Field brief · hardware · October 2026
Which GPUs to buy, how many, and every part in the box, ranked by months to pay back on Vast.ai and the other open rental marketplaces. Every price links to where it came from.
Every build is one 8-GPU server. Click a build name for its full bill of materials.
| # | Build | GPUs | GPU unit price | All-in cost | Host revenue / mo | Payback (live rates) | Payback (Q1-26 low) | Purchase → ROI* | Price sources |
|---|---|---|---|---|---|---|---|---|---|
| 1 | HGX B200 — Central Computer bundle | 8× B200 SXM 180GB | in server ($365,699) | $425,660 | $32,377 | 13.1 mo | 48.2 | 16.8 mo | [9] [28] [29] [32] |
| 2 | HGX B200 — Thinkmate (firm order) | 8× B200 SXM 180GB | in server ($493,433) | $498,541 | $32,377 | 15.4 mo | 56.5 | 18.8 mo | [10] [29] [32] |
| 3 | 8× RTX 4090 server | 8× RTX 4090 24GB, 2-slot blower | $3,199.50 | $36,261 | $1,845 | 19.6 mo | 33.6 | 21.1 mo | [15] [18] [19] [25] [26] [29] [30] [31] [32] |
| 4 | HGX H200 — Bizon X9000 G3 | 8× H200 SXM 141GB | in server ($363,097) | $428,044 | $16,785 | 25.5 mo | 59.8 | 26.7 mo | [12] [28] [29] [32] |
| 5 | 8× RTX 5090 server | 8× RTX 5090 32GB, 2-slot Turbo | $6,525 | $62,865 | $2,403 | 26.2 mo | 49.3 | 27.6 mo | [15] [16] [17] [25] [26] [29] [30] [31] [32] |
| 6 | 8× RTX 3090 server | 8× RTX 3090 24GB, 2-slot Turbo (refurb) | $1,799.97 | $25,065 | $927 | 27.1 mo | 54.7 | 28.3 mo | [15] [22] [25] [26] [29] [30] [31] [32] |
| 7 | 8× RTX PRO 6000 Max-Q server | 8× RTX PRO 6000 Max-Q 96GB (300W blower) | $16,499.99 | $151,434 | $5,082 | 29.8 mo | 57.2 | 31.0 mo | [15] [20] [25] [27] [29] [30] [31] [32] |
| 8 | 8× RTX 6000 Ada server | 8× RTX 6000 Ada 48GB (refurb) | $7,999.99 | $74,665 | $2,209 | 33.8 mo | n/a | 35.0 mo | [15] [21] [25] [26] [29] [30] [31] [32] |
| 9 | 8× L40S server | 8× L40S 48GB | $8,595 | $79,425 | $2,280 | 34.8 mo | n/a | 36.0 mo | [15] [23] [25] [26] [29] [30] [31] [32] |
| 10 | HGX H100 — Viperatech (new) | 8× H100 SXM 80GB | in server ($283,000) | $313,032 | $8,125 | 38.5 mo | 60.6 | 40.0 mo | [13] [28] [32] |
| 11 | HGX A100 — Viperatech (used) | 8× A100 SXM 80GB | in server ($108,000) | $115,948 | $2,964 | 39.1 mo | n/a | 40.3 mo | [14] [27] [29] [32] |
| 12 | 8× A100 80GB PCIe server | 8× A100 80GB PCIe (refurb) | $18,900 | $170,634 | $2,058 | 82.9 mo | n/a | 84.1 mo | [15] [24] [25] [27] [29] [30] [31] [32] |
*Purchase → ROI = lead time + ~0.7 mo ramp + payback. Q1-26 low = that GPU's lowest monthly median in Jan–Mar 2026, at 70% occupancy; n/a = no history for that GPU. Revenue is gross host revenue before power/colo.
