jay goldberg

End demand for Nvidia chips is becoming harder to finance as AMD closes the hardware/software gap (fully competitive by 2027), hyperscalers deploy custom silicon like Google TPU which may be superior for large model classes, and Nvidia's $134B in prepaid cloud compute services represents backstop guarantees to neoclouds that effectively discount pricing despite claimed 100% utilization.

OpenAI represents a disproportionate share of Microsoft and Oracle's AI data center capacity, yet remains a cash-burning startup with uncertain trillion-dollar funding plans, making Nvidia's long-term revenue outlook heavily dependent on an unreliable counterparty.

End demand is weakening as it gets harder to finance chip purchases; AMD is closing the hardware and software gap with competitive silicon and ROCm by 2027; hyperscalers are designing custom silicon (Google TPU arguably superior for large workload classes); $134B prepaid cloud services suggest backstop discounts to neoclouds; demand concentrated in few players like OpenAI with uncertain funding; macro/China trade risks and economy's dependence on Nvidia earnings create fragility.

AMD's roadmap shows competitive silicon and rack-level solutions; ROCm software stack has closed the gap considerably from years behind to near parity; by 2027 AMD will be fully competitive with Nvidia across the stack.

Google's custom TPU chips present a credible alternative to Nvidia GPUs for significant categories of AI training and inference workloads, representing a structural competitive threat as hyperscalers vertically integrate silicon.

End demand for Nvidia chips is getting harder to generate and finance; $134B in neocloud bookings come with demand guarantees that function as discounts, and OpenAI's uncertain trillion-dollar spending plan bakes unreliable long-term demand into Nvidia's outlook.