Model watch · This week (event date 08-17)
From the Aug 18, 2026 daily brief
A team reported on the blog of the model hosting platform Hugging Face that their constraint-aware GPU allocator — their own product — was compared against the simplest first-come-first-served scheduler across seven scenarios. Hardware identical, workload identical, and GPU utilisation rose by as much as 33 percentage points (the control baseline sits around 50%, which is half a cluster idling). Priority-weighted output rose in all seven, by as much as 105%. That second ruler measures completed work weighted by job priority, against first-come-first-served in the same scenario. ⚠️ The two are different rulers and neither converts into "105% more efficient." The shape of the problem is worth understanding. Four kinds of work compete for one pool of GPUs. Training, batch inference and quantisation are batch-shaped — once started, each wants a whole GPU continuously until it finishes. Real-time inference is the opposite, with a demand curve that moves at every timestep. Real-time inference cannot wait, so the only way to guarantee availability is to reserve for the daily peak all day long. An application that needs six GPUs at midday and two at 4am holds six for twenty-four hours, and those four idle GPUs are unavailable to any batch job all day; they are not being used, and they are not free either. Read alongside main line item 2, it stings: on one side, 4.25 gigawatts of compute requires a US$105 billion guarantee; on the other, the same hardware yields another third of itself from a change of scheduling. How much of the "compute shortage" is real scarcity and how much is scheduling nobody fixed is a question with teeth for the capital narrative. ⚠️ That juxtaposition is our inference, not this post's claim. ⚠️ Four more limits travel with it: this is the publishing party's own measurement; the baseline is one they set themselves; there is no independent replication; and there is no peer review.
Just an email address, unsubscribe anytime. This is the only thing we ask of you.
What follows is not a preprint. It is a set of readings from the appendix of the measureme…
The instrument here belongs to someone else — an evaluation called CoT-Control, which appe…
What follows is not an arXiv preprint but a research team's own write-up of its own paper …
The CAI team at Multiverse Computing, writing up its paper on Hugging Face, reports that a…