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

Summary

• SemiAnalysis built its consulting arm to bridge AI infrastructure and capital allocation—what Abhilash Jain calls “tech finance co-design.” Dedicated teams turn its 13-14 institutional models and technical network into custom work lasting from three to four weeks up to a year, serving investors considering $1 billion-$2 billion investments and industry clients making product, site and go-to-market decisions.
• The clearest example is a global token-supply model delivered as living software rather than a slide deck. It combines the current chip install base and its expected five-year growth, training-versus-inference allocation, throughput per GPU and price per million tokens; clients can change model, chip, context and cache-hit assumptions in a dashboard and see “literally every number changing.”
• Strategy work maps changing workloads into product specifications and customer priorities. One memory-and-storage company replays agentic traces to test bandwidth-versus-capacity trade-offs, while a semiconductor vendor with roughly 100 possible buyers uses market interviews to select the 10-15 worth pursuing when each qualification takes “at least three to four months.”
• A $1 billion neocloud investment requires both technical proof and commercial benchmarking. SemiAnalysis reviews SLAs, OEM warranties and support, tests clusters, identifies what they are good and bad at, and builds 100- or 180-day remediation plans, while its TCO data benchmarks CapEx and OpEx against the market.
• SLA structure is financing infrastructure, not contractual housekeeping. As operators expand from hyperscalers and specialists to real-estate developers and family offices, Abhilash’s blunt assessment is that “without strong SLAs, you won’t be funded, period”; a failure can put lender payments at risk, trigger credits or termination, and leave a 15-20-year data-center lease to be paid despite a five-year GPU contract.
• Risk assessment must trace each customer promise through the data center, OEM, hardware and software stack. Reviews can range from three core contracts to L3-L5 readiness and bare-metal failure simulations, because electricity, cooling, racks, monitoring or software can each break performance—and “only one thing has to go bad” for the operator to owe SLA credits.
• Build-versus-rent recommendations turn primarily on time, operating capability, willingness to pay and access to capital. A client seeking 100 MW “in literally the next six months” was told to rent because building was infeasible; clients able to wait until 2028 or 2029 received the opposite recommendation.

Deep dive

1. Consulting exists where off-the-shelf research stops

• SemiAnalysis had always done consulting, but established a dedicated team only in the prior year. Much of its 13-14 institutional-model and open-research work, including projects such as Infant X and ClusterMAX, was off the shelf; deeper client-specific work lacked an operating model.

• Engagements receive dedicated staffing for anywhere from three to four weeks to a year. Abhilash’s framing is “tech finance co-design”: “We need to teach tech bros more finance” and finance professionals more technology.

• Financial clients include large private-equity and hedge funds evaluating billion-dollar chip, neocloud and data-center investments or developing the “next wave” after themes such as memory and CPUs. Industry clients span hyperscalers, AI labs and semiconductor companies planning products three to five years out, remote sites in India or go-to-market execution.

2. Token-supply forecasts become living software, not static decks

• One hedge fund asked for total worldwide token supply implied by the chip base. The model links installed capacity and its five-year growth, the share assigned to training versus inference, daily throughput per GPU, and pricing in dollars per million tokens.

• The difficult variable was future throughput: SemiAnalysis had to extend its public inference data to chips and models that did not yet exist. The simulator tests changing silicon and memory specifications alongside possible 10-trillion-, 15-trillion- or 20-trillion-parameter models.

• Too many permutations made “a static Excel” impractical, so the deliverable became an editable dashboard. Inputs include inference allocation, AMD/TPU/NVIDIA specifications, input-output ratios, agentic harnesses, sequence and context lengths, data types, and cache-hit rates of 80%, 85% or 90%.

• Jordan’s correction is worth keeping: ClusterMAX did not originate as consulting, but the AI cloud TCO model did. The inference simulator followed a similar path—from internal “science project,” to one client, to several clients, and ultimately toward becoming a consulting product by itself.

3. Strategy work links workload physics to products and customers

• A memory-and-storage client uses the simulator to replay agentic traces across model-and-chip combinations. The product question is concrete: how should bandwidth-versus-capacity trade-offs change when serving real production inference workloads?

• Another semiconductor company had a product and perhaps 100 potential buyers, but no clear answer to “who to sell to.” Because qualifying one customer takes at least three to four months, SemiAnalysis narrowed the field to 10-15 priorities and shaped the technical sales conversations for each.

• That prioritization required speaking with 20, 30 or 40 prospective customers and aggregating their requirements into a market view. Abhilash said the firm avoids expert-call agencies here, relying instead on direct ecosystem relationships.

4. A $1 billion neocloud check demands technical and commercial proof

• Against Dan’s discussion of $7 trillion in debt issuance to fund AI infrastructure, Abhilash described diligence as increasingly necessary for both established investors and newer entrants. A private-equity fund placing $1 billion into a neocloud needs separate technical and commercial answers.

• Technical diligence asks whether the operator can meet its promised SLAs, whether OEM support, warranties and contracts leave hidden exposure, and how the cluster actually performs. ClusterMAX-inspired testing then supports a 100- or 180-day improvement plan.

• Commercial diligence reconstructs the build’s CapEx and OpEx and benchmarks them against current market costs through the AI cloud TCO model. The framework changes for chip companies, model companies, data centers and powered land, but the aim remains measurable downside analysis.

5. SLAs turn one weak link into a financing risk

• Jordan separated two major termination paths: late delivery and insufficient availability. If a cluster arrives months late, the customer may cancel; if availability is under 90%—roughly three unavailable days in 30—the remedy can include credits or termination.

• Abhilash called SLAs “really for the bad times.” Lenders care whether downtime could stop an interest payment, while operators face reputational damage that may prevent a second customer deal or third lease even as a 15-20-year data-center obligation remains.

• SLA insurance can cover specified penalties in exchange for a premium measured in basis points, potentially making a contract more investable. It is not a substitute for operations: “If you have bad performance, you won’t even get insurance,” because insurers underwrite the neocloud itself.

• Risk assessment can begin with the offtaker, OEM and data-center contracts, then expand to L3, L4 and L5 readiness, rack delivery, power and cooling integration, and bare-metal testing. SemiAnalysis deliberately simulates failures and checks monitoring and recovery because “only one thing has to go bad” to trigger liability.

6. The capacity clock decides whether to build, buy or partner

• Tight colocation supply, chip availability and labor mean building is often incompatible with near-term demand. The decision turns on time to market, willingness to pay a rental premium, in-house cluster-operations capability and credibility in capital markets.

• A company needing 100 MW within six months was advised to rent: “There was no way” it could build that quickly. Where a client could wait until 2028 or 2029, the recommendation was “completely opposite”; these assessments typically take three to four weeks.

• Repeated neocloud diligences have become “muscle memory”: information requests on day two, interviews on day four, five anticipated follow-up questions on day six and a midpoint readout on day eight. When Jordan objected that conclusions cannot be predetermined, Abhilash joked, “You’d be surprised how often they are.” Jordan then asked whether recurring neocloud contracts make results clear quickly, and Abhilash agreed.