One of the hardest things for investors to recognize during an exponential technological transition is the difference between new information and delayed understanding. Markets do not always move because the facts have changed. Often, they move because investors finally accept facts that had been hiding in plain sight. When you combine that delayed recognition with unprecedented revenue and earnings reality, you get parabolic moves. The intersection of being unprepared and then being hit with a historic surprise can be explosive. That is what appears to have happened in the first five months of the year. The AI story did not radically change between January, March, and June. What changed was the sell-side and buy-side willingness to believe it.
As we approach the 250th birthday of America, the best way to think about this period is through the image of a fireworks show. At first, the sky is dark. People are waiting, but they are not yet sure what they are going to see. Then the first explosion lights up the sky. A few people notice the shape. Then more fireworks follow, each one confirming the pattern. By the finale, everyone can see it. The show is spectacular, but once it is over, the surprise is gone. We move on until the next fireworks show.
That is where I think we are in AI. The fireworks show is over, and the last couple of weeks have been the climax of an incredible display. But the agentic AI buildout has just begun.
I have covered this journey in my writings and in the conversations I have had with people. Whether it was debating inference, tokens, and memory last year against the cries of ROIC or bubble, or watching everyone finally get there, the pattern has been the same. Whether it was the hate mail after my Marvell paper in March or the two-month echo chamber of “How can I buy these up here?” as earnings season began, the last two weeks have felt to me like the climax of a fireworks show. Maybe this is my time to be caught going against the exponential tide, but I think the warning signs are clear. More people are now fighting against the bottleneck and shortage side than against the agentic AI buildout itself. For me, that usually means the all-you-can-eat AI buffet has left everyone full and in need of a digestion period.
To begin the story of why, let’s go through my version of the investor journey toward realizing the importance of agentic AI. It can be seen through three major speaking gatherings in the AI world: Jensen Huang’s presentation at the Consumer Electronics Show, the Morgan Stanley Technology Conference in March, and the most recent Computex gathering in Taiwan. Each event told the same story with slightly more confirmation. At CES, the message was still early and uncomfortable for many investors. At Morgan Stanley, it became broader and more institutional, with actual numbers from the companies. By Computex, it had become consensus. The market did not discover a new AI world in June. It spent the first five months of the year catching up to the agentic world Jensen had already described in January.
At CES, Jensen effectively laid out the next phase of AI infrastructure. The key point was not simply that NVIDIA had another chip cycle. The key point was that Vera Rubin represented a new architecture for a different kind of workload. Investors were still largely thinking about AI through the lens of training large language models and scaling data centers for chatbots. The mental model was still “more GPUs for more ChatGPT.” But the world Jensen was describing was much more dynamic. It was about inference, agents, tokens, memory, networking, latency, power efficiency, and the movement of data back and forth through optical fiber across massive AI factories. He followed up CES at Davos a couple of weeks later, where he emphasized, in an interview with Larry Fink, the now-understood five-layer cake for the $90 trillion AI buildout.
That distinction around inference matters. A chatbot is episodic. You ask a question, it gives an answer, and the session ends. What everyone learned after OpenAI was that an agentic system is persistent. It plans, reasons, remembers, uses tools, calls other models, interacts with software, and keeps working. That creates a completely different infrastructure problem. It is not just about training a smarter model once. It is about serving intelligence continuously. The bottleneck shifts from pure compute to system throughput. Tokens per watt matter. Networking matters. Optical connections matter. Memory matters. Storage matters. Latency matters. The entire factory matters. Investors were still questioning the importance of chatbots while missing the transition to the agentic rise that had begun in late 2025.
This is why Vera Rubin was so important. It was not just a new product announcement. It was a map of the world NVIDIA believed was coming. The platform was designed around the reality that agentic AI would create enormous demand for inference and system-level coordination. Investors who were still debating whether AI was a chatbot bubble were being shown the early blueprint for an economy where software starts acting on behalf of humans. In hindsight, CES was not the beginning of the story. It was the first flare in the sky.
