NVIDIA RTX PRO 6000 Blackwell Server Edition
600 W passive PCIe Blackwell card with 96 GB GDDR7, priced and ordered like a component rather than a baseboard.
NVIDIA RTX PRO 6000 Blackwell Server Edition is one of 24 data center GPUs tracked in the SecondWatt catalogue, each with specifications, lead times and indicative secondary-market pricing.
NVIDIA RTX PRO 6000 Blackwell Server Edition specs: 96 GB GDDR7, 600 W configurable, PCIe Gen5 passive, four MIG instances, and three seller prices.
The RTX PRO 6000 Blackwell Server Edition is the one Blackwell part a buyer can price and order like an ordinary component. It is a PCIe Gen5 x16 card with 96 GB of GDDR7 at 1,597 GB/s, rated up to 600 W configurable, and passively cooled to run continuously in server airflow. NVIDIA positions it for the data center under a heading of universal compute and visual computing, and it sits in the enterprise software support lineup beside B200 and B300. There is no NVLink on it at all, so multi-GPU scaling runs over PCIe and partitioning is the scaling story instead. Multi-Instance GPU splits one card into as many as four isolated instances of 24 GB each. That, plus 600 W in a standard slot, is why it lands in ordinary enterprise racks rather than forcing a 57 kW rebuild. Lenovo ships it passive, dual-slot, full-height and full-length, and allows eight in an SR675 V3 or four in an SR650a V4 only with a 450 W power cap. Eight cards at 600 W is 4.8 kW of GPU a server, and 3.6 kW at the cap. Do not carry the Workstation Edition figures across. That part is also 96 GB GDDR7 but runs 1,792 GB/s with active double flow-through cooling, so capacity matches while bandwidth and cooling do not. On export status this card sits below the 6,500 GB/s memory-bandwidth gate in the January 2026 rule at 1,597 GB/s, but NVIDIA and its distributors publish no ECCN for the card as of 10 September 2026, which makes its export status a question for counsel rather than a settled one.
Why it matters
Every other part in this category is a baseboard sold inside a server, priced by division and quoted on allocation. This one has three published prices from three sellers, an authorized distribution path and a 3 to 8 week quoted lead time. At 600 W in a standard slot it also fits a 10 to 20 kW enterprise rack, which is the whole point of it.
Who buys it
Enterprises that want Blackwell inference, rendering and video in existing 2U to 4U servers without rebuilding a rack, virtual desktop and visualization estates, and inference operators whose models fit in 96 GB and need no NVLink. Also anyone who would rather have four isolated 24 GB instances a card than one large NVLink domain.
Role in the data center
Inference, rendering, virtual desktop and video, with no training role because the part carries no NVLink. It lands in the ordinary 10 to 20 kW enterprise rack band at 4.8 kW of GPU per eight-card server, which is the reason to choose it over an SXM baseboard.
Power envelope
- Per accelerator
- 600 W — Verified: Maximum configurable board power, manufacturer specification. source
- Cooling class
- Air or DLC — Derived: Deployable air-cooled; liquid cooling raises achievable rack density.
No published node-level power rating exists for this part, so node, rack and per-MW figures are not stated. Multiplying accelerator watts by an assumed overhead under-provisions a building by roughly 45% against measured systems.
Node figures are published OEM system ratings; rack counts are arithmetic against a stated rack budget and move with your own cooling design. Assumption set version 2026-09-11a.
Key specifications
| GPU memory | 96 GB GDDR7 with ECC |
|---|---|
| Memory bandwidth | 1,597 GB/s |
| Memory bus width | 512 bit |
| Max power consumption, configurable | up to 600 W |
| OEM power-capped variant | 450 W |
| Form factor, air-cooled | dual-slot full-height full-length passive PCIe card |
| Form factor, liquid-cooled | single-slot full-length PCIe card |
| Host interface | PCIe Gen5 x16 lanes |
| NVLink support | none on this part |
| Multi-Instance GPU partitions | up to 4 of 24 GB each |
| CUDA cores | 24,064 cores |
| Peak FP32 throughput | 120 TFLOPS |
| FP4 throughput, vendor figure | 4 PFLOPS |
| Cards per server, Lenovo SR675 V3 | 8 cards |
| GPU power a server at eight cards | 4.8 kW |
| Observed new price range | 15,410.53 to 17,585.64 USD per card |
Technical summary
Architecture: Blackwell, RTX PRO line, 24,064 CUDA cores and 752 fifth-generation Tensor cores. Form factor: dual-slot full-height full-length passive air, or single-slot full-length liquid. Memory: 96 GB GDDR7 with ECC on a 512-bit bus at 1,597 GB/s. Max power: up to 600 W configurable, with a 450 W OEM power-capped variant available. Host interface: PCIe Gen5 x16, single 16-pin power connector. Interconnect: no NVLink, so multi-card scaling runs over PCIe only. Partitioning: Multi-Instance GPU to as many as four isolated instances of 24 GB each. Throughput: 120 TFLOPS peak FP32 and 4 PFLOPS FP4, vendor figures. Server density: up to 8 per Lenovo SR675 V3, 4 per SR650a V4 at the 450 W cap, 2 per SR650 V4. GPU power a server: 4.8 kW at eight cards and 600 W, or 3.6 kW at the 450 W cap. Display: 4x DisplayPort 2.1, which the SXM parts do not carry.
Major variations
The Workstation Edition is the variant that causes real errors: the same 96 GB of GDDR7 but 1,792 GB/s against this card's 1,597 GB/s, and an active double flow-through cooler rather than a passive one. Same capacity, different bandwidth, cooling and product. Within the Server Edition there are three axes. Cooling splits into a dual-slot full-height full-length passive air card and a single-slot full-length liquid card. Power splits into the standard up-to-600 W configurable part and an OEM 450 W capped variant under its own feature code. And the part number appears as 900-2G153-0000-000, 900-2G153-0000-200 and a Lenovo-labelled 900-2G153-2700-030, two priced 2,175 dollars apart with no published explanation of the suffix. NVIDIA also lists an RTX PRO 4500 Blackwell in the same supported lineup, which is the step down rather than a variant of this card.
Configurations and options
Passive dual-slot air in mainstream 2U to 4U enterprise servers, or single-slot full-length liquid where a cold-plate loop already exists. Density follows the OEM thermal matrix rather than the card, with Lenovo allowing 8 in an SR675 V3, 4 in an SR650a V4 only at the 450 W cap, and 2 in an SR650 V4. Partitioning to four instances of 24 GB is the common serving configuration.
Compatibility and dependencies
Host platform is any qualified server with a PCIe Gen5 x16 slot, two slots of clearance and 600 W of slot and cable power through a single 16-pin connector. There is no baseboard, no NVSwitch and no NVLink, so multi-card scaling runs over PCIe and partitioning to four 24 GB instances is the alternative to a large domain. Eight cards at 600 W is 4.8 kW of GPU a server, or 3.6 kW at the 450 W cap, which fits an ordinary 10 to 20 kW enterprise rack without new busway or a coolant loop. Plan alongside racks and PDUs for that envelope, and liquid cooling only where the single-slot liquid variant raises slot density. Cooling is a host dependency rather than a card feature: the air card is entirely passive, so chassis fan curves, baffles and inlet temperature are the thermal design, and an OEM support entry at 600 W is not the same as one at the 450 W cap.
Pricing and availability
Published price range: $15,410.53 - $17,585.64 per card, new through authorized distribution, September 2026, with a third observation at 17,400 euro ex VAT in the EU. Evidence: asking price · new condition · Single card · observed Sep 2026. Trend: stable No five-year series exists, because partner platforms shipped from May 2025 and the card had about 16 months of market history on 10 September 2026. The three observed prices span 2,175 dollars across two US sellers quoting two different SKU suffixes, which reads as SKU and channel variation rather than as a trend. Recorded as flat until a second dated observation of the same suffix exists. Lead time, new: 3 to 8 weeks estimated OEM lead time through authorized US distribution, with one seller quoting stock arrival between 1 October and 5 November 2026. One US seller showed out of stock and one EU seller quoted an offer within the hour.. Warranty, new: Channel-dependent rather than part-dependent: 3-year limited through a US authorized distributor, and 5 years with onsite diagnostics and repairs from an EU seller. Read the actual offer on the quote..
Lifecycle and maintenance
NVIDIA publishes no MTBF or service-life figure for this card, and it is current rather than deprecated in NVIDIA's enterprise software support lineup alongside B200, B300 and the RTX PRO 4500. Partner platforms shipped from May 2025, so the oldest fielded cards had roughly 16 months of service on 10 September 2026. Warranty as sold varies by channel rather than by part, with a US distributor quoting 3-year limited and an EU seller quoting 5 years with onsite diagnostics and repairs. Maintenance is ordinary PCIe card practice with one exception: the cooler is passive, so chassis fan and baffle condition is a GPU maintenance item rather than a chassis one.
Common failure points
- Passive cooling in an under-specified chassis — Sustained thermal throttling and lost throughput on a card that passes a short benchmark
- Wrong edition purchased — Delivered card reads 1,792 GB/s and carries an active double flow-through cooler
- Undeclared 450 W power cap — Throughput lands short of the datasheet with no fault logged
- GDDR7 memory errors — Correctable ECC counts climb and inference workers restart under sustained load
- PCIe-only scaling assumed to behave like NVLink — Multi-card jobs scale far below expectation on collective operations
- Slot and airflow budget exhausted before the power budget — A server accepts fewer cards than the power supply would allow
Inspection checklist
SKU suffix on the quote: match 900-2G153-0000-000 or 900-2G153-0000-200 exactly, because two US sellers priced those two suffixes 2,175 dollars apart and NVIDIA publishes no explanation of the difference. Bandwidth check against the Workstation Edition: this card is 1,597 GB/s, and a 1,792 GB/s figure on the listing means the seller has quoted the Workstation Edition instead. Memory capacity and type: confirm 96 GB GDDR7 with ECC on a 512-bit bus, not the HBM used by the SXM parts. Cooling variant: confirm passive dual-slot for air or single-slot full-length for liquid, since the passive card has no fans and depends entirely on chassis airflow. Power cap setting: the card ships up to 600 W configurable and OEMs sell a 450 W capped variant, so read the strapped limit rather than assuming 600 W.
Procurement channels
Authorized distribution, which no other part in this category offers. Named sellers with published prices on 10 September 2026 were Uniqcli and SabrePC in the United States and Servermall in the EU, and HPE and Lenovo both publish OEM-branded part numbers for it. This is the only Blackwell part here that can be quoted, ordered and received as a component rather than as a server line item.
Regional notes
Classification is open here: no export classification appears on the NVIDIA or HPE pages published as of 10 September 2026, so whether the card falls under 3A090 at all is a question for export counsel. BIS guidance of 31 May 2026 turns on the counterparty: a licence is required for any entity headquartered in Country Group D:5 or Macau, or with an ultimate parent there, from every destination outside the US, regardless of location. 91 FR 1684 of 15 January 2026 gates case-by-case China and Macau review at both total processing performance under 21,000 and DRAM bandwidth under 6,500 GB/s. It publishes 1,597 GB/s, under the bandwidth gate, and no TPP figure.