Nvidia crowned AI networking’s top dog, but its throne isn’t quite safe

Gartner named Nvidia the 'company to beat' in AI network fabric, citing its AI accelerator dominance and broad data center networking portfolio, but warned that a shift from training to inference and the rise of agentic use cases could change the vendor landscape. The analyst firm also flagged Nvidia's lack of alignment with mainstream enterprise markets and the growing demand for open Ethernet-based alternatives from hyperscale customers, potentially opening doors for rivals like Arista, Cisco, and Marvell.

Nvidia’s claim to the world’s top networking company crown just got a Gartner-shaped boost after the analyst firm anointed it the "company to beat" in the AI network fabric space. Nvidia CEO Jensen Huang proudly proclaimed the chip giant was the “ largest networking company in the world https://www.sdxcentral.com/news/nvidia-thinks-it-has-dethroned-cisco-as-the-worlds-biggest-networking-company/ ” during the GTC event earlier this year. Gartner’s latest report said the vendor “stands out” for its AI accelerator dominance paired with its broad portfolio of data center networking offerings. Gartner heaped praise on Nvidia’s use of proprietary protocols and features, such as NVLink, which it touted as providing “high performance and reliability” across scale-out and scale-up networking for AI clusters. Nvidia’s front-runner status might not be as safe as it would like, though. Gartner’s report warned that the demand shift from training to inference, combined with the rise of agentic use cases, could well change the vendor landscape. The analyst firm also pointed to the rise of big hyperscale customers wanting more open Ethernet-based alternatives. Nvidia does, of course, offer some Ethernet-based offerings, like its Spectrum-X switches, sales of which are believed to be “ on par https://www.sdxcentral.com/news/nvidia-earnings-soar-as-chip-giants-networking-segment-continues-to-drive-ai-boom/ ” with those of its InfiniBand lines. IDC only recently suggested that Nvidia has strong-armed a 21.5% slice https://www.sdxcentral.com/news/nvidia-tops-ethernet-data-center-switch-charts-by-revenue-as-sales-soar/ of the Ethernet switch data center segment, with Spectrum-X demand seeing “significant traction” from hyperscalers and cloud providers building so-called AI factories. Gartner also billed Nvidia’s product and go-to-market GTM strategy as “lack ing alignment” with the mainstream enterprise market. Should it lose the hyperscale race to more open alternatives, that could open the door to a number of traditional networking rivals already circling, like Arista https://www.sdxcentral.com/news/arista-networks-joins-the-16t-race-with-ai-centric-switch-platform/ , Cisco https://www.sdxcentral.com/analysis/cisco-touts-networking-as-central-to-success-key-for-platform-growth/ , and Marvell https://www.sdxcentral.com/news/marvell-debuts-16t-coherent-pluggables-to-scale-ai-data-center-interconnects/ . Of that rival trio, Marvell was listed as the company to beat for AI data center optical connectivity. Huang himself crowned it the “next trillion-dollar company,” with Gartner adding to the pile of praise, describing Marvell’s end-to-end offerings as making it the brand to beat. Marvell’s offerings already span digital signal processor DSP -based pluggables https://www.sdxcentral.com/news/marvell-updates-16t-optical-dsp-line/ and analog optical components, and the vendor is planning to expand into scale-up co-packaged optics CPO for scale-up networking, an effort set to be aided by its $3.25 billion Celestial AI tech acquisition https://www.sdxcentral.com/news/marvell-completes-325b-acquisition-of-photonic-startup-celestial-ai/ . To close the gap, Gartner said competitors like Ayar Labs https://www.sdxcentral.com/news/ayar-labs-joins-nvlink-fusion-bringing-co-packaged-optics-to-nvidias-ai-infrastructure/ , Nvidia https://www.sdxcentral.com/news/lambda-offers-glimpse-into-nvidias-cpo-switch-enabling-more-tokens-through-substantial-power-savings/ , and even Broadcom https://www.sdxcentral.com/news/broadcom-ceo-hock-tan-cautious-on-silicon-photonics-cpo-need/ need to target integration or offer differentiated architectures. A potential challenge to Marvell’s success, though, could be CPO execution risk. With the market still emerging, it