A price comparison of on-demand H100 rentals finds a 4.57-fold difference for the same H100 SXM 80GB, depending on provider. RunPod Community lists the GPU at $2.69 per hour, or about $1,964 for a continuously running 730-hour month; RunPod Secure is $3.99, AWS’s p5.48xlarge works out to $6.88 per GPU-hour, and Azure’s ND96isr_H100_v5 costs $12.29, or $8,972 monthly. The comparison deliberately separates H100 SXM, PCIe and NVL models, and uses each vendor’s cheapest route to a given GPU, rather than combining unlike hardware or counting within-vendor options as market differences.
The price gap is not presented as a pure markup: the Azure ND instance includes 400 Gbps InfiniBand and GPU Direct RDMA for multi-node training, as well as enterprise reliability, compliance and capacity offerings. Those features may be unnecessary for single-node fine-tuning but important for large distributed workloads, making the cheapest option unsuitable for some users. The figures use on-demand list prices for one region each - AWS us-east-1 and Azure eastus - and exclude reserved, committed-use and spot discounts, which could narrow the gap. GPU counts are inferred from vendor documentation; availability is not reflected. Lambda is excluded from the headline comparison because its published rates were not yet in effect and current rates require a login.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
The UK’s planned 2G and 3G shutdown by 2033 could disrupt 12-13 million smart meters, forcing replacements. The upgrade could cost about £4 billion, adding to the expense of the government’s troubled rollout. (theregister.com)
Honda plans to build a $3 billion auto plant in Ohio, creating about 2,000 manufacturing jobs. The investment will expand Honda’s production presence in the state. (reuters.com)
LiteDRAM’s guide explains how FPGA-to-DRAM signals, strobes, board layout, termination, reference voltages, and clocks affect memory communication. DDR3 and newer interfaces need boot-time tuning to align signal timing for reliable operation. (github.com)
Alice Huang reportedly helped a network of companies secure more than $7.5 billion in Nvidia GPU imports and move 240,000 restricted chips, mainly H100s and H200s, into China. The claims are based on an investigation tracing corporate links to Megaspeed and Shanghai Century Litong, alongside reports of her personal ties to Nvidia CEO Jensen Huang. (tomshardware.com)
Tendrils raised $30 million in a Benchmark-led round to develop a processor with thousands of parallel cores, designed to move beyond traditional CPU architectures. The company is also building Vine, an open-source programming language tailored to the new chip design. (tendrils.co)
Toyota is reviving Scion as “Scion by Toyota,” a vehicle-design initiative whose first model is the KEI-01, a rebadged kei truck planned for U.S. sale. Limited to under 25 mph, the right-hand-drive truck is intended for low-speed roads and commercial or agricultural use. (theautopian.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.