NVIDIA (2022-2024): the AI-chip rocket from $400 billion to $3 trillion in 18 months

Where NVIDIA was going in

Before late 2022, NVIDIA was understood primarily as a graphics-processing-unit company with a growing data-center accelerator business. The CUDA programming model (introduced in 2006-2007) had built up over 15 years a near-monopoly developer-software stack for general-purpose GPU computation. Through the 2010s NVIDIA had become the de facto default GPU for deep-learning research and training because the CUDA ecosystem worked and AMD/Intel alternatives did not have comparable software.

When ChatGPT launched on November 30, 2022 and demonstrated that large language models had reached commercial-usability quality, the implication for compute demand was immediate: training and inference of LLMs at scale required massive concentrations of GPU compute, and the GPUs that the hyperscalers could deploy at scale were predominantly NVIDIA. NVIDIA was the supply-side bottleneck of the AI build-out before most market participants had priced that.

The revenue acceleration

NVIDIA’s fiscal 2023 (ended January 29, 2023) had data-center revenue of approximately $15.0 billion — already large, growing, and structurally important, but not category-defining. Then fiscal 2024 (ended January 28, 2024): data-center revenue tripled to $47.5 billion, total revenue grew 126% to $60.9 billion, and the data-center segment alone became larger than the entire previous NVIDIA business. Quarter-by-quarter the acceleration was visible: Q1 FY2024 data center $4.3B, Q2 $10.3B, Q3 $14.5B, Q4 $18.4B. Year-over-year growth rates on the data center segment peaked at +427% in Q1 FY2025.

The product line that drove the revenue was the H100 (announced March 2022, shipping 2023) and its successors. The H100 became the standard chip for training and serving large language models. Sales of H100 and A100 exceeded half a million units in Q3 2023 alone. Hyperscaler buying (Meta, Microsoft, Google, Amazon, plus emerging neo-cloud providers like CoreWeave and Lambda) became the primary demand driver. NVIDIA’s gross margins expanded substantially as the data-center mix shifted toward the highest-end accelerators, with non-GAAP gross margin reaching above 75% in some quarters.

Why competitors did not catch up

AMD, Intel, and a number of AI-chip startups (Cerebras, Graphcore, SambaNova, Groq) all offered alternative hardware in the 2022-2024 period. None of them captured material share. The structural reasons were threefold. First, the CUDA software stack had 15+ years of compounding investment from NVIDIA and from the developer ecosystem; rewriting training and inference workloads to run on alternative hardware required substantial engineering investment that hyperscalers were unwilling to fund when NVIDIA hardware was available. Second, NVIDIA’s networking acquisition (Mellanox, completed 2020) and NVLink interconnect created GPU-to-GPU communication advantages at the cluster level that competitors could not match. Third, the supply contracts with TSMC, the memory partners (SK Hynix, Samsung), and the systems partners (Dell, HPE, Supermicro) had been built around NVIDIA timelines and product specs for years.

The combination of CUDA, networking, and supply-chain integration produced what economists call a winner-take-most market structure. The 98% data-center GPU revenue share by 2023 was the visible outcome of the structural moats. As of mid-2024 the main competitive threat came not from alternative-vendor GPUs but from hyperscaler-custom-silicon (Google TPU, Amazon Trainium, Microsoft Maia), where the hyperscaler has the engineering capability to do the CUDA-equivalent work for its own workloads.

How RGM thinks about platform-capture moments

When clients ask about how to recognize and respond to platform-capture moments, the NVIDIA 2022-2024 sequence is the clearest current example. Three structural lessons. First, platform-capture rarely shows up suddenly — the CUDA ecosystem had been building for 15 years before the ChatGPT moment crystallized the value. The right time to invest in platform-software-and-developer-ecosystem advantages is decades before the demand catalyst arrives. Second, the demand catalyst (ChatGPT and the LLM compute build-out) was not predictable from any single quarter’s data; it was an exogenous event that re-priced an already-strong position. Companies in adjacent semiconductor or platform-software positions need to be ready for catalysts they cannot forecast but can prepare for. Third, the speed and magnitude of value capture (market cap from $400B to $3T in 18 months) reflects how concentrated the gains can be when a platform-capture moment arrives. Markets in this kind of dynamic do not give second prizes.

For most clients, the practical takeaway is that platform-capture moments are not generic strategy advice — the necessary preconditions (developer ecosystem, supply-chain integration, complementary asset moats) take a decade to build. But for clients positioned in long-cycle semiconductor, infrastructure-software, and developer-platform categories, the question of which adjacent positions could become platform-capture-capable over the next decade is worth direct strategic attention. The companies that benefit most from the next platform-capture moment are the ones building the preconditions now.

Frequently asked questions

Related