For years, co-founder and chief executive officer Jensen Huang and other higher-ups at Nvidia have been banging on the message that the company is more than its GPUs, that the chips that have become crucial to the rapid expansion of AI over the past several years are key parts of a larger platform story that also comes with other processors, networking, storage, software, reference architectures, AI models, and more.

Huang drove home that point several times during this two-plus hour keynote at Nvidia’s GTC 2026 conference last month in San Jose, California, and the mantra could be heard in other rooms and hallways of the San Jose McEnery Convention Center. That said, it’s not difficult to understand why, when thinking about Nvidia, the focus tends to fall on the GPUs. The company started out in 1993 as a GPU maker, and for more than a decade has led the charge to the top of the AI heap with those chips firmly in hand

Nvidia’s CUDA computing platform and programming model opened the door for the GPUs to move beyond its graphics rendering roots and into general-purpose computing in the datacenter, the NVLink interconnect enabled fast communication directly between GPUs, and newer software innovations like Dynamo – an open, distributed inference framework – all play key roles in Nvidia’s platform push, but it’s the compute complexes like Grace Blackwell today and the upcoming Vera-Rubin compute complex slated for later this year – and eventually the follow-on to that, the Vera-Feynman platform – that gets people jazzed.

They are also the key drivers behind the eye-watering revenue numbers Nvidia generates. The company ended its FY 2026 with $215.9 billion in revenue, a 6.5 percent year-over-year jump, and include $68.1 billion in the FY fourth quarter that ended in January. In all, Nvidia’s datacenter revenue was $193.7 billion and $62.3 billion, respectively.

But behind those GPU-driven numbers are the innovations behind the GPUs themselves, and Nvidia executives are pushing the point again this week in the wake of what they say are record-setting AI inference performances on MLPerf benchmarks.

The vendor’s co-design strategy that includes hardware, software, and AI models is what drove the AI inference performance numbers in such measures as minimum token rate and shorter time to the first token, according to Dave Salvatore, director of accelerated computing products for Nvidia. The company no doubt delivers powerful GPUs to run power-hungry AI and agentic AI workloads, but it’s the advances in software that is making much of the different, he told The Next Platform.

“People point at Nvidia and go, ‘Well, Nvidia because they have these great GPUs. We do have amazing GPUs,” Salvatore said. “The technology and the architectural innovations that we are making in our GPUs really represent the leading edge of what's possible for AI. There is a tendency with Nvidia to just think about our GPUs. But we are a datacenter platform company, which means GPUs are just the beginning.”

Nvidia also has brought in outside help, through its $20 billion “acquihire” in December of startup Groq’s development team and the licensing deal for its LPU engines for AI inferencing, a deal that began to show its fruits at GTC.

MLPerf is an industry standard run by the MLCommons Consortium, and the latest results were run on the MLPerf Inference v6.0 benchmark suite, which includes datacenter tests that are either new or updated. The tests in green below are new, according to Salvatore.

The new tests included DeepSeek-R1 Interactive that tested token deliver speed and reduced time to first token, the GPT-OSS-120B, a mixture-of-experts reasoning model, and Qwen3-VL-235B-A22B, a multimodel vision-language mode. Some were tested in multiple environments, as seen below: offline, server, and interactive.

“These different workloads really are a good representation of a lot what's happening out there in the market in terms datacenter AI,” he said. “An important thing that the consortium understands the speed which this market is moving and understand, In version 6, there was a concerted effort to update the workloads and bring them forward to better align to where the market is today.”

The tests largely address inferencing, which Nvidia executives arguing has overtaken training as the primary AI and agentic AI workload. For large language models (LLMs) and reasoning models, the coin of the realm are tokens – Huang last month called them the “new commodity” – and token generation, both the speed and the cost, are bottom line for vendors. Salvatore echoed Huang’s argument made at GTC, that Nvidia’s platforms – while expensive – improves token generation.

“Increases in token generation or increases in performance basically generate more revenue, they reduce costs, they get you more value from the same infrastructure,” he said. “This is the 'so what' of the performance from these latest MLPerf results.”

Nvidia used systems built on its Blackwell Ultra GPUs – and with 14 partners, from OEMs like Dell Technologies and HPE to cloud services providers such as Google Cloud – submitting results. The Nvidia-powered systems delivered the highest token throughput in a range of workloads.

The company also noted speed improvements with the GB300 NVL72 v6.0 over v5.1, ranging from 1.21 times in the Llama 3.1 405B offline benchmark to 2.77 times for DeepSeek-R1 server test.

“In just the past six months, we have been able to nearly triple our performance on DeepSeek-R1, which is a very popular reasoning model being used in quite a number of places,” Salvatore said, noting that such performance improvements translate into reduced costs and better scale. “Being able to perform at scale, it becomes hugely important in terms of thinking about token cost, because in order to service that many users, you have to be cranking out a whole lot of tokens. In order to do that, you have to do it in a way that's cost-effective.”

Nvidia’s work in software is key to the improved results, Salvatore said. That includes both what Nvidia is doing in-house as well as its embrace of open inference frameworks, like TLLM and SGLang.

Inside Nvidia, he pointed to Dynamo, the inference framework that allows for disaggregated serving, which splits the prefill and decode stages of inferencing among multiple GPUs to optimize resource utilization, which drives down the cost of tokens. In the DeepSeek-R1 Interactive benchmark, the Nvidia-powered system reached 250,634 tokens generated per second, which brought down the cost to 30 cents per 1 million tokens generated.

He also stressed improvements in TensorRT-LLM, an open library that accelerates LLM inferencing on its GPUs through such capabilities as parallelism techniques and multi-token prediction, which enables language models to learn to predict multiple future tokens simultaneously, rather than just the next single token.

In CUDA, another optimization is kernel fusion, where Nvidia is “able to take several kernels and bring them together to make one slightly larger kernel, which can dramatically speed up all the work that all those kernels would have done individually,” Salvatore said. “In a similar way, we can sometimes overlap kernels, where before one kernel completes its work, another one can kick its work off. And again, bringing them together a little bit helps to speed up end-to-end processing time.”

Nvidia will always be first and foremost, known for its big, powerful, and expensive GPUs. That said, executives will continue to tout the co-design story, noting that ongoing software improvements and such offerings as reference architectures and AI factory designs will enable those chips and – despite their cost – make using them a cost advantage.