CASPA 2026: Patel and Rasgon on AI buildout financing, substrates, and pricing power
At the CASPA 35th Annual Conference & Dinner Banquet — The CEO Summit (Santa Clara, theme: Silicon, Systems, and Intelligence), Austin Lyons (Chipstrat) moderated a Semiconductor Industry Outlook panel with Dylan Patel (CEO, SemiAnalysis) and Stacy Rasgon (Managing Director, Bernstein). The conversation looked ahead to 2027–2028 across financing, supply-chain bottlenecks, pricing power, GPUs vs XPUs, and enterprise AI. CASPA posted the full video; Chipstrat published a lightly edited transcript.
On financing, Rasgon framed cost of capital as one of several constraints moderating the buildout — alongside power, supply, and tools — and said hyperscalers "were the most cash generative businesses in the history of humankind" but "are burning cash now and probably going to be raising equity," pointing to Meta's new share-class structure and CoreWeave unsecured debt trading near 15% yields in the market. He argued constraints can also be constructive, producing "stronger for longer" dynamics rather than a tech-bubble-style overbuild ("there's dark fiber in the ground today that was overbuilt and never used").
Patel said neoclouds without high-grade offtake increasingly cannot get financed, with Anthropic/OpenAI-backed debt priced around 13–15% (citing Volta's Anthropic cluster: ~12% coupon sold at 95 cents ≈ 15% all-in). He described Nvidia revenue-share structures with Sharon AI, Fermi, and Lambda (10% on an Anthropic cluster) and Broadcom rent-to-own chip deals as ways to backstop capacity. He put Meta's on- and off-balance-sheet compute and data-center commitments at about $1.4 trillion, and called financing part of a "holy trinity" with land/power/shell and chips.
On bottlenecks into 2027–2028, Patel ranked substrates (and related PCB/materials, especially Japan supply chains) as "the biggest bottleneck of all the bottlenecks," ahead of memory and co-packaged / near-package optics; he argued power is less binding than markets assume because nearly 50 GW of data centers are being built globally next year (~35 GW in the US), with on-site turbines, engines, and fuel cells easing the crunch. Rasgon expects tightness through 2028 (2029 as a demand question), flagged memory as 70–80% of silicon area in an AI rack, and noted WFE and clean-room capacity as further constraints.
On pricing, Rasgon said the industry is on track for roughly $1.5 trillion this year (up ~100%), with 70–80% of growth from memory ASPs, after historically destroying value for customers until post-COVID pricing power stuck. Patel contrasted TSMC's steady gross-margin climb (~45% to ~60% over a decade) with memory's boom-bust, and said the all-in cost to build a gigawatt of compute rose from roughly $40 billion in 2024 to north of $60 billion. Later in the panel he argued Nvidia is losing share of CoWoS and HBM even as earnings can still double, and that AI is making custom XPU design easier for hyperscalers and radical startups.
The panel is a clear sell-side / research read on whether AI infrastructure can keep scaling through 2028: financing and materials (not only power) may cap neocloud and non-hyperscaler build, while substrate, memory, and optics bottlenecks keep pricing power with suppliers. For Dealroom readers tracking SemiAnalysis, Bernstein coverage, Nvidia, Meta, CoreWeave, and TSMC, it is a primary source on how two of the most-watched semiconductor voices are framing the next two years.
CASPA YouTube: Semiconductor Industry Outlook Panel | CASPA 2026 CEO Summit (published 24 Oct 2026).
Chipstrat transcript: My Conversation with Dylan Patel and Stacy Rasgon (Austin Lyons, 24 Oct 2026).
Event: CASPA CEO Summit agenda.
