# SF Compute (GPU market + Autoresearch sandboxes)

Official pricing: https://autoresearch.sfcompute.com/  
Category: gpu-cloud · Isolation: vm

## Pricing regimes (raw)

- **Autoresearch RL sandbox (billed per GiB of RAM, auto-park)** (sizes): sandbox-sm 1 vCPU/2 GiB $0.09144/h; sandbox-md 4 vCPU/8 GiB $0.73152/h; sandbox-lg 8 vCPU/32 GiB $1.46304/h; sandbox-xl 16 vCPU/64 GiB $2.92608/h
- **Autoresearch CPU node (cpu-2 / cpu-8 container, persistent disk)** (sizes): cpu-2 4 vCPU/8 GiB $0.18/h; cpu-8 16 vCPU/32 GiB $0.4824/h [alt]
- **Autoresearch interactive H100 node, on demand (h100-1 = 1 GPU $3.96/h; h100-8 = 8 GPUs)** (resource): $0/vCPU-h, $0/GiB-h [stock]
- **Autoresearch batch jobs on H100 (queued, preemptible)** (resource): $0/vCPU-h, $0/GiB-h [alt, spot]
- **SF Compute market: reserved 8-GPU node time (orderbook buy)** (resource): $0/vCPU-h, $0/GiB-h
- **SF Compute market: spot deployment (auto-buys at or below max rate)** (resource): $0/vCPU-h, $0/GiB-h [spot, beta]

## Features

Yes: snapshots, memory snapshots, fork/clone, pause/resume, persistent disk, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, your own Docker/OCI image, full VM (own kernel), root, HTTPS preview URLs, egress allowlist, open internet, private networking, SOC 2, GPU, Python, Node.js, wake on request, live fork (no pause), memory fork, audit logs, EU data residency, scoped API keys, spend limits, MCP server, snapshots on demand, hosted agent API (their own agent), fast boot (benchmarked), gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables)

No: volumes, Docker inside, nested virtualization, browser, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, preinstalled agents, SSH, public IPv4, custom domains, raw TCP inbound, static egress IP, self-hosting, BYOC, open source, HIPAA, SSO, arm64, Windows, macOS, GPU desktop, HTTP method/path egress rules, live resize, automatic snapshots, agent harness API, secret proxy, secret proxy for any API, extra volumes, shared volumes

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/sfcompute.json → feature_evidence

## Caveats

- Two products under one card: (1) the SF Compute GPU market (8-GPU node-hours bought/resold on an orderbook; reserved or spot; no list price, so GPU rates are null), and (2) Autoresearch/givemeanode (fixed per-minute list prices for sandboxes, CPU and H100 nodes, batch jobs). Benchmarks measured the Autoresearch sandboxes.
- Sandbox rate is per GiB of RAM; vCPU is priced implicitly via a 4 GiB/core floor, so CPU-heavy/low-RAM shapes are expensive (4 vCPU/8 GiB bills as 16 GiB = $0.73/h flat-out). Parking can cut this by up to ~99% for idle-heavy agent loops (knob).
- Autoresearch H100 list rate is $3.96/GPU-h ($0.066/min) on the live pricing table; llms.txt still implies $3.60 ("$0.90 at the H100 list rate" for a 15-min window) and the CLI example shows $0.0666/min - stale docs.
- cpu-node vcpu = 2 x physical cores (assumption). cpu-8 "33% off (CPU node launch pricing)" - unclear whether $0.4824/h is pre- or post-discount.
- Sandbox egress price unpublished; SF Compute specs page says "No ingress/egress fees" for GPU clusters, applied card-wide (egress_gib 0). Public IPv4 on market VMs is optional at create; price unpublished.
- Sandbox snapshots: $0.10/GiB-month resident (deduplicated), $0.05/GiB-month cold in object store (first fork after cold pays a one-time rehydrate delay); live until deleted unless an expiry is set. Encoded at $0.10.
- Market VMs: no persistent storage (terminate deletes data), boot up to ~10 min, billed per node-minute including boot; compute limits in node-hours per account. Market prices vary by SKU, zone and window; third-party reports ~$1.49-1.68/GPU-h H100 averages in Feb-Mar 2026 (spheron.network) - not official.
- No free credit found (promotional credit exists but amount unpublished).
- Billing migration is date-gated: the prepaid-credit model was announced 2026-09-22 for new accounts; existing monthly-invoice accounts move on 2026-10-01, and pre-October usage stays invoiced under the old rules. The July 29 $10-$10,000 prepayment option was optional. Cold $0.05 snapshot tier introduced 2026-08-22; node snapshot cut $0.11 -> $0.10 on 2026-07-29. (patch v8-history, 2026-09-28)
- SSH: Autoresearch docs say "There is no SSH. Commands run in the node's container" and "Nodes have no public address and no open ports"; sandboxes are exec-over-API only, so features.ssh=false. The SF Compute market VMs (reserved/spot modes) keep their ssh:true override. (verified 2026-09-28, autoresearch.sfcompute.com/docs)
- Sandbox disk: no included sandbox disk size and no running-disk rate are published (only the parked rate ~$0.000139/GiB-h on what a parked sandbox holds). included_disk_gib / disk_gib_month stay null. Nodes (cpu/h100) have a 250 GiB persistent volume (node_volume_gib default), parked at $0 when stopped. (2026-09-28)
- Autoresearch stopped nodes are deleted, with their disk, after 60 days stopped with no activity (emails at 30 and 45 days; any wake/boot command/snapshot/network change resets it). Snapshot first to keep files (snapshot storage $0.10/GiB-month past 250 GiB free).
- Market compute limits are in node-hours (compute_max_total_node_hours, account-wide and per-SKU; resting buy orders count before they fill). Raising them is not self-serve: verify the org, then contact SFC. The card's fleet block describes Autoresearch limits only.
- Market buy orders hold credit for the worst case (limit price x full window x node count) until filled; unused hold released after fill/cancel.
- Elastic InfiniBand Partitions for market VMs entered public preview 2026-09-25 (opt-in 'sf preview set infiniband-preview=true'); specs page still says IB on VMs 'in Q3 2026'. No separate price published.
- Autoresearch queue has paid priority tiers ('folks who buy priority are ordered as if they had been waiting a while'); tier prices are not published. Every account's first node gets the largest head start.

