AI model
Nvidia's Grace Blackwell GB300 is a GPU-CPU server platform combining Blackwell Ultra GPUs with Grace CPUs, used for large-scale AI training and inference.
Brightray has signed a nonbinding letter of intent for a 10MW Indiana AI data center that it aims to deliver within a year using factory-built modules. The twelve months run from a definitive contract that has not been signed, so the date cannot yet go into a plan.
Reality
Nvidia licensed Poolside's Model Factory and hired 109 of a team its CEO put at fewer than 115 engineers and researchers. The corporate shell survives; the roadmap does not.
Reality
Jalapeno's lead is measured against last generation and the volumes are tiny, but a first-pass ASIC clearing Nvidia, AMD and Google parts reprices the design barrier, not the supply chain.
Perspective Coverage
7 publishersReality
Eight weeks after Nvidia started lending its balance sheet to small AI clouds, internal warnings about antitrust and about how much it was dictating to customers have stopped parts of the programme, and the operators who needed it most have no obvious substitute.
Reality
Kimi K3's 1.4 terabytes of weights take eight Nvidia GB300s just to sit in memory. Export controls keep those chips away from Moonshot. Modal, Fireworks and Baseten price the hosted result at $3 in and $15 out.
Reality
The 1.7x to 3.6x latency range is set by the baseline systems, not the chip, and the report's own publication date is unsettled. Read it as direction, not evidence.
Perspective Coverage
9 publishersReality
The deal is the largest slice of more than $33bn Microsoft has committed to specialist GPU clouds. It leaves the fleet, the debt behind it and the 18-month depreciation clock on someone else's balance sheet.
Reality
Bloomberg says executives have discussed an American majority shareholder. G42 already pulled Huawei gear out of its racks and wired US-specified monitoring into its own data centres, and each concession came with an expiry date.
Reality
The release attaches speculative decoding to weights it publishes under MIT, which lowers what a self-hosting threat costs to stand up, even though the performance claim behind it is still the vendor's own.
Reality
The zettaFLOPS figure is an estimate. The measured one, sitting in a GB300 example, is a 25 percent density gain, and it is paid for out of the safety margin.
Reality
Nvidia's newest racks draw more than utilities planned for, and grid equipment is the queue that matters. But the headline gigawatt figures do not add up to the shortfall being advertised.
Reality
NVIDIA's tutorial post-trains a 4B model into a Franka manipulation policy that runs on Jetson Thor with about 0.6 seconds of latency slack per cycle. Closed-loop success is 22.9%.
Reality
Chinese AI firms can reportedly rent restricted Nvidia capacity from data centers abroad. Bank of America is unmoved, keeping a $350 target on a stock near $225.
Reality
Horizon 1 is delivered at Childress, Texas. Billing only starts after a five-day GPU validation, which is now the number that matters more than the headline contract value.
Reality