B300 supply is no longer early. It is still scarce. NVIDIA B300 specs vs B200, a dated October 2026 price band, power and cooling facts, and what to prepare before you list.
B300 supply is no longer early. It is still scarce, and the window for being ready is open. If you are weighing whether to buy, host or list the NVIDIA B300 GPU, the useful questions are what the card changes for buyers, how crowded the market already is, what a B300 rents for, and what your facility needs. This guide covers the B300 against the B200, a dated October 2026 b300 price band with every source and its definition, the power and cooling facts, the risks of being first or late, and how to list. One disclosure: packet.ai lists supplier capacity, so we have an interest in this topic. The prices are market figures from public trackers, not what any supplier earns, and this guide sets out no packet.ai payout or fee terms.
Key takeaways
The NVIDIA B300 is the Blackwell Ultra GPU, and its headline change is memory. The chip is specified at up to 288 GB of HBM3e, and NVIDIA lists 270 GB per GPU for HGX B300 servers against about 180 GB for HGX B200, so b300 memory is about 1.5x higher with similar bandwidth. Integrator summaries of NVIDIA's datasheet also put dense FP4 throughput about 1.5x higher. For buyers, b300 vs b200 is mostly a question of whether the workload is limited by memory capacity, which is the same logic as in Got H200s? Where inference demand is heading.
Sources: NVIDIA product pages for the H100 and H200, NVIDIA DGX B200 and B300 documentation, and integrator documentation for B200 and B300 power settings, retrieved October 29, 2026. B300 memory and power follow NVIDIA's HGX and GB300 NVL72 documentation; some OEMs advertise liquid-cooled HGX B300 configurations at higher power, and figures can differ by chassis.
One documentation wrinkle is worth knowing before you compare quotes. NVIDIA specifies the Blackwell Ultra chip at up to 288 GB of HBM3e and says the capacity varies by product: 270 GB per GPU (2.1 TB per eight) for HGX B300 and 279 GB for GB300 NVL72, while some DGX B300 documents quote 2.3 TB for eight GPUs. When a vendor quotes b300 288gb, ask which number they mean, and plan with the figure for the exact system. A weights-only calculation using 270 GB shows what the extra memory buys.
Arithmetic only: weights at 1 byte per parameter for FP8 and 2 for FP16, ignoring activations, runtime overhead and KV cache, using about 180 GB for an HGX B200 and 270 GB for an HGX B300. It is not a benchmark.
The B300 also comes in three shapes that matter commercially. The DGX B300 is NVIDIA's own eight-GPU appliance. The HGX B300 is the eight-GPU baseboard that OEMs build into their own b300 server designs. And GB300 NVL72 is the liquid-cooled rack-scale system covered in GB300 NVL72 vs GB200 NVL72 and GB200 NVL72 explained. Which one you can host depends on your facility, not just your budget.
The common pitch is that early B300 supply gets noticed and late supply gets compared. By late October 2026 the market has moved past that. One GPU price tracker listed 38 providers renting the b300 gpu on October 28, and NVIDIA's own filings show the ramp: its quarterly report for the period ended April 26, 2026 describes the ramp of its Blackwell 300 products, and the report for the period ended July 26, 2026 says Data Center revenue of $89.0 billion was driven by the ramp of Blackwell Ultra infrastructure. Being first no longer sets a supplier apart.
What still sets a supplier apart is being ready and being easy to compare. Trackers list B300 offers side by side, so buyers see price, region, cluster size and availability together. Demand is also visible: packet.ai's demand board currently lists buyer requests for B300 capacity, which is one signal and not a market survey. The b300 gpu provider that wins on that comparison is the one that can deliver full eight-GPU nodes, in a stated region, with the power and networking already in place.
Our view
Our reading is that the premium for B300 supply comes from scarcity of ready capacity, not from the label. Scarcity, readiness and commercial advantage are three different things: scarcity is how much capacity exists, readiness is whether you can deploy it, and advantage is whether it earns more than it costs. The first two can support the third but do not guarantee it, which is why verified, deployable nodes with clear specifications matter more than timing.
As with the H200, the b300 price depends on what a tracker counts, so the table gives each reading with its type and date. Retrieved October 29, 2026, and the snapshots are not all from one day.
Read together, they show a wide spread. On-demand list prices run from about $6.60 to $17.80 per GPU-hour, spot offers start near $4.30 on one tracker, and the median has risen about 14% in three months, which is consistent with tight supply, although price alone does not prove scarcity; that would take evidence on inventory and lead times. A rising median does not prove rising demand on its own, because supply and the provider mix also move it, and some trackers are run by providers that sell B300 capacity, so treat each as one vendor's view. Buying is the other option: system-level guides put an eight-GPU DGX B300 at about $300,000 to $350,000, which is a different calculation from renting.
