AI and digital platforms: Value chain, from upstream to the customer
Model builders and apps selling AI as assistants, APIs and copilots, not hardware
Value chain, from upstream to the customer
Criticality to the chain: 5 · Chokepoint 4 · Critical 3 · Important 2 · Supporting 1 · Commodity service Stage 1 AI chipsDesign the GPUs and custom accelerators that train and run models, and sell them with networking and software stacks to clouds and labs.5
Why chokepoint: Nvidia took an estimated 80-90% of AI accelerator revenue in 2025, and it, AMD and most custom chips are fabricated and CoWoS-packaged by one foundry, TSMC in Taiwan, so supply is rationed with no near-term substitute. Technologies GPU and TPU architectures High-bandwidth memory (HBM) Advanced packaging (CoWoS) High-speed interconnects (NVLink, InfiniBand, Ethernet) CUDA and other software stacks Depends on Commodities Si SiliconCu CopperAu GoldHe HeliumTa Tantalum Chemicals ESG Electronic specialty gasesPR PhotoresistsUPW Ultra-pure wet chemicalsEPX Epoxy resins Energy Dist Distribution grids, mini-grids and interconnectors Industries SemiconductorsExamples: Nvidia, AMD, Google (TPU), BroadcomBuild and run data centres full of accelerators and rent training and inference capacity to labs, enterprises and governments.1,3,6,7
Why critical: Three hyperscalers held about 63% of world cloud infrastructure spending in early 2026, and only a few firms can finance and power gigawatt-scale campuses, which wait years for grid connections and gas turbines. Technologies GPU clusters with liquid cooling High-voltage grid connections Software-defined networking Model serving platforms On-site storage and backup power Depends on Commodities Cu CopperAl AluminiumSt SteelH₂O WaterGas Natural gas Chemicals HFC Fluorochemicals and refrigerantsWT Water treatment chemicals Energy DC Data centresHVAC AC transmissionCCGT Gas turbines and combined cycleLFP Lithium iron phosphate batteries Industries Data centres and cloudComputers and electronicsElectrical equipmentExamples: Microsoft Azure, Amazon Web Services, Google Cloud, CoreWeave, Yotta Data Services · ≈ 55% of AI spending in 2025 (calculated from)1License text, images, video and code for training, and supply human labelling and expert feedback to tune models.
Why important: Many sources and labelling firms exist, but high-quality licensed and expert data is scarce and copyright disputes can cut off supply. Technologies Data licensing and rights management Annotation platforms Reinforcement learning from human feedback Synthetic data generation Depends on Energy DC Data centres Industries Media and entertainmentSoftware and IT servicesExamples: Scale AI, Reddit, Shutterstock, iMeritTrain AI researchers and engineers, publish methods and benchmarks, and build open models and datasets, including for Indian languages.
Why important: Talent pipelines take years, but many universities worldwide train AI researchers. Technologies Open-source models and datasets Benchmarks and evaluations Academic compute clusters Indian-language corpora Depends on Energy DC Data centres Industries Education and researchExamples: Stanford University, Mila, IIT Madras (AI4Bharat), IISc BengaluruFund model labs, chip start-ups and AI applications with equity, compute credits and debt for data centres.2
Why important: Corporate AI investment hit US$ 581.7 billion in 2025, but it depends on a few very large backers and on capital market sentiment. Technologies Venture and growth equity Compute-for-equity deals Project finance for data centres Sovereign AI funds Depends on Industries Finance: banking, insurance, marketsExamples: Microsoft, SoftBank Vision Fund, Sequoia Capital, Peak XV PartnersTrain large language, image and multimodal models and sell access through apps and APIs to consumers, developers and enterprises.2
Why critical: A handful of labs build frontier models, with annual compute bills of US$ 16.3 billion for OpenAI and US$ 6.8 billion for Anthropic in 2025, though open models give partial substitutes. Technologies Transformer and mixture-of-experts architectures Large-scale distributed training Reinforcement learning and alignment Inference optimisation Safety evaluations Depends on Energy DC Data centres Industries Data centres and cloudEducation and researchMedia and entertainmentFinance: banking, insurance, marketsExamples: OpenAI, Anthropic, Google DeepMind, Meta, Sarvam AIProvide the data warehouses, model hubs, fine-tuning and monitoring tools companies use to build their own AI systems.
Why supporting: Many vendors and open-source tools compete, and switching takes weeks for most workloads. Technologies Data lakehouses Vector databases Model hubs and fine-tuning Evaluation and monitoring Agent frameworks Depends on Energy DC Data centres Industries Software and IT servicesData centres and cloudExamples: Databricks, Snowflake, Hugging Face, Fractal AnalyticsIntegrate models into enterprise systems, build agents and copilots for clients, and run AI operations under services contracts.1
Why supporting: Thousands of integrators compete for AI services work, which Gartner puts at about US$ 439 billion in 2025, so clients can switch within weeks. Technologies Retrieval-augmented generation Workflow and agent integration Legacy system modernisation AI governance and testing Depends on Energy Bld Energy in buildings Industries Software and IT servicesExamples: Accenture, Tata Consultancy Services, Infosys, IBM ConsultingOffer assistants, search, copilots and AI features in social, productivity and vertical apps, earning subscriptions, usage fees or ads.
Why supporting: Thousands of apps compete and users can switch easily, though platforms with large user bases have strong distribution advantages. Technologies Chat and voice assistants Coding and office copilots AI search Recommendation feeds Agents that act in software Depends on Energy DC Data centres Industries Software and IT servicesData centres and cloudFinance: banking, insurance, marketsExamples: OpenAI (ChatGPT), Microsoft (Copilot), Perplexity, Meta (Meta AI), Zoho (Zia)