My view is that investors were not ready. They heard the words, but they did not fully process the implications. They were still anchored to the prior phase of AI. They understood the data center buildout, but they had not yet internalized the traffic patterns of an agentic world. They understood GPUs, but the parabolic rise of GPU demand was seen with one company. The agentic rise brought a broadening out of beneficiaries. Investors had not yet focused enough on memory bandwidth, optical interconnects, networking, storage, and power. They understood model training, but they had not yet shifted their thinking to always-on inference. In January, the message was there. The market was still catching up, and due to many companies already moving in parabolas, the cries of bubble were loud. Market bubblistas need many examples rather than one GPU company at a 10+ bagger over three years.
By the time we got to the Morgan Stanley conference in March, just ahead of earnings season, the story had expanded. It was no longer just Jensen Huang making the case. The broader technology ecosystem was now reinforcing it. Leaders from NVIDIA, AMD, Intel, Dell, and other infrastructure companies were all discussing AI not as a narrow chip cycle but as a full-stack infrastructure transition. The language had shifted from models to systems, from training to inference, from software tools to agents, and from data centers to AI factories. This was the second stage of recognition.
That mattered because investors often need repetition from multiple sources before they are willing to reprice a theme. One CEO can sound promotional. One keynote can be dismissed as visionary. But when the same message appears across semiconductors, cloud platforms, software companies, networking providers, and hardware manufacturers, it becomes harder to ignore. Morgan Stanley gave investors the ecosystem confirmation that CES had only started to reveal.
The key realization was that the agentic world was not theoretical. It was already being built into product roadmaps, capital spending plans, supply agreements, and customer conversations. The implications were enormous. If agents become the next interface for software, then compute demand does not merely grow linearly. It becomes tied to activity. Every workflow becomes a potential source of token demand. Every enterprise application becomes a possible agent platform. Every person could eventually have multiple digital workers operating in the background. That means AI infrastructure is not a one-time buildout. It becomes a recurring industrial system.
That is why the March conference was so important for the market. It broadened the narrative from “NVIDIA has a new platform” to “the entire technology industry is organizing itself around the agentic transition.” The investor base began to recognize that the winners would not only be the obvious GPU companies. The opportunity set included networking, optical, memory, storage, power, cooling, servers, security, edge devices, and software platforms that could monetize agentic workflows. The market began to reprice the picks and shovels of the new intelligence economy.
Then came Computex in Taiwan this past week. What was most interesting for me about Computex was not how much changed. It was how little changed. The same core message was repeated again. Vera Rubin. AI factories. Agentic AI. Local AI PCs. Inference. Networking. Power efficiency. Taiwan as the center of the supply chain. The world presented at Computex was the same world described earlier in the year. The difference was that investors were now ready for it. What sounded aggressive in January sounded obvious by June.
That is a crucial shift. At CES, the market was still learning the language. At Morgan Stanley, the market was translating the language into earnings estimates. At Computex, the market already knew the language. That means the burden of proof has changed. Earlier in the year, companies could surprise investors simply by confirming that agentic AI demand was real. Now the market assumes it is real. Earlier in the year, the upside came from recognition. Now the upside has to come from execution.
This is where the fireworks analogy becomes useful. In January, the sky was still dark. Jensen’s CES presentation was the first major burst of light. In March, the Morgan Stanley conference added more explosions, showing that the pattern was not isolated to one company. By June, Computex became the finale. The sky was fully lit. Everyone could finally see the same picture. But once everyone sees it, the market changes. The crowd stops being amazed and starts asking what comes next.
But Computex itself is not the finale for me. There have been other signs, and that is why the recent market evidence is so important. Broadcom reported very strong AI-related results last night, but the stock still underwhelmed because expectations had already moved. That is what happens when a theme goes from underappreciated to widely accepted. Good news is no longer enough. Investors begin asking whether growth is faster than expected, whether guidance is being raised enough, whether supply is available soon enough, whether margins can hold, and whether the next leg of demand is already reflected in the price.
