Frequently asked questions

Five reconciled modules: accelerator shipments, ASPs, and specifications by SKU, optical engine and co-packaged optics attach, supply chain and capacity orders, HBM and memory tracking, and customer-level installed base with FLOPS. Forecasts run to 2030 with history back to 2021, on a quarterly basis.

As an Excel workbook with dashboard access for your team.

Yes. Because production is built up from equipment requirements, wafer starts, and packaging orders, revenue for many upstream and downstream companies can be estimated directly from the data.

Four reconciled modules: cluster network configurations for every major buyer, bottom-up networking builds and spend, top-down market volumes and vendor share, and a master pricing and power table. Data runs quarterly with forecasts through 2026.

As an Excel workbook with dashboard access for your team.

Yes. Unit volumes, pricing, and wallet share by customer mean revenue for switch, optics, and interconnect vendors can be estimated directly from the build.

Installed wafer capacity and utilization by fab, vendor, and node, DRAM node transitions across 1α, 1β, and 1γ, NAND layer counts from 176L to 300L and beyond, die size and yield analysis with gross and net good die, bits per wafer, bit supply and demand by vendor and end market, contract pricing with forecasts, HBM pricing and shipments by generation and supplier, vendor margins from die-level cost modeling, and a DRAM and NAND capex tracker. Quarterly data runs from 2022, through 2030 for DRAM and 2027 for NAND.

All major and emerging vendors. For DRAM: Samsung, SK Hynix, Micron, CXMT, Nanya Tech, Winbond, JHICC, and YMTC. For NAND: Samsung, SK Hynix including Solidigm, Kioxia-Sandisk, Micron, and YMTC.

HBM pricing across multiple generations and suppliers, both backward- and forward-looking, paired with vendor-level shipment volumes expressed in stacks, bits, and wafers, plus HBM and SOCAMM supply tracked in wafers and bits. The analysis is aligned with the SemiAnalysis Accelerator Model and its global AI accelerator shipment data.

As an Excel model with quarterly data from 2022, through 2030 for DRAM and 2027 for NAND, one year of quarterly updates with additional features and enhancements, an initial onboarding call on the model structure and methodologies, ongoing ad-hoc support calls, and timely updates from the lead analyst on key industry developments.

Four reconciled modules: per-accelerator performance and cost benchmarks, the all-in GPU cost of ownership and rental pricing, a cluster finance suite through a three-statement model, and workload and LLM economics.

As an Excel workbook with dashboard access, including one year of quarterly updates, an onboarding call with the team to walk through the model and methodologies, and ad-hoc calls for questions that come up in use.

That is what the finance suite is for: net present value and residual value from future cash generation, cumulative project and equity cash flow, IRR and the full returns stack, and a three-statement model whose capital structure, contract, and lifetime assumptions you control.

Nvidia and AMD accelerators plus the major custom silicon programs, TPU and Trainium among them, each with throughput, power, cost of ownership per hour, inference cost per million tokens, and training cost per FLOP.

The full cost stack of a space datacenter built from first principles, total cost of ownership for space and terrestrial deployments bridged to a levelized cost of compute, launch economics for Falcon 9 and Starship, terrestrial supply and silicon constraints as five layers of incremental power, demand and installed base by scenario, cost crossover and parity outputs with sensitivity tables, orbital mechanics, and 3-D interactive CAD models of major components. Annual from 2026 to 2050.

A Base Case with a bullish chip and Earth datacenter capacity ramp, and an “Elon Musk” Case with constrained terrestrial supply and aggressive launch and Terafab assumptions. Both are fully adjustable, and every scenario is driven by user-controlled assumptions.

Total cost of ownership in dollars per GPU-hour, per PFLOP-hour, and per billion tokens, bridged to a levelized cost of compute via radiation availability and GPU redundancy and SLA provisioning, plus cost crossover and parity-ratio outputs and full launch-cost and terrestrial-cost sensitivity tables.

Solar arrays, radiators and the thermal transport loop, battery and power management, bus hardware from ADCS to reaction wheels, propulsion and propellant, radiation and MMOD shielding, communications and the ground segment, and structure, integration, and test. Every line is adjustable for cost and technology scaling.

Installed base and investment forecasts for hyperscalers, foundation labs, and neoclouds, bottoms-up token throughput by hardware system, model architecture, and workload, addressable market analysis of the token economy, tracking of SaaS disruption by token consumption pricing, ROIC of AI deployments, and aggregate inference and training hardware demand with revenue and profit forecasts across rental, model, and software lines.

Demand side first. Token consumption is tracked across applications, API endpoints, and token-native software, and set against bottoms-up token production per hardware system and model architecture. Usage and revenue growth then translate into future hardware demand and ROIC by deployer, closing the loop between AI spending and AI income.

