Deep|LLM: RSI Is the Most Important Variable, and Compute Is the Deepest Moat July 2026 produced the sharpest drawdown of this cycle for the AI trade. Semiconductors posted their worst month since 2002, and more than $1 trillion of AI-linked market value came off. The concern behind the move is straightforward: coding looks like a special case that is hard to replicate, enterprise adoption is running behind expectations, and the “next coding” that would justify hundreds of billions of dollars of capex has not yet appeared. We think the debate is framed around the wrong question. The market is evaluating AI through an application lens, looking for the next killer use case and checking whether current revenue covers capital expenditure. The industry’s own objective, however, is AGI, and the mechanism most labs point to on that path is recursive self-improvement (RSI). A more useful way to think about it is as a sequence of capability milestones, each with its own commercial payoff. Coding capability has already been monetized, in the form of this year’s first-half revenue inflection. The next milestone is continual learning, which lets models accumulate knowledge in deployment the way an employee does; combined with deployment mechanisms such as forward-deployed engineering, it will open up most of the enterprise opportunity. Full RSI, in which models participate in and accelerate their own development, sits furthest out and will determine the competitive structure that follows. Progress is visible at each stage, and in several places it is accelerating. We make three claims. First, the “context problem” the market worries about is a capability question rather than a use-case question, and it is what continual learning is designed to address. Once models can accumulate tacit organizational knowledge the way employees do, the search for a single “next coding-like” use case dissolves: enterprise use cases open up progressively. The TAM reference then shifts from a software budget measured in hundreds of billions to a labor market measured in tens of trillions. Second, early signals of RSI are already observable. Model iteration has moved from annual to monthly. GPT-5.5 took roughly one month from the end of pre-training to release. OpenAI cut prices by up to 80% three weeks after GPT-5.6 shipped, against a history in which price cuts followed months of inference optimization. Lab code is now written almost entirely by models, and AI’s measured contribution to internal R&D velocity rose from 5% to 15–20% over six months. Third, if RSI crosses its threshold, catch-up strategies built on distilling frontier model outputs are likely to become progressively harder, and compute becomes the least compressible constraint and the deepest competitive moat. When the frontier iterates monthly, a follower may still hold a fixed time lag, but it cannot readily replicate a frontier lab’s full learning, data, and experimentation stack. Native self-improvement capability requires proprietary experimental infrastructure and the compute to run it. This is also the view inside China’s leading labs: China has not hit a scaling wall, but does not yet have enough compute to reach one. Capex, on this reading, should not be dismissed as a bubble. The latest earnings season shows real current returns from AI-driven cloud growth, and beyond that, capex is what ensures a player has sufficient compute when the RSI threshold arrives. Strategic necessity is not the same thing as shareholder return, and we treat that distinction explicitly in a scenario framework. The variable that matters most is less “what is the next application” than when RSI arrives and who gets there first. Detailed Report

Jukan is right. Korea has seen some weakness recently. A certain Korean equity analyst has also expressed a cautious view, drawing much of his research from information in the Chinese market. This pattern has persisted for several quarters now — this quarter is certainly not the first time. This is largely because China’s inference demand has yet to take off, while the smartphone market there remains concentrated in lower-end models. However, you can also see that his US-based colleagues, the reports we have published, the detailed notes in our latest interview database, as well as surveys such as the most recent Cleveland Research survey, all indicate that the US has broadly accepted a new round of price increases, with the scope of the increases already negotiated and agreed upon. This has also made China's influence on pricing increasingly limited, and in reality it is not nearly as significant as this Korean equity analyst believes. The analyst says Americans are too bullish, but that's simply because the reality is exactly that. If China's inference market were to improve substantially, Chinese buyers would change their stance as well.

