Databricks charges $7 per H100 GPU-hour, 2.4 times the market rate, and caps a node at eight GPUs

DatabricksPVT$216.58B — Databricks, private, latest valuation $216.58B, the private data-platform company valued at $190 billion by its August investors↗ ¹, moved AI Runtime to general availability on October 21, 2026↗ ². The service rents serverless NVIDIA▲+2.27% — NVIDIA, up 2.27 percent today GPUs for fine-tuning open models inside the lakehouse, the governed store where its enterprise customers already keep their data, with NVIDIA's NeMo containers and the new Nemotron 3.5 models available from day one ³. The Inference tracked Databricks' paper here on October 2; the new fact this week is the service's first public unit price, and the fine print under it ³.

The showcase ran 16 GPUs. The docs cap a node at eight

The showcase customer made the pitch. Kostia Omelianchuk, an applied research scientist at the email app maker Superhuman, moved a training run from one node of eight A100 GPUs to 16 H100s; training time fell from about seven days to 48 hours, a 3.5-times speedup ². The quote does not separate the newer chip's gain from the second node's ⁴, and the second node is the product's open seam. Databricks' documentation, updated October 20, the day before launch, says all AI Runtime accelerators provision a single node, capped at eight GPUs ⁴.

Seven dollars an hour, on hardware Databricks does not own

List price on AWS▲+0.34% — AWS, up 0.34 percent today, Premium plan, US East: $7.00 per H100 GPU-hour and $2.50 per A10, both still tagged Preview after general availability ⁴. At list, Superhuman's run is 768 GPU-hours, about $5,376, with two nodes billing $112 an hour ⁴. The open market rents the same chip for less: Silicon Data's standardized index of specialized GPU clouds↗ read $2.89 per H100-hour on October 25 ⁵. Databricks' list price is about 2.4 times what the open market charges for identical silicon, and 52 cents below the market rate for NVIDIA's newer B300, $7.52 ⁵.

The docs, not the blog, describe the product

The launch post says workloads "scale to hundreds of GPUs reliably", backed by 100,000-plus GPU-hour runs at hundreds of customers including Robinhood▼−2.29% — Robinhood, down 2.29 percent today, Merck and FactSet▼−3.03% — FactSet, down 3.03 percent today ². The hardware table has no row beyond eight GPUs ⁴. The docs warn capacity "may be constrained or unavailable in your region", say high-demand runs can land in other regions at data-transfer cost, and allow 14-day CLI jobs whose restart needs hardware again ⁴.

Databricks does not own the silicon: its chief executive said in August it carries multibillion-dollar cloud commitments with AWS, Microsoft▲+1.10% — Microsoft, up 1.10 percent today and Google ¹. Rental rates are already moving, with the H100 index up 6.2 percent in a week and the B300 up 10.6 percent ⁵. If wholesale hardens, Databricks absorbs the spread, re-tags the price, or declines supply under the footnote that displayed pricing "does not guarantee product availability"↗ ⁴.

Tubi is happy. The analysts are split

The counterweight is real usage. Tubi's principal machine-learning scientist is "very happy with the compute availability" ². Analysts credit AIR's integration with Unity Catalog↗, Databricks' permissions layer, as the actual differentiator, while flagging that available GPU capacity is the open question given limited supply, and that most enterprises may never need training compute at this scale ⁶.

Watch the Preview tag

Two signals say whether the $7 anchor holds: whether the Preview tags come off the prices, and the B300 waitlist, the one inventory line in the docs ⁴. A two-node job started on launch day that runs the full 14-day cap hits it on November 4, and the restart is another bet on capacity being there ⁴.