Intel’s Arc Pro B70 has reportedly outrun NVIDIA’s flagship RTX 5090D on a DeepSeek R1 inference workload while costing roughly a quarter as much, according to a benchmark write-up published by Wccftech. The result, which Wccftech says sees the Arc Pro B70 delivering over 2,000 tokens per second on DeepSeek’s reasoning model, lands at a delicate moment for the AI accelerator market: buyers are actively hunting for cheaper ways to serve open-weights models, and DeepSeek itself is reportedly working on custom silicon to reduce its dependence on both NVIDIA and Huawei.
Key takeaways
- Wccftech reports Intel’s Arc Pro B70 beat NVIDIA’s RTX 5090D running the DeepSeek R1 large language model, clocking over 2,000 tokens per second in the tested configuration.
- The Arc Pro B70 reportedly costs about a quarter of the RTX 5090D, sharply changing the price-per-token equation for on-premise DeepSeek inference.
- The result matters most for developers running open-weights reasoning models locally, where memory-bound inference has historically favoured NVIDIA’s top-tier consumer and workstation GPUs.
- DeepSeek is separately reported by Wccftech and capacityglobal.com to be developing its own inference chip, aimed at reducing reliance on NVIDIA and Huawei.
- Neither Intel nor NVIDIA has issued a formal response to the benchmark at the time of writing.
What Wccftech reports about the DeepSeek R1 benchmark
According to Wccftech, Intel’s Arc Pro B70 has beaten NVIDIA’s RTX 5090D specifically on DeepSeek’s R1 reasoning model, and did so while offering over 2,000 tokens per second in the tested setup. Wccftech frames the result as notable not only because Intel’s professional Arc card came out ahead of NVIDIA’s China-market flagship, but because the RTX 5090D reportedly costs around four times as much. That ratio, if it holds up in independent testing, would represent a significant shift in the price-per-token metric that increasingly dictates GPU choice for serving open-weights models such as DeepSeek R1.
The Wccftech headline focuses on a single, narrow claim: that in this particular DeepSeek R1 configuration, the Arc Pro B70 both outran the RTX 5090D and did so far more cheaply. It does not claim that the Arc Pro B70 is faster than the RTX 5090D across the board, nor across other models, precisions or batch sizes. Readers evaluating the result for their own AI models database planning should treat it as a single data point in one workload, pending broader third-party benchmarks.
Why DeepSeek R1 is the workload to watch
DeepSeek R1 has become one of the most-watched open-weights reasoning models on the market, and its inference profile is unusual: long chains of thought, heavy key-value cache usage, and a strong preference for GPUs with generous memory bandwidth. That combination is precisely where the balance between raw compute and memory subsystem design matters most, and it is why a mid-tier professional card can sometimes surprise a nominally more powerful consumer flagship. Wccftech’s write-up positions the Arc Pro B70’s result in that context, arguing that DeepSeek R1’s memory-heavy behaviour rewards Intel’s architectural choices.
For teams sizing hardware for local DeepSeek deployments, the practical takeaway is that headline FLOPS matter less than sustained tokens per second on the actual model. Our free VRAM calculator is designed for exactly this kind of planning, letting readers cross-check whether a given card can even hold the DeepSeek R1 weights and cache at their target context length before worrying about throughput.
How the two cards compare on the reported numbers
The Wccftech piece frames the story primarily as a price-performance upset. Only a subset of specifications is directly referenced in the source, so the table below sticks strictly to what Wccftech reports and what is publicly known about the products by name; anything not stated in the source is left blank rather than guessed at.
| Card | Reported DeepSeek R1 throughput | Relative cost (per Wccftech) |
|---|---|---|
| Intel Arc Pro B70 | Over 2,000 tokens/s | Roughly a quarter of the RTX 5090D |
| NVIDIA RTX 5090D | Beaten by the Arc Pro B70 in the same test | Baseline (roughly 4x the Arc Pro B70) |
For deeper cost modelling around models like DeepSeek R1, our AI price-performance index tracks how these ratios shift across generations and workloads, and our best GPUs for AI roundup covers the wider set of contenders that developers are weighing this year.
