FlopSource

Daily News

AI and Compute: Daily Briefing (Nov 04, 2026)

By Tyler Babin · FlopSource Research

The H100 market floor falls to $1.79, AMD data center revenue nearly doubles to $8.5bn, and Hetzner ships its first four-GPU server. Plus: a UK energy company opens B300 reservations, and Supermicro pairs Nvidia prefill with SambaNova decode in one inference rack.

Published Nov 04, 2026

1. H100 market floor falls to $1.79 at Vast.ai; Google holds at $5.38, Oracle at $10.00

As of November 4, the cheapest on-demand H100 SXM anywhere in FastGPU's live feed is $1.79 per GPU-hour at Vast.ai, down from the $1.99 floor recorded on November 3. Google Cloud is unchanged at $5.38 on demand and $1.09 on Spot, so Google's on-demand rate is now 3.0x the floor, up from 2.7x yesterday on the floor move alone. Oracle Cloud sits further out: $10.00 per GPU-hour on demand for an H100, with no preemptible H100 offered at all, which is 5.6x the $1.79 floor. The B200 floor is unchanged at $4.89 at DeepInfra, against Google's $1.74 Spot rate and Oracle's $14.00 on-demand rate. Operator read: the hyperscaler premium did not move today; the open market underneath it got cheaper. If your workload can run on a marketplace host, the gap you are paying for a hyperscaler badge widened again.

Source: FastGPU, Google Cloud GPU Pricing in 2026 · Nov 04, 2026; FastGPU, Oracle Cloud GPU Pricing in 2026 · Nov 04, 2026

2. AMD data center revenue nearly doubles to $8.5bn on EPYC and Instinct MI450 demand

AMD reported third-quarter results on November 3: total revenue was a record $13.4 billion, up 45% year over year, and Data Center revenue was a record $8.5 billion, up 97%, driven by strong demand for EPYC processors and the launch of the Instinct MI450 Series GPUs. Data Center was $8.546 billion against $4.341 billion a year earlier, about 64% of total sales. AMD guided fourth-quarter revenue to approximately $17 billion, plus or minus 3%, and CEO Lisa Su told analysts the company expects total revenue to grow more than 80% in 2027, with data center chips making up more than 80% of sales. Operator read: this is segment revenue, not AI GPU revenue alone - EPYC CPUs are in the same number - but a near-doubling at this scale is direct evidence that second-source accelerator demand is real and still accelerating into 2027.

Source: DQ India, reporting AMD's third-quarter results release · Nov 04, 2026 (AMD release Nov 03, 2026)

3. Hetzner launches its first multi-GPU server: four Blackwell cards, 384 GB, from EUR 3,599 a month

Hetzner has introduced the GEX430, its first multi-GPU dedicated server, built around four NVIDIA RTX PRO 6000 Blackwell Max-Q GPUs with a combined 384 GB of GDDR7 ECC memory - 96 GB per card. The rest of the machine is an Intel Xeon Gold 5412U (24 cores), 128 GB of DDR5 ECC memory and two 1.92 TB Gen4 NVMe SSDs, sold as a dedicated, single-tenant machine rather than shared cloud. Pricing starts at EUR 3,599 per month with a one-off setup fee of EUR 3,599; Hetzner also lists a US price of $4,199 per month plus a $4,199 setup charge. That works out to roughly $1,050 per GPU per month before setup, with no per-hour option. Operator read: this is the inference-hosting end of the market formalizing - fixed monthly bare metal with EU data location for GLM, Mistral and Qwen-class open-weight models, priced to compare against reserved GPU cloud, not against spot.

Source: Data Centre News UK, Hetzner launches first four-GPU server for AI work · Nov 04, 2026

4. Fuse Energy opens UK B300 reservations: 128 GPUs in December, 1,024 by Q2, 116 MW from 2028

Fuse Energy, a UK energy supplier founded by Revolut alumni, has opened reservations for Fuse Compute, a bare-metal, single-tenant NVIDIA GPU service hosted in UK data centres under UK jurisdiction. The first 128 NVIDIA B300 GPUs are due online in December, expanding to 1,024 GPUs in the second quarter of the following year, then a further 22 MW in the fourth quarter and 116 MW from 2028. Fuse says it supplies 470,000 UK homes, holds supplier and meter operator licences, and has a data centre connection pipeline of 142 MW across 15 projects. No pricing is published: pricing, terms and delivery are agreed per contract. Operator read: another energy-first entrant selling power access as the product, with GPUs attached. The test is delivery - December's 128 GPUs are a reservation book today, not live capacity.

Source: Data Centre News UK, Fuse launches UK AI compute service with NVIDIA GPUs · Nov 04, 2026

5. Supermicro and SambaNova pair Nvidia prefill with SN50 decode, planning 10,000 chips

SambaNova announced on November 3 that Supermicro will build and sell AI inference systems that pair NVIDIA GPUs with SambaNova SN50 Reconfigurable Dataflow Units, in a collaboration initially planning for 10,000 SN50 chips. The split is explicit: GPUs run the prefill phase, reading the prompt, and the SN50 RDUs run decode, writing the response. The systems join Supermicro's Data Center Building Block Solutions range; Supermicro will integrate, test, deliver and support them, including on-site service, while SambaNova engineers tune models on SN50. The first sales effort targets neoclouds and sovereign clouds, then enterprises running AI in their own data centres. No pricing, rack configuration or shipment schedule was disclosed. Operator read: disaggregated inference is becoming a buyable rack, not a research architecture - but plan on the 10,000-chip figure as a planning volume, not shipped systems, until deployments are named.

Source: SambaNova, SN50 chips to power Supermicro's new inference racks · Nov 03, 2026

Sources

Compare providers for your workload

Live pricing, hardware specs, and availability across the GPU clouds, neoclouds, and bare-metal providers in our directory.

Open the Directory