Lenovo Blackwell lineup: ThinkSystem SR680a and SR780a V3 systems.

Most OEMs ship a Blackwell server. The interesting question for an enterprise buyer is what happens to the heat. An eight-GPU HGX B200 board now draws 1,000W per GPU, and a GB300 NVL72 rack pulls past 135 kilowatts in a single footprint. The thermal envelope, not the silicon, is what increasingly decides where a system can be installed, how much it costs to run, and whether a data center needs a chiller plant it does not have. That is the lens through which to read Lenovo's ThinkSystem GPU lineup, because cooling is the place Lenovo Infrastructure Solutions Group has spent two decades building a real engineering moat.

This is a spotlight on Lenovo as one of the qualified OEMs Rillor lists and can source against a forward contract, not a partnership announcement. Lenovo builds the hardware; Rillor standardizes the complete OEM system into a forward SKU, runs the contract, deposit, escrow, and NVIDIA channel compliance, and delivers the physical machine. The point is to help a buyer evaluating Lenovo ThinkSystem GPU servers understand what the SR680a V3, SR780a V3, and GB300 NVL72 rack are, where Neptune warm-water cooling genuinely changes the math, and how each maps to a standardized Rillor SKU.

Two HGX servers, one chassis decision

Lenovo's HGX-class Blackwell servers come in an air-cooled and a liquid-cooled variant, and the split is the most useful thing about the lineup. A buyer is not forced into liquid cooling to get density, and is not forced into a sprawling air-cooled box to keep deployment simple. The two systems carry the same NVIDIA GPU board and the same class of head node; what differs is how they reject heat and, as a direct consequence, how much rack space and facility infrastructure they demand.

ThinkSystem SR680a V3, air-cooled HGX B200

The ThinkSystem SR680a V3 is the air-cooled member of the family. It is an 8U server carrying eight NVIDIA B200 SXM6 GPUs (180GB HBM3 each, 1,000W per GPU) on a single HGX baseboard, with two fifth-generation Intel Xeon CPUs as the head node, up to 32 DDR5 DIMMs for a 4TB memory ceiling, up to 16 PCIe 5.0 NVMe bays, and ten PCIe 5.0 x16 slots that support 400Gb/s InfiniBand. The original SR680a V3 product guide documents the same chassis with DDR5-5600 RDIMMs and ConnectX-7 NDR400 adapters.

This is the workhorse configuration for a site that has not yet committed to facility water. It drops into a conventional high-density air-cooled hall, draws cooling from the room, and uses standard rear-door or in-row containment. The cost is footprint and fan power: 8U per node and a meaningful share of system wattage spent moving air. For a buyer doing a first Blackwell deployment in an existing colo, the SR680a V3 is the path of least resistance, and on Rillor it standardizes to the B200 system SKU, RIL-GX-B200-2T.

ThinkSystem SR780a V3, Neptune liquid-cooled

The ThinkSystem SR780a V3 is the same GPU complement in a fundamentally different thermal design. It is a 5U server, three rack units shorter than the air-cooled SR680a V3, running eight NVIDIA B200 1,000W SXM6 GPUs (180GB HBM3e each) and dual fifth-generation Intel Xeon CPUs with 32 DDR5 DIMM slots and ConnectX-7 NDR400 networking. The density gain comes entirely from cooling. The SR780a V3 uses Lenovo's 6th-generation Neptune warm-water cooling with four independent loops, one each for the CPUs, the front GPUs, the rear GPUs, and the NVLink switches, and removes up to 75% of total server heat directly into warm water rather than into the room.

That 75% figure is the part worth internalizing. When three quarters of the heat leaves the chassis in liquid, the remaining air-handling burden collapses, fans shrink, and the same eight 1,000W GPUs that need 8U in air fit into 5U in liquid. For a buyer measuring deployments in GPUs per rack and watts per square foot, the SR780a V3 is the denser, more efficient expression of the identical compute, and on Rillor it standardizes to the B300-class forward SKU, RIL-GX-B300-2T, reflecting its position as Lenovo's flagship liquid-cooled HGX node.

The Neptune differentiator: warm water, no chiller

Plenty of vendors now offer direct-liquid cooling. What distinguishes Lenovo is that Neptune is an open-loop, warm-water design engineered to run without a chiller, and that design predates the current generation by years. The relevant specification lives in Lenovo's SC777 V4 Neptune product guide: the platform supports direct warm-water cooling at inlet temperatures up to 45C (113F), which means most customers need no chillers at all, only a dry cooler or the building's existing facility water loop.

The reason 45C inlet matters is economic, not academic. A chilled-water plant is expensive to build, expensive to run, and a constraint on where a high-density deployment can physically go. If a system tolerates 45C supply water, the data center can reject heat with ambient air across most climates for most of the year. Lenovo quantifies the payoff: up to 40% data-center energy savings from reduced air conditioning, an additional 7 to 10% direct energy savings from removing system fans, and near-100% heat recovery, since the warm water leaving the racks is hot enough to reuse. The October 2024 next-era Neptune announcement makes the headline claim plainly: 6th-generation Neptune lets 100kW-plus server racks run without specialized data-center air conditioning while powering the NVIDIA Blackwell platform, including GB200 Grace Blackwell.

For a buyer, the differentiator is a concrete planning question. If a deployment is power-constrained or chiller-constrained, and most modern AI deployments are one or both, the warm-water envelope decides whether the racks can be installed where they are needed without a facility upgrade. That is the case for the SR780a V3 over the SR680a V3 on its merits, and it is why we treat the air-versus-liquid decision as a first-class procurement variable in air versus direct-liquid cooling for Blackwell systems.

Rack scale: Lenovo GB200 and GB300 NVL72

Above the single-node servers, Lenovo builds NVIDIA's rack-scale Grace Blackwell platforms, where the whole rack behaves as one accelerator and where Neptune stops being a feature and becomes a requirement.

GB300 NVL72 on Neptune open-loop DLC

The Lenovo NVIDIA GB300 NVL72 is a rack-scale platform with 72 NVIDIA Blackwell B300 GPUs (279GB HBM3e each) and 36 NVIDIA Grace CPUs across 18 compute trays, all interconnected by fifth-generation NVLink at 1.8 TB/s in a full mesh, with ConnectX-8 adapters reaching up to 800 Gb/s. The power and thermal profile is where the engineering shows: the rack carries a 135 kW standard TDP and up to 155 kW peak, fed by six to eight power shelves at 33 kW each over a 50V DC bus-bar architecture. CPUs, GPUs, and NVSwitches are liquid-cooled, while the OSFP optical modules, drives, and power-distribution-board components are air-cooled, a hybrid warm-water-and-air design rather than a fully immersed one.

A 135 kW rack is not something you cool with room air. At that density the warm-water loop is not optional, and Lenovo's two decades of Neptune deployment experience is the reason the GB300 NVL72 is a system a buyer can install rather than a thermal science project. On Rillor, this rack maps to RIL-NVL72-GB300.

GB200 Grace Blackwell on the SC777 V4 / N1380

Lenovo's GB200 Grace Blackwell node, the SC777 V4 Neptune, shows where the warm-water platform began for this generation. Each SC777 V4 tray pairs two NVIDIA Grace processors (72 Arm Neoverse V2 cores each at a 3.1 GHz base clock) with four NVIDIA Blackwell B200 GPUs (186GB HBM3e each), and up to eight of those nodes slot into the 13U ThinkSystem N1380 Neptune enclosure, all on open-loop direct warm-water cooling for the full GB200 Grace Blackwell architecture. The N1380 is the enclosure Lenovo introduced alongside 6th-gen Neptune in late 2024, and it is the blueprint the GB300 NVL72 builds on.

For the rack-scale tradeoffs themselves, independent of vendor, our comparison in GB200 NVL72 versus GB300 NVL72 at rack scale covers what changes between the two generations at the rack level: GPU memory, NVLink generation, networking step from ConnectX-7 to ConnectX-8, and the power envelope that follows.

The components inside the chassis

A complete OEM GPU system is more than its accelerators, and the head-node and fabric choices materially affect what a buyer is taking delivery of. Lenovo's Blackwell systems standardize on a recognizable component set.

The CPU pairing is worth a buyer's attention. Lenovo's published HGX guides specify fifth-generation Intel Xeon for the SR680a V3 and SR780a V3 head nodes, and the broader ThinkSystem line supports Intel Xeon 6 Granite Rapids and AMD EPYC Turin (9555 and 9655). The choice between Granite Rapids and Turin is a real one, with implications for memory bandwidth, core count, and PCIe lanes feeding the GPUs, and we work through it in Granite Rapids versus EPYC Turin for GPU server head nodes.

On the fabric, the generational story is the move from ConnectX-7 at 400G NDR (the standard on the HGX servers) to ConnectX-8 at 800G (on the GB300 NVL72), paired with either Quantum-2 InfiniBand or Spectrum-X Ethernet switching. NVIDIA's own GB200 NVL72 documentation frames the same split: HGX B200 systems reach up to 400Gb/s through Quantum-2 InfiniBand and Spectrum-X, while GB200-class rack systems can step up to Quantum-X800 and Spectrum-X800 at up to 800Gb/s, with 1.8 TB/s of NVLink per GPU binding the rack into a single logical accelerator that NVIDIA cites at up to 30X faster real-time trillion-parameter LLM inference. The interconnect generation is the difference between a cluster of servers and a rack that trains as one machine.

Reliability, support, and consumption

Two operational factors round out the Lenovo case for an enterprise buyer, beyond silicon and cooling.

The first is consumption flexibility. Lenovo offers its TruScale model, a pay-for-what-you-use consumption framework, alongside outright purchase, and its hybrid AI portfolio with NVIDIA was set out in the March 2024 Smarter AI for All announcement. For a buyer modeling capex versus opex, the existence of a consumption path alongside a capital purchase widens the financing options. It is worth being precise about scope here: a Rillor forward contract is always a physical-delivery contract for the complete system, not a consumption lease. The TruScale option lives on Lenovo's side of the relationship; what Rillor standardizes and delivers is the machine.

The second is global support and supply reliability. Lenovo is a top-tier global infrastructure vendor with a worldwide service footprint, which matters for a multi-site or sovereign deployment where on-site break-fix coverage is part of the buying decision. Supply reliability is the variable a forward market is built to price: when a buyer signs a Rillor forward for a Lenovo system, the seller posts a performance bond and the balance sits with an independent escrow agent until delivery, so the OEM's ability to ship on the contracted date is backed by a real obligation. We explain what that bond does and does not cover in what a seller performance bond covers, and what it does not.

How the Lenovo lineup maps onto Rillor

Bringing it together, here is the practical mapping a buyer needs. Each Lenovo system below is standardized into a Rillor forward SKU: a bilateral OTC forward with physical delivery intended, a 10% deposit at execution, the balance in independent escrow until delivery, a seller performance bond, NVIDIA channel compliance with the end-customer of record captured, and the option to transfer the contract pre-delivery to another KYC'd buyer with Rillor and OEM approval.

The reason a buyer would commit forward rather than chase allocation is the same reason any serious procurement function uses a forward market: it locks delivery at a transparent, contract-backed price instead of negotiating one machine at a time against an opaque queue. The full catalog of standardized systems sits at the SKU index, the contract path for buyers is laid out on the buyers page, and live forward activity is visible on the marketplace. Because Lenovo is one of several qualified OEMs Rillor lists, a buyer can also compare the same B200 SKU across vendors, the Lenovo SR680a V3 against the Gigabyte G893 and G894 systems or the Dell PowerEdge XE9680L and XE9780, and avoid single-OEM lock-in across a multi-rack buildout. The standardization is the point: the forward contract resolves to a qualified configuration on delivery, and the buyer is never trapped in one vendor's queue.

For an enterprise buyer whose deployment is power- or chiller-constrained, the Lenovo case comes down to one sentence: Neptune warm-water cooling lets you run dense Blackwell, including a 135 kW GB300 NVL72 rack, without building a chiller plant, and Rillor is how you contract that system forward at a price you can defend.