Weekly|Kimi K3 "DeepSeek Moment" Jitters, TSM and $ASML Beat-and-Raise, $ATI & the Starship Supply Chain, $INTC , $AXTI , $NBIS Semis got smoked this week, with the indexes closing out weekly losses on fears of another “DeepSeek moment” after Kimi K3’s release. We think that read gets K3 exactly backward. K3 is a 2.8-trillion-parameter model priced at $3/$15 per million tokens - in line with Claude Sonnet, not the cut-price disruption playbook - and it currently runs only at maximum reasoning effort, which means it very likely burns more compute per task than its US frontier peers. A Chinese lab scaling up aggressively and pricing at frontier levels is evidence that scaling still works and that demand for inference compute keeps climbing. That is not a bearish datapoint for the compute chain. The fundamentals told a story opposite to the tape. TSMC and ASML - the two most important prints of the season - both delivered beat-and-raise quarters. TSMC lifted its full-year 2026 USD revenue growth guidance to above 40%, raised CapEx again, and C.C. Wei described AI demand as getting “stronger and stronger” through 2029-30. ASML raised its 2026 revenue guide to EUR 43-45bn (midpoint 12% above consensus) and is expanding both EUV and immersion DUV capacity by 30% in 2027, with another 30% under study for 2028. Yes, TSMC’s 3Q gross margin guide came in light - but the culprit is a faster-than-expected 2nm ramp, which is the kind of margin problem you want to have. When the supply chain’s two most important companies are guiding like this while the market frets over a model release that actually reinforces the compute-demand thesis, we are comfortable treating this week’s selloff as positioning, not fundamentals. We stay constructive on the AI compute chain. This Week’s Reports Kimi K3 - scaling still works, and it is not cheap. K3 lands in the global top tier on coding, agents, and long-horizon tasks. However, prices at Claude Sonnet levels and runs only at max reasoning effort - more proof that frontier capability still costs frontier compute. TSM & ASML - the supply chain says demand is accelerating, not slowing. Both quarters were beat-and-raise: TSMC surprised on CapEx and lifted full-year growth to 40%+, while ASML guided 2026 revenue 12% above consensus with 30% EUV capacity expansion ahead. The 2nm-driven margin dilution at TSMC is noise next to the demand signal. ATI - the materials chokepoint of the Starship era. Superalloy content per Starship stack is 25-30x a Falcon 9, and specialty melting is the one layer SpaceX cannot vertically integrate - yet ATI trades at the lowest multiple of the Big Three melt names—one of the few structural public-market entries into the SpaceX supply chain. Premium Report Snapshot A portion of our research is reserved for Premium subscribers and is not distributed via Substack. Below is a snapshot of what Premium subscribers received this week beyond the Substack feed. ... Selected Weekly Expert Interviews A snapshot of the expert calls conducted for Premium subscribers this week. Full transcripts and takeaways are available on the FUNDA platform. ... Claude Deployment Spend, ROI and Governance Constraints Employee of AI Department, Aviva Canada Aviva’s Canadian Claude rollout is seeing strong consumption in coding and insurance workflows, but ROI remains uneven, and governance limits broader adoption. AI Agent usage Former Engineer, Google Google DeepMind’s Gemini app commercialization is framed as advancing through harness-system buildout, but consumer monetization, execution discipline, and coding competitiveness remain challenges. Detailed Report

Deep| $SPCX : Grok 4.5 Brings SpaceXAI Back at the Frontier-Lab Table; Remain Bullish on Harness Data Flywheels and Compute Demand What matters most about the Grok 4.5 release is that xAI / SpaceXAI is back at the table among North American frontier labs. At the beginning of the year, the North American frontier-model race still looked like a five-player game: OpenAI, Anthropic, Google, Meta, and xAI. But for a stretch, xAI had clearly fallen behind, and the market was more willing to treat it as a new neocloud, so the narrative shrank from five players to four. More recently, reports that Meta wanted to build “Meta Compute” and sell spare AI compute were read as frontier-lab consolidation and compute oversupply, bearish for neoclouds and AI infrastructure. Detailed Report

FUNDA now offers the Market Top and Bottom Indicator, a mid to long-term contrarian signal system. Low scores signal buy zones, high scores signal sell zones. Top and Bottom Score A 0 to 10 contrarian composite weighted across 11 sub-components, updated daily. Low readings correspond to fear, high readings to greed. As a high-conviction indicator, it triggers approximately 2 signals per year. Component Breakdown 11 dimensions scored on a 0 to 10 scale, covering credit spreads, yield curve, equity flows, bond flows, hedge fund positioning, CTA positioning, long-only positioning, sentiment, greed, equity breadth, and credit technicals. Each score reflects that dimension's current contrarian reading. Market Top and Bottom Indicator is available to all paid Substack subscribers. Start here:

Research| $META : Building a NeoCloud and Continuing to Rent Capacity are not Contradictory Not a contradiction: Meta considering external commercialization of surplus AI compute does not mean its AI compute demand has peaked. Meta is still locking in major new capacity, including roughly 1.6 GW from Crusoe data centers in Childress, Texas, and Warrenton, Missouri. Generational compute shift: Meta’s large H100/H200 fleet remains valuable for inference, fine-tuning, enterprise model serving, image/video generation, and traditional ML. But for 3T+ parameter MoE models, long context, multimodal training, and RL-heavy post-training, GB200/GB300 and future Vera Rubin systems offer better economics for frontier training. Tiered AI infrastructure: The AI compute market is moving from a single GPU shortage into multi-generation, tiered pricing and usage. GB300/Rubin-class systems remain scarce for frontier model training, while H100/H200 can shift toward inference, hosted models, agent workloads, and external compute monetization. Supply chain implication: Meta building a NeoCloud is not a bearish signal that demand has collapsed. It suggests GPU fleets are becoming financialized, multi-generation assets: older GPUs do not go to zero, and next-generation training compute remains scarce. Detailed Report

