July was a historic month for factor markets. The major indexes remained relatively calm, while momentum suffered one of its most violent unwinds on record, led by technology and, more specifically, the crowded AI-infrastructure theme. Factor volatility had been rising for months and exploded to record levels in July. As AI fears and leverage turned uncertainty into forced selling, situational awareness became the casualty of a duration mismatch: investors held long-duration AI views inside portfolios that required short-term liquidity.
The month also reinforced a larger lesson. AI progress continues to accelerate, compressing economic time along with it. Positioning, liquidity, and rapidly changing assumptions can now collide far faster than human expectations, which are still built around a familiar linear world.
During the emotional final two weeks of the month, the major hyperscalers, companies driving much of the capex anxiety around the AI trade, reported earnings. In the middle of a historic momentum unwind, it was difficult to separate genuine signals from the noise created by forced selling and mean reversion. Yet one thing stood out to me. Even as the hyperscalers reported accelerating cloud revenue, record backlogs, capacity constraints, and increasingly visible AI monetization, many investors and commentators on X remained focused almost exclusively on CapEx risk, rising CDS costs and free cash flow.
I heard it repeatedly from people who reached out with some variation of the same question: “With negative free cash flow, aren’t you worried about circular AI revenues, CapEx cuts, chip obsolescence, open-source competition, data-center delays, or the next version of the same fear that has been written about for the last year?”
Price leads the narrative, especially during a momentum crash. As the market’s leaders fell, investors searched for an explanation that matched the price action. A necessary correction in an AI buildout trend that had been in place for a year failed to satisfy that narrative need. The familiar AI fears were already available, so they became the explanation for a decline that also reflected crowded positioning, factor deleveraging, and forced selling.
I understand the AI-fear conversations because they will continue forever; they are all debates about the future which none of us know the ending. However, the degree to which investors dismissed the importance of real-time data in the hyperscaler earnings and commentary struck me. It reminded me of similar conversations after the COVID market lows. Investors found hundreds of reasons why the market should not rally, extrapolation fears that ignored the printing press. During the early stages of the recovery, I would ask people anchored to the bearish distribution of outcomes, “How much money needs to be printed before you change your view?” The phone would usually go quiet. The answer was clear: the narrative had become the conclusion, and the conclusion had become the anchor and no math could change their anxiety extrapolated minds.
Investing demands a Bayesian process. We make decisions about an uncertain future, then update our probabilities as the evidence changes. I cannot emphasize that belief enough. We must think in probabilities and recognize the hundreds of possible paths ahead. Taking in new information is critical to success. My own view remains that compute is currently the world’s most valuable resource because it is the scarcest asset relative to a demand curve growing exponentially. AI agents are expanding that demand curve, while compute capacity is built at the pace of physics, power, permitting, construction, supply chains, and human coordination.
That mismatch is why I believe supply will remain behind demand for a long time, supporting continued revenue and earnings growth for infrastructure companies. AI is the only way I see this getting back into line. AI agents will eventually help solve elements of the supply problem through algorithmic efficiency, better design, engineering, permitting, grid management, and construction. However, we are still in the 1st inning of AI agent demand and reaching that solution point requires more compute. The system needs the scarce resource in order to create the intelligence that can eventually expand its supply.
I therefore follow the data points that reveal whether the thesis is strengthening or weakening: hyperscaler earnings and commentary, cloud backlogs, capacity-reservation trends, Anthropic and OpenAI annualized revenue, DRAM pricing, and compute-pricing trends. I combine those with the concerns I hear from investors because those views set the odds on the tote board. Like every investor, I will be wrong at points along the way. The relevant question is reward relative to risk. When the data moves away from sentiment and positioning, I become more excited.
In response to the recent suggestion that companies will move away from Anthropic and OpenAI because of price, I find it difficult to accept the argument that Fortune 500 companies will simply avoid them in favor of internally built or open-source intelligence. AI-native startups and technology companies represent a different case. They can often move faster because they have technical talent, fewer legacy systems, and the ability to build directly on open models or create their own tooling. Most large enterprises operate very differently. They have bureaucracies, legacy systems, security requirements, compliance obligations, and predominantly nontechnical users. They need more than an open router connecting intelligence models. They need workflow integration, permissions, governance, support, reliability, interfaces that allow nontechnical employees to use AI productively, and strong indemnification protections.
