B200 Vs H200: Specs, Benchmarks and Rental Prices Compared
Summary
B200 leads H200 on every performance spec and roughly triples throughput in a same-software, same-weight-format MLPerf result.
It has 180 GB of memory against 141 GB, 7.7 TB/s of bandwidth against 4.8 TB/s and NVLink at 1.8 TB/s against 900 GB/s. B200 costs more to run, drawing up to 1,000 W per GPU against 700 W and about 40% more power per DGX system. It also rents for roughly 1.4–1.6x the H200 rate in November 2026.
Choose B200 for new capacity that serves large models at high concurrency, long context or FP4. Choose H200 to expand an existing Hopper deployment, to fit a site sized for Hopper power, or to serve low-concurrency, short-context workloads.
For an operator, the decision comes down to three numbers: throughput per server, power per rack and revenue per GPU-hour.
Comparison Table
The table sets the H200 SXM against the HGX B200, both in the 8-GPU format, and adds the two rows operators plan around. Every hardware figure is NVIDIA's, sparse figures are labeled, and the rental row uses on-demand prices from Nebius for November 2026.
| Specification | H200 SXM | HGX B200 |
|---|---|---|
| Architecture | Hopper | Blackwell |
| Die design | Single die | Two reticle-limited dies joined at 10 TB/s; 208 billion transistors |
| Memory per GPU | 141 GB HBM3e | 180 GB HBM3E |
| Memory per 8-GPU system | 1,128 GB | 1.4 TB |
| Memory bandwidth per GPU | 4.8 TB/s | 7.7 TB/s |
| NVLink per GPU | 900 GB/s (4th generation) | 1.8 TB/s (5th generation) |
| FP8 Tensor Core per GPU | 3,958 TFLOPS (sparse) | 9 PFLOPS (sparse); 4.5 PFLOPS dense |
| FP4 | Not supported | Supported (NVFP4); 144 PFLOPS sparse, 72 PFLOPS dense per 8-GPU system |
| Max power per GPU | Up to 700 W | Configurable up to 1,000 W |
| 8-GPU DGX system power | 10.2 kW max | About 14.3 kW max |
| On-demand rent, Nebius, November 2026 | $5.40 per GPU-hour | $8.50 per GPU-hour |
The H200 column comes from NVIDIA's H200 specifications, which list 141 GB of HBM3e at 4.8 TB/s, up to 700 W and 900 GB/s NVLink. The B200 per-GPU memory, bandwidth and power figures come from NVIDIA's Blackwell datasheet. NVIDIA's HGX page puts the B200 system total at 1.4 TB across eight GPUs, and the HGX B200 reference page lists the rest of the system specification.
Architecture and Memory: What Blackwell Changes
B200 is a full generation ahead of H200, with 28% more memory, about 60% more bandwidth and a 4-bit number format driving its inference gains. Memory and precision decide how much paid work a server holds at once, which is what an operator sells.
The Blackwell architecture joins two reticle-limited dies with a 10 TB/s chip-to-chip link. Each die is the largest a lithography tool can print in one exposure, and together they form one GPU with 208 billion transistors. H200 is a single Hopper die.
Memory sets the ceiling on concurrent work. To generate each token, the GPU reads the model weights and the KV cache, the stored context of every active request. Bandwidth decides how fast those reads happen, and capacity decides how many long requests fit in memory together. Moving from 141 GB at 4.8 TB/s to 180 GB at 7.7 TB/s lets a B200 server hold more simultaneous long-context requests.
NVFP4 is Blackwell's 4-bit floating-point format, which scales small blocks of values to preserve accuracy. Blackwell's fifth-generation Tensor Cores run it in hardware, while Hopper's fourth generation stops at FP8. A headline multiple that sets B200 at FP4 against H200 at FP8 compares two different precisions. It overstates the gain a customer sees running the same model.
Sparse and dense labels need the same care. NVIDIA's sparse figures assume 2:4 structured sparsity, where two of every four values are zero, and come out at twice the dense rate. Dense numbers are the safer basis for pricing an offtake contract against realistic throughput. The step from B200 to B300 is a separate comparison, covered in its own article.
Benchmarks: About 3X on the Same Software, More at High Concurrency
In the newest same-software, same-weight-format result, an 8-GPU B200 system delivered about 3x the throughput of its H200 equivalent. The source is the MLPerf Inference v6.0 results from MLCommons: closed division, available systems, the gpt-oss-120b model at MXFP4 precision. Red Hat submitted both systems on LLM-D v0.5.0 and Red Hat Enterprise Linux 9.6, with slightly different OpenShift versions.
| MLPerf v6.0 scenario | 8x H200 (tokens/s) | 8x B200 (tokens/s) | B200 multiple |
|---|---|---|---|
| Offline | 28,680.0 | 93,070.7 | 3.2x |
| Server | 24,103.2 | 71,588.1 | 3.0x |
Offline measures batch throughput with every query available at once. Server sends queries at random intervals under latency limits, closer to live customer traffic.
