From Discovery to Digestion
In macro, a mid-cycle slowdown is the pause inside an expansion. Growth is still positive and the cycle is still intact, but the rate of change cools, the early leaders get expensive, and investors trade an easy question for a harder one. The question shifts from “is this real?” to “what do I pay for the next dollar of it?” AI just ran through its early-cycle catch-up at a speed no normal economy could match. We went from skepticism to acceptance to a full capital rush in roughly six months. That catch-up is now behind us. What comes next looks a lot like a classic mid-cycle slowdown. The story is still powerful, but the easy money in the obvious trade is no longer sitting on the table.
The clearest evidence of how far the early cycle carried us is the cadence of the models themselves. The releases have been arriving faster than the market can digest them, each one landing before the last has been fully priced. That cadence culminated last week in Fable 5, the most capable model ever made broadly available, posting state-of-the-art results across software engineering, scientific research, vision, and long-horizon agentic work. Then, late Friday, it was pulled back and restricted on national-security grounds. A commercial product was released to the public and then partially withdrawn. Regardless of the reason, this is not a normal software cycle. It is one of the clearest markers we have of how far the underlying technology has traveled, and it tells you the binding constraint in this market is no longer whether the models are good enough. They are.
The Model Curve Keeps Moving
It is worth being precise about where this model progress came from, because it changes how you handicap the next leg. A meaningful share of the recent jump in capability has come from algorithmic and post-training advances: explicit reasoning and test-time compute, reinforcement learning in its various forms, including RLHF and the newer push into reinforcement learning on verifiable rewards and AI feedback, distillation, and better, increasingly synthetic training data. The models got smarter by thinking longer and being trained better, rather than merely being trained bigger.
That is the setup for what comes next, because we have spent this stretch improving the software of intelligence while the next generation of hardware was still being built. Now both arrive at once. Blackwell is shipping in volume, the Vera Rubin platform has just entered production, and Rubin Ultra sits on the 2027 roadmap behind it, all of it feeding far larger data centers. Stack continued algorithmic progress on top of that step-change in compute and the model curve should keep climbing. In my view, it points toward something that looks like Einstein-level problem-solving sooner than investors are positioned for. That is the whole case for the application layer. Every gain in the underlying models flows downhill to whoever turns them into products.
Two Horizons
It is worth separating two horizons, because they trade very differently. Einstein-level intelligence is more than a sharper answer to a question you already know how to ask. Paired with agents that can plan, use tools, and act with minimal supervision, it becomes something closer to autonomous problem-solving: systems that can take an open-ended, genuinely hard problem and work it toward a solution largely on their own. That is the disruptive frontier, and it is enormous. Whole categories of progress that have been rate-limited by the scarcity of human genius, in materials, drug design, engineering, and the harder corners of finance, could come unstuck in a way that reorders entire industries.
This is also where a handicapper has to stay honest about the odds. That frontier is uncertain in both timing and shape. It will mint winners and losers the market cannot price today, and it is a multi-year disruption rather than a clean near-term trade. So I hold it as the bigger call to position into as it arrives, rather than the one to underwrite right now. For the next year, the part you can actually model is more workmanlike and more bankable: productivity. Enterprise adoption, efficiency, and margin gains are finally beginning to show up in real financials. Get paid for the productivity wave that is already here, and stay positioned for the problem-solving disruption that is coming behind it.
There is a sting in the tail of this for the hardware trade, too. The faster intelligence improves through software, the louder the question grows of whether every announced dollar of capex is really needed. Cheaper, more efficient models also tend to set off price wars that commoditize the very intelligence the buildout was meant to produce, leading to a shift to the edge. That does not mean the data centers stop getting built. Demand for compute can keep climbing even as the price of a token falls. It means pricing power and margin migrate away from manufacturing raw intelligence and toward applying it. That is one more reason the easy money has left the obvious trade and the durable value is moving up the stack.
Infrastructure Endures, But the Basket Trade Matures
That brings us to the part of the trade that is now fully understood, though understood is not the same as finished. The infrastructure buildout, the bottom of Jensen Huang’s five-layer cake, was the early-cycle winner for good reason. Demand was real, compute was scarce, backlogs were visible, and earnings revisions did the heavy lifting. None of that is a secret anymore. Here is the part the bears get wrong. In absolute terms this buildout is still in the early innings. The scale of what has to be constructed in power, data centers, and silicon over the coming years is so large that the theme does not roll over. It endures.
