The Drum · Weekly AI news, freeSubscribe →
Timeline
The race for AI chips represents one of the most consequential hardware competitions in computing history. What began as a niche market for specialized processors has evolved into a global struggle for technological dominance, with companies like NVIDIA, AMD, Intel, and emerging players from China and other nations competing fiercely for market share. The competition intensified dramatically after 2017, when transformer-based models demonstrated that specialized silicon could dramatically accelerate AI workloads. Today, the stakes extend beyond profit margins: control over AI chip design and manufacturing capacity shapes geopolitical influence, national security strategies, and the pace of AI development itself. The competition remains far from settled, with new architectures, manufacturing breakthroughs, and strategic partnerships constantly reshaping the landscape.
The Story
For decades, GPUs designed for gaming graphics processing proved surprisingly effective for machine learning workloads. NVIDIA capitalized on this accident of history, building a near-monopoly in AI accelerators by the early 2020s. The narrative shifted when cloud providers, chip startups, and international competitors recognized the strategic importance of custom silicon. Google developed TPUs, Amazon invested in Trainium and Inferentia, and Chinese firms pursued indigenous alternatives to circumvent export restrictions. The unresolved tension remains: can anyone dethrone NVIDIA's entrenched position, or will the company's software ecosystem and manufacturing partnerships prove insurmountable? Simultaneously, questions linger about whether the current trajectory of ever-larger chips and models is sustainable, or whether efficiency-focused alternatives will eventually prevail.
1997
NVIDIA Founded, Graphics Processing Focus
NVIDIA was established to design graphics processing units for consumer gaming and professional visualization. The company's early focus on parallel processing architecture would later prove foundational for machine learning acceleration.
2006
CUDA Platform Launched by NVIDIA
NVIDIA introduced CUDA, a parallel computing platform that allowed developers to harness GPU power for general-purpose computing beyond graphics. This framework became critical infrastructure for scientific computing and would later enable GPU-accelerated machine learning.
2012
Deep Learning Breakthrough Uses GPUs
Geoffrey Hinton's team won the ImageNet competition using deep convolutional neural networks trained on NVIDIA GPUs. This watershed moment demonstrated that GPUs could dramatically accelerate neural network training, sparking widespread adoption in AI research.
2013
Google Begins TPU Development
Google started designing Tensor Processing Units, custom silicon optimized specifically for machine learning inference and training. This marked the beginning of major cloud providers pursuing proprietary chip strategies.
2015
NVIDIA Launches Maxwell GPU Architecture
Maxwell represented a significant efficiency improvement over previous generations, offering better performance-per-watt for AI workloads. The architecture solidified NVIDIA's position as the dominant player in AI acceleration.
2016
Google TPU v1 Deployed in Production
Google's first-generation TPU entered production use for inference tasks in Google's data centers. The chip demonstrated that custom silicon could outperform general-purpose GPUs for specific AI workloads.
2017
Transformer Architecture Published
The attention mechanism paper introduced transformers, which became the foundation for modern large language models. This architectural shift increased demand for massive parallel processing power, accelerating the AI chip race.
2017
NVIDIA Launches Volta GPU Architecture
Volta introduced tensor cores specifically designed for matrix operations central to deep learning. The architecture represented NVIDIA's explicit pivot toward AI as a primary use case.
2018
Google TPU v2 and v3 Announced
Google released second and third-generation TPUs with significantly increased performance and memory capacity. These chips powered Google's internal AI services and became available through Google Cloud.
2018
NVIDIA Becomes AI Chip Market Leader
NVIDIA's data center revenue surged as AI adoption accelerated across industries. The company captured an estimated 80-90 percent market share in AI accelerators.
2019
AMD Launches MI100 GPU for AI
AMD introduced its MI100 accelerator to compete with NVIDIA's offerings in the data center AI market. The chip offered competitive performance but struggled to gain market traction due to software ecosystem disadvantages.
2019
Amazon Announces Trainium and Inferentia
Amazon Web Services unveiled custom chips designed for training and inference workloads. AWS positioned these chips as cost-effective alternatives to NVIDIA GPUs for customers using AWS infrastructure.
