Hyperscalers spend on custom silicon
- Broadcom agreed to support Meta's next-generation AI chip development under a multiyear deal, reflecting heavy hyperscaler silicon bets. - Meta has paid roughly $2.3 billion to Broadcom recently and plans $115bn–$135bn in AI infrastructure spend this year, per coverage. - Custom chips, Google’s new silicon announcements, and increased IT spending forecasts highlight infrastructure and power constraints shaping AI deployment ( )
Broadcom and Meta expanded their AI chip partnership on April 15, locking in a multiyear effort to build Meta’s next generation of custom silicon. (investors.broadcom.com) Broadcom said the work will support Meta Training and Inference Accelerator chips through 2029, starting with an initial commitment above 1 gigawatt and scaling into a multi-gigawatt rollout for Meta data centers. (investors.broadcom.com) Meta said the chips are part of a “portfolio approach” to AI hardware, with custom accelerators handling some workloads inside Facebook, Instagram, WhatsApp, and other services instead of relying only on Nvidia graphics processors. (about.fb.com) Custom silicon means a cloud company designs chips for its own jobs, the way a warehouse builds a conveyor belt for one package flow instead of buying a general-purpose machine. Meta said its MTIA chips are meant to improve performance and lower total cost for specific AI tasks. (about.fb.com) The spending behind that strategy is already visible. Meta disclosed it paid Broadcom $2.3 billion in 2025 for chip design, engineering, research and development, and related component purchases. (theinformation.com) Meta’s broader infrastructure budget is even larger: the company told investors in January that 2026 capital expenditures would land between $115 billion and $135 billion, up from $72.22 billion in 2025. (datacenterdynamics.com) Google is making the same turn toward in-house hardware. At Cloud Next on April 22, it introduced separate eighth-generation tensor processing units for training AI models and for inference, the step where models answer prompts. (finance.yahoo.com) The money is spreading beyond a few companies. Gartner said on April 22 that worldwide information technology spending will reach $6.31 trillion in 2026, up 13.5% from 2025, with AI infrastructure and advanced memory driving the increase. (financialcontent.com) The bottleneck is not only chips. The Register reported on April 23 that AI demand is pulling memory supply toward servers and straining power and management-chip capacity needed to build and run data centers. (theregister.com) That leaves hyperscalers chasing the same outcome from different angles: design more of their own chips, reserve more electricity, and build more data center capacity before the next wave of AI models arrives. (investors.broadcom.com, finance.yahoo.com, theregister.com)
