Neocloud Market Heterogeneity: Why Are AI Silicon Vendors Becoming Clouds?

Silicon Valley-based fabless Artificial Intelligence (AI) compute vendor Groq has entered into a deeper partnership with co-location player Equinix to build inference capacity for its European customers. This will serve Groq’s European customers, as well as Equinix Fabric users, who will be able to deploy inference workloads to GroqCloud, which is already available in the United States, Canada, and Saudi Arabia. Similar cases include:

• SambaNova: SambaCloud is a low-power Application-Specific Integrated Circuit (ASIC) inference cloud platform targeting open-source models.
• Cerebras: Inference Cloud is based on the wafer-scale CS-3 chip with a presence in North America and France.
• Recogni: Partnership with DataVolt to integrate its inference-optimized ASIC into an AI cloud in Saudi Arabia.

Other silicon vendors, like Positron, offer developer cloud access to test their alternative AI inference platform before investing in hardware. ABI Research forecasts that, once the vertically integrated concept is proven by Groq, for example, and reaches a wider customer base, other players may follow down the route of expanding beyond developer clouds to also offer full-stack cloud services and monetize their AI accelerators.

The uptake of challenger hardware by cloud providers has been limited, thus far, as incumbent cloud infrastructure from NVIDIA and, to a lesser extent, Intel and AMD has continued to take the lion’s share of orders. This was demonstrated by the Initial Public Offering (IPO) documentation submitted by Cerebras, which revealed a high level of concentration of one sovereign AI initiative in the Middle East. Existing platforms are deemed as lower risk, and there is a high degree of confidence that AI workloads will continue to run well on incumbents’ hardware. Neoclouds answer customer demand in an agile manner and, despite access to cheap venture capital seeking the next wave of AI returns, must still operate lean business models—especially given the Capital Expenditure (CAPEX) required to spin up NVIDIA or AMD infrastructure. The wider developer ecosystem is operating within NVIDIA’s CUDA moat, and this has proven remarkably watertight to date—nonetheless, neoclouds.

Challenger AI compute, such as Groq’s Language Processing Units (LPUs), Cerebas’ WS-3 chip, and SambaNova’s Reconfigurable Dataflow Unit (RDU) offer compelling alternatives to Graphics Processing Unit (GPU) compute from an efficiency perspective as they have been designed from the ground up for AI workloads, which is particularly important for this stage of AI’s inference-focused scaling. The prediction is that NVIDIA will remain on top, but for challengers offering cloud instances beyond simple developer access, there is a compelling path to monetization for AI silicon startups for several reasons, which can augment direct chip sales:

• Low Risk Access for Developers: Instead of insisting on risky (vis-à-vis GPUs) CAPEX in new compute hardware, developers and enterprises can test, run Proofs of Concept (PoCs), and quickly scale their Generative Artificial Intelligence (Gen AI) applications on alternative silicon yet to reach a critical mass of adoption.
• Neocloud Playbook: Inference platforms with easy access to open-source models have been implemented by many neoclouds to date, such as Sesterce, NexGen Cloud, and ClusterPower, and this is within reach of AI silicon challengers with significant AI software talent in-house.
• Preferential Hardware Procurement: Silicon vendors will be able to procure their own chips at cost, which means higher margins for their cloud operations, setting them apart from neoclouds procuring similar silicon.
• A Faster Path to Monetization: Platforms can generate revenue as soon as they are released, whereas silicon sales are generally associated with longer lead times, negotiations, and the timing of replacement cycles.

The massive inference scaling of Agentic AI and reasoning AI models will create disruption and also provide an opportunity for those able to offer the most efficient infrastructure catering to a diverse customer base.

AI compute challengers should look to replicate the success of Groq and others—especially those struggling with making inroads into enterprise and cloud markets with their alternative to NVIDIA GPUs. Important considerations include the following:

• Energy Procurement Has Become a Limiting Factor: Lead the value proposition with efficiency credentials and placate the Operational Expenditure (OPEX) fears of those looking to expand their AI compute stack.
• Openness and Upgradeability: Ensure that designs conform to Open Compute Project (OCP) standards and emphasize the simplicity of integrating server designs into existing enterprise and data center infrastructure. Modular designs will also simplify future scaling requirements.
• Collaborate with Software Efforts: Contribute to the efforts of the UXL Foundation and the PyTorch Foundation to simplify efforts to port software from proprietary ecosystems directly to alternative silicon offerings. Alignment with emerging Agentic AI standards Model Context Protocol (MCP) and Agent-to-Agent (A2A) is also important for tomorrow’s workloads.
• Out-of-the-Box Support: Any budding challenger AI silicon neoclouds must replicate NVIDIA-based neoclouds software platforms and offer simple support for popular open-source models like Llama, Mistral, DeepSeek, Qwen, and Stable Diffusion.
• Sovereign AI: Neoclouds, especially in Europe, have become synonymous with sovereign AI infrastructure. It is vital to offer instances in multiple regions with adherence to local data laws.

Neoclouds have, to date, mainly relied on NVIDIA hardware to build out their AI-first offerings. However, these players should watch the progress of new silicon (including the above challenger AI silicon neoclouds) and contemplate future hardware procurement roadmaps beyond NVIDIA, AMD, and Intel. Important considerations include the following:

• Infrastructure-First Neoclouds: Any inference-based solutions on highly-efficient AI compute hardware will also be highly competitive on price. Neoclouds with limited platform stickiness and basic software offerings will quickly lose out to alternative silicon platforms offering instances at a fraction of the price.
• Compute Mix Strength: Different challenger ASICs, while all able to process Gen AI, excel at different tasks, whether that is sequential processing or handling ultra-large models. Consider a blend of challenger silicon that can cater to different developer and enterprise needs, which can be accessed via a smart orchestration platform to squeeze out the most performance gains for customers.
• Consider Regional or “Sovereign” Silicon: There is a growing demand for silicon that is not only an alternative to NVIDIA and AMD, but also regionally diverse and not as tightly integrated in the same supply chains (foundries notwithstanding). This includes European silicon such as Tachyum and SiPearl, as well as Korean challengers such as Rebellions and FuriosaAI.