SemiAnalysis — searchable YouTube transcripts & AI Q&A

Ep. 043 - The Midterms and the 2028 Datacenter Buildout (Datacenters)

The 2027 data center buildout is largely locked in, but 2028 capacity is at risk from state-level moratoriums and conditions that could follow the US midterms. Texas is the biggest risk, where a freeze on air permits could delay projects that lack final approvals, potentially impacting even 2027 deliveries. The speaker argues that while both parties support tougher conditions, a full pause is less likely, and developers can adapt by meeting stricter requirements.

Ep. 042 - GDPx: The Benchmark for Frontier AI | Andrew Megalaa, Jordan Nanos, Max Kan

The GDPx benchmark is designed to measure frontier AI capabilities in domains that can absorb increasingly intelligent models, such as chip design, biology, and cybersecurity, rather than day-to-day tasks. The creators argue that many popular benchmarks are flawed due to unfair grading or lack of realism, and they partnered with specialized data companies to create tasks that are both difficult and fair. Initial results show that models like Astra and Opus perform well, but safety classifiers can artificially lower scores in certain domains, and the benchmark aims to provide a more accurate measure of model capabilities.

Ep. 041 - EMERGENCY: Inside the Anthropic’s S-1 (Tokenomics/ Markets)

Anthropic's S-1 reveals a highly profitable business with 85% inference gross margins and 41% compute margins, already profitable for two quarters. The speakers argue that despite large compute commitments, the company's flexible deals and strong balance sheet make them manageable, and they see Anthropic as well-positioned to dominate the AI market, potentially surpassing OpenAI. They also note that enterprise adoption is still early, with top customers spending significantly more than average users, indicating substantial growth runway.

Ep. 040 - ASML Is Building a Hybrid Bonding Tool, and We Have the Patents (AI Supply Chain)

ASML is developing a hybrid bonding tool that uses a conveyor belt system, where dies remain on a track and the wafer moves to meet them, a departure from traditional pick-and-place tools. The speaker argues this approach leverages ASML's metrology and wafer-handling strengths for higher alignment accuracy, but faces challenges in cleanliness and maintaining precision in a moving system. Hybrid bonding is primarily for logic chips, not HBM, which is unlikely to adopt it before 2029 due to cost and current demand.

Ep. 039 - [EMERGENCY] Brad Gerstner Cut Our 2027 Forecast in Half. We Raised It. (Datacenters)

The SemiAnalysis team defends their 2027 data center forecast of 50 gigawatts against Brad Gerstner's public criticism, arguing that their bottom-up, site-by-site tracking of construction, permits, and power allocations makes their number more reliable than headline-driven haircuts. They explain that while moratoriums, grid constraints, and labor shortages are real, they affect only a small fraction of projects, and modular construction is accelerating build times. They also warn that by 2028, data center supply could outpace chip supply, flipping the current constraint.