What changes in a mid-cycle slowdown is the rate of change and the texture of the trade. Growth decelerates off a torrid pace, the tape gets rockier, and the one-decision basket stops doing the work for you. I expect the tight correlations that carried the whole complex higher to break down. Names and sub-themes that moved together on the way up will start to diverge, and leadership will rotate within the buildout as the bottlenecks themselves keep migrating from raw compute toward power, cooling, optical, memory, and advanced packaging. Each bottleneck hands the baton to a different set of winners. The capital intensity is rising, and the financing is increasingly showing up as large debt and equity raises rather than as free cash flow thrown off by the business. The bar for earnings has moved from beating estimates to beating and raising, or getting punished. So the message is not to abandon the infrastructure theme. I still remain steadfast in my view that this is a multi-year rotation from the spenders to the infrastructure receivers. The message is that the receiver theme has matured from an easy basket directional trade into more of a stockpicker’s market.
From Building Intelligence to Applying It
Here is the synthesis the mid-cycle forces on you. If the models are now this good, and if the agentic world has officially arrived, then the scarce resource is no longer simply the factory that produces intelligence. It is also not raw intelligence itself, which is likely to get cheaper over time. The scarce resource becomes the trusted application of intelligence to solve problems: proprietary data, workflow integration, distribution, identity, settlement, domain expertise, and real-world outcomes that customers can actually rely on.
The first phase of AI funded the compute, the data centers, the power, and the model labs. The next phase has to prove the return on all of that spending, and that proof lives in the application layer that has lagged the entire way up. My base case for the next year is that this is where the market’s attention migrates. As enterprise adoption broadens from pilots to production, the productivity benefits stop being a slide in a keynote and start showing up in real company financials: in margins, in headcount that no longer grows in lockstep with revenue, and in work that gets done faster and at lower cost. That is the moment the application layer stops being a promise and becomes a print. Markets pay for prints.
Within that application layer, two areas I have talked about have my focus. The first is drug discovery. The same generation of models that got restricted for being too powerful is also the first to generate novel, testable scientific hypotheses at scale. That capability points straight at the lab, where the prize is owning the compression of time and cost involved in finding a molecule that actually works and proving it in the clinic. Healthcare around it benefits in the same motion, from diagnostics to clinical workflow to trial design to the administrative back office that swallows a third of every dollar spent on care. These are domains where specialized AI applied to proprietary, hard-to-replicate data produces outcomes rather than just throughput, and outcomes are what enterprises will pay for.
More broadly, this is where the agentic shift earns its keep: software where agents execute multi-step work instead of autocompleting a sentence, financial and back-office automation, legal, compliance, and customer operations. The reason this is investable now and was not eighteen months ago is simple. The models finally clear the capability bar, and agents can finally act rather than merely suggest. That is the difference between a demo and a deployment, and deployment is where measurable enterprise value gets created and, eventually, captured.
The Financial Guardrails of the Agentic Economy
The second area is the financial guardrails of the agentic economy, an application layer that almost nobody files under applications. The common thread between drug discovery and crypto is not that they look similar on the surface. It is that both sit where agentic intelligence leaves the chat box and enters the real world. In science, agents need to generate, test, and validate hypotheses. In finance, agents need to authenticate, transact, settle, and prove permission. Both require guardrails.
Knowing the skepticism on this crypto angle, I will begin a new YouTube weekly after the summer to coincide with my belief that broad investor focus will move here the following year driven by the rise of tokenization and the parabolic growth of the agentic economy. Once these Einstein-level IQ AI agents begin to transact, authenticate, and coordinate on their own, they need rails for settlement, identity, and trust that do not depend on a human clicking approve. That is the native use case for crypto infrastructure, stablecoins, and programmable settlement, which is why I treat these monetary guardrails as part of the application layer rather than as a separate macro bet.
On timing, consider the line widely attributed to Charlie Munger: if all you ever did was buy quality when it traded down to its 200-week moving average, you would do very well over time, and almost no one has the discipline to do it. Bitcoin tagged its 200-week moving average for the first time this cycle in early June, near $61,000, with sentiment about as washed out as it gets. That line has marked the major bottoms before, with 2022 the honest exception, so I am not promising it holds. I am pointing out that the financial layer of the agentic economy is going on sale at the exact moment the crowd has given up on it. That is usually when the handicapper starts paying attention.
The Transition
That is the whole mid-cycle thesis in a single frame. The boom is not ending. The buildout is not ending. The leadership is rotating. It is rotating within the infrastructure trade as the old correlations break and new names take the baton. It is also rotating over the next year toward the application layer built on top of it. Capital chased the easiest part of the story first: the chips, the power, and the scarcity. That was what investors could see and underwrite.
The next phase moves toward the parts of the stack where the models, now extraordinary and now agentic, produce measurable value in the real economy, with drug discovery and the financial guardrails of the agentic economy at the top of my list. That is the transition from discovery to digestion, and from digestion to application. The fireworks phase is likely behind us. The more durable phase is the one that has to pay for everything that came before it.
