· via dev.to (home feed)
Post-SC26 analysis: racks, watts and water displace GPU counts in AI capacity planning
A dev.to analysis of post-SC26 announcements argues that rack power, cooling loops and serving-engine versions — not accelerator counts — now determine how much AI capacity a platform team can actually deploy.

An analysis published on dev.to after Supercomputing 2026 in Chicago argues that the standard way of sizing AI clusters — pick an accelerator, count units, multiply by a utilisation estimate — no longer reflects how hardware is actually sold or deployed. The author's core claim is that the useful unit of planning is now the rack, its power feed, its cooling loop, and the exact software build running on the silicon.
Vendors shipped racks, not components
The pattern the analysis highlights sits across several announcements rather than in any single one. On 18 November, Compal announced a rack-scale liquid-cooled system built on NVIDIA's HGX Rubin NVL8 platform, and Giga Computing showed rack-scale infrastructure with liquid cooling at the same show. A day later, according to the post, an HPCWire roundup collected new systems from Compal, Penguin Solutions, ASUS and Dell — rack-scale machines sold as integrated products with liquid cooling and storage, rather than components left for an integrator to assemble.
The most telling figures, the author argues, came from a vendor that sells no accelerators. On 16 November, nVent said its liquid-cooling portfolio has more than four gigawatts deployed, demonstrated a 2.3-megawatt modular coolant distribution unit, and unveiled an 800-volt DC power rack in its MAVERICK line. Installed cooling measured in gigawatts means the thermal problem has moved from an engineering afterthought into the capital plan. The voltage figure matters because moving distribution to 800 V DC strips out conversion stages between the utility feed and the accelerator — fewer transformers, rectifiers and UPS blocks, less copper and less waste heat in the power room.
Retrofit paths for existing halls
Most teams inherit rooms with a fixed power allowance, low ceilings and no water near the rows, and the analysis flags the announcements aimed at that installed base. Nexalus and Renda launched the Nova8, a 4U chassis holding eight NVIDIA Blackwell GPUs with CPU, memory, storage, power and cooling integrated — a product built to fit an existing rack rather than require the room to be rebuilt around it.
Fujitsu took a different route. According to the post, it began global sales in November of its 144-core MONAKA processor and server, built from a 2 nm core die stacked with 5 nm SRAM and I/O dies, and showed it at SC26 as a sovereign-AI server aimed at inference in air-cooled facilities. The author's point is that a fast CPU in an existing hall can beat an accelerator that needs a construction project — a claim about facilities and scheduling, not raw benchmarks.
Throughput is a property of the software build
The argument's less obvious half is that the work extracted from a rack depends on the software version. The post cites SGLang v0.5.23, released 2 November with 1,007 pull requests from 257 contributors. Its Paged Experts feature extends paging beyond KV caches to a model's expert weights, letting a mixture-of-experts model larger than a single GPU's memory run on one consumer card at up to 30 times the speed of CPU offload. The same release added fused persistent megakernel paths for NVIDIA and AMD hardware, making single-user decode up to 44% faster on a GB300 and 75% faster on an MI350X.
The vLLM 0.34.0 release shows the operations side of the same story. Extensible KV-cache sizing grew a MiniMax-M3 cache on an H200 from 433k to 672k tokens, but at concurrency 128 some models paid 6–18% higher per-token decode latency — a throughput-for-context trade-off shipped inside a release note. The release also added checkpoint/restore through CRIU and the ability to drop a failed rank from a data-parallel/expert-parallel group and redistribute its experts without restarting the service.
Why it matters
For infrastructure teams the consequence is that capacity is no longer a single figure derived from a GPU datasheet; it has to be tracked per facility and per software release. The analysis proposes a checklist:
- Audit what the building accepts today: kilowatts per rack, ceiling height, and the presence or absence of water. A 4U eight-GPU server may be the only option that fits.
- Decide the distribution voltage before ordering hardware, since retrofitting the electrical chain afterwards is expensive.
- Treat a multi-megawatt CDU as a facility asset with a vendor, an SLA, spares and a maintenance contract.
- Record the serving-engine version, attention backend and cache policy alongside tokens per second, and re-measure after every upgrade rather than assuming each release is faster.
- Test rank-failure recovery under hours of sustained load rather than a short smoke test.
- Weigh CPU inference for low-traffic, local or sovereign workloads.
The broader shift the author names is that silicon now comes in more forms than ever, while the hall, the power feed and the software build decide how much of it is usable. Plans still written purely in GPU counts may look exact on paper; the ones written in racks, watts and software versions are the ones that survive commissioning day.
- #datacenters
- #liquid-cooling
- #ai-infrastructure
- #inference
- #capacity-planning