The H200 system ran on an IBM Cloud instance with 80 host CPU cores, against 112 in the B200 system. The submitter, model, weight format and LLM-D stack match, while the host hardware differs. Because Hopper has no FP4 hardware, part of the gap likely reflects Blackwell computing natively in 4-bit.
A 2026 Dell engineering study, with NVIDIA's performance engineering team credited as a contributor, tested both chips on the same vLLM and NVIDIA AIPerf stack across three models. They were Llama 3.3 70B at FP8, Qwen3.5 122B MoE at BF16 and Mistral Medium 3.5 128B at FP8. At a concurrency of 128, B200 delivered three to eight times the H200's throughput. It cut inter-token latency by up to 88% and returned the first token up to 18 times faster on long-context workloads.
Concurrency is the number of requests a server handles at the same time, and the gap widens as it rises. Under heavy load with long prompts, H200 runs short of memory and bandwidth first, while B200 keeps more requests in flight. Longer context makes each request's KV cache larger, so the memory advantage counts for more. At short context the gain is smaller: Dell measured +173% for Llama 3.3 70B with 256-token prompts.
Agentic workloads compound the difference, because an agent chains many model calls and waits on each one. In Dell's test, a 10-step reasoning agent took about 36 minutes on H200 and under five minutes on B200. Dell's own conclusion is fair to the older chip: "The H200 remains a sound choice for legacy estates and low-concurrency, short-context workloads".
These results are best read as the throughput an operator can sell under its customers' real load. A customer sending short prompts at low concurrency won't see 3x, so the offtake contract should be priced on the workload it covers.
Power, Site and Rental Price: What B200 Costs to Run
A DGX B200 is specified at about 14.3 kW maximum against 10.2 kW for a DGX H200, about 40% more power per 8-GPU server. A site sized for H200 fits fewer B200 servers per rack. Rack density and rent per GPU-hour are the two numbers behind both the site plan and the offtake contract.
Both system figures come from NVIDIA's user guides for the DGX H200 and the DGX B200. Per GPU, B200 is configurable up to 1,000 W against up to 700 W for H200. HGX B200 servers from other builders follow their own power and cooling design, so check the builder's specification.
Take an illustrative rack with a 40 kW power budget, before networking and overhead. It holds three DGX H200 servers (30.6 kW) or two DGX B200 servers (28.6 kW), meaning 24 H200 GPUs against 16 B200 GPUs. On the MLPerf result above, two B200 servers would likely still out-produce three H200 servers. The site's cooling and power contract decide what is actually buildable.
Rental prices tell the revenue side. The table compares each provider's own H200 and B200 list rates in November 2026.
| Provider (November 2026) | H200 per GPU-hour | B200 per GPU-hour | B200 premium |
|---|---|---|---|
| Nebius, on-demand | $5.40 | $8.50 | 57% |
| Together AI GPU Clusters, pay-as-you-go | $5.99 | $8.19 | 37% |
Together AI's GPU Clusters price the gap lower than Nebius. Across both providers, B200 rents for roughly 1.4–1.6x the H200 rate. On throughput, MLPerf v6.0 shows about 3x, and Dell measured three to eight times at a concurrency of 128. Whether that spread becomes margin depends on your offtake contract, utilization and power cost, which no public price list captures.
Blackwell is also where NVIDIA's sales sit. NVIDIA's CFO Commentary for Q2 fiscal 2027 says "our Blackwell architecture remains the vast majority of our revenue," a statement that covers every Blackwell product, B200 and Blackwell Ultra included. For data center operators hosting either generation, power per server is the planning number that changes most.
Financing Either Generation: What a Lender Looks At
Both generations are financeable, and one March 2026 financing in Spring, Texas covered 10 H200 servers and 38 B200 servers in a single deal. That entry sits in USD.AI's loan history alongside two April 2026 financings. One covered 75 H200 servers in North Kansas City, Missouri, and the other 256 B200 GPUs across 32 servers in Spring, Texas.
As market practice, lenders underwrite the revenue behind the servers: the operator, the customer buying the compute and the offtake contract. A well-drafted contract with a creditworthy customer carries the deal for either GPU model. That is why operators finance both Hopper and Blackwell, sometimes in the same transaction.