Pulled from the Vast.ai marketplace API on 17 Oct 2026: every on-demand offer, split into rented and available [1]. 6-month occupancy and Q1 lows come from 500.farm [2]. Occupancy counts online machines only, so real utilization is probably 5–10 points lower.
| GPU | Available | Rented | Live occupancy | Median rented price (renter $/GPU-hr) | Host $/GPU-hr (×0.8) | Verified occ. | Unverified occ. | 6-month avg occ. | Q1-2026 low (renter) |
|---|---|---|---|---|---|---|---|---|---|
| RTX 5090 | 588 | 3,598 | 86% | $0.67 | $0.53 | 90% | 73% | 77% | $0.39 (Jan 26) |
| RTX 4090 | 172 | 1,239 | 88% | $0.51 | $0.41 | 94% | 67% | 77% | $0.33 (Mar 26) |
| RTX 3090 | 310 | 227 | 42% | $0.27 | $0.21 | — | — | 74% | ~$0.14 |
| RTX PRO 6000 Max-Q | 46 | 346 | 88% | $1.38 | $1.10 | 91% | 74% | 79% (WS) | $0.81 (Jan 26) |
| RTX PRO 6000 WS | 85 | 424 | 83% | $1.47 | $1.17 | 85% | 75% | 79% | $0.81 (Jan 26) |
| RTX PRO 6000 Server | 140 | 467 | 77% | $1.47 | $1.18 | 85% | 55% | 76% | $0.81 (Jan 26) |
| RTX 6000 Ada | 13 | 104 | 89% | $0.64 | $0.51 | — | — | 74% | — |
| L40S | 60 | 191 | 76% | $0.80 | $0.64 | — | — | 61% | — |
| A100 PCIe 80GB | 114 | 170 | 60% | $0.73 | $0.59 | — | — | — | — |
| A100 SXM4 80GB | 135 | 299 | 69% | $0.87 | $0.70 | — | — | 73% | — |
| H100 SXM | 65 | 103 | 61% | $2.35 | $1.88 | 86% | 23% | 74% | $1.58 (Mar 26) |
| H200 | 68 | 160 | 70% | $4.61 | $3.68 | 90% | 47% | 78% | $2.19 (Feb 26) |
| B200 | 42 | 185 | 81% | $9.00 | $7.20 | 79% | — | 77% | $2.70 (Feb 26) |
Verified / unverified split computed from the same pull. — = too few machines to split, or no history.
Every line, quantity, unit price and live listing. The consumer builds share one chassis platform. The HGX servers ship as complete systems with GPUs, CPUs, memory, NICs and power supplies inside the listed price; the extra lines are what you add so they pass Vast verification.
| Component | Qty | Unit | Extended | Status on 17 Oct / notes | Live listing |
|---|---|---|---|---|---|
| Central Computer BUNDLE00079: ASRock 8U8X-TURIN2, 8× B200 SXM (1.44TB HBM3e), 2× EPYC 9655, 128GB DDR5, 1× 3.84TB U.2, $499 build fee, 1-yr warranty | 1 | $365,698.94 | $365,698.94 | Orderable (cart works), but the barebone inside is “Special Order”, so no ship date yet | [9] |
| 64GB DDR5-4800 ECC RDIMM ×22 | 22 | $2,493.30 | $54,852.60 | New. Brings RAM to 1.54TB — Vast requires RAM ≥ 1.44TB VRAM | [28] |
| 7.68TB U.2 NVMe — WD Ultrastar DC SN200 | 4 | $1,249 | $4,996 | Refurbished. Adds ~4TB per GPU of renter storage | [29] |
| StarTech C19→C20 14AWG power cord | 6 | $18.71 | $112.26 | New | [32] |
| Total | $425,659.80 |
8 × $7.20/hr host × 730 h × 77% occupancy = $32,377/mo → 13.1 months to pay back $425,660. Lead time ≈ 3.0 mo.
| Component | Qty | Unit | Extended | Status on 17 Oct / notes | Live listing |
|---|---|---|---|---|---|
| Thinkmate GPX GXS6-2808-8HGX-B2: 8× B200 SXM, 2× Xeon 6740P 48C, 1TB DDR5-6400, onboard 2× 10GbE, no drives, 5-yr parts replacement | 1 | $493,433 | $493,433 | Add To Order verified at this price; no lead time stated. 1TB RAM is below Vast's RAM ≥ VRAM rule — fine for training and other platforms, but upgrading RAM pushes it over $500k | [10] |
| 7.68TB U.2 NVMe — WD Ultrastar DC SN200 | 4 | $1,249 | $4,996 | Refurbished. Adds ~4TB per GPU of renter storage | [29] |
| StarTech C19→C20 14AWG power cord | 6 | $18.71 | $112.26 | New | [32] |
| Total | $498,541.26 |
8 × $7.20/hr host × 730 h × 77% occupancy = $32,377/mo → 15.4 months to pay back $498,541. Lead time ≈ 2.8 mo.
