Wrapping up 2025: AI Chips

^{The content below is provided for information purposes only. It is not intended to amount to advice on which you should rely.}

^{Third Bridge does not make any warranties or representations, expressed or implied as to the currency, accuracy, completeness, timeliness or fitness for any particular use of the content below. Third Bridge accepts no liability for the content or for any errors, omissions or inaccuracies in respect of it.}

Google – TPU Update, GPU System Consumption Trends & CAPEX Outlook

Key insights 

• Expert anticipates Google will deploy 2-3 million TPUs in 2026 and 3-4 million in 2027, at lower cost per unit than Nvidia's premium pricing but, software ecosystem is Google's biggest challenge, with PyTorch/Cuda dominating while Google's XLA/Jax expertise stays largely internal and not widely open-sourced for external developers
• Google currently allocates 75-80% of TPU capacity internally, with only 10-20% available externally, but may deliberately increase external allocation to build ecosystem despite higher internal margins
• MediaTek partnership represents supply chain diversification rather than Broadcom replacement, with Google seeking to reduce Broadcom's "hefty tax" on system-in-package margins while maintaining the primary partnership
• Expert believes Google's TPU market share growth will come at AMD's expense rather than Nvidia's, as hyperscalers abandon internal ASIC programmes in favour of Google's proven seventh-generation technology
• Google may capture 50% of AI compute by dominating the model layer rather than competing directly with Nvidia's hardware ecosystem and developer tools

Already a client? Access the full transcript.

Broadcom – Custom ASIC Momentum, Hyperscaler Risks & AI Infrastructure Growth

Key insights 

• Node transitions from 5nm to 2-3nm increase costs, but unlock higher ASP premiums. Broadcom's margins may initially dip, but should recover to c60%
• 35-45% of Broadcom's custom AI revenue comes from Google and 25% from Meta. Broadcom plans to diversify its AI customer base to 3-4 major clients by 2027. Specialist expects quarterly AI semiconductor revenues could reach ≥USD 5bn, with 6-8 tape-outs per quarter and advanced packaging comprising 30-40% of AI revenues
• Broadcom maintains a one-year lead in 200-plus Gbps SerDes, winning c80% of high-speed switching customers with 10-20% pricing premiums vs competitors such as Marvell
• HBM4 technology could increase Broadcom's content and engagement by 50%, with potential for custom interface designs and 20-30% incremental content across hyperscalers
• Broadcom lacks in-house liquid cooling technology, which could be a significant barrier as hyperscalers adopt liquid cooling for high-performance systems

Already a client? Access the full transcript.

Merchant GPUs vs Custom ASICs – Token Factory TCO Analysis, ASIC vs GPU Market Share Outlook & Unit & Revenue Forecasts

Key insights 

• Nvidia's GB300 experiencing delays due to reliability issues from enhanced HBM3E 12-Hi load on same package size, breaching safe reliability envelope
• AMD's MI400 will outperform Nvidia's Rubin in H1 2026, featuring N2 vs N3 nodes and 432GB vs 288GB memory capacity for first performance reversal
• Microsoft diversified Maia 200 development between GUC for inference and Marvell for training products after initial reliability issues caused schedule delays to Q1 2026
• Training compute share is shifting mid-teen percentages from GPUs to ASICs in 2026 as multiple ASIC products achieve better technology nodes than Nvidia
• Revenue share between GPUs and ASICs in 2025 stands at 73% vs 27%, respectively, dramatically different from widely reported 90% GPU dominance figures

Already a client? Access the full transcript.

Not yet a Third Bridge client?  Request a free trial to experience the expert insights relied upon by the world's decision makers.