2020
NVIDIA Acquires Arm Holdings Bid Announced
NVIDIA announced plans to acquire Arm Holdings for 40 billion dollars, seeking to expand its chip design capabilities. The deal faced regulatory scrutiny and ultimately failed, but signaled NVIDIA's ambitions for vertical integration.
2020
NVIDIA Launches A100 GPU
The A100 became NVIDIA's flagship data center GPU, offering unprecedented performance for large-scale AI training. Demand for A100s far exceeded supply, creating a bottleneck for AI development globally.
2020
China Faces US Export Restrictions on AI Chips
The US government imposed export controls on advanced semiconductors to China, citing national security concerns. This policy accelerated China's efforts to develop indigenous AI chip capabilities.
2021
SambaNova and Cerebras Secure Major Funding
AI chip startups SambaNova and Cerebras raised hundreds of millions in venture funding, betting on novel architectures to challenge NVIDIA's dominance. Both companies pursued dataflow-based approaches distinct from traditional GPU designs.
2021
Apple Announces M1 Pro with Neural Engine
Apple integrated specialized neural processing capabilities into its M1 Pro chip for consumer devices. This move demonstrated that AI acceleration was becoming a standard feature across computing platforms.
2022
NVIDIA Launches H100 GPU
The H100 represented a major generational leap, offering 3x the performance of A100 for transformer training. Demand immediately exceeded supply, with customers waiting months for delivery.
2022
Intel Launches Gaudi Accelerator
Intel entered the AI accelerator market with Gaudi, targeting training and inference workloads. The chip struggled to gain adoption against entrenched competitors.
2022
US Tightens Export Controls on Advanced Chips to China
The Biden administration expanded restrictions on semiconductor exports, blocking sales of advanced chips and chip-making equipment to China. The policy aimed to slow China's AI development but raised concerns about global chip supply chains.
2023
ChatGPT Launches, Driving GPU Demand Surge
OpenAI's ChatGPT became a cultural phenomenon, demonstrating large language models' capabilities to mainstream audiences. The explosion in demand for LLM inference and training created unprecedented pressure on GPU supply.
2023
Meta Announces Custom AI Chip Strategy
Meta revealed plans to develop custom silicon for AI workloads, reducing dependence on NVIDIA. The company aimed to optimize chips for its specific inference and training needs.
2023
NVIDIA Launches H200 GPU
The H200 offered increased memory bandwidth and capacity compared to H100, addressing bottlenecks in large model training. Supply constraints persisted despite production increases.
2023
Graphcore Faces Restructuring Amid Competition
Graphcore, a prominent AI chip startup, announced significant layoffs and strategic pivots after struggling to compete against NVIDIA. The company's challenges highlighted the difficulty of breaking into the market.
2023
China Launches Huawei Ascend 910B Chip
Huawei unveiled its Ascend 910B processor, designed to circumvent US export restrictions and provide indigenous AI acceleration. The chip represented China's progress toward semiconductor self-sufficiency.
2024
NVIDIA Launches Blackwell GPU Architecture
Blackwell represented the next major generational leap, offering significant efficiency improvements and increased memory capacity. The architecture was designed for both training and inference at massive scale.
2024
AMD Launches MI300 Series Accelerators
AMD released MI300 and MI300X chips with competitive performance against NVIDIA's latest offerings. The chips featured increased memory and improved software support through ROCm ecosystem improvements.
2024
Google Announces TPU v5e and v6e
Google released new TPU generations with improved efficiency and cost-effectiveness for inference workloads. The chips reinforced Google's strategy of vertical integration for AI infrastructure.
2024
US Semiconductor Manufacturing Incentives Expand
The CHIPS Act funding accelerated domestic semiconductor manufacturing capacity in the United States. Investments targeted both advanced logic production and specialized AI chip fabrication.
2024
Startup Cerebras Pivots to Inference Focus
Cerebras shifted strategy from training-focused systems to inference optimization, acknowledging market realities. The pivot reflected broader industry recognition that inference represents a larger long-term market.
2024
NVIDIA Maintains 80+ Percent Market Share
Despite increased competition, NVIDIA retained dominant market position in AI accelerators. The company's software ecosystem and manufacturing partnerships proved difficult to displace.
© 2026 Cedar & Bloom. All rights reserved.
