Coatue: "Agents Launching Agents" Is Creating a Digital Population Explosion — and a 16x CPU Opportunity Nobody Is Pricing
Sourcery podcast, May 15, 2026 — Coatue CIO Jaimin Rangwalla's spring investor update unpacked
The Most Underappreciated Unlock in AI
Coatue's Chief Investment Officer of Public Investments, Jaimin Rangwalla, sat down with Sourcery to walk through the firm's spring investor update, and the single most important idea he put forward is one that has not yet been adequately reflected in market pricing: agents that spawn their own agents are creating what he calls a "digital population explosion" that will multiply every person's semiconductor and power footprint by roughly 1,000 times.
"Just imagine if we're all going to be doing things where we're running hundreds of agents, and it can be simple tasks — it doesn't have to be a very complex programming task — but the footprint of each individual in terms of their semiconductor content, their power content, is all going to increase significantly because of those behavioral changes," Rangwalla said. The logic is arithmetically blunt: if seven billion people each run thousands of virtual agents, the effective digital population of the planet multiplies to something in the trillions. Each of those agents requires CPUs, GPUs, and memory.
The specific catalyst Rangwalla points to is Anthropic's Claude Opus 4.5, which introduced the ability for a single agent to spawn multiple subordinate agents autonomously. The human, previously a bottleneck who forced agents to check in at every intermediate step, is now largely removed from the loop. Rangwalla described the practical consequence: "I open a new one, I want this done. I open a new one, I want this done. They're all working concurrently — pools of agents also launching multiple agents below." The exponential compounding of work output is, in his view, the most underappreciated development in the current cycle.
The CPU Trade: A 16x Market Size Expansion Hidden in Plain Sight
The most actionable structural call Coatue is making stems directly from the agent thesis. For the past three to four years, data center architecture ran at roughly one CPU for every eight to sixteen GPUs — all the heavy parallel computation sat on the GPU side, with the CPU reduced to a final execution role. Coatue's analysis shows that ratio has already shifted to one CPU for every four GPUs, an improvement of roughly 2x. But Rangwalla believes the ratio is on course to invert entirely, moving to four CPUs per GPU, and aggressive estimates put it at eight CPUs per GPU.
"You have a 4:1 going to 1:4, which just mathematically is a 16x improvement in a market size," Rangwalla said plainly. The reason is the nature of agent workloads. When an agent books a dinner reservation or sends a calendar invite, that task is serial and sequential — exactly what CPUs do. It does not require the massively parallel floating-point compute that GPUs excel at. As agentic use cases proliferate beyond deep-learning training and inference into everyday task execution, the demand mix tilts sharply back toward CPUs.
The primary beneficiaries Coatue identifies are Intel, AMD, and ARM on the CPU side, with Amazon's custom silicon (developed through its Annapurna acquisition) a quietly important winner inside the AWS ecosystem. On Intel specifically, Rangwalla made what amounts to a straightforward recovery thesis: "Some of the best theses are the simple theses. You have a 4:1 going to 1:4, which is a 16x improvement, and then you have Elon Musk giving a stamp of approval to Intel and increasing the pace of innovation." He acknowledged Intel has had a large move in 2026 but noted it remains a significant laggard over any multi-year period relative to peers in the semiconductor complex.
Anthropic Is Adding $2.5 Billion in ARR Every Single Week
Coatue's data on the private AI leaders is striking in its scale. Rangwalla noted that Anthropic's annualized recurring revenue stood at roughly $9 billion at the end of December 2024, and had reached approximately $30 billion by the time of the update — implying an addition of around $20 billion in ARR in a matter of months, or approximately $2.5 billion per week. "Most of the companies in the SaaS universe don't even have $2.5 billion of ARR annually," he said. "They're adding that in a week."
He further noted that when OpenAI and Anthropic are considered together, their combined ARR already exceeds that of ServiceNow and Salesforce — platforms that took fifteen to twenty years to build. These companies are reaching $25 billion in annualized revenue at between one-third and one-half the time it took the hyperscalers and original Mag 7 to reach equivalent scale.
The valuation context is equally extraordinary. OpenAI's most recent financing round was done at over $800 billion. SpaceX, including the XAI transaction, is valued at $1.25 trillion. Anthropic's last round was in the high $300 billion range, with the next round widely rumored to be materially higher. Rangwalla's framing is that these are no longer merely large private companies — they have broken into the top 25 companies in the world by value before filing for an IPO, something without historical precedent. "It doesn't even matter whether they're public or private. They've already reached that milestone."
Follow the Gigawatts: Coatue's Organizing Framework
Rangwalla explained that Coatue has progressively refined the lens through which it maps the AI supply chain. The firm began with "follow the GPU" when it first backed Nvidia, then evolved to what he now calls "follow the gigawatts." The gigawatt, in his framing, has become the atomic unit of AI growth and simultaneously one of the most binding constraints in the system.
