Weekly|Singapore Gathering, Malaga ECOC Happy Hour, AI Labs Keep Training, Optics Supply Tightens, AIDC Power, CRM Dreamforce, PLTR Sovereign AI, BE First, thank you to everyone who joined our gathering in Singapore this week. Registration and attendance were well beyond our expectations, and the atmosphere was amazing. Members of both our analyst and developer teams were present and had great conversations with investors and industry experts. Next, our analyst Fabian will attend ECOC in Malaga, Spain on Sep 22-23 and will host a happy hour there. Please RSVP if you are interested. Coming back to the market, the AI hardware selloff after Amodei’s call for restraint looked overdone to us, and the subsequent rebound suggests investors are reconsidering what a slowdown would actually mean. We spoke with people at three frontier labs: internal training and compute expansion are continuing, while additional safety checks are more likely to lengthen evaluation and public-release timelines. Safety risks deserve attention, but our base case remains limited disruption to near-term infrastructure demand. The enterprise side helps explain that view. In our 41-company Tokenomics panel, less than 5% of AI spending goes to the most advanced frontier models. Existing deployments can keep expanding as models become cheaper, more reliable, and easier to integrate. Greater scrutiny also creates work for security platforms and identity providers: enterprises need to monitor agents in production, govern what they can access, and retain logs they can trust. Drug discovery gives us another reason to be constructive on AI demand beyond the current enterprise rollout. The Morgan Stanley Healthcare Conference commentary points to early spending across the workflow: TMO is already seeing AI-driven high-throughput experiments lift reagent consumption, CRL expects more shots on goal to support preclinical demand, and DHR is seeing early lab-automation adoption. SDGR says pharma AI budgets are increasingly incremental, with BMS an early proof point. Morgan Stanley’s healthcare specialist Dan Bone also highlighted Novo Nordisk’s use of Claude Science for drug-discovery challenges, GenScript’s completion of Lilly’s TuneLab validation loop for wet-lab services, and $290mn of financing for Anew Labs, which spun out of ByteDance’s internal drug-discovery team in June. The rally in names such as TWST shows investors are taking notice. For us, the demand case is that more AI-generated candidates can support repeated rounds of modeling, screening and physical validation, expanding compute use alongside laboratory spending. It is early to size that compute opportunity, but the evidence of incremental budgets and reagent consumption makes drug discovery a credible additional source of AI demand. Physical delivery remains the harder constraint across the infrastructure names we cover. CIOE orders and production schedules extend into 2027, but 3nm DSPs, lasers, and PCBs determine which optics vendors can ship. On power, our follow-up channel work supports 2.4–3.0MW of turbine nameplate per MW of IT for the off-grid designs we examined; that ratio doesn’t apply to the entire turbine market. Oracle’s results reinforce the value of earlier energization for Bloom, with permitting, manufacturing, and service capacity still the key checks. This Week’s Reports AI labs keep training despite the pacing debate. Our interviews point to longer evaluation and release timelines with limited near-term disruption to training investment or enterprise inference demand. Optics: orders extend into 2027, and access to supply determines who delivers. Our conference and CIOE notes track higher laser pricing, tight 3nm DSP capacity and the NPO/CPO ramp, with leading vendors best placed to turn strong demand into shipments. AIDC: nameplate capacity needs to be translated into dependable IT power. We answer six investor questions with new channel checks and a rebuilt off-grid sizing framework, separating redundancy, site derating and project readiness from headline turbine supply. Cybersecurity: AI safety concerns strengthen the case for platform and identity security. Our note maps the gaps exposed by agent incidents to runtime monitoring, non-human identity governance, isolation, vulnerability management and tamper-resistant logs. 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. Research | CRM: Dreamforce Clarifies Claudeforce Monetization as Salesforce Cedes the UI Deep | PLTR: Sovereign AI Expands the Opportunity; Partner Delivery Sets the Long-Term Ceiling Research & Roadshow | BE: Oracle Reinforces the Time-to-Power Thesis; Delivery Remains the Test Weekly Expert Interviews Summary A snapshot of the expert interviews we conducted during the past week is below; full transcripts and takeaways are available on the FUNDA platform. Meta: Muse Monetization and Advertising Share (META, GOOGL) Enterprise AI: Token Budgets and Model Routing (GOOGL) Microsoft: Internal AI Spend and Usage Mix (MSFT) RoboSoft: Workflow Automation and AI Spend Growth LendingClub: Production AI and Model Routing (LC, AMZN) Delta: Vertical Power Supply Bottlenecks and 800V Architecture (FLEX, GOOGL, MPWR, META) Vicor: Vertical Power Economics at Very High Current (VICR, NVDA) Detailed Report