{ "features_resolved": { "snapshot_any": true, "snapshot_mem": true, "fork": true, "pause_resume": true, "persistent_disk": true, "whole_disk": false, "persist_scope": true, "volumes": false, "auto_stop_idle": true, "long_sessions": true, "custom_image": true, "image_snapshot": true, "image_oci": true, "vm_isolation": true, "docker_inside": false, "nested_virt": false, "root": true, "systemd": null, "browser": false, "desktop": false, "computer_use": false, "code_interpreter": false, "browser_control": null, "desktop_control": null, "browser_and_desktop": false, "anti_bot": false, "captcha_solving": false, "residential_ip": false, "agent_harnesses": false, "ssh": false, "public_ipv4": false, "inbound_https": true, "custom_domain": false, "raw_tcp_inbound": false, "egress_allowlist": true, "open_internet": true, "static_egress_ip": false, "private_network": true, "self_host": false, "byoc": false, "open_source": false, "soc2": true, "hipaa": false, "sso": false, "gpu": true, "arm64": false, "windows": false, "macos": false, "lang_python": true, "lang_node": true, "lang_go": null, "lang_rust": null, "lang_java": null, "gpu_desktop": false, "credential_injection": null, "egress_http_rules": false, "wake_on_request": true, "live_resize": false, "fork_running": true, "memory_snapshot_fork": true, "webhooks": null, "scheduled_wakeups": null, "audit_logs": true, "eu_data_residency": true, "zero_data_retention": null, "scoped_api_keys": true, "spend_limits": true, "mcp_server": true, "usage_api": true, "snapshot_auto": false, "snapshot_on_demand": true, "harness_api": false, "own_agent_api": true, "fast_boot": true, "strong_isolation": true, "ingress_rules": true, "guest_firewall": true, "secret_proxy": false, "secret_proxy_any": false, "volume_attach": false, "volume_shared": false, "sidecar": null, "fuse": true, "kernel_features": true, "burst": null }, "company": { "id": "sfcompute", "checked": "2026-09-30", "company": "The San Francisco Compute Company (SF Compute)", "ownership": { "value": "independent", "parent": null, "source": "https://www.datacenterdynamics.com/en/news/sf-compute-raises-40m-for-ai-compute-marketplace-offering/", "quote": "Founded in 2023, the company recently hired Eric Park, former CEO of AI cloud provider Voltage Park, as CTO." }, "stage": { "value": "series-a", "source": "https://www.datacenterdynamics.com/en/news/sf-compute-raises-40m-for-ai-compute-marketplace-offering/", "quote": "SF Compute has secured $40 million in a Series A equity financing round for the expansion of its AI compute marketplace.", "confidence": "high", "note": "No accelerator found for SF Compute itself; searches of the YC company directory turned up no match (an earlier Evan Conrad startup, not SF Compute, had informal ties to Y Combinator people — rejected as not applicable to this provider)." }, "accelerators": [], "total_funding_usd": { "value": 52000000, "status": "confirmed", "basis": "Sum of the $12M seed (company's own LinkedIn/hiring post, Jan 2025) and the $40M Series A (WSJ/DataCenterDynamics, Nov 2025). Note: Tracxn's aggregate figure states total funding of only \"$40M over 2 rounds,\" which appears to undercount the seed and disagrees with the primary-source seed amount; we use the sum of the two independently confirmed rounds.", "source": "https://www.linkedin.com/posts/andyrapista_were-hiring-at-san-francisco-compute-company-activity-7287572464002314240-SdKr", "quote": "We raised a $12 million seed round led by Jack Altman (Alt Capital) with participation from Akshay Kothari (Co-founder & COO, Notion), Christina Cacioppo (Co-founder & CEO, Vanta), Terrence Rohan (Otherwise Fund), Noah Smith, and many others." }, "last_round": { "type": "series-a", "amount_usd": 40000000, "date": "2025-11", "leads": [ "DCVC", "Wing Venture Capital" ], "status": "confirmed", "source": "https://www.datacenterdynamics.com/en/news/sf-compute-raises-40m-for-ai-compute-marketplace-offering/", "quote": "The funding round was led by DCVC and Wing Venture Capital. It also saw participation from existing backers, including Electric Capital and Alt Capital." }, "investors": [ "DCVC", "Wing Venture Capital", "Electric Capital", "Alt Capital (Jack Altman)", "Akshay Kothari", "Christina Cacioppo", "Terrence Rohan (Otherwise Fund)", "Noah Smith" ], "portfolio_listings": [ { "source": "Electric Capital", "company": "SF Compute", "batch": null, "how": "domain" } ], "revenue": { "value": null, "status": "unknown" }, "employees": { "value": 30, "source": "https://www.datacenterdynamics.com/en/news/sf-compute-raises-40m-for-ai-compute-marketplace-offering/" }, "note": "Independent, venture-backed startup founded in 2023 by Evan Conrad (CEO) and co-founders including Ethan W. Anderson (COO); two products under one brand — the SF Compute GPU orderbook/marketplace and the newer Autoresearch sandbox/node platform this card's pricing centers on. Series A (Nov 2025) valued the company at $300M post-money per WSJ. No Y Combinator or other accelerator affiliation found for SF Compute itself." }, "id": "sfcompute", "name": "SF Compute (GPU market + Autoresearch sandboxes)", "url": "https://autoresearch.sfcompute.com/", "category": "gpu-cloud", "isolation": "vm", "modes": [ { "key": "sandbox", "label": "Autoresearch RL sandbox (billed per GiB of RAM, auto-park)", "flags": [], "os": [ "linux" ], "regions": null, "gpu_only": false, "features": { "snapshot": "mem", "fork": true, "pause_resume": true }, "pricing": "sizes", "vcpu_h": null, "ram_gib_h": null, "cpu_basis": "alloc", "ram_basis": "alloc", "active_floor": 0, "ram_per_vcpu": [ null, null ], "ram_fixed_per_vcpu": null, "vcpu_options": null, "min_vcpu": null, "max_vcpu": null, "max_ram_gib": null, "sizes": [ { "name": "sandbox-sm", "vcpu": 1, "ram_gib": 2, "disk_gib": null, "hour": 0.09144, "month_cap": null }, { "name": "sandbox-md", "vcpu": 4, "ram_gib": 8, "disk_gib": null, "hour": 0.73152, "month_cap": null }, { "name": "sandbox-lg", "vcpu": 8, "ram_gib": 32, "disk_gib": null, "hour": 1.46304, "month_cap": null }, { "name": "sandbox-xl", "vcpu": 16, "ram_gib": 64, "disk_gib": null, "hour": 2.92608, "month_cap": null } ], "pool_tiers": null, "requires_always_on": false, "always_on_month_per_instance": null, "commit_note": null, "min_billed_seconds": 0, "granularity_s": 60, "start_fee": 0, "boot_overhead_s": null, "multiplier": 1, "windows_vcpu_h": null, "gpu": null, "note": "$0.04572 