- Nvidia $NVDA still rules, for now: Nvidia $NVDA controls about 81% of the AI data center chip market and recently posted a record $75.2 billion data center quarter. Buyers want cheaper options. - OpenAI builds its first chip: OpenAI unveiled Jalapeño on June 24, their first inference chip made with Broadcom $AVGO. Their own AI models helped design it. Early tests show better efficiency. - Google $GOOGL leads with Ironwood: Google $GOOGL started building TPUs in 2015. Their 7th generation Ironwood chip hits 4.6 petaflops each, and Anthropic plans to use up to 1 million of them. - Amazon $AMZN scales Trainium: Amazon $AMZN launched Trainium3 in December 2025, claiming 50% lower costs. Anthropic committed over $100 billion across 10 years to run Claude on these chips. - Microsoft $MSFT answers with Maia: Microsoft $MSFT revealed Maia 200 in January 2026 built on 3nm with 140 billion transistors. The company claims 30% better performance per dollar than their current hardware. - Meta $META goes all in: Meta $META is scaling their MTIA inference chip with Broadcom $AVGO. The company expects to deploy over 1 million of their own chips in 2026, cutting Nvidia $NVDA spend. - Apple $AAPL wrote the playbook: Apple $AAPL proved the model when they dropped Intel $INTC for their own silicon. Now reports say they are working with Broadcom $AVGO on a data center AI chip. - Spacex $SPCX takes it furthest: Spacex $SPCX and Tesla $TSLA are building Terafab, a $55 billion chip plant in Texas. Their orbital A|1 satellite spans 70 meters (230 feet) and runs swappable chips. The single supplier era ends. Custom chip shipments grow about 45% each year, nearly triple merchant GPUs. Nvidia $NVDA still rules training, but inference is now the battleground everyone wants.
