The AI
Ecosystem
Supply Chain
Stack

NVIDIA designs GPUs (Graphics Processing Units) that have become the dominant hardware for training and running AI models. Originally built for gaming graphics, NVIDIA's chips turned out to be well suited for the parallel processing AI demands. Today, virtually every major AI model — including ChatGPT, Gemini, Grok and Claude — was trained on NVIDIA hardware.

Blackwell Is the Engine. Blackwell and Blackwell Ultra are now in volume production and deployed across all major cloud providers. Demand continues to outstrip supply. Networking inside the data centre reached $11 billion in Q4 FY2026 alone, more than 3.5x the prior year.

Rubin Is Already Lined Up. NVIDIA has unveiled the Rubin platform, comprising six new chips that aim to deliver up to a 10x reduction in inference token cost compared with Blackwell. AWS, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure are among the first announced deployers.

Mega-Deals and Sovereign AI. NVIDIA announced a strategic partnership with OpenAI to deploy at least 10 gigawatts of NVIDIA systems. Anthropic announced it will run on NVIDIA infrastructure, initially adopting around 1 gigawatt of compute capacity. The first Blackwell wafer was produced on US soil at TSMC's Arizona facility. The US Department of Energy is building Solstice with Oracle and 100,000 NVIDIA Blackwell GPUs.

Broadcom designs custom AI chips (ASICs) and networking semiconductors. While NVIDIA sells general-purpose GPUs, Broadcom builds bespoke chips for individual customers — chips optimised for exactly what Google, Meta or other hyperscalers need. Their networking chips are also the backbone that connects servers inside virtually every major data centre on Earth.

AMD designs CPUs and GPUs and is NVIDIA's primary like-for-like competitor in AI accelerators. Its Instinct MI300X and MI350 series chips are winning meaningful contracts at Microsoft, Meta and Oracle as customers seek alternatives to a single-source supply. AMD also dominates roughly half of the global data-centre CPU market.

AMD's CEO Lisa Su is widely regarded as one of the strongest executives in semiconductors. She turned AMD around from near-bankruptcy in 2014 into a credible challenger across CPUs, GPUs and AI accelerators — a track record that has earned the company significant credibility with hyperscale buyers who want a second source.

ARM does not make chips. ARM does not even design complete chips end-to-end. ARM designs the architecture — the fundamental instruction set and core designs that other companies license to build their own chips. It is one of the most distinctive business models in technology.

ARM charges a licensing fee when a company uses its architecture to design a chip, then takes a royalty on every chip sold. That means ARM earns money every time Apple sells an iPhone, every time a server ships, every time any ARM-based device is manufactured — without building anything physical itself.

TSMC is the world's most important manufacturer of advanced semiconductor chips. They do not design chips. They build chips designed by others. Apple brings them iPhone designs. NVIDIA brings them GPU designs. AMD brings them CPU and accelerator designs. TSMC builds all of them — better than anyone else on Earth at the leading edge.

Every NVIDIA AI GPU is manufactured by TSMC. Apple's entire chip supply, from iPhones to MacBooks, is made by TSMC. AMD's MI-series accelerators, Broadcom's custom ASICs, and many of Google's, Amazon's and Meta's in-house silicon designs are all manufactured by TSMC.

Taiwan is claimed by China, and TSMC sits at the centre of the resulting geopolitical tension. Taiwan calls TSMC its "Silicon Shield": the theory being that global dependence on TSMC makes any conflict prohibitively costly for everyone involved. If Taiwan-based capacity were disrupted, the global tech economy would effectively halt within months.

SK Hynix is the world's leading producer of HBM — the specialised memory chips that sit directly alongside NVIDIA's AI GPUs and allow them to process data fast enough to run modern AI models. SK Hynix has been the primary memory supplier for every major NVIDIA AI generation: A100, H100, H200 and the Blackwell family.

Micron is the only US-headquartered manufacturer of DRAM, NAND flash and HBM at scale. As the AI memory market expands, Micron sits in a strategically important position: it has US government support via the CHIPS Act, growing HBM market share, and a domestic-champion role that makes it politically important well beyond its size.

Samsung is simultaneously the world's largest consumer electronics company and one of the three dominant memory chip manufacturers. It produces DRAM, NAND and HBM, while also operating its own foundry business that competes with TSMC for advanced chip manufacturing.

Samsung has demonstrated up to 11 Gbps signalling for HBM4, ahead of SK Hynix's 10 Gbps reference, and is in HBM4 talks with NVIDIA, Broadcom and Google. Volume HBM4 supply is expected to begin in the first half of 2026.

Equinix is the world's largest data centre company, structured as a REIT. It owns and operates the physical buildings where the internet, cloud computing and AI infrastructure actually live. Microsoft, Google, Amazon, Meta and Apple all pay Equinix to house servers and to interconnect with each other inside its facilities. In effect, Equinix is the landlord of the internet.

AI workloads consume far more power per server rack than traditional cloud workloads — often 10x or more. That requires entirely different cooling, power and design specifications. AI demand is, in effect, creating a data centre capacity crunch.

Microsoft is best known for software, but it is also one of the largest builders and operators of AI data centres in the world. Its Azure cloud is the primary commercial deployment platform for OpenAI's products. Microsoft's data centre and energy commitments are arguably more important to its AI economics than any single chip purchase.

Constellation Energy is America's largest nuclear energy producer, operating a fleet of around 21 reactors with approximately 22 gigawatts of total nuclear capacity. Nuclear power is emerging as the preferred energy source for AI data centres because it provides 24/7, carbon-free, uninterruptible electricity — exactly what AI workloads demand.

Constellation is also pursuing the $26.6 billion acquisition of Calpine and signing additional long-term clean power deals with data-centre operators. Together with the Three Mile Island restart, this positions the company as the largest single source of contracted, carbon-free baseload power for the AI buildout in the United States.