per GiB-hour of billed RAM, per minute, while RUNNING; each core comes with 4 GiB so sandbox-md (4 vCPU/8 GiB) bills as 16 GiB = $0.73152/h. Cores are shared. Sandbox parks after ~1.5 s idle, handing back spare memory, which then bills only at the parked ~$0.000139/GiB-h disk rate (use the running-share knob). Forks free (CoW); baking bills as sandbox time. sandbox_vcpus limit starts at 4 (lg/xl need a raise); workspace RAM ceiling 8 GiB by default (64 GiB platform cap). Calendar discounts do NOT apply (pricing table shows none for sandboxes). [basis verified 2026-09-28] Docs: \"a sandbox bills per GiB of its RAM per minute while it is RUNNING\"; \"A core comes with 4 GiB\"; \"A sandbox with no command in flight PARKS itself after ~1.5s idle: it hands back the memory its guest can spare, so you stop paying the RAM rate on what was handed back\" (autoresearch.sfcompute.com/llms.txt). Billed RAM = configured size (>= 4 GiB/core) while running; kept sizes + sandbox_running_share knob for parking.", "cpu_class": "shared" }, { "key": "cpu-node", "label": "Autoresearch CPU node (cpu-2 / cpu-8 container, persistent disk)", "flags": [ "alt" ], "os": [ "linux" ], "regions": null, "gpu_only": false, "features": { "snapshot": "fs", "fork": false, "pause_resume": true, "persistent_disk": true, "gpu": false, "fork_running": false, "memory_snapshot_fork": false }, "pricing": "sizes", "vcpu_h": null, "ram_gib_h": null, "cpu_basis": "alloc", "ram_basis": "alloc", "active_floor": 0, "ram_per_vcpu": [ null, null ], "ram_fixed_per_vcpu": null, "vcpu_options": null, "min_vcpu": null, "max_vcpu": null, "max_ram_gib": null, "sizes": [ { "name": "cpu-2", "vcpu": 4, "ram_gib": 8, "disk_gib": null, "hour": 0.18, "month_cap": null }, { "name": "cpu-8", "vcpu": 16, "ram_gib": 32, "disk_gib": null, "hour": 0.72, "month_cap": null } ], "pool_tiers": null, "requires_always_on": false, "always_on_month_per_instance": null, "commit_note": null, "min_billed_seconds": 0, "granularity_s": 60, "start_fee": 0, "boot_overhead_s": null, "multiplier": 1, "windows_vcpu_h": null, "gpu": null, "note": "cpu-2 = 2 reserved PHYSICAL cores + 8 GiB ($0.003/min = $0.18/h); cpu-8 = 8 physical cores on one NUMA node + 32 GiB ($0.012/min = $0.72/h; the earlier \"33% off (CPU node launch pricing)\" $0.00804/min rate is gone from the live table, checked 2026-10-14). vcpu encoded as 2 x physical cores (assumption). A node keeps billing through a 15-min idle window after its last command unless stopped ($0.045 / $0.18 at most); stopped nodes park their disk and cost nothing. Friday 32% (cpu-2) / 33% (cpu-8), weekend 20% off per calendar knob. Container nodes for batch/ingest work, not a sandbox API.", "cpu_class": "dedicated" }, { "key": "h100-node", "label": "Autoresearch interactive H100 node, on demand (h100-1 = 1 GPU $5.04/h; h100-8 = 8 GPUs)", "flags": [ "stock" ], "os": [ "linux" ], "regions": null, "gpu_only": true, "features": { "gpu": true, "persistent_disk": true, "pause_resume": true, "snapshot": "fs", "fork": false, "fork_running": false, "memory_snapshot_fork": false }, "pricing": "resource", "vcpu_h": 0, "ram_gib_h": 0, "cpu_basis": "alloc", "ram_basis": "alloc", "active_floor": 0, "ram_per_vcpu": [ null, null ], "ram_fixed_per_vcpu": null, "vcpu_options": null, "min_vcpu": null, "max_vcpu": null, "max_ram_gib": null, "sizes": null, "pool_tiers": null, "requires_always_on": false, "always_on_month_per_instance": null, "commit_note": null, "min_billed_seconds": 0, "granularity_s": 60, "start_fee": 0, "boot_overhead_s": null, "multiplier": 1, "windows_vcpu_h": null, "gpu": { "H100": 5.04 }, "note": "Per GPU: h100-1 $0.084/min = $5.04/h; h100-8 $0.672/min = $40.32/h (= $5.04/GPU-h). CPU/RAM included. Idle window keeps billing 15 min (1x, $1.26 max) / 20 min (8x, $13.44 max) after the last command unless stopped; no minimum charge. Stopped node parks its encrypted disk at $0. Friday 33% / weekend 20% off (calendar knob), applied to minutes run on those days (US Pacific). [live table checked 2026-10-14: \"h100-1 $0.084/min ($5.04/hr)\", \"h100-8 $0.672/min ($40.32/hr)\"]", "stock_note": "queued provisioning: create_node joins a queue and quotes an ETA; not an instant-capacity guarantee", "cpu_class": "shared" }, { "key": "batch-h100", "label": "Autoresearch batch jobs on H100 (queued, preemptible)", "flags": [ "alt", "spot" ], "os": [ "linux" ], "regions": null, "gpu_only": true, "features": { "gpu": true }, "pricing": "resource", "vcpu_h": 0, "ram_gib_h": 0, "cpu_basis": "alloc", "ram_basis": "alloc", "active_floor": 0, "ram_per_vcpu": [ null, null ], "ram_fixed_per_vcpu": null, "vcpu_options": null, "min_vcpu": null, "max_vcpu": null, "max_ram_gib": null, "sizes": null, "pool_tiers": null, "requires_always_on": false, "always_on_month_per_instance": null, "commit_note": null, "min_billed_seconds": 0, "granularity_s": 60, "start_fee": 0, "boot_overhead_s": null, "multiplier": 1, "windows_vcpu_h": null, "gpu": { "H100": 3.78 }, "note": "$0.063/min per GPU = $3.78/GPU-h, \"25% below interactive\" (live table, checked 2026-10-14). Queued time free; checkpoints survive preemption. Calendar discount reaches jobs only when cheaper still (Friday interactive $3.38 < $3.78 moves jobs; weekends $4.03 do not - matches llms.txt \"At today's numbers Friday moves a job and the weekend does not\"). Batch jobs capped at 85% of fleet.", "cpu_class": "shared" }, { "key": "reserved", "label": "SF Compute market: reserved 8-GPU node time (orderbook buy)", "flags": [], "os": [ "linux" ], "regions": [ "us", "eu" ], "gpu_only": true, "features": { "snapshot": "none", "pause_resume": false, "fork": false, "persistent_disk": false, "ssh": true, "public_ipv4": true, "raw_tcp_inbound": true, "inbound_https": false, "private_network": true, "docker_inside": true, "root": true, "systemd": true, "gpu": true, "fork_running": false, "memory_snapshot_fork": false }, "pricing": "resource", "vcpu_h": 0, "ram_gib_h": 0, "cpu_basis": "alloc", "ram_basis": "alloc", "active_floor": 0, "ram_per_vcpu": [ null, null ], "ram_fixed_per_vcpu": null, "vcpu_options": null, "min_vcpu": null, "max_vcpu": null, "max_ram_gib": null, "sizes": null, "pool_tiers": null, "requires_always_on": false, "always_on_month_per_instance": null, "commit_note": "You buy a node x time window up front on the orderbook (whole minutes, min 1 node x 1 minute, 8 GPUs per