AI Could Turn Software Into “Dumb Data Pipes” | Dom Rizzo

Dom Rizzo manages the 8.7 billion dollar T. Rowe Price global technology fund under ticker PRGTX and a 300 million dollar ETF launched in October 2024. The fund has returned 43.6 percent per annum since he took over on December 1, 2022, outperforming its benchmark by 5 percent annually. Early semiconductor coverage beginning in 2015 led to sizable positioning in Nvidia and AI infrastructure before most investors recognized the opportunity.

Rizzo places the current AI CapEx cycle at roughly inning four or five of nine and argues overspending has not yet occurred. His core principle is that compute equals revenue, meaning the more compute a company deploys the more revenue it can generate, and this game theory is driving sustained capital commitments across major players. Google raising approximately 85 billion dollars in equity is his primary evidence that the CapEx boom is larger in size, scale, and scope than most investors expect, since the third most profitable company in the world would not raise equity unless the opportunity was unusually large. Compute is expected to remain very hard to obtain through 2026, 2027, and potentially 2028.

Rizzo predicts ChatGPT and Claude will eventually sit on top of the entire enterprise software stack, reducing most other applications to dumb data pipes. Users will begin their day inside those interfaces rather than in tools like Microsoft Word or Outlook. Enterprise software companies became complacent by raising prices and overselling seats under low-churn recurring models rather than innovating, and AI attacks zero incremental cost products like software first as a horizontal technology. Salesforce is described as essentially a customer relationship management database whose core functionality could be replaced by a ChatGPT-style interface, and Rizzo says if forced to wager he would expect its roughly 7 percent growth rate to decelerate rather than accelerate. Approximately 45 billion dollars of AI spending is coming from existing IT budgets, creating a direct crowding out effect for traditional software vendors.

Software companies also face a business model shift from recurring revenue to usage-based revenue where they must fight to stay in the token path of AI systems. Microsoft faces a difficult internal trade-off over allocating GPU capacity between its first-party applications and its cloud business. Azure grew 40 percent year over year in its last reported quarter. Rizzo cautions that AI is not a normal technology because it is intelligence itself, making good enough a more complicated competitive position than in prior technology cycles.

Agentic AI, where models complete tasks rather than answer questions, is shifting hardware demand significantly. In a training compute world the GPU to CPU ratio is eight to one, but in an agentic world that ratio shifts toward parity or two CPUs per GPU. This explains strong demand for AMD and Intel processors, which have been among the largest portfolio positions since the start of the year. The data center CPU market has been around 25 billion dollars for some time and Rizzo believes it is heading toward 125 billion dollars. Nvidia remains approximately 18 percent of the fund and is consistently the largest position. He describes Nvidia as a systems company integrating GPUs, CPUs, and networking, and calls the Mellanox acquisition likely the best in semiconductor history. AI chips were roughly 45 billion dollars in 2023 and are projected to reach one trillion dollars by 2030, and Rizzo believes the market will be large enough for Nvidia, AMD, and Intel to all do well simultaneously.

Memory is described as a commodity in the sense that chips from Samsung, Hynix, and Micron are interchangeable in a system, though Rizzo calls DRAM the hardest commodity in the world to make. In an agentic AI system memory consumption is five to ten times higher than in prior architectures. Memory sector revenue growth went from approximately zero percent to approximately 500 percent over the past twelve months. Hynix and Samsung trade at roughly four to five times earnings while Micron trades at eight to nine times. Rizzo describes himself as memory curious but remains humble about the deceleration ahead as pricing normalizes from peak growth rates.

Anthropic's run rate revenue was 5 billion dollars when Rizzo participated in a funding round last summer and has since been announced at 47 billion dollars roughly nine months later. OpenAI and Anthropic combined are on a path to approximately 200 billion dollars of annualized revenue by year end at a combined valuation of roughly 2 trillion dollars, implying around 10 times revenue. The steel man against the compute equals revenue thesis is that models one generation behind frontier are roughly 90 percent cheaper and nearly as capable, though Rizzo reads software CEO claims of shifting to cheaper models as a desire to avoid capital intensity rather than a description of current practice. He also expects Apple to face a serious problem if it cannot transition to an agentic operating system by around 2027, and notes that Apple's services revenue depends heavily on a roughly 25 billion dollar annual payment from Google whose balance of power could shift as Google delivers more AI directly to users.