Broadcom was not an isolated example. Cerebras came public with enormous fanfare because it represented one of the clearest public-market expressions of the search for alternatives and complements to the GPU model, particularly with regard to memory. In a world where memory bandwidth, inference speed, and system architecture are becoming central bottlenecks, the IPO attracted attention because it fit directly into the market’s new obsession: not just who wins AI, but who solves the infrastructure constraints created by inference and agents. It has not traded well post-IPO. At almost the same time, Google moved to raise roughly $85 billion to fund its AI ambitions, a financing amount larger than the market value of a majority of S&P 500 companies. I will add that Berkshire Hathaway, which has been raising cash throughout the last few years, bought $10 billion of a company that has more than doubled over the last year. Really, now? That can only be late agentic understanding. Then came the continued wave of trillion-dollar-plus AI capex-related IPO announcements, the reality of Vera Rubin moving from roadmap in January to launch, and the completion of another earnings season in which every major company had the opportunity to update investors on AI demand. Now that is a finale!
Most importantly for me, this investor acceptance and the obvious capital needs are now combined with the reality of bottlenecks, shortages, hoarding, dramatic price increases, and the continued uncertainty around the Strait of Hormuz. It is simple: the odds have changed since CES. Stocks are up. Sentiment no longer doubts. Signs of “easy money” being made are showing up in Korea, while many of the surprises and new pieces of information are now appearing on the other side of the agentic reality. At a minimum, I see a period of two-sided consolidation needed to burn off the excesses and bring sentiment back to a place of balance. If the Strait of Hormuz follows the expectations of prediction markets, with less than a 50% chance of normal flows before October, I think the surprise will now be about which part of the agentic buildout suffers the most. This is a complex buildout, and if any one part sees delays, expect it to show up across many of the companies.
When you put all of these events together, and add in the divergences I have been highlighting in the sectors related to AI and factor volatility, the message to me is clear: the inference buildout is no longer a hidden or misunderstood theme. It is now broadly recognized, broadly financed, and much more fairly reflected in prices than it was at the start of the year. The first five months were not just about AI getting bigger. They were about investors rapidly responding to the shift from chat-based AI to agentic AI. The market did not slowly adjust. It sprinted. As my father taught me, the odds on the tote board are now fair, so let’s sit out this race until we see better odds.
This does not mean the AI cycle is over. I believe the agentic thematic portfolio I put together will continue to outperform in the months and years to come. However, I believe the joy of watching the fireworks show is over, and now the hard work of navigating a more challenging long/short environment, with a significant macro supply chain risk ahead of midterm elections, will be an overhang to expectations. The agentic buildout continues. The real economic impact of agentic AI may still be in its earliest stages. Enterprises are only beginning to understand how digital workers will change workflows. Consumers are only beginning to understand what always-on personal AI will mean. The hardware supply chain is still scaling. The power grid is still adapting. The software layer is still being rewritten. The long-term transformation remains enormous.
But the stock market is not the technology. The technology can keep accelerating while the stocks become more two-sided. That is the point investors need to understand now. In January, disbelief created opportunity. By June, belief had created a higher bar. Earlier in the year, investors were paid to recognize the agentic shift before others did. Going forward, they will need to separate the companies that can actually deliver from the companies merely attached to the story. I am in a good seat to hear and see the sentiment shift, and we are at a very different place today.
We have entered a different phase of the AI trade. The first phase was about whether AI was real. The second phase was about whether infrastructure spending would be large enough to matter. The third phase was about whether agentic AI would require a new architecture. Those questions have largely been answered. The next phase is about expectations, timing, bottlenecks, and valuation.
That makes the market more challenging. It is now equally possible for companies to underwhelm as it is for them to overwhelm. A company can report extraordinary growth and still disappoint. A company can be central to the AI buildout and still see its stock fall. A company can have the right long-term story but the wrong short-term setup. That is what happens when prices catch up to the narrative.
The lesson from CES to Morgan Stanley to Computex is not that the AI story changed. The lesson is that investor perception changed. The agentic world was introduced in January, validated in March, and accepted by June. That acceptance is bullish for the long-term transformation but more complicated for the stocks. The easy part of the repricing may be behind us. From here, the winners will need to prove not only that the agentic world is coming, but that they can translate that world into revenue, margins, supply, and earnings faster than the market already expects.
The fireworks show is over. The market has seen the light. Now comes the harder part: digestion, execution, and selectivity.
That is the new setup. AI has moved from hidden to known, from discovery to digestion, from skepticism to consensus. The buildout is still real. The agentic world is still coming. But the market has caught up to the story. Now the story has to keep outrunning the market.