node); prepaid credit must cover the full contract. Unused time can be resold (seller pays a fixed per-node-hour + percentage platform fee; buyers pay no fees). SFC reports a 25% average realized discount via resale (2026-09-09).", "min_billed_seconds": 60, "granularity_s": 60, "start_fee": 0, "boot_overhead_s": null, "multiplier": 1, "windows_vcpu_h": null, "gpu": { "H100": null, "H200": null, "B200": null }, "note": "No fixed list price: price per node-hour is set by the orderbook per SKU and delivery window, so GPU rates are null. Illustrative official figures only: homepage resale widget \"Reserved price $3.00/gpu/hr\" (3-month example); docs orderbook example ~$17.42/node-h (~$2.18/GPU-h); old.sfcompute.com/prices \"H100 nodes from $1.50/gpu/hr\". Billed per node-minute including boot/idle. VMs (bare metal / managed Slurm via sales). No ingress/egress fees. Instances have no persistent storage.", "gpu_min_count": 8, "cpu_class": "dedicated" }, { "key": "spot", "label": "SF Compute market: spot deployment (auto-buys at or below max rate)", "flags": [ "spot", "beta" ], "os": [ "linux" ], "regions": [ "us", "eu" ], "gpu_only": true, "features": { "snapshot": "none", "pause_resume": false, "fork": false, "persistent_disk": false, "ssh": true, "public_ipv4": true, "raw_tcp_inbound": true, "inbound_https": false, "private_network": true, "docker_inside": true, "root": true, "systemd": true, "gpu": true, "fork_running": false, "memory_snapshot_fork": false }, "pricing": "resource", "vcpu_h": 0, "ram_gib_h": 0, "cpu_basis": "alloc", "ram_basis": "alloc", "active_floor": 0, "ram_per_vcpu": [ null, null ], "ram_fixed_per_vcpu": null, "vcpu_options": null, "min_vcpu": null, "max_vcpu": null, "max_ram_gib": null, "sizes": null, "pool_tiers": null, "requires_always_on": false, "always_on_month_per_instance": null, "commit_note": null, "min_billed_seconds": 60, "granularity_s": 60, "start_fee": 0, "boot_overhead_s": null, "multiplier": 1, "windows_vcpu_h": null, "gpu": { "H100": null, "H200": null, "B200": null }, "note": "Spot is a buying strategy on the same orderbook (immediate-or-cancel buys at or below --max-rate $/node-hour); you pay the matched seller price per node-minute for the whole contract block (min runtime 1 min-24 h), including boot/idle minutes. 2-minute preemption notice via metadata endpoint (none if min runtime <= 2 min); preemption adds no charge. Spot deployments are in public preview (available to everyone, best-effort support, no SLA). Prices market-set, so null.", "gpu_min_count": 8, "cpu_class": "dedicated" } ], "storage": { "included_disk_gib": null, "disk_gib_month": null, "disk_billed_when_stopped": true, "snapshot_gib_month": 0.1, "paused_ram_billed": false }, "network": { "egress_free_gib": null, "egress_gib": 0, "ipv4_month": null, "ipv4_available": true, "internet": { "default": "configurable", "evidence": [ { "quote": "A sandbox CAN reach the internet: git clone, pip install, DNS and outbound HTTPS all work, up to 1 Gbps per sandbox.", "url": "https://autoresearch.sfcompute.com/llms.txt" }, { "quote": "pass egress: 'none' to create_sandbox_env / create_sandbox: no network device at all", "url": "https://autoresearch.sfcompute.com/llms.txt" } ] } }, "plans": [ { "name": "Prepaid credit (pay as you go)", "fee": 0, "flags": [], "fee_is_credit": false, "included_usd": 0, "per_seat": 0, "concurrency": null, "max_session_h": null, "max_vcpu": null, "max_ram_gib": null, "sizes_allowed": null, "note": "No plan fee. Credit bought $10-$10,000 (whole cents), valid 12 months; promotional credit expires after 3 weeks; auto top-up default \"below $20 add $100\". Card never charged for usage after the fact; work parks/stops at $0. Monthly-invoice accounts move to prepaid credit on 2026-10-01. Spend caps count credit-paid usage." }, { "name": "Enterprise / bare metal / managed Slurm", "fee": null, "flags": [ "sales" ], "fee_is_credit": null, "included_usd": null, "per_seat": null, "concurrency": null, "max_session_h": null, "max_vcpu": null, "max_ram_gib": null, "sizes_allowed": null, "note": "Contact sales: bare metal, managed Slurm with InfiniBand, dedicated solutions engineer." } ], "extras": [ { "item": "Job builds", "price": "$0.01/min, 300 min/month free" }, { "item": "Logs / traces", "price": "$0.50/GB ingested, 5 GB/month free each" }, { "item": "Metrics", "price": "$8 per 1,000 active series/month, first 10,000 free" }, { "item": "Node snapshot storage", "price": "$0.10/GiB-month, 250 GiB free" }, { "item": "Object storage", "price": "first 100 GB free, $0.10/GB-month to 10 TB, $0.07 to 100 TB, $0.05 to 1 PB, $0.035 beyond (prorated daily)" }, { "item": "Market resale platform fee", "price": "fixed $/node-hour + % of sale value, set per org, unpublished" }, { "item": "Autoresearch managed researcher (private preview)", "price": "intelligence units $0.01 (i1) / $0.03 (i2) / $0.06 (i3) per unit; experiment jobs at job rates" } ], "free": { "monthly_credit": 0, "one_time_credit": 0, "note": "Referral: a friend who signs up via a referral link and adds a card gets $50 of credit (promotional credit expires 3 weeks after grant). Conditional, so one_time_credit stays 0." }, "features": { "linux": true, "windows": false, "macos": false, "arm64": false, "gpu": true, "gpu_types": [ "H100", "H200", "B200", "B300" ], "snapshot": "mem", "pause_resume": true, "fork": true, "persistent_disk": true, "volumes": false, "auto_stop_idle": true, "max_session_h": null, "custom_image": true, "docker_inside": false, "nested_virt": false, "root": true, "systemd": null, "browser": false, "desktop": false, "computer_use": false, "code_interpreter": false, "agent_harnesses": false, "ssh": false, "public_ipv4": false, "inbound_https": true, "custom_domain": false, "raw_tcp_inbound": false, "egress_allowlist": true, "static_egress_ip": false, "private_network": true, "self_host": false, "byoc": false, "open_source": false, "regions": [ "us", "eu" ], "soc2": true, "hipaa": false, "sso": false, "lang_python": true, "lang_node": true, "lang_go": null, "lang_rust": null, "lang_java": null, "gpu_desktop": false, "credential_injection": null, "egress_http_rules": false, "wake_on_request": true, "live_resize": false, "fork_running": true, "memory_snapshot_fork": true, "webhooks": null, "scheduled_wakeups": null, "audit_logs": true, "eu_data_residency": true, "zero_data_retention": null, "scoped_api_keys": true, "spend_limits": true, "mcp_server": true, "snapshot_auto": false, "snapshot_on_demand": true, "own_agent_api": true, "harness_api": false, "ingress_rules": true, "guest_firewall": true, "secret_proxy": false, "volume_attach": false, "volume_shared": false, "isolation": "firecracker", "image_oci": true, "anti_bot": false, "captcha_solving": false, "egress_ip_type": "datacenter", "max_session_unlimited": true, "fast_boot": true, "open_internet": true, "usage_api": true, "persist_scope": "folder", "persist_path": "the home folder" }, "perf": { "cpu_runs_s": null, "cold_start_ms": null, "burst_tti_ms": 362.87, "dax_s": 44.36 }, "perf_source": "computesdk (repo key \"givemeanode\" = Autoresearch sandboxes): burst_tti median 362.87 ms at 100 concurrent (2026-09-25), DAX total 44.36 s (2026-09-25, 8 vCPU / 31.4 GiB x86_64 env). No cold_start_tti row, no hpc CPU run. GPU VM boot (up to ~10 min) not benchmarked.", "knobs": [ { "id": "sandbox_running_share", "label": "Share of sandbox time not parked (holding RAM)", "kind": "slider", "min": 0.05, "max": 1, "step": 0.05, "default": 1, "unit": "%", "options": null, "effect": "hours_factor", "modes": [ "sandbox" ], "why": "Autoresearch sandboxes park after ~1.5 s idle and stop paying the $0.04572/GiB-h RAM rate on handed-back memory (parked disk ~$0.000139/GiB-h, ~330x cheaper). Vendor example: 13% tool use / 87% thinking. Only memory actually handed back is discounted." }, { "id": "calendar_discount", "label": "Which days the work runs (US Pacific calendar discount)", "kind": "select", "min": null, "max": null, "step": null, "default": 1, "unit": null, "options": [ { "value": 1, "label": "Mon-Thu only (no calendar discount)" }, { "value": 0.934, "label": "22 weekdays incl. ~4.4 Fridays at 33% off (x0.934)" }, { "value": 0.895714, "label": "Spread evenly over 7 days (Fri 33%, Sat/Sun 20% off, x0.896)" }, { "value": 0.8, "label": "Weekends only (x0.80)" }, { "value": 0.67, "label": "Fridays only (x0.67)" } ], "effect": "rate_factor", "modes": [ "cpu-node", "h100-node" ], "why": "Autoresearch nodes: Fridays 33% off (cpu-2: 32%), Sat/Sun 20% off, applied to the minutes that run on those days. Not applied to sandboxes. Batch jobs only move on Fridays at current rates (not modeled; batch-h100 not in this knob)." } ], "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 rose to $5.04/GPU-h ($0.084/min) on the live pricing table (was $3.96, $0.066/min); batch jobs to $3.78 ($0.063/min) and cpu-8 to $0.72/h ($0.012/min, launch discount ended). llms.txt now agrees (\"$13.44 at the H100 list rate\" for the h100-8 20-min window) and the home-page CLI sample shows \"tulip running, $0.084/min\". (checked 2026-10-14)", "cpu-node vcpu = 2 x physical cores (assumption).", "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.", "Fork fan-out: the home page says \"One fork per sample, up to a thousand at once. Forking is free. 50ms per fork, 100ms per delete.\", while llms.txt still documents fork_sandbox as \"clone 1..=256 running sandboxes\" per call - likely ~1000 concurrent forks reached over several calls. Not a priced field. (checked 2026-10-14)" ], "sources": [ "https://autoresearch.sfcompute.com/", "https://autoresearch.sfcompute.com/llms.txt", "https://autoresearch.sfcompute.com/docs", "https://sfcompute.com/", "https://sfcompute.com/specs", "https://sfcompute.com/news/realized-discount", "https://old.sfcompute.com/prices", "https://docs.sfcompute.com/llms-full.txt", "https://docs.sfcompute.com/preview/spot-deployments", "https://docs.sfcompute.com/preview/orders", "https://docs.sfcompute.com/preview/guides/reselling", "https://www.computesdk.com/benchmarks/sandboxes/", "https://www.spheron.network/blog/spheron-vs-sf-compute/" ], "fleet": { "default_max_instances": null, "default_max_vcpu": null, "default_max_gpus": null, "raise": "self-serve", "raise_time_days": 1, "verification": "Billing by saved card or prepaid credits; org admin must submit limit increase with reason.", "tiers": "Autoresearch live limits are readable through list_limits/autor limits. Docs illustrate default 64 GPU nodes and 256 CPU nodes per member, not aggregate organization entitlement.", "notes": "Requests up to raisable_to apply immediately; larger requests require explicit escalation and operator review, usually within one business day, no SLA. Scalar time describes human review, not self-serve. Organization spending cap and hard_max also apply; GPU node count is not GPU count.", "sources": [ "https://autoresearch.sfcompute.com/docs" ], "researched": "2026-09-28" }, "audit_2026_09_29": { "applied": 6, "source": "checked against the provider's live pricing and docs pages", "reviewed": 5 }, "feature_evidence": { "snapshot_auto": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "your persistent disk mounts at the `home` parameter (an absolute path; default the image's WORKDIR, else /root), and ONLY that path survives stops", "note": "Autoresearch nodes auto-stop after an idle window and the disk parks (persistent disk, not a state snapshot): 'Packages install outside /home/dev, so a stop resets them'. The platform may move a parked disk into a stored copy for placement, still disk-only. Sandboxes auto-PARK after ~1.5 s idle but that only hands back spare memory; 'Processes you left running keep running' on the host (no checkpoint/restore). Market VMs: 'Instances have no persistent storage.'", "checked": "2026-09-29" }, "snapshot_on_demand": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "snapshot_sandbox(sandbox): capture a RUNNING sandbox's current state as a new forkable snapshot, without stopping it", "note": "Sandbox snapshots include memory ('snapshots the fully initialized machine - memory, processes, page cache, disk') and fork_sandbox clones 1..256 running sandboxes from one, each resuming exact state. Nodes: snapshot_node is disk-only, needs the node stopped, and create_node from_snapshot boots new nodes from it (compatible node type/image). Market VMs: custom images only, no snapshots.", "checked": "2026-09-29" }, "own_agent_api": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "steer_research(name, expected_version, ...): one action per call. The action is a plan, an instruction to the researcher, a question_id + answer, a planner handoff", "note": "Provider's own agent only: the managed 'researcher' (private preview, off by default) is