What the result means for local and on-prem DeepSeek deployment
Cost per token is now the primary driver of many hardware decisions in the open-weights ecosystem, particularly for teams that have chosen to self-host DeepSeek models rather than hitting an API. If Wccftech’s numbers are borne out by independent benchmarks, the Arc Pro B70 could shift the calculus for small studios, research labs and enterprise pilots that were previously assuming they needed NVIDIA’s top-tier silicon to hit interactive tokens-per-second targets on DeepSeek R1.
Even accepting the caveat that this is one benchmark on one model, the reported four-to-one price gap is large enough that even a much smaller performance win in Intel’s favour would still translate into materially cheaper inference. Teams weighing whether that changes their build-versus-buy decision can model both sides using our self-hosting vs API calculator, which contrasts the amortised cost of on-prem GPUs against DeepSeek’s hosted pricing.
The wider DeepSeek hardware picture
The Arc Pro B70 story arrives alongside another shift in DeepSeek’s own hardware strategy. Wccftech reports that DeepSeek is building its own inference chip to break free from both NVIDIA and Huawei, and capacityglobal.com similarly reports that the Chinese lab is developing an in-house AI chip to cut its reliance on Nvidia and Huawei. Neither outlet, in the snippets provided, states a shipping date or detailed specification.
Read together, these strands point in a consistent direction: the DeepSeek ecosystem is diversifying its silicon options at both ends. Intel is emerging as a credible cheaper alternative for running DeepSeek’s models externally, while DeepSeek is reportedly building its own chip to serve them internally. For developers, both trends widen the set of viable inference targets beyond the NVIDIA-only default that dominated the previous cycle. Readers tracking the model side of that ecosystem can follow updates on our DeepSeek V4 page.
Caveats and what still needs verifying
A single-workload benchmark, however striking, is not a general verdict. Wccftech’s report focuses on DeepSeek R1 in a specific configuration; it does not, in the snippet provided, detail quantisation level, context length, batch size, or the software stack used on either card. All of those variables can swing tokens-per-second results substantially, and Intel and NVIDIA drivers continue to evolve. Until independent testers replicate the result on the same model and disclose their setup, the safest reading is that the Arc Pro B70 is a serious contender for DeepSeek R1 inference at its price point, not that it has generally overtaken the RTX 5090D.
It is also worth noting that the RTX 5090D is a China-market variant of NVIDIA’s flagship, subject to export-driven design constraints. That context is relevant to the price comparison Wccftech draws, since pricing and availability for the 5090D are shaped by policy as well as by market forces.
Frequently asked questions
What did Intel’s Arc Pro B70 reportedly do in the DeepSeek R1 test? According to Wccftech, the Arc Pro B70 beat NVIDIA’s RTX 5090D running DeepSeek R1 and delivered over 2,000 tokens per second in the tested configuration.
How much cheaper is the Arc Pro B70 than the RTX 5090D? Wccftech reports that the Arc Pro B70 costs roughly a quarter of what the RTX 5090D does, though exact regional pricing was not detailed in the snippet.
Does this mean the Arc Pro B70 is faster than the RTX 5090D overall? No. The reported result is specific to DeepSeek R1 in one configuration. Wccftech does not claim general superiority across other models, precisions or workloads.
Is DeepSeek really building its own chip? Both Wccftech and capacityglobal.com report that DeepSeek is developing an in-house AI inference chip aimed at reducing its reliance on NVIDIA and Huawei. Neither snippet cites a launch date.
What should developers do with this information? Treat it as a strong signal that non-NVIDIA hardware is becoming competitive for DeepSeek inference, and re-run price-per-token calculations for planned deployments once independent benchmarks appear.
The bottom line
If Wccftech’s numbers survive independent scrutiny, the Arc Pro B70 will have shown that a professional Intel card can not only keep up with but out-serve NVIDIA’s China-market flagship on one of the most influential open-weights reasoning models, at roughly a quarter of the cost. Combined with the separate reports that DeepSeek is building its own inference chip, the overall picture is one of a maturing, more competitive silicon landscape around DeepSeek’s models. For anyone planning on-prem deployments over the next few quarters, that shift is worth pricing into hardware decisions today, even before the wider benchmark community weighs in.
Sources: news.google.com. Reported July 08, 2026.
