Semiconductors & Data Centres in AI Ecosystem

There is no denying the fact that for some time now, money is chasing economies where AI induced capex super cycle is holding sway. Whatever may be our apprehensions regarding AI, we live in a world where AI matters, because AI will increasingly shape economic activity, public services, defence capabilities and access to knowledge. True self sufficiency in AI rests on a strong presence in the following interconnected layers.

1. Energy & Compute (Hardware Foundation)

At the very base of the ecosystem is raw power and processing. AI models require massive amounts of electricity to run, and the specialized processing power (provided by GPUs and TPUs) which relies on foundries and semiconductor fabricators to design and produce hardware at scale.

• Key Players: NVIDIA, AMD, Intel, TSMC, and global energy infrastructure providers.

2. Infrastructure & Cloud

This layer serves as the “factory floor” where the hardware lives. It consists of datacentres, server racks, high-speed networking, cloud storage, and the platforms used to execute computations and train algorithms.

• Key Players: Amazon Web Services (AWS), Microsoft Azure, and Google Cloud.

3. Data

High-quality, diverse, and clean data is the fuel that allows AI systems to learn, identify patterns, and drive decisions. This layer involves data pipelines, vector databases, and knowledge graphs that prepare data for use.

• Key Players: Snowflake, Databricks, Pinecone, and MongoDB.

4. Foundation Models (Intelligence Layer)

This is the “brain” of the ecosystem. It consists of sophisticated algorithms and Large Language Models (LLMs) that have been trained on vast datasets. These models perform reasoning, generate content, and analyse data.

• Key Players: OpenAI, Anthropic, Google (Gemini), and Meta (Llama).

5. Agents & Applications (User Experience)

At the top of the stack is where humans and businesses interact with AI. It includes both user-facing applications (chatbots, copilots, automation tools) and autonomous agents that can independently set goals, analyse situations, and use tools to perform multi-step tasks.

• Key Players: OpenAI (ChatGPT), Microsoft (Copilot), Salesforce, and enterprise-specific automation platforms.

US based Space X, Anthropic and Open AI have launched or are planning to raise resources through IPOs, at combined valuations of trillions of dollars. While OpenAI and Anthropic operate mainly in the Intelligence and Agentic (User Experience) layers, Space X covers a wider gamut of AI spectrum, including space-based AI infrastructure. As of 16^th June, that is, within three days of its listing, Space X market capitalisation exceeded $2.7 trillion, displacing Amazon in the fifth place in terms of market capitalisation in the listed space in the US and making Elon Musk the founder, the first trillionaire in the world. The current investment money trail suggests that India appears to have missed the AI bus. But it has come to realise the importance AI sovereignty. Under the GOI funded AI mission, India has built a multilingual, Large Language Model (LLM); Sarvam.ai. The semiconductor production has begun to take shape with the Tata Group preparing to roll out made-in-India chip sets. But India remains far behind global leaders in most layers of the AI ecosystem, though it is attracting billions of dollars in investments in data centres, strengthening infrastructure backbone required for AI. To appreciate the data centre ecosystem, it is necessary to understand the type of semiconductors required in a data centre and major global semiconductor suppliers.  

AI is supposed to be the next-generational leap in technology. But AI models need bleeding edge semiconductors working non-stop in data centres – both for training and inference. Semiconductor is a broad term and they are ubiquitous as they are found in numerous devices that we use daily. But for AI/data centre applications the relevant ones are the ones required for GPUs, CPU, NANDs, DRAMs and HBMs. They are briefly described below.

GPU or Graphic Processing Units that do the parallel processing, that is doing thousands of calculations simultaneously, that is, at one and the same time. These lend themselves for training AI models.

CPU or Central Processing Units handle ‘logistics’ in data centres. They fetch massive datasets to and from the storage for the GPUs to process. They are also capable of light inference tasks.

NAND flash memory acts as the data warehouse and is non-volatile in nature, that is, data stay after shutting down.

DRAM or Dynamic Random Access Memory. It is like a kitchen with ingredients retrieved from the warehouse (NAND) for the CPU to cook. DRAMs are volatile and the data are erased when the device is switched off. They hold active operating systems, apps and hot datasets in memory for the CPU to process without latency (delay).

HBM Or High Bandwidth Memory chips where multiple DRAMs are vertically stacked and placed next/close to the GPU to ensure faster speeds.

Broadly while GPU and CPU are known as logic chips, NANDs, DRAMs and HBMs are called memory chips.

Although there are many players in the chip space the following eight more or less cover the bulk of the design(fabless) and manufacturing(fab) of semiconductors.

Nvidia (US): Designs, sells GPUs. Fabless; outsources manufacturing to foundries like TSMC. 90% revenue relates to data centres.

Broadcom (US): Designs, supplies semiconductors (60% revenue) used in data centres, network hardware. Manufacturing mostly outsourced to TSMC, Foxconn & others. Rest of the revenue from data centre related software.