started with start_research(name, objective, ...), polled with get_research (state, open question, full record of plans/observations/decisions), steered with follow-up instructions, stopped and listed; HTTP /preview/research and `autor research start|get|steer|stop|ls`. It runs platform-side and launches experiment jobs, not inside a user sandbox. No API to drive Claude Code/Codex/OpenCode inside a sandbox (`autor mcp install` only wires those local agents to the platform MCP server).", "checked": "2026-09-29" }, "harness_api": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "steer_research(name, expected_version, ...): one action per call. The action is a plan, an instruction to the researcher, a question_id + answer, a planner handoff", "note": "Provider's own agent only: the managed 'researcher' (private preview, off by default) is started with start_research(name, objective, ...), polled with get_research (state, open question, full record of plans/observations/decisions), steered with follow-up instructions, stopped and listed; HTTP /preview/research and `autor research start|get|steer|stop|ls`. It runs platform-side and launches experiment jobs, not inside a user sandbox. No API to drive Claude Code/Codex/OpenCode inside a sandbox (`autor mcp install` only wires those local agents to the platform MCP server).", "checked": "2026-09-29" }, "egress_allowlist": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "For a HERMETIC sandbox, pass egress: \"none\" to create_sandbox_env / create_sandbox: no network device at all, so nothing inside can reach anything.", "note": "Binary per-sandbox posture only: egress \"open\" (NAT'd internet) or \"none\" (no NIC), fixed at bake time and inherited by forks. No domain or IP allowlists. Nodes/jobs: default open internet with a fixed platform blocklist (SMTP 25, mining-pool ports, cloud metadata IP). Market VM firewalls are ingress-only: 'Firewalls do not filter outbound traffic' (egress rules 'reserved for future use').", "checked": "2026-09-29" }, "egress_http_rules": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "DNS is unfiltered and TLS is never intercepted", "note": "Node/job egress policy is port/IP-level only; sandbox egress is open/none. No HTTP method/path/header rules documented anywhere.", "checked": "2026-09-29" }, "ingress_rules": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "The URL IS the secret (an unguessable hostname): treat it like a password, or pass auth: \"bearer\" for a token on top, returned ONCE in that response.", "note": "Sandboxes: per-port expose/unexpose (ports 1024-65535), unguessable HTTPS URL plus optional bearer token. Nodes: no public address by default, expose_port gives a public HTTPS URL. Market VMs: public IPv4 only if enabled at create, workspace firewalls with allow-only ingress rules (protocol, port range, source CIDR), https://docs.sfcompute.com/preview/networking/firewalls: 'Each rule specifies a direction, protocol, port (or port range), and source CIDR'.", "checked": "2026-09-29" }, "guest_firewall": { "value": true, "source": null, "quote": null, "note": "Sandboxes are microVMs with their own guest kernel and commands run as root; market instances are full GPU VMs with root. Autoresearch nodes are containers (dev user with passwordless sudo); CAP_NET_ADMIN there is not documented.", "checked": "2026-09-29", "derived": true }, "secret_proxy": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "secrets go in env and file bytes go in write_file, never the command string", "note": "Secrets for running code are env vars on run_command (encrypted at rest, redacted in logs). Org 'connections' (HF, S3 keys, registries) are used host-side for import_data/export_data and image pulls: 'A connection's credential is exercised on the host - it never enters the node'. That is a managed data-transfer path, not a proxy injecting credentials into the machine's own outbound requests.", "checked": "2026-09-29" }, "volume_attach": { "value": false, "source": "https://docs.sfcompute.com/preview/instances", "quote": "Instances have no persistent storage.", "note": "Market VMs have only local disk. Autoresearch nodes get one fixed persistent volume (org node_volume_gib, 250 GiB default) and 'you cannot change the size of the node's own disk or its scratch after create'; no extra attachable volumes. Object storage buckets exist but are copied in with import_data, not mounted.", "checked": "2026-09-29" }, "volume_shared": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "hf_cache is a per-MACHINE shared cache at HF_HUB_CACHE (best-effort, LRU-evicted)", "note": "No shared filesystem between machines: files move between nodes via export_data/import_data or copy_file; 'there is no node-to-node path by default'. The only shared mount is the platform-managed Hugging Face cache per host for jobs.", "checked": "2026-09-29" }, "isolation": { "value": "firecracker", "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "Boots a small isolated microVM from a curated image (\"sbx-base\"...)", "note": "Autoresearch (the sandbox/node product this card centers on) explicitly calls its per-sample compute an isolated 'microVM' ('its own forked microVMs'), the industry term almost exclusively associated with Firecracker-class microVM hypervisors; the underlying hypervisor is never named explicitly, so this is inferred rather than confirmed. The separate SF Compute market product uses full bare-metal/VM instances instead (qemu-kvm-class), per sfcompute.com/specs: 'Bare metal and managed Slurm for the lowest-level access, and VMs for flexibility.'", "checked": "2026-09-30" }, "image_oci": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "from_image takes a container image of YOUR own instead of a curated name - an RL task image, a framework's official image, a SWE-bench instance - so an environment that already exists as a Docker image is imported rather than re-expressed as a warm script. Digest-pinned (\"ghcr.io/acme/task@sha256:<64 hex>\")", "note": "Applies to sandboxes (from_image) and nodes (image takes 'YOUR OWN container image as a registry ref'). linux/amd64 only.", "checked": "2026-09-30" }, "docker_inside": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "A node is a container on an H100 with a persistent, encrypted disk... Commands run as the dev user with passwordless sudo", "note": "not offered: nodes are themselves containers (shared kernel) and sandboxes are microVMs where commands run as root through your own shell; nowhere does the 3000+ line API/CLI reference mention a Docker/container runtime, privileged mode, or dockerd available inside a node or sandbox. Derived from the absence of any such capability in an otherwise exhaustive doc.", "checked": "2026-09-30" }, "nested_virt": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "A node is a container on an H100... Boots a small isolated microVM from a curated image", "note": "not offered: nodes are containers and sandboxes are microVMs; no /dev/kvm, nested virtualization, or emulator capability is mentioned anywhere in the API/CLI docs. Derived from absence of any such capability.", "checked": "2026-09-30" }, "anti_bot": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "On-demand GPU nodes and asynchronous GPU jobs for AI agents... Use Autoresearch when a task needs a real GPU machine and there is no cluster to put it on.", "note": "not offered: outside the product's scope. SF Compute/Autoresearch is a GPU-compute and RL-sandbox platform, not a browser-automation or scraping product; no stealth/fingerprinting feature exists.", "checked": "2026-09-30" }, "captcha_solving": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "On-demand GPU nodes and asynchronous GPU jobs for AI agents... Use Autoresearch when a task needs a real GPU machine and there is no cluster to put it on.", "note": "not offered: outside the product's scope. There is no browser or CAPTCHA-solving capability in this GPU-cloud/sandbox API.", "checked": "2026-09-30" }, "egress_ip_type": { "value": "datacenter", "source": "https://docs.sfcompute.com/preview/networking/ip-addresses", "quote": "SF Compute assigns the public IP automatically and releases it when you delete the instance. No two instances share an IP.", "note": "Market VM instances get a dedicated (non-shared) but ephemeral datacenter IPv4, assigned at create and released on delete - not a fixed/reserved address you keep. Autoresearch sandboxes/nodes reach the internet outbound from shared host NAT (no per-machine public IP for outbound); no residential/mobile IP option exists anywhere in the docs.", "checked": "2026-09-30" }, "fork_running": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "fork_sandbox(from, count?): clone 1..=256 running sandboxes from a snapshot, each resuming its exact state. Forks are FREE (copy-on-write)", "note": "Applies to the Autoresearch sandbox product; the sandbox that is forked continues running unaffected while forks are created (snapshot_sandbox captures a RUNNING sandbox 'without stopping it'). Node and market-VM modes already carry fork_running:false in the card (no fork capability there).", "checked": "2026-09-30" }, "memory_snapshot_fork": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "snapshots the fully initialized machine - memory, processes, page cache, disk. Returns a snapshot id to fork forever... fork_sandbox(from, count?): clone 1..=256 running sandboxes from a snapshot, each resuming its exact state.", "note": "Sandbox snapshots/forks include memory and running processes, resuming exact state - a true memory-snapshot fork. Node/market-VM modes are disk-only snapshots (already false in the card for those modes).", "checked": "2026-09-30" }, "wake_on_request": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "stop_node parks the disk with files intact and the next run_command wakes it.", "note": "Also true for sandboxes: 'an inbound request wakes a parked sandbox, so traffic puts it back on the active RAM tier' (expose_sandbox_port section) - both a wake-on-command (node) and wake-on-inbound-HTTP-request (sandbox) path exist.", "checked": "2026-09-30" }, "live_resize": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "The size is fixed AT THE BAKE and every fork inherits it, so bake the shape you mean to fork.", "note": "Sandbox vCPU/RAM shape is fixed at bake time; nodes have a fixed chip/cgroup budget for their lifetime (create_node's resources), with no documented live-resize API for a running node or sandbox.", "checked": "2026-09-30" }, "lang_python": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "Boots a small isolated microVM from a curated image (\"sbx-base\", the default: python3, node 24, git, curl, a compiler; ... \"sbx-task\": python3 + numpy)", "note": "Python3 ships in the default sandbox image and node images ('image_nvcc', torch-enabled catalog images); PYTHON_CPU_COUNT/os.cpu_count() env handling throughout the docs also assumes Python is the primary language.", "checked": "2026-09-30" }, "lang_node": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "Boots a small isolated microVM from a curated image (\"sbx-base\", the default: python3, node 24, git, curl, a compiler...)", "note": "Node 24 ships in the default sandbox image (sbx-base); npm/bun package installs and an npm pull-through cache are documented for sandboxes.", "checked": "2026-09-30" }, "inbound_https": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "expose_sandbox_port(sandbox, port, auth?)... a PUBLIC HTTPS URL for a server running inside a sandbox - the preview link for a dev server... auth: \"bearer\" for a token on top", "note": "Sandboxes and nodes both support expose_port giving an unguessable HTTPS URL (optionally bearer-token protected) to reach a listening port; streaming/SSE/WebSocket pass through.", "checked": "2026-09-30" }, "static_egress_ip": { "value": false, "source": "https://docs.sfcompute.com/preview/networking/ip-addresses", "quote": "SF Compute assigns the public IP automatically and releases it when you delete the instance.", "note": "not offered: public IPv4 (market VMs only) is allocated per instance and released on delete, not a reserved/static address you keep across instances; Autoresearch sandboxes/nodes have no public IP at all for outbound and no static-egress-IP feature is documented anywhere.", "checked": "2026-09-30" }, "open_source": { "value": false, "source": "https://sfcompute.com/", "quote": "GPU clusters you can sublease", "note": "not offered: outside the product's scope. SF Compute/Autoresearch is a closed, proprietary commercial cloud platform (CLI/API + web console); no OSI-licensed core product, and no GitHub org shipping the platform itself was