remains to be seen whether the technology is fully ready to make the leap to actual deployments. Microsoft is among the hyperscale players still toying with photonic interconnects, with the cloud giant on a hiring spree https://www.sdxcentral.com/news/microsoft-ramps-hiring-for-photonic-ai-interconnect-push/ to build out related engineering teams. The potential rise of alternative photonic architectures, such as near-packaged optics NPO , where the optical engines are embedded on the printed circuit board PCB immediately adjacent to the chip rather than on the same substrate or package, could also hamper Marvell’s CPO efforts, the analyst firm warns. AMD the brand to beat for CPUs, Broadcom the custom silicon king If Nvidia is winning on the network side, then AMD is the king of server central processing units CPUs , the report suggests. The chip giant was touted as the company to beat for enterprise AI processors due to its consolidation capabilities, agentic AI orchestration tools, and I/O bandwidth. Gartner praised the Lisa Su-led outfit ahead of its Advancing AI event in San Francisco for “consistent roadmap execution” and broad ecosystem support. Having eaten into a portion of the once Intel-dominated market, AMD’s EPYC line has time and again been popular among virtual machine VM vendors, finding homes in units like Microsoft Azure’s Da/Ea/Fasv7-series https://www.sdxcentral.com/news/microsoft-rolls-out-new-azure-vms-for-increased-network-performance/ VMs that debuted earlier this year. Its standard server-grade lines are also routinely supported by edge-centric subsets to bring inference and computing closer to applications, like via the Embedded 4005 series https://www.sdxcentral.com/news/amd-aims-for-the-edge-with-eypc-embedded-4005-processors/ . AMD finds itself as the Gartner-crowned frontrunner for an enterprise AI server CPU market that’s becoming increasingly diverse. The vendor and x86 stalwart Intel find themselves facing rising competition from Nvidia, through that vendor's earlier Grace https://www.sdxcentral.com/news/nvidias-grace-cpu-targets-ai-bottlenecks/ and forthcoming Vera https://www.sdxcentral.com/news/nvidia-vera-cpu-enters-full-production-pitched-at-agentic-ai-workloads/ CPUs, and Arm’s AGI chip https://www.sdxcentral.com/news/arm-follows-nvidia-into-a-crowded-enterprise-cpu-race/ . AMD also looks set to face challenges from hyperscalers wanting their own custom silicon, with Meta's MTIA https://www.sdxcentral.com/news/meta-could-start-production-of-iris-ai-chip-in-september-report/ , OpenAI’s Jalapeño https://www.sdxcentral.com/news/openai-gets-spicy-with-broadcom-via-ai-data-movement-optimized-jalape%C3%B1o-chip/ , and Amazon Web Services AWS Graviton https://www.sdxcentral.com/news/aws-launches-graviton5-powered-ec2-instances/ among the high-end hardware set to form the basis of integrated stacks. Helping develop a sizable portion of those custom silicon projects is Broadcom, with Garner lauding it as the company to beat for AI general processing units XPUs . The likes of Anthropic https://www.sdxcentral.com/news/broadcom-bulges-with-anthropic-ai-orders/ , Google https://www.sdxcentral.com/news/broadcom-secures-deal-with-google-for-next-gen-ai-racks-and-custom-tpus/ , and Apple https://www.sdxcentral.com/news/apple-working-with-broadcom-to-develop-ai-specific-server-chip-report/ join OpenAI https://www.sdxcentral.com/analysis/openais-broadcom-deal-signals-the-end-of-nvidias-ai-chip-monopoly/ and Meta https://www.sdxcentral.com/news/meta-partners-with-broadcom-for-development-of-multiple-generations-of-chips/ as having signed mammoth contracts with Broadcom to design custom AI accelerators. The analyst firm touted its leading-edge process nodes, advanced packaging, memory tech, and die-to-die interconnects as means for boosting its application-specific integrated circuit ASIC design efforts. For competitors to compete with Broadcom CEO Hock Tan’s custom silicon empire, Gartner said rivals would have to improve “weaknesses” in their intellectual property portfolio and offer flexible engagements with system-level designs to close the gap. “To better compete, tech provider executives should also address AI portfolio shortcomings by collaborating with third-party IP suppliers, open industry consortiums, or acquiring the necessary technology,” the report reads.