As an Excel workbook with dashboard access, including one year of quarterly updates, an onboarding call with the team to walk through the model and methodologies, and ad-hoc calls for questions that come up in use.

Yes. It is the first demand-side model of AI, built to translate revenue and usage growth into future hardware demand across Nvidia GPUs, AMD GPUs, Google TPUs, Amazon Trainium, and more, split into inference demand from adoption and training demand from model development.

Four reconciled modules: site-level capacity by operator across more than five thousand tracked facilities, power demand and supply by geography, industry capex with the equipment stack behind it, and the economics and energy analysis on top. Data runs from 2017 with forecasts through 2032.

With publicly available evidence: property records, permits, power usage, FOIA requests, and frequent satellite imagery, with computer vision models accelerating the reads on size, capacity, timelines, and construction progress.

As an Excel workbook with dashboard access, including one year of quarterly updates, an onboarding call with the team to walk through the model and methodologies, and ad-hoc calls for questions that come up in use.

Yes. The supply-demand analysis sets each operator’s facility pipeline against the power its estimated accelerator shipments will draw, by hyperscaler and by region, including the ones that come up massively short.

More than a thousand datacenter buildings in mainland China across over sixty players, with an annual capacity ramp from 2017 to 2023 and a quarterly ramp from 2024 to 2032. Colocation, telecom, hyperscaler self-build, and financial-institution capacity are all in scope. Each row is one development phase, carrying its owner, tenant where attributable, location, EDWC hub, and construction and operations dates.

Same bottom-up approach, and the same core tabs, metrics, and views. Key differences: 1) this model is supply-side focused; 2) it does not include some of the new functions we introduced specifically for the global Datacenter Industry Model (the neocloud tabs, the HPC Lease Tracker, Datacenter Developers, Cryptominers, Datacenter Backstops, the Vision Model, or the grid view on the map); 3) it adds a country-specific EDWC hub analysis tab, and a hyperscaler commitment tracker that compares leasing and self-build start trends with reported capex; 4) locations are given at the district level, and there is no satellite view.

Small projects under 30 megawatts, captive government and enterprise datacenters, edge nodes, and very early announcements. Financial-institution self-builds are not covered comprehensively. Announced phases with no construction visibility are recorded but labelled speculative and kept out of every chart.

A major update every quarter. As with our global Datacenter Industry Model, we will keep adding layers with every update. Subscribers get the map today and ongoing tracking of China’s AI buildout.

We present a live dashboard with supply analytics, company and hyperscaler views, leasing trackers, EDWC hub analysis, capex and load growth, an interactive map, and an explorer table, plus a full Excel model builder and the methodology notes.

Top-down utility-level datacenter load forecasts adjusted for PUE and reserves, a bottom-up national inventory of generation, storage, and onsite energy translated to firm capacity, tracking of new generation and transmission projects, order-by-order tracking of behind-the-meter generation and storage at US datacenter sites, regulatory-framework coverage of markets like ERCOT and PJM, and supply-demand gap analysis across major US electricity markets.

A national inventory of electric generators, storage, and onsite energy solutions is integrated with SemiAnalysis datacenter IT capacity projections and the integrated resource plans of electric utilities, RTOs, and ISOs. Demand is then compared against each utility’s own forecasts, so the gaps are visible before they show up in filings.

As an Excel workbook with dashboard access, including one year of quarterly updates, an onboarding call with the team to walk through the model and methodologies, and ad-hoc calls for questions that come up in use.

Yes. The model provides a forward-looking view of which utilities face the most severe supply-demand gaps, where and how new generation capacity comes online, and where power supply and transmission constraints may delay load growth, along with the solutions that may bridge each gap.

Three connected pieces: demand-side analysis built from Bills of Materials across 20+ datacenter designs with capex per megawatt for 75+ equipment lines, a supply-side tracker covering 550+ suppliers with revenues built bottom up by equipment line, and the supply-demand reconciliation with imbalance analysis by segment. Update notes, research notes on key Industrials topics, and calls with the research team ship with the model.

The industrial side of the datacenter buildout across five segments: cooling, electricals, construction, other hardware, and services & digital. Coverage runs from UPS, backup gensets, switchgear, and transformers to chillers, CDUs, the shell, MEP contractors, and commissioning, mapped across 550+ tracked suppliers.

Demand first: a unique dataset of Bills of Materials across 20+ datacenter designs sets capex per megawatt for 75+ equipment lines, and those BoMs are mapped onto the 6,000+ individual sites of the SemiAnalysis Datacenter Model. Supply is then built from the bottom up, starting from the revenues of each industrial equipment line, and the reconciliation between the two is where the imbalance analysis lives.

