NVIDIA's Blackwell GPU architecture has now been deployed in more than 10,000 data center configurations worldwide, according to figures shared at the company's GTC conference, cementing the chipmaker's position at the center of the AI infrastructure build-out that is reshaping the technology industry.
The B100 and B200 chips at the heart of Blackwell represent a significant step up from the H100 generation that powered ChatGPT's explosive growth. NVIDIA claims Blackwell delivers up to 30 times the performance on transformer inference tasks compared to its predecessor, a figure that matters enormously to hyperscalers trying to serve millions of simultaneous AI queries at manageable cost.
Microsoft, Google, Amazon, and Meta have all committed to Blackwell deployments at scale. Microsoft's Azure AI infrastructure team confirmed it is integrating B200 clusters into new data center regions in Arizona and Sweden, part of a $13.75 billion capital expenditure program announced earlier this year. Google's TPU roadmap is developing in parallel, but analysts note that Blackwell's software ecosystem — particularly through CUDA and the NVIDIA AI Enterprise suite — remains a moat that is difficult to replicate quickly.
The supply chain picture has also stabilized. TSMC's CoWoS advanced packaging, which was a bottleneck for H100 production in 2023, has expanded capacity significantly. NVIDIA is now shipping in volumes that were not possible eighteen months ago, though lead times for large cluster orders still run three to six months for most enterprise customers.
Competition is intensifying. AMD's MI300X accelerator has found genuine traction at several cloud providers, with Microsoft confirming MI300X clusters for Azure, and Oracle deploying them in dedicated AI infrastructure. Intel's Gaudi 3 is targeting more cost-sensitive workloads. But analysts at Bernstein estimate NVIDIA still holds approximately 85% of the AI chip revenue market, a share that has proven more durable than initially expected given the competitive announcements of the past two years.
The downstream effect on enterprise AI is tangible. As Blackwell supply increases and pricing pressure from competition mounts, inference costs for large language models have dropped by roughly 80% since mid-2023. This cost reduction is enabling new use cases that were previously economically unviable — real-time video analysis, per-user personalization at internet scale, and AI-native applications in healthcare and manufacturing.
For the semiconductor industry broadly, the AI cycle has reinflated capital spending in a way not seen since the mobile boom of the early 2010s. TSMC's revenue grew 34% year-over-year in Q1 2026, driven almost entirely by AI chip demand. Samsung and SK Hynix are both running HBM3E memory production near capacity.
What remains uncertain is the demand ceiling. Models are getting more efficient — GPT-4o processes tokens at a fraction of the cost of GPT-4 — and if that efficiency curve continues, raw chip demand could plateau before supply does. For now, though, the Blackwell deployment wave shows no signs of slowing.