Our view
New GPU classes tend to price at a premium and then converge as supply arrives. SemiAnalysis reported that its H100 one-year contract index fell to a low in October 2025 and then rose almost 40% by March 2026, so the direction can reverse in either direction. Use a conservative realized rate for any b300 per hour plan, not a list price, and read GPU utilization explained and what an idle GPU costs before you model revenue.
Power. NVIDIA's DGX B300 user guide lists about 14.5 kW for the system, with 15 kW as the system maximum for the version with power supplies, in a 10U chassis with twelve 3.2 kW supplies in an N+N arrangement on 200 to 240 V input. A busbar version uses 54 V DC. In rack terms, three nodes is about 43.5 kW and four is about 58 kW, against a standard colocation rack that one operator's guide puts at 5 to 15 kW, which is where what colo owners earn as landlords versus operators matters. Cooling and the b300 liquid cooling question. Many guides say the B300 requires liquid cooling. That is only partly true. GB300 NVL72 racks are fully liquid-cooled and some OEM configurations run the GPUs at up to 1,400 W with direct liquid cooling, but NVIDIA lists HGX B300 at up to 1,100 W, integrator documentation describes NVIDIA's DGX B300 as air-cooled with front-to-back airflow, and Supermicro lists an 8U air-cooled HGX B300 with the GPUs at about 1,100 W. Either way, the per-rack power density is the real constraint. Keep three numbers apart: per-GPU power (up to 1,100 W for HGX B300), server power (about 14.5 to 15 kW for a DGX B300) and rack power, which is what your feeds and cooling must deliver for every node.
Networking. A DGX B300 pairs its eight GPUs with eight ConnectX-8 adapters at up to 800 Gb/s, so size your fabric for the nodes you plan to sell. That is a DGX figure, and the top NVIDIA-documented option is up to 800 Gb/s per GPU, so check the adapters and fabric in the exact server you are buying. Monitoring and benchmarks. Track allocated, busy and billable hours from the first day and benchmark with the batch sizes and context lengths buyers will test. Licensing, warranty and isolation. Confirm the software terms and that your warranty covers multi-tenant, around-the-clock use, and keep tenant workloads separate. Lead times. One provider guide reports 8 to 12 weeks for DGX B300 systems, so confirm current lead times with your vendor before you commit to a go-live date.
The next generation. NVIDIA's CEO said in the Q2 fiscal 2027 results on August 26, 2026 that Vera Rubin is now in full production. A new generation does not erase demand for Blackwell Ultra, but it changes the buyers' comparison set over the life of a loan or lease. An announced production milestone is not a delivery date for any particular customer or configuration, so the real comparison is B300 capacity on a known timeline against what a given buyer can actually get later. Price cycles. A premium paid for scarce capacity can fall as supply arrives, as the H100 index shows. Financing. If you borrow against the hardware, lenders will look at contracts, utilization and residual value, as covered in GPU financing and utilization. Utilization and commitments. A B300 that sits idle costs more per hour than an H100 because the fixed cost is higher, and if you hold reserved capacity you do not use, see six options for reserved capacity you do not use. Concentration. Big B300 requests tend to be large and lumpy, so a few buyers can matter a lot.
Whether you plan to rent out b300 capacity directly, list b300 nodes on a platform, or offer b300 gpu hosting for a client, the routes are the same four: a marketplace or supplier program, direct sales of reserved capacity to a buyer, a white-label cloud under your own brand, or colocation. The five ways to turn idle GPUs into revenue compares them, and sibling guides cover other cards: Got H200s? Where inference demand is heading, Rent out RTX PRO 6000 and Rent out A100 and L40S. The b300 cloud market is crowded enough that buyers can rent b300 time from many places, so the questions to answer are what you can deliver, in which region, and how fast.
If you are ready, the B300 supplier page describes listing B300s, and you can apply with your B300s. If you are weighing a purchase first, the B200 vs H200 ROI framework covers the economics, and the GPU provider glossary defines terms such as SXM, NVLink, HBM and NVL72. You may also be asking whether to buy b300 systems outright or rent, and the price table above is the starting point for that comparison.
Retrieved October 29, 2026. Prices change daily. Several trackers are run by providers that sell B300 capacity, and product documentation varies by chassis.
Last reviewed: October 29, 2026. This is general information, not financial or engineering advice. It contains no earnings figures and no guarantee of income, and the prices are market figures that change daily. For packet.ai's supplier information, see the supplier page.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start building →