Ep. 035 - Tech DD’s, Performance Projections, Supply Chain, Investment Thesis (Consulting)

Summary

Abhilash Jain and Jordan Nanos explain how SemiAnalysis Consulting evolved from ad hoc work into a dedicated team serving two main client groups: financial investors and industry operators. On the investor side, the firm helps with diligence, thesis building, and AI infrastructure investment decisions around chips, cloud, and data centers; on the industry side, it advises hyperscalers, AI labs, chip companies, and semiconductor firms on roadmaps, siting, and go-to-market. A recurring theme is that the work is deeply technical and highly custom, often requiring dedicated teams for weeks or months and sometimes evolving into reusable internal tools. One example is a token-supply model that depended on installed base, inference/training split, GPU productivity, and token pricing, with the hardest part being simulation of unreleased chips and models. The episode also frames SLAs, insurance, and contract structure as central to neo-cloud financing and long-duration AI infrastructure risk management.

Notes

• SemiAnalysis says consulting existed informally before, but a dedicated consulting team was formed only in the previous year.
• The firm already spans newsletters, institutional models, and open research projects, so consulting was created to handle bespoke asks.
• The stated purpose of the consulting arm is to support deeper requests from industry participants and investors that internal operating models could not satisfy.
• The consulting motion is fully bespoke rather than productized work.
• Engagements typically use a dedicated team for 3 to 4 weeks, and sometimes up to a full year.
• The company frames its broader mission as bridging technology and finance.
• Financial clients are the first major bucket: large PE funds and hedge funds making large AI-infrastructure bets.
• The investor diligence use case includes evaluating billion- or multi-billion-dollar commitments to chips, cloud, and data centers.
• Hedge-fund work includes investment-thesis development and combining multiple data sources and narratives to find the next wave.
• The second major client bucket is industry: hyperscalers, AI labs, chip companies, and semiconductor companies.
• Product roadmap and engineering advice are part of the work, especially for semiconductor clients planning 3 to 5 years ahead.
• The team also advises on data center siting, including remote regions in India.
• Go-to-market support is offered to companies with strong products but weak understanding of customer journeys.
• A recurring theme is that many consulting deliverables are closer to software products than to static reports.
• For complex questions, SemiAnalysis moved away from simple slide decks toward live dashboards.
• Static Excel was insufficient because the assumption space had too many combinations.
• The delivered artifact became a dynamic dashboard that lets clients manipulate assumptions and watch outputs update in real time.
• The company says its network allows it to interview 20 to 40 customers directly, rather than relying on an expert network.
• Some institutional and data products reportedly began as consulting projects or were inspired by a client question.
• The inference simulator is described as an internal science project that later became a product after client validation.
• An internal inference simulator is used by a memory company to replay agentic paths across different models and chips.
• The simulator is explicitly used for product-strategy questions, not just engineering analysis.
• A memory vendor uses the simulator to choose between bandwidth and capacity tradeoffs for inference workloads.
• In a separate modeling example, the team estimates global token supply with a three-part equation.
• The three pieces are installed base growth, training-versus-inference capacity split, and token pricing per million tokens.
• The hardest part of the token-supply model was estimating GPU productivity.
• Public inference data had to be combined with simulations for unreleased chips and unreleased models.
• The simulator was used to ask how Nvidia Blackwell-like specs, memory reductions, and OpenAI training runs at 10, 15, or 20 trillion parameters affect throughput.
• The consulting output often needed an interactive system because there were too many permutations for a fixed spreadsheet.
• The dashboard is intentionally visual and updateable so clients can see every number change as assumptions move.
• SemiAnalysis also supports semiconductor companies that know their product but not their customer list.
• Qualification cycles can take 3 to 4 months per customer, making broad qualification at scale unrealistic.
• The firm says it helps clients prioritize the best 10 to 15 customers rather than trying to qualify 100.
• This customer-prioritization work relies on direct outreach to 20 to 40 customers to assemble a market-wide view.
• A lot of the company’s institutional and data products are presented as a byproduct of those earlier bespoke engagements.
• The episode ties consulting tightly to productization: one-off questions can become reusable tooling if multiple clients care.
• In AI cloud diligence, the first technical questions are the SLAs and whether the provider can meet them.
• The diligence also covers OEM support structure, warranties, and contract terms as risk controls.
• Commercial diligence benchmarks capex and opex against the market, with costs measured in real time through the TCU model.
• DD has been applied across cloud companies, chip companies, model companies, datacenters, and even powered land.
• The new-cloud buyer base is expanding beyond hyperscalers to crypto miners, specialists, real-estate developers, and family offices.
• Because many new operators lack prior cloud experience, contract protections matter more.
• The hosts distinguish SLAs from broader termination rights, saying the key issue is what happens if delivery is late.
• Late server delivery by several months can trigger contract cancellation.
• The most common SLA issue discussed is uptime: buyers expect 99.9% availability on GPUs.
• Below 90% uptime is framed as effectively losing access to the GPUs you paid for.
• Buyers or lenders may seek insurance against termination and downtime-related downside risk.
• The need for insurance is framed as lender protection from not receiving scheduled payments if a cluster fails.
• SLAs matter most when things go wrong, because they determine whether parties can exit without penalty.
• GPU contracts are said to run about 5 years, while datacenter leases are often 15 to 20 years.
• A weak SLA record can damage reputation and block the next financing or leasing step.
• Insurance can be used to cover financial penalties from SLA failures, with premiums quoted in basis points.
• But the hosts stress that insurance is not a substitute for good performance.
• SLAs can attach to multiple layers: GPUs, nodes, racks, clusters, the site, and even power and cooling systems.
• The more counterparties involved, the more reasons there are to assess even a single risky contract component.