Wafer capacity forecasts for leading-edge logic and 10nm-class DRAM through 2028, lithography sales forecasts by tool model, units, and revenue, layer-by-layer lithography requirements from first EUV use through 1.4nm-class, aggregate wafer fab equipment sales across lithography, deposition, etch, wafer track, and metrology and inspection, capex by chipmaker and process, and a lithography tool database with throughput, uptime, rework, release dates, and ASPs.

Applied Materials, ASML, Lam Research, KLA, Screen, ASMI, Kokusai Electric, Tokyo Electron, and more, with sales split across leading-edge logic, specialty including trailing-edge logic, analog, power, and RF, DRAM, NAND and non-volatile, and China.

Bottoms-up: wafer capacity and process node requirements drive equipment sales. Layer-by-layer flows for advanced logic, across finFET, gate-all-around, and backside power delivery, translate capacity into tool demand, with lithography tool productivity from the database and technology shifts like high-NA EUV, pattern shaping, and directed self-assembly priced into future requirements.

As an Excel model covering 2021 through 2028 on an annual basis, with one year of monthly updates, an initial consultation call on the model and methodologies, and ad-hoc follow-up calls for questions that come up in use.

Foundry and wafer revenue and ASPs by node, wafer shipments by node, non-wafer revenue streams, installed capacity and utilization by node, capex, wafer starts by customer and end market, detailed TSMC–Apple modeling, and financial metrics including margins, R&D, depreciation, COGS, and inventory, with quarterly forecasts through 2030. The bundle also includes the Apple, Intel, and Huawei wafer-demand models, the analog and power supply-demand model, the EDA Model, the TSMC MOPS Capital Budget Model, and the live Foundry and EDA dashboards.

Operating foundries covered include TSMC, Samsung Foundry, SMIC, UMC, GlobalFoundries, Tower Semiconductor, Hua Hong, DB HiTek, Vanguard International Semiconductor, Powerchip, X-Fab, Intel Foundry, SkyWater, and other tracked manufacturers, with more than 20 years of history for the five largest and more than 10 years for the rest. Announced and emerging projects, including Rapidus, Tata Electronics, and Tesla Terafab, are monitored and clearly distinguished from operating foundries.

Bottom-up from proprietary wafers-per-tool calculations and fab build plan tracking, blended with conference materials, public filings, customs data, and primary research. Fabless revenue, COGS, and inventory reconcile the demand side, leading-edge wafer starts are built bottom-up, fab by fab, from our customer wafer-demand models, bottom-up cost models estimate margins, and everything is back-tested against reported results each quarter.

As a quarterly Excel workbook with the raw tables and near-term forecasts, alongside a quarterly PDF snapshot with charts and executive insights. Market-event addenda ship between releases, and the live Foundry and EDA dashboards support company, process-node, market, and mature-node analysis with exportable charts and tables.

A system-level bill of materials and integrated power view across six groups: compute and interconnect modules (the Strata module with Rubin GPU, Vera CPU, and SOCAMM memory, plus BlueField-4 and ConnectX-9), NVLink and switching infrastructure, system and tray infrastructure, thermal and liquid cooling, PCB, substrate, and materials content, and power delivery and supply.

The compute building block of the platform: Rubin GPU, Vera CPU, and SOCAMM memory on one module. The model prices it with the VRMs and substrate content around it, and follows the rest of the tray from there.

Yes. The model is a bill of materials with an integrated power view: power shelves, busbars, connectors, VRMs, and power delivery modules are priced as components, and the power budget reads alongside the six groups.

The Vera Rubin NVL72 rack generation: Rubin GPUs, Vera CPUs, and the NVL72 architecture around them, from Strata modules to NVSwitch trays and the rack manifold.

As a standalone licensed model at USD 10,000, separate from the newsletter subscription. Buy it on the VR NVL72 model page: log in or create an account and complete the purchase. Once you have bought it, logging in there shows the report and the download.

A monthly AI and semiconductor industry tracker with timely and targeted insight into the overall state of the industry and specific sectors and trends. Each edition carries thousands of unique data points selected to distill clear, actionable insights for investors and company decision makers.

A 40 to 50 page deliverable with semiconductor charts and equity references, plus full access to every previous edition.

Core trackers are consistent monthly slides covering the value chain and cycle: memory trackers, smartphone and PC imports, WFE imports and exports by geography, auto exports, bare wafer and boule trade, and PCB sales. Timely datasets are rotating slides hand-picked from 250+ actively tracked datasets, such as WFE imports by tool type, auto exports by powertrain, the Taiwan CoWoS supplier revenue index, and the Nvidia server component supplier index.

Yes. The subscription includes monthly calls with the ChipBook analysts for additional and personalized insight. Questions reach the team at ChipBookSales@SemiAnalysis.com.