AMD (US): Designs, sells CPUs, GPUs for data centres & PCs. Fabless; outsources to TSMC, UMC, Samsung. 50% revenue from data centre chips.

Micron (US): Designs, manufactures memory chips. Derives major part of the revenue from data centres, hyperscalers (Google, Microsoft, Amazon, etc) and phones and PCs. Mix by tech: DRAM 80%, NAND 20%.

TSMC (Taiwan): Operates foundries (fabs) manufacturing semiconductors for fabless logic players. 58% revenue from AI & consumer GPUs, 5G/6G network chips etc. and 29% from smartphone SoCs (System on Chip).

SK Hynix (South Korea): Designs, manufactures NAND (21%) and DRAM (77%). Within DRAM – Servers25%, GPU memory including HBM 30%.

Samsung (South Korea): Memory major, but diversified electronic player. Memory 50%, Smartphones 25%, Appliances 10%, Display panels 5%, Chip design, Foundry 5%.

ASML (Netherlands): Supplies lithography systems, allied software & services to fabs. Absolute monopoly in high-end lithography systems. Lithography – printing microchip patterns on silicon wafer using light.

It is estimated that about 60% cost of a data centre goes towards computer hardware (which comprises the above chip types). India with its increasing share of renewable energy coupled with the fact that it generates an estimated 20% of the world’s data, driven by its massive population, high smartphone penetration, and large-scale digital public infrastructure (such as UPI), has attracted both domestic and global investments in data centres. Google, Amazon, Microsoft and Air Trunk (Australia) have committed to invest over $ 90 billion by 2030 to set up data centres in India. Apart from these there are Indian companies, such as TCS, Reliance and Adani that are planning investments in data centres either on their own or in collaboration with foreign partners and the aggregate investment is likely to be $ 200 billion by 2030. At 60% of $200 billion, the opportunity for the semiconductor players is huge. India’s listed AI ecosystem features companies spanning enterprise IT services, hardware infrastructure, and embedded design. But as yet they are peripheral players on the global stage. Nevertheless, investors can find opportunities galore in the so called ‘picks and shovels’ stocks in relation to data centre build up, that provide services and materials required by the data centres. There are many of them and the share price of some of them like HFCL, Sterlite Technologies,  and MTAR Technologies have run up by anywhere from 200% to 500% in a year or so, even as Nifty/Sensex going nowhere in the last couple of years. The important ones are given below, grouped into six segments.

*Approximate ranges based on market prices during the past 12 months; exact returns depend on measurement date.

The names with the highest upside potential are probably Anant Raj, Techno Electric, E2E Networks and Netweb because their earnings can change dramatically if India’s AI and hyperscale buildout accelerates. The names with the lowest execution risk are Hitachi Energy India, ABB India, Siemens India, Cummins India and Polycab, because they benefit from the same trend without needing to successfully develop and lease large data-centre capacity. And if the if AI induced capex super-cycle persists, investors in the above companies are likely to reap handsome returns.

But caveat emptor. When a seemingly pathbreaking technology takes hold of the investors’ imagination, valuations are likely to go out of whack, as investors bake in euphoric future cash flows to arrive at current valuations. Many of the assumptions made in estimating such future cash flows and discount rates to arrive at net present value can come unstuck. Overoptimistic investors might then have to face the AI bubble burst a la the dotcom burst of 2000. Also, how the accelerating AI investments would shape the economies going forward is still uncertain. It is quite possible that AI technology might turn out to be nightmare for investors as the Citrini report predicts.

Secondly, although at the moment the pace of innovation and growth prospects in the AI space look convincing, there are market veterans with counter-views. Michael Burry of the Big Short fame and Peter Berezin of BCA Research are sceptical and talk of the bull-whip effect (See the Chart). It manifests when panic/FOMO driven spike in consumer demand (like for toilet paper during lockdowns) sets off a wave of placing large orders at each node of the supply chain, eventually leading to drastic ramp up in production by the manufacturer. By the time the ramped-up supply hits the shelves, the panic subsides, leading to a flooding of inventory and price correction, which at times can be sharp.

Thirdly, the reported move by the US to restrict access to advanced AI models such as Anthropic’s Mythos and Fable 5 for non-American citizens has once again exposed the unpleasant reality that countries that do not control foundational AI technologies will remain dependent on those that do. In such a situation the burgeoning Data Centre infra will be at best be a ‘tenant’ of global giants like Nvidia, Google, Microsoft, et al and India, despite its size will be subject to what could be termed as AI Colonisation. Finally, the hype about India becoming the global hub for data centres may have to be toned down. It is pertinent to note that the AI Infra, particularly data centres are energy guzzlers and require humongous quantities of ultra-pure water thereby imposing natural limits, sooner than later, in an already frayed environment. It is quite likely that euphoria may have to confront reality sooner than investors currently expect.