| Component | Qty | Unit | Extended | Status on 17 Oct / notes | Live listing |
|---|---|---|---|---|---|
| NVIDIA RTX 4090 24GB — 2-slot blower (full AD102, not the 4090D) | 8 | $3,199.50 | $25,596 | “In Stock” at MillionMiner, but they confirm price and stock before the order is binding. Firm fallback: $4,000 on eBay | [18] [19] |
| Supermicro AS-4124GS-TNR 4U barebone — dual SP3, 8× PCIe 4.0 x16 double-width slots, 32 DDR4 DIMM slots, 4× 2000W Titanium PSUs | 1 | $3,399 | $3,399 | Used, 8 available. PSUs + fans included; confirm rails and CPU heatsinks with seller | [15] |
| AMD EPYC 7543 (32 cores / 64 threads), unlocked | 2 | $819 | $1,638 | Used, 10+ available. Must be unlocked (Dell/Lenovo-locked chips won't boot in Supermicro) | [25] |
| 32GB DDR4-3200 ECC RDIMM | 16 | $239.99 | $3,839.84 | New. 512GB total, 1 DIMM per channel on both sockets. Meets Vast RAM ≥ VRAM and RunPod's 512GB for 8×24GB | [26] |
| 7.68TB U.2 NVMe — WD Ultrastar DC SN200 (renter storage) | 1 | $1,249 | $1,249 | Refurbished, 15 left. ~1TB per GPU (TensorDock minimum) | [29] |
| Kingston NV3 1TB M.2 NVMe (OS / boot) | 1 | $156.77 | $156.77 | New | [30] |
| Supermicro CBL-PWEX-0962Y-50 — 12VHPWR to Supermicro 8-pin GPU power cable | 8 | $38.43 | $307.44 | New, in stock. Check the chassis GPU header pinout before ordering | [31] |
| StarTech C19→C20 14AWG power cord | 4 | $18.71 | $74.84 | New | [32] |
| Total | $36,260.89 |
8 × $0.41/hr host × 730 h × 77% occupancy = $1,845/mo → 19.6 months to pay back $36,261. Lead time ≈ 0.8 mo.
| Component | Qty | Unit | Extended | Status on 17 Oct / notes | Live listing |
|---|---|---|---|---|---|
| Bizon X9000 G3: 8× H200 SXM (1.13TB HBM3e), 2× EPYC 9135, 512GB DDR5, 4TB NVMe, dual 10GbE, 5-yr labor / 3-yr parts | 1 | $363,097 | $363,097 | In stock, 5 left, ships Oct 24–27 2026 | [12] |
| 64GB DDR5-4800 ECC RDIMM ×24 | 24 | $2,493.30 | $59,839.20 | New. Replaces the 512GB with 1.5TB — Vast requires RAM ≥ 1.13TB VRAM | [28] |
| 7.68TB U.2 NVMe — WD Ultrastar DC SN200 | 4 | $1,249 | $4,996 | Refurbished. Adds ~4TB per GPU of renter storage | [29] |
| StarTech C19→C20 14AWG power cord | 6 | $18.71 | $112.26 | New | [32] |
| Total | $428,044.46 |
8 × $3.68/hr host × 730 h × 78% occupancy = $16,785/mo → 25.5 months to pay back $428,044. Lead time ≈ 0.5 mo.
| Component | Qty | Unit | Extended | Status on 17 Oct / notes | Live listing |
|---|---|---|---|---|---|
| NVIDIA RTX 5090 32GB — 2-slot Turbo blower ($6,475 + $50 shipping) | 8 | $6,525 | $52,200 | Open box, 10 available, ships from Shenzhen. Alternate: MillionMiner $6,099 (non-binding until confirmed) | [16] [17] |
| Supermicro AS-4124GS-TNR 4U barebone — dual SP3, 8× PCIe 4.0 x16 double-width slots, 32 DDR4 DIMM slots, 4× 2000W Titanium PSUs | 1 | $3,399 | $3,399 | Used, 8 available. PSUs + fans included; confirm rails and CPU heatsinks with seller | [15] |
| AMD EPYC 7543 (32 cores / 64 threads), unlocked | 2 | $819 | $1,638 | Used, 10+ available. Must be unlocked (Dell/Lenovo-locked chips won't boot in Supermicro) | [25] |
| 32GB DDR4-3200 ECC RDIMM | 16 | $239.99 | $3,839.84 | New. 512GB total, 1 DIMM per channel on both sockets. Meets Vast RAM ≥ VRAM and RunPod's 512GB for 8×24GB | [26] |
| 7.68TB U.2 NVMe — WD Ultrastar DC SN200 (renter storage) | 1 | $1,249 | $1,249 | Refurbished, 15 left. ~1TB per GPU (TensorDock minimum) | [29] |
| Kingston NV3 1TB M.2 NVMe (OS / boot) | 1 | $156.77 | $156.77 | New | [30] |
| Supermicro CBL-PWEX-0962Y-50 — 12VHPWR to Supermicro 8-pin GPU power cable | 8 | $38.43 | $307.44 | New, in stock. Check the chassis GPU header pinout before ordering | [31] |
| StarTech C19→C20 14AWG power cord | 4 | $18.71 | $74.84 | New | [32] |
| Total | $62,864.89 |
8 × $0.53/hr host × 730 h × 77% occupancy = $2,403/mo → 26.2 months to pay back $62,865. Lead time ≈ 0.8 mo.