A financing review for either system asks for the same documents:
- Equipment configuration and quantity
- Supplier and purchase order, or ownership records for GPUs already in service
- Site and colocation agreement
- Power and cooling readiness for the chosen servers
- The offtake contract or, for GPU-as-a-Service providers selling on-demand capacity, documented on-demand revenue
- System configuration and quantity. The server model, GPU count and build, such as HGX B200 or H200 SXM in an 8-GPU server.
- Supplier and purchase order. For new hardware, the order with the supplier; for servers already running, records for deployed servers.
- Deployment site and colocation agreement. Where the servers will run and the terms of the space.
- Power and cooling readiness. Evidence that the site supports the system's draw, about 14.3 kW per DGX B200 or 10.2 kW per DGX H200.
- Offtake contract or documented on-demand revenue. The revenue that repays the loan, with the customer and terms behind it.
The hardware choice shows up in two places: the purchase order and the power plan. A B200 deployment needs a site that delivers about 40% more power per server, and the review checks that readiness before funding. An H200 expansion at a Hopper site usually clears the power check more easily, because the site was built for that draw.
When to Choose Which
The right GPU follows from your contracts and your site. Three scenarios cover the common cases.
- New high-load capacity: B200. Customers running 70B+ models, long context, agentic workloads, high concurrency or FP4 get the full 3x gain or more.
- Hopper expansion: H200. It suits an existing Hopper deployment, a site sized near 10 kW per server, or short prompts at low concurrency.
- Mixed plan: both. Add B200 for new high-load contracts and keep H200 on existing ones. The March 2026 Spring, Texas financing covered both generations in one deal.
Each scenario starts from the offtake contract and the power the site can deliver.
Conclusion
B200 delivers about 3x H200's throughput on the same software stack and weight format, and more under heavy load. It rents for 1.4–1.6x as much and draws about 40% more power per server. H200 remains a sound fit for Hopper sites and lighter, short-context workloads. The choice turns on the load your customers send and the power your site can deliver.
GPU financing through USD.AI covers new deployments and refinances GPUs already in service, for either generation. Loans cover up to 80% of deployment cost, with the operator contributing at least 20% as equity, at fixed or floating rates over 3–5-year terms. Every deal is scored by Compute IQ on a 100-point scale, using the operator, the customer's credit and the offtake contract. The score sets financeability, the pricing tier, and the loan's size and structure.
Both the hardware choice and the financing start from the offtake contract, so bring your equipment plan and that contract to the first conversation. Talk to USD.AI
FAQ
Is the B200 better than the H200?
On performance, yes: an 8-GPU B200 system delivered about 3x the throughput of its H200 equivalent in the same-software, same-weight-format MLPerf Inference v6.0 result. B200 also draws more power and rents for more. H200 still fits Hopper sites and low-concurrency, short-context work.
How much power does a B200 use compared with an H200?
A B200 is configurable up to 1,000 W per GPU, against up to 700 W for an H200. At the system level, NVIDIA specifies a DGX B200 at about 14.3 kW maximum and a DGX H200 at 10.2 kW maximum.
What is NVFP4, and can the H200 run it?
NVFP4 is Blackwell's 4-bit floating-point format, run in hardware by the B200's fifth-generation Tensor Cores. The H200 has no FP4 hardware; its Tensor Cores support FP8, FP16, BF16, TF32 and INT8.
Where does the H100 fit next to the H200 and B200?
The H100 and H200 are both Hopper GPUs, and the H200 adds memory: 141 GB of HBM3e at 4.8 TB/s. The B200 is the next generation, built on Blackwell.
Editor Notes
- Unpublished posts to link when live. B300 vs B200 article (Architecture section, the sentence on the B300 comparison); GPU server and rack density article (Power section); GPU depreciation article (Financing section).
- DGX power links. The outline's
nvidia.com/en-us/data-center/dgx-h200/now redirects to NVIDIA's DGX platform page. The Power section links the DGX H100/H200 user guide (10.2 kW max) and the DGX B200 user guide (~14.3 kW max) instead of the gated datasheet. - Weighted-average APR by server type. Held back from USD.AI's 2026 YTD report under the Approved exceptions rule on rate figures and the Two tracks and Loan terms rules.
- 2026 YTD principal by server type. B200 $50M and H200 $20M held back because the report is internal and not published on usd.ai; the Blog posts rule allows only published deals.
- B200 server price. Held back because USD.AI's pricing reference says no current B200 server price may be stated.
- Borrower names. Names in the loan history held back under the Borrower names rule.
- Nebius pre-October 2026 prices. Held back because they couldn't be verified on the live page.