## How this provider charges


SF Compute is a GPU cloud built around an **orderbook**. You buy (and resell) time on 8-GPU nodes, as reserved windows or as
spot deployments, and there is no list price. Its sister product **Autoresearch / "givemeanode"** is an agent PaaS with
fixed per-minute list prices for RL sandboxes, CPU nodes, H100 nodes and batch jobs. The computesdk benchmark ("givemeanode")
measured the Autoresearch sandboxes. The main sfcompute.com/pricing is a login wall and sfcompute.com/prices now serves the homepage.
The live Autoresearch pricing table is at https://autoresearch.sfcompute.com/ (saved to raw/autoresearch_sfcompute_com_home_pricing.md).

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Sandbox running | Autoresearch RL sandbox with a command in flight | Per GiB of **RAM** per minute. Each core comes with 4 GiB, so billed GiB = max(RAM, 4 × vCPU) | $0.04572/GiB-h. sm 1/2 → 2 GiB $0.0914/h. **md 4/8 → 16 GiB $0.7315/h**. lg 8/32 $1.4630/h. xl 16/64 $2.9261/h | https://autoresearch.sfcompute.com/, /llms.txt |
| Sandbox parked | Automatically after ~1.5 s idle | Spare memory handed back stops billing at the RAM rate. The held disk bills at the parked rate. Processes keep running, wake takes tens of ms | ~$0.000139/GiB-h (~330× cheaper) | same |
| Sandbox fork / bake | `fork_sandbox` (1-256 per call), `create_sandbox_env` | Forks are free (CoW) and each fork pays its own RAM-minutes. A bake bills as sandbox time | $0 fork | same |
| Sandbox snapshots | Stored bake/branch snapshots | Deduplicated GiB-month. Resident on host vs cold in object store (automatic, one-time rehydrate) | $0.10 resident / $0.05 cold per GiB-month. Live until deleted or expiry | same |
| Sandbox limits | New workspace | `sandbox_vcpus` starts at 4 (lg/xl need a raise). RAM ceiling 8 GiB (64 GiB platform cap) | - | /llms.txt |
| H100 interactive node | `create_node --chip h100` / `8xh100` | Per minute while running **plus an idle window** after the last command unless stopped. Stopped means disk parked at $0 | h100-1 $0.066/min = **$3.96/h** (15-min window, $0.99 max). h100-8 $0.528/min = $31.68/h (20-min, $10.56 max) | https://autoresearch.sfcompute.com/ |
| CPU node | chip `cpu-2` / `cpu-8` | Per minute, 15-min idle window | cpu-2 (2 physical cores, 8 GiB) $0.18/h. cpu-8 (8 cores, 32 GiB) $0.4824/h ("33% off, CPU node launch pricing") | same |
| Calendar discount | Minutes that run on Fri / Sat / Sun (US Pacific), **nodes only** | Applied per minute, never raises a quoted rate | Fri 33% off (cpu-2 32%), Sat/Sun 20% off. **Sandboxes: none** | same |
| Batch jobs (H100) | `submit_job`, sweeps, tasks | Per GPU-minute. Queued time free. Preemption-tolerant (checkpoints survive) | $0.0495/min = **$2.97/GPU-h** (25% below interactive). Friday interactive ($2.65) beats it | same |
| Market: reserved | `sf buy`, orderbook limit order on an 8-GPU SKU (H100/H200/B200) and delivery window | Prepaid node-minutes for the whole window, whole minutes, min 1 node × 1 min. Includes boot/idle. Buyers pay no fees | **No list price.** Illustrations: homepage "Reserved $3.00/gpu/hr" (3-mo), docs example ~$17.42/node-h (~$2.18/GPU-h), old prices page "H100 from $1.50/gpu/hr" | https://docs.sfcompute.com/preview/orders, https://sfcompute.com/ |
| Market: resale | Selling unused reserved time | Fill price minus platform fee (fixed $/node-h + %, per org, unpublished), credited | Vendor reports a 25% average realized discount (2026-09-09): $4.50 → $3.375 realized | https://sfcompute.com/news/realized-discount |
| Market: spot deployment | Preview access | Immediate-or-cancel buys at or below max $/node-h. Pays the seller price for each contract block (min runtime 1 min-24 h). 2-min preemption notice | market price | https://docs.sfcompute.com/preview/spot-deployments |
| Bare metal / managed Slurm (InfiniBand) | Contact sales | Negotiated | null | https://sfcompute.com/specs |
| Egress / ingress | GPU clusters | "No ingress/egress fees" | $0 (sandbox egress not separately stated) | https://sfcompute.com/specs |
| Credit | All usage is prepaid | Buy $10-$10,000. Purchased credit valid 12 months, promo credit 3 weeks. Auto top-up default <$20 → +$100. Invoice accounts move to prepaid on **2026-10-01** | - | /llms.txt |
| Other meters | - | Builds $0.01/min (300 min free). Logs/traces $0.50/GB (5 GB free). Metrics $8/1k series (10k free). Node snapshots $0.10/GiB-mo (250 GiB free). Object storage 100 GB free then $0.10→$0.035/GB-mo tiers | - | https://autoresearch.sfcompute.com/ |

