader market still has earnings, breadth, and momentum on its side. But with margin debt up 55% year-over-year, private credit above $2.1 trillion globally, and AI capex increasingly tied to debt and off-balance-sheet structures, discipline matters. The winners may still be extraordinary businesses, but this week’s message was simple: even in AI, price and funding costs still count.