A few thoughts about AI infra/semis in the wake of $NVDA’s report and big FY28 guide: 1) Along with valuations, sentiment/positioning look much more favorable than they did a couple months ago. Goldman’s data points to HF gross and net exposures being near 12-month lows and Nasdaq-100 futures short exposure being at multi-year highs; put/call ratios for semis and semi ETFs are often much higher than they were a couple months ago; and some frustrated momo-chasers have even resorted to buying cryptos again :). Just as investors were trying too hard in the spring to rationalize weak price action in tech stocks that were being sold to chase semis, I think investors have been trying too hard lately to rationalize AI infra price weakness caused by a momo unwind and rotations to former laggards. 2) One thing those expecting AI capex to peak soon might be overlooking is that – judging by commentary/reports from lab execs, research firms, etc. – the frontier labs could be close to delivering LLMs that are much better at handling long-horizon agentic tasks, including possibly for knowledge/enterprise work involving non-verifiable domains and tasks that require high accuracy. There are a lot of rumblings right now about the labs getting close to achieving continual learning and initial forms of RSI (some have argued the latter is already here). And while there are longer-term risks to the capex trade related to continual learning/RSI potentially enabling giant LLM training and inference efficiency gains, in the short-to-medium term they’re likely to drive another inflection in token consumption and ARR. Some of Jensen and Colette’s earnings call commentary – about expected agentic growth and 2027 supply/demand, as well as their prediction that the frontier labs will become the world’s largest tech companies – might have been colored by expectations of such LLM advances. And in the meantime, OpenRouter’s data suggests token consumption is still soaring; GPU pricing remains strong; and offerings such as Copilot, ChatGPT Work and Grok Bot are already driving strong growth in non-coding agentic workloads. 3) I said in June that I’m wary of AI infra plays with valuations that require capex to keep growing strongly beyond 2028 for the stocks to work. But right now, it’s not hard to find names with valuations that are arguably pricing in a 2028 capex decline. Along with memory stocks trading at mid-single-digit forward P/Es, you have everything from $NVDA, $TSM and $AVGO, to high-growth semicap subsystem plays ( $UCTT, $MKSI, $AEIS), to ODMs with big exposure to GPU and custom ASIC programs ( $CLS, $FLEX, $SANM), to data center industrials and BTM power plays with large backlogs ( $MOD, $FTAI, $BW, $GNRC), sporting 2027E EPS multiples that are in the teens (sometimes low teens). Not to mention a lot of foreign-traded AI infra plays that also look pretty inexpensive. There are still some richly-valued AI infra plays out there (e.g. $ARM, $CIEN, a few semicaps and industrials). And higher long yields and a tougher credit environment are risk factors for neoclouds and data center leasing firms with big 2027/2028 capital-raising needs. But unless one expects capex to roll over relatively soon, the risk/reward for a lot of other names seems decent here.

$NVDA sharing a bunch of details about its purchase commitments and guarantees in its CFO commentary. Among them: Supply/capacity commitments rose by $160B Q/Q to $279B, "primarily" due to memory.

A couple of thoughts on the memory debate: 1) I think those expecting ASPs/GMs to soon fall sharply because memory makers are signaling that – amid the signing of LTAs with price caps/floors and large volume commitments – GMs won’t go any higher often don’t appreciate how much unmet data center memory demand there is right now. With memory capacity directly or indirectly affecting LLM size, context length, TPS and the ability to handle long-horizon agentic tasks, the likes of $NVDA and $AMD would undoubtedly attach more HBM to their accelerators, and both they and hyperscalers would include more main memory and flash storage in their server designs, if supply wasn’t an issue and they could be assured that prices would be around current levels or lower going forward. All of that’s worth keeping in mind when trying to gauge the longer-term impact of LTAs. Yes, LTAs have been broken before and it’s possible the ones memory makers are now signing are eventually broken as well. But given all that unmet demand, it’s also possible they lead to far more memory being attached to accelerators and servers in 2028/2029 (as supply opens up), and with the memory being sold at ASPs that yield healthy GMs by historical standards. 2) At current valuations, memory makers might be trading at single-digit multiples of what they’ll earn in a couple of years even if DRAM/NAND ASPs drop by 30% or so, especially after accounting for volume growth, cost/bit declines and buybacks. For memory stocks to look truly expensive here, ASPs would have to implode, and -- given all that unmet data center demand, as well as price elasticity for consumer memory products -- that seems unlikely to me unless AI capex meaningfully declines. And if that’s the scenario one is betting on, then memory stocks are far from the only AI infra plays one should be selling here.

A couple of notable disclosures from this afternoon's earnings calls (transcripts via Quartr): 1) $KLAC is hiking its 2026 WFE forecast to the low-$150B range from $140B plus (compares with 2025 WFE of ~$120B), and now sees its advanced packaging revenue growing 70%+ Y/Y (up from a prior high-50s guide). 2) $BE says its backlog is growing faster than revenue. For context, backlog was $20B at the end of 2025, and revenue was up 165.5% Y/Y in Q2.