Weekly|GLM 5.2, Memory Crowding Out Others, LLM Deep Dive, $MU Super-Cycle, $NBIS Re-Rating, Optics Moat Intact, Orbital DC Reality Check One of the questions that set the tone this week was whether China’s open-source models have started to hollow out frontier pricing power. We’d resist the framing. The recent risk-off across AI names traced a leverage unwind that rolled from Korea into Japan and then the US far more cleanly than it traced any GLM headline, and GLM has so far drawn a fraction of the hedge-fund attention DeepSeek once commanded. We also won’t wave the model away. On our internal data, GLM 5.2 lands around Opus 4.6, with a few benchmarks closer to 4.7–4.8, and the gap that actually matters has quietly compressed as third-party token routers turn cleaned inference traffic into RL datasets that open-source labs can finally buy. The cadence still reads as roughly six months behind the frontier, inside the market’s six-to-nine-month consensus, yet “good enough for the agent workload” is becoming the operative phrase for a widening set of tasks. A more important narrative shift drew heavy market attention this week: rising memory prices have begun to squeeze the margins of other semiconductor and hardware names, including AI beneficiaries. The market ran this logic once before, in 4Q last year, but it stayed contained within consumer electronics. This time, it has spread across AI, semi, and hardware beyond memory itself. We still see the trend in LLM ARR growth as the variable that sets market confidence. Yet regulation is keeping the latest Anthropic and OpenAI models off the market, Meta and Google have been slow to show clear progress, and Chinese open-source models keep pressing in. On balance, we think the AI trade, which has run strong for months now, sits in a fragile window, and investors should take note. We will keep tracking both the LLM side and the supply chain to read where the sector turns next. This Week’s Reports H1 2026 LLM Update (Parts 1 & 2). Our latest read on the model layer: the Agent Scaling Law is the real story, and token maxxing has moved into its ROI-discipline phase. Part 1 frames the industry dynamics and why a token-price pullback doesn’t mean peak AI demand; Part 2 goes lab by lab across the frontier and Chinese model vendors, landing on compute as the binding constraint of the value chain. Orbital Data Centers. Lower launch costs get the headlines, yet the orbital data-center case lives or dies on heat rejection, asset life, and replacement economics. We lay out why GW-scale stays a distant option and which engineering milestones actually de-risk it. Premium Report Snapshot - Institutional Report Review|MU FY26Q3: Clearing a High Bar; Memory Super Cycle Endures Micron printed $41.5bn revenue and $25.11 EPS against buy-side at $38.1bn / $22.38, and guided the August quarter above. The memory super-cycle keeps building, and the print does little to argue otherwise. Deep|NBIS: From Scarce GPU Capacity to Full-Stack AI Cloud Optionality An updated Nebius deep dive that reframes the story from scarce GPU supply to a full-stack AI cloud, with Token Factory and utilization as the re-rating levers, and stronger open-source models as a tailwind for token-factory demand. Deep|LITE & COHR: China Laser Capacity Doesn’t Dent the High-End AI Optics Moat Channel concern over second-tier Chinese laser and InP capacity is overdone; the high-end AI optics moat at Lumentum and Coherent holds because the roadmap moves too fast for mature-node-style commoditization. Detailed Report

We published our $MU preview for institutional clients last week. The EPS guidance MU announced today was almost exactly in line with our preview. Memory supercycle: AI-driven demand is shifting negotiating power toward memory vendors, with CSP customers appearing more passive on both short-term contract pricing and LTA terms. We believe the cycle could peak in 2H27 but remain at elevated levels for some time. DRAM and HBM pricing: We estimate traditional DRAM prices will rise about 45% in FYQ3, while overall DRAM including HBM rises around 40%. For FY4Q26 and FY1Q27, we model overall DRAM price increases of 27% and 10%, and assume HBM prices at end-CY2027 are 80% higher than end-2026. Agentic AI demand: Intel and AMD server CPU shipments are expected to grow 30% and more than 50%, respectively, in 2027. Agentic AI workloads require real-time caching of execution states, Chain of Thought steps, tool calls, and agent interactions in system DRAM, adding incremental data center memory demand. Financial outlook: We forecast FYQ3 revenue of $37bn and EPS of $21.6, above company guidance of $33.5bn ± $750mn and $19.15 ± $0.40. For FYQ4, we project revenue of $49bn and EPS of $30.1; for CY2027, assuming an 80% HBM price hike, Micron revenue could reach $279bn with EPS of $179.2. Read the full report: