FlopSource

Daily News

AI and Compute: Daily Briefing (Nov 03, 2026)

By Tyler Babin · FlopSource Research

Google Spot undercuts the H100 floor, OpenAI promises to shed up to 1 GW when the grid is strained, and Nvidia reportedly draws a China line for its cloud partners. Plus: Rubin production specs are lower than the preview numbers, and an India cloud turns inference into a pay-per-token product.

Published Nov 03, 2026

1. Google Cloud H100: $5.38 on demand, $1.09 on Spot, against a $1.99 market floor

As of November 3, an NVIDIA H100 on Google Cloud costs $5.38 per GPU-hour on demand and $1.09 as a Spot VM, each in the cheapest region FastGPU's live feed finds. The cheapest on-demand H100 SXM anywhere in that feed is $1.99 at Voltage Park, so Google's on-demand rate is 2.7x the market floor. Spot flips the comparison: Google's cheapest B200 Spot rate of $1.74 sits below the cheapest on-demand B200 anywhere at $4.89, and its H100 Spot rate of $1.09 sits below the $1.99 on-demand floor. The caveat is in the product: a Spot VM can be stopped at any time, and Google sells its H100 and B200 machines on demand only as whole eight-GPU VMs, so the smallest uninterrupted H100 bill is eight times the per-GPU figure. Operator read: for interruptible work, hyperscaler Spot is now price-competitive with the open market; for steady work, the premium is still roughly triple.

Source: FastGPU, Google Cloud GPU Pricing in 2026 · Nov 03, 2026

2. OpenAI launches Grid Protectors: up to 1,000 MW of curtailable demand in Georgia

OpenAI announced on November 2 a program it calls Grid Protectors, committing campuses it directly develops to reduce electricity use during grid emergencies, starting with a planned Georgia campus. Project Camellia, a planned 3,200 MW campus in Effingham County, Georgia, will make up to 1,000 MW available for curtailment under a 25-year agreement with Georgia Power. That is about 31% of the campus peak load and roughly the average demand of 700,000 homes. OpenAI says it will pay the full infrastructure and electric-service costs for the site, and that flexibility arrangements will vary by campus. Read it plainly: the campus is planned, not built, so this is a contract commitment about future demand cuts, not a demonstrated grid service today. The supporting evidence OpenAI points to is a December 2025 trial in the London area, run with National Grid, Emerald AI, EPRI, Nebius and NVIDIA, where a Blackwell Ultra cluster cut power by up to 40% across more than 200 simulated scenarios.

Source: Dataconomy, reporting OpenAI's announcement · Nov 03, 2026 (OpenAI announcement Nov 02, 2026)

3. Nvidia reportedly tells data center partners to keep China exposure below half

Nvidia has instructed its data center partners to reduce exposure to Chinese customers, according to a Financial Times report published Tuesday and reported by Invezz on November 3. Partners were reportedly told to ensure at least 51% of their services are sold to non-Chinese companies. The guidance was communicated verbally, applies to future contracts, and does not affect existing agreements, the report said. Nvidia has not officially confirmed the policy and neither confirmed nor denied the report; a follow-up report by The Information said the instructions were intended to encourage a broader customer mix and were not being enforced as a mandatory restriction. Operator read: if you rent Nvidia capacity through an overseas cloud to serve China-linked workloads, that route is getting politically harder, even where no formal rule has been published.

Source: Invezz, reporting the Financial Times · Nov 03, 2026

4. Rubin production specs land lower than the preview: HBM4 at 19.2 TB/s, NVLink at 3.0 TB/s

NVIDIA has clarified its revised Vera Rubin specifications, saying the final figures reflect what the system sustains at qualified, volume production, in a statement reported by Wccftech on October 29. Per Rubin GPU, HBM4 capacity is unchanged at 288 GB, but HBM4 bandwidth is now listed at 19.2 TB/s, down 12.7% from the 22 TB/s preview figure, and NVLink 6 bandwidth at 3.0 TB/s, down 16.7% from 3.6 TB/s. Networking bandwidth moved the other way, up 12.5%. NVIDIA says individual component specs do not determine application performance, that Vera Rubin NVL72 is in full production and shipping, and that third-party benchmarks published so far were already measured using the production specifications. Operator read: budget and compare Rubin on the production datasheet, not the GTC preview numbers, and judge it on measured tokens per dollar, not on bandwidth headlines.

Source: Wccftech, reporting NVIDIA's statement · Oct 29, 2026

5. E2E Networks launches Blaze, a pay-per-token inference service, on a 1,024-GPU Blackwell base

E2E Networks has launched Blaze, a pay-per-token inference service for open AI models, with a single OpenAI-compatible endpoint and customers paying by token consumption, YourStory reported on November 3. The service runs on NVIDIA GPU infrastructure with components from NVIDIA DSX OS, is hosted on E2E's India-based cloud, and is aimed at voice, coding agents, document AI, OCR, and image and video generation, with data sovereignty as part of the pitch. E2E says it is expanding its AI infrastructure capacity to approximately 14.6 MW, including a 1,024-GPU NVIDIA Blackwell cluster that has been operational since May 2026. Operator read: this is the inference economy arriving as a product in a sovereign market: no GPU-hour to manage, just a per-token price to compare against your own serving cost.

Source: YourStory, daily roundup · 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