## Gotchas

1. **Sandboxes are priced on RAM only, with a 4 GiB/core floor.** CPU-heavy shapes are expensive: 4 vCPU / 8 GiB bills as 16 GiB = $0.73/h flat-out, 3.3× Isorun.
2. **Parking is the whole economics.** Agent loops that are idle 87% of the time pay about 13% of the flat-out figure, but only for memory actually handed back. A sandbox holding a loaded model keeps paying.
3. **Calendar discounts do not apply to sandboxes.** They apply to Autoresearch nodes only. Our earlier record assumed they might, and the live table shows "-" for sandbox rows.
4. **Idle windows on nodes**: a node keeps billing for 15-20 min after its last command unless you stop it.
5. **Stale docs**: llms.txt implies an H100 list price of $3.60/h ("$0.90 at the H100 list rate" for 15 min), and the CLI example shows $0.0666/min. The live table says $0.066/min = $3.96/h.
6. **Market GPUs have no price list.** Prices vary by SKU, zone and window, and are sold per 8-GPU node. Market VMs have no persistent storage and can take ~10 min to boot, and boot minutes are billed.
7. **Everything is prepaid**, and spend caps count credit-paid usage, so a cap can stop nodes while credit remains.

## Worked example

Workload: 4 vCPU / 8 GiB, 50 concurrent × 8 h/day × 22 days = **8,800 sandbox-hours**, 30% CPU, 50 GiB snapshots retained, 100 GiB egress.

- Sandbox-md flat-out (never parked): 8,800 × 16 GiB × $0.04572 = **$6,437.38**.
- Parked 70% of the time (using the 30% CPU figure as a proxy for "command in flight"): 2,640 h × $0.73152 = **$1,931.21**, plus a small parked-disk charge (unknown disk size, ~$0.000139/GiB-h).
- Snapshots: 50 GiB × $0.10 = **$5.00** resident ($2.50 cold). Egress: $0 (not charged per specs page; unverified for sandboxes).

| Regime | Feasible? | Monthly total |
|---|---|---|
| Sandbox-md, never parked | Yes (sandbox_vcpus limit starts at 4 per workspace, so 50 × 4 vCPU needs a raise) | **$6,442.38** |
| Sandbox-md, parked 70% | Yes | ≈ **$1,936** + parked disk |
| cpu-2 node (2 physical cores / 8 GiB), stopped between shifts, weekdays | Alt product (container node, not sandbox API) | 8,800 × $0.18 = $1,584. With ~4.4 Fridays at 32% off: **$1,482.62** |
| H100 / market GPU | Not a GPU workload | n/a |