found.", "checked": "2026-09-30" }, "eu_data_residency": { "value": true, "source": "https://docs.sfcompute.com/llms-full.txt", "quote": "sf infiniband-partitions create --name training --fabric europe-north1-a.fab1", "note": "SF Compute market VMs/reserved capacity are offered in an EU region (europe-north1 / 'EU North'), matching the card's regions:[\"us\",\"eu\"] for the reserved/spot modes; a snapshot's region also identifies the datacenter it is stored in, so EU-region workloads can keep storage in-region too.", "checked": "2026-09-30" }, "max_session_h": { "value": "none", "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "stop_node parks the disk with files intact and the next run_command wakes it... A node that stays stopped with no activity for 60 days is deleted with its disk", "note": "No maximum runtime is imposed on a running node, sandbox, or market VM - a session runs until stopped/deleted by the customer or a health/idle event, not a platform-enforced session ceiling. (Idle nodes auto-stop after a short grace window, and stopped-and-abandoned disks are deleted after 60 days, but neither caps how long an ACTIVE session may run.)", "checked": "2026-09-30" }, "ssh": { "value": false, "source": "https://autoresearch.sfcompute.com/docs", "quote": "No inbound by default. Nodes have no public address and no open ports. There is no SSH. The one way in is an exposed port: HTTPS at our edge, per port, with the URL or a bearer token as the credential.", "note": "Confirmed for the Autoresearch nodes/sandboxes (the card's flagship 'stock' mode, h100-node): exec is run_command over the API only, no SSH. The separate SF Compute GPU-market VMs (reserved/spot modes) DO offer SSH (docs.sfcompute.com: 'sf instances ssh ') and already carry ssh:true on those two modes; the top-level features.ssh models the primary Autoresearch product, so false stands.", "checked": "2026-09-30" }, "desktop": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "You want an interactive shell to sit in. Exec runs through run_command over the API and logs are the interface; if the work cannot be expressed as commands, this is the wrong platform.", "note": "not offered: outside the product's scope. Autoresearch nodes/sandboxes and the SF Compute GPU-market VMs are headless API/CLI compute for training, RL sandboxes and batch jobs; no VNC/RDP/streamed desktop is documented anywhere on autoresearch.sfcompute.com, sfcompute.com or docs.sfcompute.com.", "checked": "2026-09-30" }, "browser": { "value": false, "source": "https://sfcompute.com/", "quote": "We build data centers, the GPU clusters inside them, & the cloud on top. Get VMs, Bare Metal, & managed Slurm from the only cloud designed to let you sublease your reserved capacity when plans change.", "note": "not offered: outside the product's scope. The product is bare GPU VMs/bare-metal/nodes and coding-agent sandboxes for training/inference workloads; no managed/preinstalled browser is documented. The only 'browser' mentions in the docs are the human's own browser used for OAuth sign-in, not a machine feature.", "checked": "2026-09-30" }, "computer_use": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "How to call it: an agent connects over MCP (below) and uses the tools; a script or CLI uses the HTTP API with a bearer token.", "note": "not offered: outside the product's scope. Interaction is exclusively via MCP tools / HTTP API (run_command, create_node, create_sandbox_env, etc.); no screenshot+click driving API exists for any mode (sandbox, node, or market VM).", "checked": "2026-09-30" }, "volumes": { "value": false, "source": "https://autoresearch.sfcompute.com/docs", "quote": "Every node has one persistent volume, and its size is the node_volume_gib limit for your org (250 GiB by default). ... There is no per-node volume size to ask for", "note": "Autoresearch nodes get exactly one fixed persistent volume per node (default 250 GiB, raisable org-wide, not per-node extra volumes). Market VMs have no persistent storage at all ('Instances have no persistent storage' - docs.sfcompute.com/preview/instances). No extra attachable/shared volume product exists anywhere; matches card's own volume_attach:false / volume_shared:false evidence.", "checked": "2026-09-30" }, "public_ipv4": { "value": false, "source": "https://autoresearch.sfcompute.com/docs", "quote": "No inbound by default. Nodes have no public address and no open ports. There is no SSH. The one way in is an exposed port: HTTPS at our edge, per port, with the URL or a bearer token as the credential.", "note": "Confirmed for the Autoresearch nodes/sandboxes (flagship 'stock' mode): no public IP, access only via per-port HTTPS edge URL. The separate GPU-market VMs (reserved/spot modes) DO support an optional public IPv4 ('Public IPv4 is available only on SKUs that support it, and only when enabled at create time' - docs.sfcompute.com/llms-full.txt) and already carry public_ipv4:true on those two modes plus network.ipv4_available:true card-wide; top-level features.public_ipv4 models the primary Autoresearch product, so false stands for it.", "checked": "2026-09-30" }, "agent_harnesses": { "value": false, "source": "https://autoresearch.sfcompute.com/docs", "quote": "Claude Code, Codex, Cursor, opencode, Amp, oh my pi, and Zed are supported. If the install cannot write a config file safely, it prints the block for you to paste.", "note": "Claude Code/Codex/opencode/etc. are supported as MCP CLIENTS on the user's own machine that call INTO the Autoresearch platform (`autor mcp install` wires local agents to the platform's MCP server); no coding-agent harness is preinstalled inside a sandbox/node/VM itself, so agent_harnesses (baked-in inside the machine) is false.", "checked": "2026-09-30" }, "windows": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "On-demand GPU nodes and asynchronous GPU jobs for AI agents ... A node is a container on an H100 with a persistent, encrypted disk", "note": "not offered: outside the product's scope. Nodes are Linux containers on Linux hosts; market VMs boot from custom Linux images (docs.sfcompute.com). No Windows image/OS is documented anywhere on the card's sources.", "checked": "2026-09-30" }, "macos": { "value": false, "source": "https://sfcompute.com/", "quote": "We build data centers, the GPU clusters inside them, & the cloud on top. Get VMs, Bare Metal, & managed Slurm", "note": "not offered: outside the product's scope. SF Compute is an NVIDIA GPU (H100/H200/B200/B300) cluster/cloud on Linux hosts; no Apple hardware or macOS VM/host is mentioned anywhere in the docs, specs or pricing pages.", "checked": "2026-09-30" }, "arm64": { "value": false, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "chip \"cpu-8\" (8 reserved physical cores on one NUMA node, 32 GiB RAM) for parallel CPU work", "note": "not offered: outside the product's scope. All published node/instance images and GPU boxes (H100/H200/B200/B300 x86 hosts) are described only in x86/CUDA terms (e.g. image names like 'vllm-0.26-cuda12.9'); no ARM/aarch64 chip, image or SKU is documented anywhere on the card's sources.", "checked": "2026-09-30" }, "self_host": { "value": false, "source": "https://sfcompute.com/", "quote": "We build data centers, the GPU clusters inside them, & the cloud on top. Get VMs, Bare Metal, & managed Slurm from the only cloud designed to let you sublease your reserved capacity when plans change.", "note": "SF Compute sells access to compute on ITS OWN data centers/clusters (VMs, bare metal, managed Slurm, Autoresearch nodes/sandboxes) - there is no offering to run the SF Compute software/platform on a customer's own infrastructure.", "checked": "2026-09-30" }, "byoc": { "value": false, "source": "https://sfcompute.com/", "quote": "We build data centers, the GPU clusters inside them, & the cloud on top. Get VMs, Bare Metal, & managed Slurm from the only cloud designed to let you sublease your reserved capacity when plans change.", "note": "Even the 'Enterprise / bare metal / managed Slurm' contact-sales plan (card's own plans[] entry: 'bare metal, managed Slurm with InfiniBand, dedicated solutions engineer') is delivered on SF Compute's own clusters, not deployed into a customer's own cloud account; no BYOC offering is documented anywhere on the card's sources.", "checked": "2026-09-30" }, "hipaa": { "value": false, "source": "https://trust.sfcompute.com/", "quote": "Overview Resources Controls Compliance SOC 2 Resources View all SOC 2 Type 2 Report", "note": "SF Compute's public Trust Center (Vanta-hosted) lists exactly one compliance framework, SOC 2 Type 2 - no HIPAA badge, BAA offering, or mention of HIPAA anywhere on trust.sfcompute.com, sfcompute.com, docs.sfcompute.com, or autoresearch.sfcompute.com/llms.txt. Treated as not offered rather than null since the trust page is the canonical place such a claim would appear and is silent on it.", "checked": "2026-09-30" }, "sso": { "value": false, "source": "https://sfcompute.com/", "quote": "Sign in Email Continue with email OR Continue with Google", "note": "The account sign-in page (also served identically at /pricing and /enterprise) offers only email/passwordless and 'Continue with Google' - no SAML/OIDC enterprise-SSO option is presented, and no SSO/SAML/OIDC mention exists anywhere in the docs (autoresearch.sfcompute.com/llms.txt, docs.sfcompute.com/llms-full.txt) or the trust center.", "checked": "2026-09-30" }, "cpu_class": { "value": { "sandbox": "shared", "cpu-node": "dedicated", "h100-node": "shared", "batch-h100": "shared", "reserved": "dedicated", "spot": "dedicated" }, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "A busy host can slow a large sandbox down because its cores are shared. ... chip \"cpu-2\" (2 reserved physical cores, 8 GiB RAM) ... chip \"cpu-8\" (8 reserved physical cores on one NUMA node, 32 GiB RAM) ... a GPU node's CPU limit is a CFS quota ... the quota buys burst across the whole machine rather than a slice of it", "note": "sandbox: vCPUs on Autoresearch sandboxes are explicitly shared/oversubscribed cores on a multi-tenant host, no published baseline %. cpu-node (cpu-2/cpu-8): docs say 'reserved physical cores' -> dedicated. h100-node/batch-h100: the container's CPU is a cgroup/CFS quota that 'bursts across the whole machine' (not a pinned slice), so classified shared rather than dedicated; no published baseline % so not expressed as burstable:NN. reserved/spot (SF Compute GPU market): the orderbook sells whole 8-GPU nodes (gpu_min_count 8) and docs.sfcompute.com SKU listing shows a fixed VCPUS count (e.g. 56) per whole node/SKU with an instance mapped 1:1 to a node it doesn't share ('An instance only runs on nodes of its own SKU'), so the full node's vCPUs are dedicated to the buyer.", "checked": "2026-09-30" }, "fast_boot": { "value": true, "source": "https://www.computesdk.com/benchmarks", "quote": "burst of 100: p95 507 ms, 100% started", "note": "ComputeSDK sandbox benchmark 2026-09-25", "checked": "2026-09-25" }, "usage_api": { "value": true, "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "get_usage(month?): spend in USD (\"YYYY-MM\", default current) - per node, per batch (jobs by sweep label, with the priciest named), and the builds/storage lines, summing to the month's total. The `sandboxes` block splits its charge per tier (RAM, parked disk, resident and cold snapshots) and per UTC ", "note": "granularity: per-sandbox; cost: yes; SF Compute provides get_usage() API call that returns billable sandbox usage broken down per day by tier (RAM, parked disk, snapshots). Also provides get_billing() for account-level usage and get_mission() for per-mission costs. HTTP endpoint: /v2/org/usage (REST API with bearer token auth).", "confidence": "high", "checked": "2026-10-01" }, "persist_scope": { "value": "folder", "source": "https://autoresearch.sfcompute.com/llms.txt", "quote": "your persistent disk mounts at the `home` parameter (an absolute path; default the image's WORKDIR, else /root), and ONLY that path survives stops", "note": "Only the home folder survives a stop; the rest of the filesystem (packages installed elsewhere) resets.", "checked": "2026-10-01" } }, "locations": [ { "country": "US", "state": null, "city": null, "provider_region": "us", "self_serve": true, "choosable": true }, { "country": "SE", "state": null, "city": "Stockholm", "provider_region": "europe-north1", "self_serve": true, "choosable": true } ], "locations_source": "https://docs.sfcompute.com/llms-full.txt", "locations_quote": "SF Compute market VMs/reserved capacity are offered in an EU region (europe-north1 / EU North)", "locations_checked": "2026-10-01", "audit_2026_10_14": { "applied": 4, "source": "checked against https://autoresearch.sfcompute.com/ live pricing table and llms.txt", "reviewed": 4 } }