Project
NVIDIA's program licensing NVLink interconnect so partners' custom CPUs and XPUs can connect directly to Nvidia GPUs in scale-up AI systems.
d-Matrix's ISCA paper models up to 100 TB/s per card for Raptor, a stacked-DRAM inference accelerator. The figures come from modeling and the first cards are expected in Nvidia MGX racks in late 2027, so they have no bearing on inference hardware bought in the next year.
Reality
NVIDIA is licensing the Fusion chiplet, its C2C CPU attach and NVHBM to MediaTek and buying $3.5 billion of its convertible bonds, so the quickest route to a custom accelerator now ends inside an NVIDIA rack.
Perspective Coverage
3 publishersReality
Nvidia's money buys bonds that can convert into MediaTek equity. The return that matters is NVLink Fusion, which routes every custom accelerator MediaTek designs for a hyperscaler back into an Nvidia rack. The bond terms were not disclosed.
Perspective Coverage
6 publishersReality
The saving comes off cache reads, so it lands unevenly across teams, and the model's most sensitive capability now reaches users through a credential issued by an access program, not a purchase order a budget holder can sign off alone. Both changes move the decision away from the budget holder.
Perspective Coverage
3 publishersReality
The Information reports two configurations, one with two chips and one with four, aimed at developers, businesses and governments, with a possible 2029 launch. Apple already builds servers for Private Cloud Compute and has turned partners away from them.
Perspective Coverage
7 publishersReality
Qualcomm's Tony Pialis told the AI Infra Summit in Santa Clara that tokens per watt is the new key metric in the AI war. The vendors in the room disagreed about which memory architecture gets them there.
Reality
The Information reports Apple is weighing Nvidia's NVLink Fusion for an inference server built on two or four M8 Ultra chips in 2029. The engineering question is which part of that package would join the NVLink domain.
Perspective Coverage
3 publishersReality
Apple has not confirmed the plan and the M8 Ultra is years from release. Seven other companies have already licensed the same rack interconnect from Nvidia, and their dates are the ones a buyer can use.
Reality
The one measured number behind NVIDIA's tokens-per-megawatt pitch comes from Lambda's Blackwell cluster, where power reclaimed from static provisioning ran three extra nodes. Amazon's Annapurna Labs and d-Matrix got a line each.
Reality
NVHBM offers NVLink Fusion partners 30 percent more bandwidth and 15 percent less power, in exchange for NVIDIA owning the interface between their compute and their memory.
Perspective Coverage
3 publishersReality
Contract prices roughly tripled in two quarters. Marvell is selling recycled DDR4 behind CXL, but its reference customer, Meta, did the job on an ASIC of its own design.
Reality