Nvidia, AMD, Google: The AI Chip Race by the Numbers

Nvidia held 80% of the AI chip market in 2023. Its H100 GPU sold for over $40,000 per unit on secondary markets. By 2025, the entire AI chip industry was projected to surpass $200 billion in revenue. Those numbers come from a detailed PatentPC analysis published June 1, 2026, which tracks the decade-long arc of the AI hardware race through 16 data points. The picture is stark: one company dominates, two others are making real gains, and the entire market is growing faster than any hardware segment in decades.

But the raw stats tell only half the story. The other half is about lock-in, strategy, and the quiet battle over developer ecosystems.

Nvidia’s moat is not just silicon. It is CUDA.

The company’s software ecosystem locked in over 3.5 million developers as of 2023, according to the PatentPC report. That is a barrier to entry that no competitor can leap in a single product cycle. Even as AMD’s MI300 series and Google’s TPU v5 deliver competitive raw performance, any developer who wants to switch must retrain, recompile, and reoptimize. Nvidia knows this. The company’s Hopper architecture delivered 6x faster AI training speeds than its predecessor, and the upcoming Blackwell chip, expected in 2025, promises exaflop-level performance. Each generation widens the gap in raw compute and deepens the dependency on CUDA.

Nvidia’s DGX systems accounted for 30% of AI data center deployments in 2023. That is not just a market share figure. It means nearly a third of all new AI infrastructure is a turnkey Nvidia product, not a collection of parts. The company sells the rack, the networking, the software, and the support. Competitors sell chips that must be integrated into someone else’s system.

AMD is the real challenger, and it is gaining.

AMD’s AI accelerator market share hit 15% in 2024, up from 5% in 2022. That is a tripling in two years. The company’s AI chip segment grew 50% year-over-year in 2023. Its Instinct MI250 powered Frontier, the world’s fastest supercomputer, in 2023. By 2024, AMD’s chips were deployed in more than 100 supercomputers worldwide.

The MI300 series, launched in 2024, was AMD’s most aggressive attempt yet to challenge Nvidia’s H100. The PatentPC report describes it as a “strategic push to capture market share in AI data centers, enterprise computing, and high-performance AI workloads.” AMD’s pitch is straightforward: comparable performance, lower cost, and an open software ecosystem. For enterprises that want an alternative to Nvidia’s pricing and supply constraints, AMD is the only credible option at scale.

But 15% is still a long way from 80%. AMD’s growth is real, but it is growth from a small base. The company needs to sustain that trajectory for several more years to become a true counterweight.

Google plays a different game entirely.

Google’s Tensor Processing Units are not for sale. They are for Google. The company uses them internally to run its own AI workloads. In 2023, TPUs powered more than half of Google’s AI training workloads and 90% of Google Search AI models. That is a massive internal deployment that insulates Google from Nvidia’s pricing and supply constraints.

The TPU v5, announced in 2024, promised 2x performance improvements over TPU v4. Google also offers TPU access through Google Cloud, making it available to external customers. The calculus for those customers is different from buying Nvidia or AMD hardware. TPUs are custom-built for machine learning, not general-purpose GPUs adapted for AI. For specific workloads, especially large-scale training on Google Cloud, TPUs can offer better performance per dollar.

But Google’s TPU strategy is fundamentally defensive. It protects Google’s margins on its own AI services. It does not threaten Nvidia’s market share in the broader AI chip market. Google is not trying to sell TPUs to every data center. It is trying to make its cloud the best place to train AI models.

The energy efficiency angle is underappreciated.

Between 2020 and 2025, AI chip energy efficiency improved by 40% year over year, according to the PatentPC report. That compounds to a 4.8x improvement over five years. This is the silent driver of AI’s scaling. Without these efficiency gains, the cost of training and running large models would have become prohibitive. Data center power constraints would have capped model sizes. The fact that efficiency is improving this fast means the ceiling on AI compute is still rising.

This is where the competition gets interesting. Nvidia’s H100 and Blackwell chips are power-hungry. AMD’s MI300 series emphasizes energy efficiency. Google’s TPUs are custom-designed for specific workloads and can be optimized for power consumption. The company that wins the efficiency race will win the next generation of deployments, especially as AI moves from training to inference at massive scale.

The shortage of 2023 was a stress test the industry barely passed.

The AI chip shortage caused price surges of 300% in 2023. The H100 sold for over $40,000 per unit on secondary markets. Companies that had bet their AI strategies on Nvidia hardware found themselves unable to get chips at any price. This was not a supply chain hiccup. It was a structural bottleneck that revealed the fragility of the entire AI ecosystem.

The shortage accelerated two trends. First, it pushed companies to consider AMD and Google Cloud TPUs as alternatives. Second, it drove enterprises to secure long-term supply agreements and invest in their own AI infrastructure planning. The market learned a hard lesson about single-vendor dependency.

What this means for AI builders.

The numbers tell a clear story. Nvidia is dominant but not invulnerable. AMD is the only credible challenger at the chip level. Google is building a parallel universe for its own workloads and cloud customers. The market is growing fast enough that all three can win, but the shape of that growth matters.

For AI startups and enterprises building on Nvidia hardware, the calculus is straightforward: CUDA lock-in is real, but supply constraints and pricing are risks. Diversifying across AMD and Google Cloud TPUs is prudent, even if it means some upfront engineering cost. For hardware investors, AMD’s trajectory from 5% to 15% market share in two years is the most interesting signal in the data. For policymakers, the concentration of AI compute in one company is a strategic vulnerability that export controls and domestic chip manufacturing investments are only beginning to address.

The AI chip market in 2030 will not look like it does today. But the foundations being laid now in software ecosystems, supply chains, and energy efficiency will determine who builds on top.