GPU hosting blog and changelog. Notes from the cage.

The second tranche of HGX B200 landed on 19 August, burned in for a week, and cleared the waitlist this morning. Eight 8-way nodes are live, two more racks are on order for IAD-1 with a target of November. Power is the constraint, not supply: a B200 node draws ~10.2 kW at the wall and we had to add a 415 V busway to cage 4 for them.

Single-GPU B200 instances are not something we can offer — the parts only exist on an 8-way baseboard — so the smallest unit is 2× (a half-node, NVLink-partitioned) at $7,391/mo, which is two cards at $3,890 less the 5% two-GPU discount. Live stock per region sits on the B200 rental page.

Every instance's NVMe is a LUKS2 volume with a key generated at provisioning and held in the host TPM, sealed to that instance ID. Releasing the instance asks the TPM to destroy the key — at that point the drive contents are ciphertext with no key anywhere. Then we blkdiscard the namespace and run a 30-second verify pass that samples 2,000 random LBAs and confirms they read as zero. Only after that does the slot go back to inventory. This adds about 90 seconds to the release. We think it's worth it.

People kept asking to buy retired cards from us, and we kept saying no because we didn't have a process. Now we do: a hardware page with new and fleet-retired cards, colocation in our own cages, and a buy-back rate. Colocated cards use the same host fleet and dashboard as rentals — there is no separate "colo product". The card is yours and the monthly bill drops to power and space. The per-card tiers are on the GPU colocation page.

Switching an instance from weekly to monthly (or extending a monthly one by 3/6 months at the prepaid discount) no longer touches the machine. Also: snapshot restores to a different region, TOTP recovery codes regenerable, API webhooks for renewal failures, and Ukrainian and Turkish translations of the dashboard.

One rack, four nodes, all pre-sold to waitlist. Interesting: the wall-plug efficiency versus H100 on an FP8 training benchmark we run for burn-in is about 2.1× per watt. Less interesting: the busway upgrade it needed.

One 8-way MI300X platform in DFW-1 for people who asked for 192 GB of HBM at a price that isn't H200. ROCm 6.4 template included. And the first 240 RTX 5090s spread across all three regions; blower models only, because the 575 W triple-fan retail cards do not belong in a rack.

Same HGX form factor as our H100 nodes, so provisioning was uneventful. 141 GB per card means a 70B model in BF16 fits on two cards with room for a batch. Priced at $2,290/mo per card, which is still the number on the H200 rental page.

Third region. Hydro-powered facility, cool climate, good latency to the West coast and Asia-Pacific. Opens with L40S, A100 and 4090 stock; H100 nodes follow in Q4.

We rewrote the API and then rewrote the dashboard to use only the public API. If a thing is possible in the dashboard, it's possible with curl. v1 is deprecated with a 12-month sunset.

Allocation was hard to get in early 2023. We got fifty PCIe cards for DFW-1; SXM nodes came ten months later. Every one of those first fifty is still in service.

IAD-1 opens with 300 GPUs, mostly A6000 and A40, and 3090s moved from Dallas. Delayed two days by a cross-connect; see the status page for the honest version.

The most requested feature of our first year. XMR deposits are credited after 10 confirmations at the spot rate at first confirmation, same as everything else. We run our own node, and still do: see Monero GPU hosting.

One rack, one cage, one switch (we'd learn about that), forty 3090s, and a pricing page with exactly one row. Monthly terms because that's how we'd want to rent a GPU. Bitcoin and Ethereum because that's what we had. No KYC because we don't want to hold your passport any more than you want to give it to us. Five years on, no-KYC GPU hosting still works exactly that way.