A year ago, investors focused on hallucinations and questioned whether enterprises would adopt AI at all. Today, some assume public companies will freely use any model from any provider. The ARR growth points in a different direction. Having been a partner at a Fortune 500 firm, I have a hard time seeing that outcome.
That is why Anthropic and OpenAI’s opportunity extends beyond raw intelligence. They are near-trillion-dollar companies building the enterprise tools and distribution layer that help large organizations adopt AI at scale, and their ARR growth strongly suggests they are winning with the fastest growth for a tech company in history and its both of them. For a Fortune 500 company, that layer may matter as much as the model itself.
Here is the irony: the very things investors fear, the enormous compute commitments and the seemingly circular relationships between hyperscalers and frontier-model providers, may become the source of their advantage if demand continues to exceed supply. Those commitments secure scarce capacity, power, capital, and enterprise distribution before the rest of the market can access them. Enterprises that cannot afford to fall behind may concentrate much of their spending with the large platforms because their employees need usable tools, workflow integration, governance, and support alongside inexpensive intelligence.
So let’s go to the back of the hyperscaler baseball cards for the data because the revenue math is much broader than the free-cash-flow headlines. Microsoft generated $90.0 billion of quarterly revenue, up 18%, while Microsoft Cloud reached $59.3 billion, up 27%, and Azure grew 43%. Alphabet generated $119.8 billion of revenue, up 24%, while Google Cloud grew 82% to $24.8 billion. Amazon generated $200.6 billion of revenue, up 20%, while AWS grew 37% to $42.2 billion. Meta generated $60.8 billion of revenue, up 28%, while advertising revenue grew 27% to $59.4 billion. Across the group, the core cash engines are accelerating at the same time the AI infrastructure businesses are gaining scale.
Microsoft, Alphabet, and Amazon also carry roughly $1.7 trillion of commercial RPO and Cloud/AWS backlog. Azure, Google Cloud, and AWS are accelerating while capacity remains constrained. Amazon has already reserved most of its 2027 capacity. Revenue is waiting for powered data centers, GPUs, memory, networking, and the physical infrastructure required to bring capacity online.
Meta adds another important dimension. It is keeping much of its valuable compute inside its own ecosystem because management sees a greater near-term return from improving engagement, recommendations, advertising, personal agents, business messaging, and new consumer products. Meta’s 27% advertising growth, with ad impressions up 14% and pricing up 12%, shows why. Compute carries an opportunity cost: a unit of capacity directed internally can improve the economics of a 3.6-billion-person platform before it is ever offered to an outside customer.
Andy Jassy made the return math unusually explicit on Amazon’s call. Amazon begins investing in a data center roughly two years before it opens, then expects that facility to generate revenue for around 30 years once it is live. Jassy said AI servers can pay for themselves in less than three years and continue producing profit for another two to three years. Even after raising Amazon’s 2026 capital-spending plan to $220 billion, he said AWS still lacks enough capacity to satisfy demand in 2026 and expects the constraint to continue into 2027, with meaningful reservations already extending into 2028.
That is why the free-cash-flow debate requires a longer time horizon. The hyperscalers are converting operating cash flow into assets because revenue growth, backlog, utilization, and customer demand support the decision. As long as compute supply remains behind demand, each new data center and each additional cluster can move from capital spending into revenue-producing capacity. The earnings reports materially increased my probability that the spending cycle continues for longer than consensus expects.
My preferred expression remains the AI-infrastructure buildout and being long scarcity and short abundance: the CapEx receivers rather than the CapEx spenders. The hyperscalers bear the capital intensity, depreciation, asset-refresh risk, and return-on-invested-capital debate which seems like it will never go away. The companies selling memory, networking, power equipment, cooling, electrical systems, and physical infrastructure participate in the build with cleaner exposure to the duration of spending with the scarcity of physics on their side.
Yet many current conversations focus on negative free cash flow that could end the CapEx spend without incorporating the revenue acceleration, contracted backlog, capacity reservations, and management commentary behind the spending. That takes me back to the printing-press conversations of 2020. The math is factual: demand is accelerating, capacity is constrained, and the largest buyers are committing capital because they see a path to earning it back. The fears are hypotheses about what could change. Markets require us to weigh both, yet the gap between the current facts and prevailing fear is exactly why I believe the AI-infrastructure buildout remains compelling especially after a historic position unwind.