What makes the current tightness unusual — and, in Coatue's view, structurally significant — is its breadth and persistence. "I've never seen supply agreements, guaranteed commitments through 2029, 2030. I've never seen that level of tightness," he said. Memory is one example, but the constraint is system-wide. Rangwalla enumerated: power generation is a bottleneck; transmission and distribution is a bottleneck; NAND is a bottleneck; DRAM is a bottleneck; optical components are a bottleneck; skilled construction and electrical labor is a bottleneck. "The real bottleneck is that there's so many bottlenecks. If it's just one vertical slice, you can say, okay, if they put enough dollars in, they'll solve it. But even if memory solves their bottleneck, that doesn't solve the power bottleneck, that doesn't solve the optical bottleneck."
The implication for portfolio construction is Coatue's "sellers of shortage versus buyers of shortage" framework. Companies operating with fixed capacity against surging demand — memory manufacturers, power infrastructure providers, optical component makers, TSMC — are capturing extraordinary margin expansion because revenue is being driven by price rather than volume, and when fixed costs dominate, operating profit compounds faster than revenue. Meanwhile, the buyers of that shortage — Microsoft, Amazon, Meta — are seeing multiple compression as their capex rises without a corresponding unit benefit, since higher memory or infrastructure prices do not deliver more compute, only the same compute at a higher cost.
Rangwalla called out Google and Amazon as partial exceptions to the buyer-of-shortage discount, given that both are also sellers of proprietary silicon — Google through TPUs and Amazon through Trainium — which gives them hybrid exposure on both sides of the framework. TSMC is described as one of Coatue's largest holdings, benefiting from the pricing power that comes with finite leading-edge capacity.
OpenClaw and the Super App That Doesn't Exist Yet
Rangwalla spent considerable time on what he described as the OpenClaw phenomenon — a harness that allows an agent to control a remote or virtual computer and take real-world actions on a user's behalf. He views the current state as an intermediate step toward something much larger: a single super app that connects to every service a person uses and dispatches agents across all of them simultaneously. "There's going to be a big opportunity for someone to just say, here's your phone, here's the app, this is your super app. This now allows you to connect to everything that you have and you can just ask this app questions." He noted that every major Chinese internet platform — ByteDance, Tencent, Alibaba — has already built its own equivalent, and views the speed of adoption as widely underappreciated.
The memory architecture transition complements this shift. Today's models effectively have amnesia — they do not retain context between sessions, requiring users to re-explain their situation each time, which Rangwalla compared to the Adam Sandler film "50 First Dates." The next generation of accelerators is expected to bring persistent memory, meaning agents will accumulate context over time, increasing their practical utility and further driving the per-user semiconductor footprint higher.
Optical: The Surprise Winner
Among the structural beneficiaries, Rangwalla flagged optical components as the area that surprised Coatue most. Optical had historically been characterized by cyclical oversupply and undersupply, viewed as a commodity business. The firm had been closely tracking memory, CPUs, GPUs, and accelerators, but optical's emergence as a critical infrastructure layer — driven by data transmission requirements across expanding data center interconnects — was the sector-level call that came from outside their initial framework. The shift from copper to optical interconnects within and between data centers is one that Nvidia and others are already building toward in their next-generation system architectures.
The MAG 7 Reordering and What the Market Is Actually Telling You
Despite the positive macro framing, Rangwalla was direct about the internal reordering happening within the Mag 7. Using Bloomberg data tracking performance since the October 31st peak in the NASDAQ, Coatue's analysis shows that outside of Alphabet, most Mag 7 components have significantly underperformed the NASDAQ, with three constituents negative year-to-date. His read is not that the Mag 7 is broken, but that the pecking order is shifting — a structural consequence of which companies are sellers of shortage versus buyers, and which are vertically integrated in AI model development versus purely infrastructure deployers.
On broader market sentiment, Rangwalla was dismissive of the gap between negative media and social media tone and actual market performance — noting that the NASDAQ is currently sitting better than it was in April 2020 despite pervasively negative headlines. "Fundamentals do matter more so than sentiment. The S&P 500 earnings growth for 2026 is around 15%, accelerating to 18%. During periods of that type of economic growth, the stock markets do well." He added that market multiples have not materially expanded — the earnings growth rate is doing the work, meaning valuations are effectively compressing even as prices rise.
What Keeps Coatue Up at Night
Rangwalla was candid about the primary risk scenario: a technology development that suddenly reduces the compute, power, and memory intensity of frontier AI models. He used DeepSeek as a reference point, noting that another moment of that type — where a model achieves equivalent capability at materially lower resource cost — would hit the sellers-of-shortage complex hard, even if it ultimately accelerates overall AI adoption through Jevons Paradox dynamics. He was less concerned about regulation, characterizing AI as a national security matter that makes aggressive regulatory constraint politically unlikely in the current environment. The secondary risk he flagged was the type of volatility where high-quality AI holdings fall five to ten percent in a single session with no identifiable fundamental catalyst — the kind of noise that tests conviction but, in Coatue's current assessment, does not reflect a change in underlying trajectory.