No AI fatigue yet
AI is not running out of steam. We are 3.5 years into the launch of ChatGPT, and this is still early innings. Bubble watch has gone on for nearly 2 years, but the fear has one real root: the performance gap between the hyperscalers, who spend the money, and the semiconductor makers, who cash the checks. Close that gap and the bubble worries can be set aside, at least for now.
Hyperscalers: cheap and getting better
In order of preference: Google, Amazon, Microsoft, Meta. The pick favors the ones that own the infrastructure, own the software, own the distribution, and own the customer relationship, with the most cohesive offering. Token costs are falling while token prices stay steady, which sets these firms up for a second quarter of better return on invested capital. That means the heavy CapEx will be well spent. These stocks are now cheaper on price-to-earnings than they have been in the last 3 years, which makes them a buying chance. The stance on the group stays bullish, with operating cash flow and CapEx expected to keep improving.
Semiconductors: the age of custom chips
Top semi pick is Broadcom, overlooked in all the noise. The age of custom ASIC silicon has arrived. At the end of Q4 the major US frontier labs decided to make money from their models, which flipped the industry from training into inferencing. Inferencing needs 2 to 3 times more compute and 3 to 4 times more memory than training. To keep that cheap, custom ASICs help: they run about 30 to 40% lower on compute and use roughly 50% less energy.
Broadcom's partnership with Apple, announced this week, is a bullish sign. The part people miss, because everyone watches memory, is that networking silicon has grown just as fast. The push comes from the shift from 800G to 1.6T in the scale-up and scale-out designs inside data centers. Broadcom holds a 70% share of the switching silicon business, and that revenue should grow to about 3 times this year what it was last year.
Jackson Square Capital's top 10 holdings: Nvidia, Apple, Broadcom, Google, Eli Lilly, Amazon, Microsoft, KLAC, Caterpillar, and Coherent.
China and cybersecurity
Open-weight Chinese models are closing the gap with US frontier models in finding software weak spots. Per the Brookings Institute, they sit 6 to 9 months behind the best US labs. With the top model being Claude Mythos, and the recent scramble to prepare systems for it, US companies are unlikely to cut back on cybersecurity spending. Order of preference there: CrowdStrike, then Palo Alto Networks, then the recent addition Tenable One, which sits right in vulnerability discovery and management.
Token economics
Falling token costs plus steady token prices make a good mix, close to perfect right now. Run the Moore's law idea on token costs and they keep dropping, even though compute is expensive and memory prices are higher, because each unit gets more usage, tied back to the custom ASICs Broadcom helps deliver. Token costs will fall faster than token prices. To get enterprises building their own AI applications, the model makers will eventually push token pricing back down. That is likely another year or so away.
The macro backdrop
Year-end target stays at 7840. The backdrop stays supportive on record consumer cash and continued stimulus from the one big beautiful bill. The consumer has been strong. Consumer sentiment hit a two standard deviation low last month, about as bad as it gets, so it can only rise from here, especially if gasoline prices drop. That opens a chance in consumer discretionary stocks such as apparel, though the exact play is not yet pinned down. The heaviest exposure belongs in tech, networking equipment, and semiconductors, above all the recently overlooked names. The 15% pullback is a chance to reload.
