Full-stack AI platforms become the new enterprise infrastructure fight

Competition around full-stack AI platforms is moving enterprise infrastructure vendors away from selling isolated chips, storage or cloud capacity and toward bundled systems that combine silicon, software, networking and developer tooling. Executives from TensorWave, AMD, Supermicro, Vast Data and Crusoe described that shift in interviews with theCUBE, arguing that enterprise buyers are judging AI infrastructure by usable output from power and capital spend, not by component benchmarks alone.

Ilya Tabakh, vice president of innovation at TensorWave, said customers are increasingly focused on whether AI systems produce the expected business result for each watt or dollar spent. He tied that change to newer workload patterns, including agentic AI and “deep thinking” models, which can alter the balance between GPU and CPU demand.

TensorWave has introduced ScalarLM, an open-source platform for training, fine-tuning and serving large language models. The company says ScalarLM can run across AMD and Nvidia GPUs without code changes. TensorWave did not disclose customer count, revenue, usage volume or performance comparisons for the platform.

What is a full-stack AI platform?

A full-stack AI platform is an integrated infrastructure stack for AI workloads, spanning compute, storage, networking, software and operations tooling. The point is to reduce the work required to train, tune and serve models in production, rather than forcing customers to assemble and optimize each layer separately.

Linda Yang, director of AI product management at Supermicro, said open architecture means allowing multiple hardware designs and software stacks to run effectively in the same environment. Umair Piracha, senior director of supply chain and strategic sourcing for Crusoe Cloud, said Crusoe is trying to standardize around outcomes while preserving customer choice in the underlying technology.

That positioning is tied to the rise of neocloud providers, the AI-focused cloud operators that compete on dense GPU capacity and specialized infrastructure rather than broad general-purpose cloud catalogs. TensorWave’s Tabakh said hyperscalers remain strong generalists, but dense AI data centers require more GPUs connected closely together, which changes the operating model.

AMD is pitching its roadmap into that market. Ted Marena, AMD’s director of market development for GPU AI data center OEM and ODM partners, said AMD’s current MI355X architecture uses eight tightly connected GPUs in a pod. He said the next MI455X generation scales that design to 72 interconnected GPUs, with larger scale-up and scale-out networking intended to improve communication speed for neoclouds and their customers.

Storage is another constraint in the platform pitch. John Mao, vice president of global business development at Vast Data, said AI storage has to balance reads from solid-state drives with data movement across the network, creating bottlenecks that shift as one side improves. AMD, Supermicro and Vast Data are working on a combination of AMD EPYC processors, Supermicro H14 servers and Vast Data’s AI Operating System for AI storage workloads.

AMD is also preparing its next-generation EPYC processor, code-named Venice, after announcing in May that the chip was ramping production in Taiwan. Derek Dicker, corporate vice president of AMD’s Enterprise Business Group, said Venice is expected to deliver a 70% performance increase generation over generation and support PCIe Gen 6. Penny Tseng of AMD said the company plans to bring what it describes as the first x86 PCIe Gen 6 CPU to market around November.

Supermicro plans to integrate Venice across several server lines. Vik Malyala, the company’s chief business officer, cited 1U and 2U systems, liquid-cooled FlexTwin platforms and SuperBlade systems with integrated switching as target designs. The broader signal is straightforward: AI infrastructure vendors are trying to own more of the stack because the bottlenecks have spread beyond the GPU.

This story draws on original reporting from SiliconANGLE.