| Component | Qty | Unit | Extended | Status on 17 Oct / notes | Live listing |
|---|---|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Max-Q 96GB (300W, 2-slot blower) | 8 | $16,499.99 | $131,999.92 | Sold and shipped by Newegg, in stock | [20] |
| Supermicro AS-4124GS-TNR 4U barebone — dual SP3, 8× PCIe 4.0 x16 double-width slots, 32 DDR4 DIMM slots, 4× 2000W Titanium PSUs | 1 | $3,399 | $3,399 | Used, 8 available. PSUs + fans included; confirm rails and CPU heatsinks with seller | [15] |
| AMD EPYC 7543 (32 cores / 64 threads), unlocked | 2 | $819 | $1,638 | Used, 10+ available. Must be unlocked (Dell/Lenovo-locked chips won't boot in Supermicro) | [25] |
| 64GB DDR4-3200 ECC RDIMM | 16 | $709.99 | $11,359.84 | New. 1TB total — needed because 8×96GB / 8×80GB VRAM is 640–768GB | [27] |
| 7.68TB U.2 NVMe — WD Ultrastar DC SN200 (renter storage) | 2 | $1,249 | $2,498 | Refurbished. ~2TB per GPU | [29] |
| Kingston NV3 1TB M.2 NVMe (OS / boot) | 1 | $156.77 | $156.77 | New | [30] |
| Supermicro CBL-PWEX-0962Y-50 — 12VHPWR to Supermicro 8-pin GPU power cable | 8 | $38.43 | $307.44 | New, in stock. Check the chassis GPU header pinout before ordering | [31] |
| StarTech C19→C20 14AWG power cord | 4 | $18.71 | $74.84 | New | [32] |
| Total | $151,433.81 |
8 × $1.10/hr host × 730 h × 79% occupancy = $5,082/mo → 29.8 months to pay back $151,434. Lead time ≈ 0.5 mo.
These use the consumer platform above (512GB build $10,664.89, 1TB build $19,433.89 for 96GB / 80GB cards) or a complete HGX system.
| Build | GPU line | Qty | Unit | Rest of build | All-in | Sources |
|---|---|---|---|---|---|---|
| 8× RTX 3090 server | Gigabyte RTX 3090 24GB Turbo blower — refurbished, 1-yr warranty | 8 | $1,799.97 | $10,665 | $25,065 | [15] [22] [25] [26] [29] [30] [31] [32] |
| 8× RTX 6000 Ada server | NVIDIA RTX 6000 Ada 48GB — refurbished | 8 | $7,999.99 | $10,665 | $74,665 | [15] [21] [25] [26] [29] [30] [31] [32] |
| 8× L40S server | NVIDIA L40S 48GB | 8 | $8,595 | $10,665 | $79,425 | [15] [23] [25] [26] [29] [30] [31] [32] |
| HGX H100 — Viperatech (new) | Viperatech Supermicro SYS-821GE-TNHR: 8× H100 SXM5 (640GB), 2× Xeon 8468V, DDR5 (amount not stated), 480GB SATA + 7.6TB NVMe, 400G CX-7, 3-yr return-to-base | 1 | $283,000 | $30,032 | $313,032 | [13] [28] [32] |
| HGX A100 — Viperatech (used) | Viperatech Gigabyte G492-ZD0 (used): 8× A100 80GB SXM4, 2× EPYC 7742, 512GB DDR4, 8× CX-7 200G; warranty expired Feb 2025 | 1 | $108,000 | $7,948 | $115,948 | [14] [27] [29] [32] |
| 8× A100 80GB PCIe server | NVIDIA A100 80GB PCIe — refurbished, 3-yr warranty | 8 | $18,900 | $19,434 | $170,634 | [15] [24] [25] [27] [29] [30] [31] [32] |
nvidia-smi -pl 450 (small performance loss).| Budget | Strategy | Servers | GPU count | Cost | Revenue / mo | Payback (live) | Payback (Q1-26 low) |
|---|---|---|---|---|---|---|---|
