Ep. 035 - Tech DD’s, Performance Projections, Benchmarks, Supply Chain, Investment Thesis (Consulting)| Abhilash Jain, Jordan Nanos

Summary

Abhilash Jain and Jordan Nanos describe SemiAnalysis consulting as a dedicated, custom-services business that grew out of newsletters, institutional models, and open research projects. The firm serves both financial investors and industry clients, often on engagements lasting from a few weeks to a year, and frames its approach as “tech finance co-design.” A major example is an inference simulator built to estimate world token supply, which required modeling future chips, future models, memory de-speccing, and changing assumptions like cache hit rate and context length. The conversation also covers neocloud due diligence, where SLAs, OEM warranties, cluster performance, and data center readiness can determine whether a project gets funded. The episode closes on build-vs-buy tradeoffs, emphasizing real-world constraints such as colo capacity, chips, labor, time to market, and capital-markets credibility.

Notes

• SemiAnalysis formalized consulting into a dedicated team only last year.
• Projects are highly bespoke, with dedicated staffing from roughly 3-4 weeks to a year.
• Two main client buckets: institutional investors and industry operators.
• The team describes its work as tech finance co-design.
• A hedge fund client wanted the total token supply of the world.
• The token-supply estimate decomposed into install base growth, training/inference split, throughput, and token pricing.
• The hardest part was modeling throughput on unreleased chips and models.
• The simulator explicitly contemplates Nvidia Feynman, de-specced memory, and 10T-20T parameter OpenAI models.
• Static Excel was insufficient for the number of permutations.
• Clients can interactively change assumptions and watch outputs update live.
• Key model inputs include inference-vs-training allocation, next-gen chip specs, I/O ratios, harness type, context length, and cache hit rate.
• Cache scenarios were discussed at 80%, 85%, and 90%.
• Memory/storage product planning can be driven by replaying agentic traces.
• The simulator is being used to weigh memory bandwidth versus capacity tradeoffs.
• Semis companies often need help prioritizing a small set of buyers from a much larger universe.
• Product qualification can take at least 3 to 4 months per customer.
• Consulting outputs can become products such as the data center model and AI Cloud TCO model.
• Some internal science projects appear to graduate into commercial tools.
• Neocloud diligence combines technical and commercial analysis.
• Diligence checks SLAs, cluster performance, and capex/opex versus market norms.
• Strong SLAs are portrayed as a prerequisite for funding a new cloud.
• SLA insurance can protect against penalties, but it comes at a premium.
• Risk review can range from contract review to full data-center readiness testing.
• Readiness milestones include L3, L4, and L5 testing, rack arrival, and power/cooling install.
• Bare-metal testing includes logging in, simulating failures, and checking monitoring and recovery.
• Failure diagnosis spans electricity, cooling, racks, and software.
• A single component failure can still trigger SLA credits.
• Build-vs-buy decisions are constrained by colo, chip, and labor shortages.
• Time-to-market, premiums, team expertise, and capital-markets credibility shape the choice.
• One client was advised to rent compute because it needed 100 MW in six months.
• Waiting until 2028-2029 can flip the recommendation toward building instead.
• The organization is hiring aggressively.