Editor’s Note
AI remained the market’s organizing force, but investors raised the standard of proof. Spending plans, supply agreements, and model breakthroughs still matter, yet the strongest rewards went to companies converting infrastructure investment into visible revenue, cash flow, and strategic control. The opportunity remains enormous, but price and business quality matter more than ever.
AI Demand Moves Deeper Into the Stack
AMD CEO Dr. Lisa Su described AI as a $2 trillion market opportunity through 2030, including an accelerator market that could reach $1.4 trillion. Her central thesis is that inference, particularly agentic AI, will create more queries, automation, and compute demand. AMD is positioning Helios, the MI455 rack-scale architecture, and Venice CPUs across CPUs, GPUs, FPGAs, and ASICs. It also announced a $5 billion investment in Anthropic, with capacity expected to scale toward 2 gigawatts.
The buildout is making memory strategically important. Since 2012, AI chip performance has improved by more than 100,000 times, while memory bandwidth has increased only about 10 times. HBM, DRAM, NAND, DDR5 RDIMMs, SOCAMM2, and enterprise SSDs are becoming critical to model weights, longer context windows, token generation, and KV cache.
Supply remains concentrated. Samsung, SK hynix, and Micron control roughly 90% of DRAM, while SK hynix and Samsung together control approximately 85% of HBM. Anthropic has signed memory supply agreements with both companies. Nvidia is pursuing a potential $500 billion agreement with SK Hynix, and SK Telecom’s Vera Rubin cloud buildout could require 2 gigawatts of power.
New fabs need two to three years to build and another one to two years to ramp. Meaningful capacity may not arrive until late 2027 or 2028, while AI memory demand is expected to grow above 30% annually. This supports the argument that recent weakness in Micron, Nvidia, and memory stocks reflects leverage and crowded positioning more than disappearing demand.
The opportunity extends beyond chips. AI-driven data center demand could approach 194 GW by 2035, benefiting compute, networking, optics, power, cooling, nuclear energy, and semiconductor equipment. OpenAI and Broadcom reportedly designed the Jalapeño custom AI chip end-to-end in 9 months, reinforcing Hock Tan’s view that every frontier AI company will eventually pursue custom silicon.
Earnings Separate Visible Returns From Promises
Amazon and Microsoft ended the week as the clearest examples of AI spending that investors can already connect to operating results. Amazon gained 15% Friday and recorded its best week in 10 years. Q2 revenue reached $200 billion, up 20%, while AWS growth accelerated to 37%, advertising grew 26%, and online-store growth reached 15%. AWS is now a $169 billion annualized run-rate business.
Microsoft posted its best week in 25 years. Fiscal Q4 revenue reached $90 billion, up 17.8%, and Azure & Other Cloud grew 43% year over year. Microsoft Cloud revenue was $59.3 billion, Azure surpassed $100 billion in annual revenue, and Microsoft 365 Copilot exceeded 30 million paid seats. Microsoft generated $183 billion in annual operating cash flow and spent $115 billion. CFO Amy Hood said, “Even as we invest to meet the growing demand, we expect to remain cash flow positive in fiscal year 2027.”
Google’s operating momentum was also strong: revenue grew 24%, cloud revenue rose 82%, Gemini reached 950 million monthly active users, and cloud backlog hit $514 billion. The debate is whether capex approaching $200 billion to $205 billion this year, with another increase expected in 2027, will produce adequate returns. Google posted negative $6 billion of free cash flow during the quarter, its first negative quarter since going public in 2004.
Meta reported Q2 revenue of $60.8 billion, up 28%, but operating income declined from $20 billion to $18.8 billion. Debt rose from $58 billion to $84 billion, FY26 capex guidance increased to $130 billion to $145 billion, and Reality Labs lost another $4.62 billion, bringing cumulative losses since 2020 to $87 billion.
The market’s message was consistent: large spending is acceptable when returns are visible. Amazon and Microsoft showed the receipts. Google and Meta must continue proving that infrastructure investment can become durable free cash flow.
Price Discipline Returns to Technology
The week’s selloff tested whether investors owned businesses or narratives. The Technology sector fell 12% from its peak after 56 days, while more than $500 billion was reportedly erased from semiconductors in one session. Margin debt reached approximately $1.4 trillion to $1.44 trillion, and forced deleveraging spread through memory, semiconductors, and Korea-linked exposure.
Fundamentals and valuation now require separate analysis. Micron trades at 4.8 times forward earnings in one cited comparison and 5.6 times in another, but peak cyclical earnings can make a company look cheapest before profits decline. Netflix remains an attractive business with operating leverage, yet even after a 50% downside adjustment, it did not clearly offer a sufficient margin of safety.
Alphabet presents a different profile. Berkshire Hathaway’s stake exceeds $31 billion, making it Berkshire’s seventh-largest holding. Alphabet generated $403 billion in revenue last year, held $127 billion in cash or Treasuries after Q1, and produced $73 billion in 2025 free cash flow. Buffett still qualified the position: “I don’t like it as well as at least four or five other businesses that we own.”
As Adam Khoo put it, “Cheap crap is still crap.” Predictable profits, competitive advantages, balance-sheet strength, and sustainable cash flow come first. Valuation determines when those qualities become investable.
AI Becomes Cheaper and More Agentic
Model economics improved sharply. GPT-5.6 Luna pricing fell 80% to $0.20 per million input tokens and $1.20 per million output tokens. GPT-5.6 Terra fell 20% to $2/$12, while GPT-5.6 Sol added an API Fast mode offering up to 2.5x the speed for 2x the price.
Lower inference costs could accelerate AI agents in finance and commerce. Virtuals Protocol has facilitated about $500 million in agent transactions and $2.5 million in agent profits. Yet completion rates remain only 80%-plus, and hacking, prompt injection, and ambiguous instructions remain serious limitations. For now, the practical standard remains: AI informs, but the human decides.
Counter-Thesis and Risk Watch
@geopoliticaleconomyreport repeatedly warned that conflict involving Iran could disrupt Hormuz, Yemen, the Red Sea, and Saudi oil exports. The cited estimate says the U.S. used at least 1,500 of roughly 2,500 Patriot interceptors since February 28, at $4 million to $5 million each. Oil finished July above $85 per barrel, up 21%, keeping energy-driven inflation and margin pressure relevant.
The same channel argued that U.S.-China AI competition is becoming “a new Cold War” shaped by monopoly protection and military integration. Separately, @wallstreetmillennial highlighted agentic-AI security risk following the July 21st, 2026 OpenAI-Hugging Face incident, and warned that Paramount’s Warner Brothers bid could leave the combined company with about $79 billion of debt and $4.5 billion of annual interest expense.
Looking Ahead
Watch whether semiconductor deleveraging stabilizes, whether Google and Meta can defend their capex economics, and whether falling model prices translate into profitable agent adoption. The AI cycle still appears intact, but the next stage will reward companies that pair scarce infrastructure with strong balance sheets, measurable returns, and disciplined valuations.