| ~$100k | Fastest | 2× 8× RTX 4090 server | 16 GPUs: 16× RTX 4090 | $72,522 | $3,691 | 19.6 mo | 33.6 mo |
| ~$100k | Fewer gray-market cards | 1× 8× RTX 4090 server + 2× 8× RTX 3090 server | 24 GPUs: 8× RTX 4090, 16× RTX 3090 | $86,390 | $3,699 | 23.4 mo | 43.3 mo |
| ~$250k | Fastest | 6× 8× RTX 4090 server | 48 GPUs: 48× RTX 4090 | $217,565 | $11,073 | 19.6 mo | 33.6 mo |
| ~$250k | Mixed | 3× 8× RTX 4090 server + 2× 8× RTX 5090 server | 40 GPUs: 24× RTX 4090, 16× RTX 5090 | $234,512 | $10,343 | 22.7 mo | 40.5 mo |
| ~$500k | Fastest (special order) | 1× HGX B200 — Central Computer bundle + 2× 8× RTX 4090 server | 24 GPUs: 8× B200, 16× RTX 4090 | $498,182 | $36,068 | 13.8 mo | 45.3 mo |
| ~$500k | Ships next week | 1× HGX H200 — Bizon X9000 G3 + 1× 8× RTX 4090 server | 16 GPUs: 8× H200, 8× RTX 4090 | $464,305 | $18,630 | 24.9 mo | 56.3 mo |
Every plan stays under $500,000. GPU count = total cards across all servers.
The specs are built to pass the strictest open gate (TensorDock: 8 vCPU and 1TB NVMe per GPU, RAM above VRAM), so every build verifies on every open platform that takes its GPU. A machine runs one host agent at a time: list on Vast first, and use Clore as the fallback.
| Platform | Accepts outside hosts? | Verification / hardware gate | Fee / payout | Which builds qualify | Source |
|---|---|---|---|---|---|
| Vast.ai | Yes — open to anyone | Self-test: RAM ≥ total VRAM, ≥2 CPU cores/GPU, ≥500 Mbps up/down, >3 open ports, PCIe >2.85 GB/s, reliability >0.90. Datacenter badge: ISO 27001 + business + ≥5 servers | 0% host fee; renter pays ~25% on top → host keeps ~80% | All builds | [5] [7] [6] [3] |
| Clore.ai | Yes — open | NVIDIA only, ≥8GB RAM, ≥32GB disk. Bare-metal tier: Tier 3 DC, 64+ threads, 128GB+ RAM | 10% on-demand fee (+15% if renter pays crypto) | All builds; 45–52% occupancy on 4090/5090, 79% on PRO 6000 | [36] |
| TensorDock | Yes — vetted | ≥16 GB/s PCIe per GPU, ≥16GB RAM per GPU and > VRAM, ≥8 vCPU per GPU at ≥3 GHz, 1TB NVMe per GPU, 1 Gbps, static IP | 25% fee | All builds as specced (that is why each has 64 cores and ~1TB NVMe/GPU) | [35] |
| io.net | Yes | Supported list: RTX 30/40/50, A100, H100, H200, B200 | Paid in $IO; fee unpublished | All except RTX PRO 6000 (not on list) | [38] |
| Salad | Consumer PCs only | Consumer RTX cards | Host gets ~60% of customer price: 5090 $0.175–0.30/hr, 4090 $0.112–0.198/hr | 4090 / 5090 / 3090 only, at about half of Vast pay | [37] |
| Akash | Yes — run a Kubernetes provider | Subsidy programs want an 8-GPU node, redundant DC, 99.7% uptime | 1–2% take rate | HGX builds; H100 averages $2.71/hr | [39] |
| RunPod | No | Community hosting closed. Secure Cloud partners: ≥100 kW, Tier 3/4, SOC 2 / ISO 27001 | — | None at this scale | [33] [34] |
| Lambda, Shadeform, Prime Intellect | No | Run their own capacity or broker Tier 3/4 clouds only | — | None | — |
Numbers you will find online that are not used here, and why.