Nvidia's rack-scale AI server platform combining 72 Blackwell Ultra GPUs with NVLink interconnects for large-scale training and inference clusters.
Bitdeer signed a 10-year agreement for 67 MW of Malaysian data center capacity that starts arriving in Q1 2027, and no customer has signed for it. For a buyer weighing GB300 capacity there, the site is an option to negotiate for, and 2026 workloads will have to run somewhere else.
Reality
ChronoScale says its signed AI compute contracts now annualize to $1 billion, against $71.6 million of revenue in its last fiscal year. The figure counts signed agreements, so the stock depends on capacity still being built turning them into cash.
Reality
Nvidia says DSX MaxLPS power sharing fits up to 40% more GPUs in the same approved power budget, citing an Nscale trial of 192 GPUs against 140. The gain holds only while GPUs rarely peak together, and each operator has to measure that on its own fleet before buying hardware against it.
Reality
PyTorch says vLLM's frontier models now ship as hardware-specific flat definitions that torch.compile cannot trace, and the new HW agnostic layers are what users on other accelerators get instead. The overhead figure came from an H100.
Reality
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
MLPerf Inference v6.1 preview submissions put Vera Rubin NVL72 at up to 3.7x GB300 on Qwen3-VL and up to 2.5x on DeepSeek-R1, on two different inference frameworks. The four-rack 99% scaling result is an offline number.
Reality
The $875 million round prices an inference chip whose performance figures come from cycle-accurate simulations and whose production date has moved from early 2027 into the second half. The system Positron ships today runs on HBM.
Reality
Qwen3.8-Flash-Next puts 36 Gated DeltaNet layers and 12 sparse-attention layers on Hugging Face, which means the retrieval budget Qwen4 will inherit is something you can measure against your own traces now.
Perspective Coverage
5 publishersReality
SemiAnalysis has pushed fixed-length serving into maintenance mode. On replayed Claude Code traffic, NVIDIA's GB300 is credited with 15x Hopper, against 40x over H200 on the retired static test.
Reality
A five-year, $400 million offtake covers roughly 4.75MW at Bitdeer AI's Malaysian A102 facility. The contracted-before-energization structure is becoming the price of building at all.
Reality
NVIDIA says Alibaba's largest open-weight model serves over 4K tokens/sec/GPU and 350 tokens/sec/user in FP8 on a GB300 NVL72. That figure is the self-hosting floor, not a benchmark.
Reality
The company says it needs $500m of capex just to replenish capacity it has already committed. Treat advertised GPU availability at alternative clouds as a lagging indicator.
Reality