CoreWeave Sandboxes Launch; Cerebras Hits $100B Market Cap

CoreWeave today launched Sandboxes, a new execution layer that provides secure, isolated environments for reinforcement learning, agent tool use, and model evaluation, marking a significant shift in how AI infrastructure is consumed. The product is available on a customer's own CoreWeave infrastructure or serverless through Weights & Biases, with access models including on-cluster via the CoreWeave Kubernetes Service and serverless via W&B. Separately, Cerebras Systems saw its stock nearly double on its first trading day, with the IPO priced at $185 and opening at $350, giving the AI chipmaker a market capitalization exceeding $100 billion. The company reported $510 million in revenue for 2025, up 76% year-over-year, driven by major partnerships with OpenAI and Amazon Web Services. OpenAI has committed to purchasing 750 megawatts of Cerebras inference compute in a deal valued at over $20 billion, with potential to expand to 2 gigawatts total. Together, these two developments signal that the AI infrastructure market is rapidly diversifying away from a pure GPU focus toward specialized chips, orchestration layers, and CPU-heavy architectures for agentic workloads. This shift will reshape spending patterns across the entire tech stack.

The OpenAI Deal Anchors Cerebras' Valuation

Cerebras' market cap surge reflects investor conviction that specialized AI inference chips will capture a growing share of the $500 billion-plus AI infrastructure market. The company's $510 million in 2025 revenue, representing 76% growth, is anchored by the OpenAI deal. That deal is a commitment to purchase 750 megawatts of Cerebras inference compute valued at over $20 billion, with a potential expansion to 2 gigawatts. This is not a speculative bet: OpenAI is a paying customer with real compute demand. The partnership with AWS further validates Cerebras' wafer-scale architecture for inference workloads, which the company claims offers superior performance per watt compared to Nvidia's H100 and B200 GPUs. Cerebras CEO Choi and Hock have positioned the company as the primary alternative to Nvidia for inference, arguing that the industry's obsession with training compute has overlooked the coming inference boom. The $100 billion valuation, while rich at roughly 196 times trailing revenue, reflects the market's expectation that Cerebras will capture 10% or more of the inference chip market within three years, especially as agentic AI drives demand for low-latency, high-throughput inference.

The deal structure matters as much as its headline number. A 750-megawatt commitment valued at over $20 billion prices Cerebras inference compute at roughly $27 per watt annually, a premium that reflects the latency and throughput advantages of wafer-scale architecture over conventional GPU clusters. If OpenAI exercises the full 2-gigawatt option, the total contract value could approach $55 billion, more than 100 times Cerebras' 2025 revenue. That kind of backlog transforms the company from a promising chip startup into a critical infrastructure provider with visibility that rivals hyperscaler capex plans. The AWS partnership adds a distribution channel Cerebras could not have built independently: AWS Marketplace gives enterprise customers a procurement path they already trust, lowering friction to choosing Cerebras over Nvidia for inference workloads. For investors, the combination of a locked-in anchor customer, a major cloud distribution deal, and 76% revenue growth justifies a premium multiple even before Cerebras proves it can manufacture at full scale.

CoreWeave's Layered Pricing Model

CoreWeave Sandboxes creates a new revenue stream by charging for secure execution environments on top of its existing GPU infrastructure. The product is available on a customer's own CoreWeave infrastructure or serverless through Weights & Biases, with access models including on-cluster via the CoreWeave Kubernetes Service and serverless via W&B. This layered pricing model allows CoreWeave to capture margin on both compute and orchestration, similar to how AWS charges for EC2 instances and then separately for services like Lambda and SageMaker. For enterprise customers running reinforcement learning workloads or agentic AI systems, Sandboxes eliminates the need to build and maintain their own secure sandboxing infrastructure, which typically requires dedicated engineering teams and complex Kubernetes configurations. The serverless option through Weights & Biases lowers the barrier to entry for smaller teams, potentially expanding CoreWeave's addressable market beyond the hyperscaler and large enterprise segment. Morgan Stanley analyst Shawn Kim noted in a recent report that orchestration increases system complexity, and products like Sandboxes that abstract away that complexity command premium pricing. CoreWeave's bet is that as agentic AI proliferates, the demand for secure, isolated execution environments will grow faster than the underlying GPU compute market itself. The company is effectively creating a new category of infrastructure software that sits between raw compute and the application layer.

The Competitive Reshuffle: Nvidia vs. The Field

The combination of CoreWeave Sandboxes and Cerebras' $100 billion valuation signals a fundamental reshuffling of the AI infrastructure competitive landscape. Nvidia has dominated the training phase with its CUDA ecosystem and H100/B200 GPUs, but the shift toward agentic AI and inference is opening doors for competitors. Cerebras is now a credible alternative for inference workloads, backed by OpenAI's $20 billion commitment and AWS distribution. CoreWeave, which built its business on Nvidia GPUs, is hedging its bet by adding an orchestration layer that works with any underlying hardware. The most significant competitive threat to Nvidia, however, comes from the CPU side. Meta is using tens of millions of Amazon Graviton CPUs for AI inference, and AMD has signed a $60 billion deal with Meta to supply 6 gigawatts of chips over five years. This CPU pivot is driven by the reality that agentic AI workloads require a different compute mix: more CPUs for orchestration, more memory bandwidth, and more networking. Nvidia's GPU-centric architecture is optimized for dense matrix math, not the branching logic and tool-calling patterns that characterize agentic systems. The market is signaling that the winner in AI infrastructure will be the company that can provide the full stack, not just the fastest GPU.

The memory layer adds another dimension to this shift. Morgan Stanley analyst Shawn Kim noted that orchestration increases system complexity, driving demand for high-bandwidth memory and networking components that Nvidia's current roadmap does not fully address. Agentic AI requires rapid context retrieval, tool-calling, and multi-step reasoning, all of which strain memory bandwidth in ways that GPU-centric designs were not optimized to handle. This is why companies like Texas Instruments are boosting in-house chip output to serve the AI infrastructure boom: the full stack increasingly resembles a telecom network, with specialized components for compute, memory, and interconnect rather than a monolithic GPU doing everything. The competitive implication is that Nvidia must either broaden its portfolio into memory and networking or accept margin compression as specialists capture each layer of the stack.

Downstream Effects on Hyperscalers, Fabs, and Enterprise Buyers

The downstream implications of these developments are rippling through the entire technology supply chain. For hyperscalers like Amazon, the CPU pivot is a direct benefit: Meta's purchase of tens of millions of Graviton CPUs validates Amazon's strategy of building custom silicon for AI inference. Amazon Web Services will also benefit from Cerebras' success, as the partnership gives AWS a differentiated inference offering against Google's TPUs and Microsoft's OpenAI-backed infrastructure. For chip fabs, the shift toward diverse AI hardware means more wafer starts for CPUs, memory, and networking chips, not just GPUs. Texas Instruments is boosting in-house chip output specifically to serve the AI infrastructure boom, indicating that the demand for analog and mixed-signal components is growing alongside digital compute. For enterprise buyers, the CoreWeave Sandboxes launch offers a new way to deploy agentic AI without building custom infrastructure. The product's availability through Weights & Biases means that data science teams can experiment with reinforcement learning and agent tool use without needing approval for large GPU clusters. This democratization of agentic AI infrastructure will accelerate adoption among mid-market companies, creating a new wave of demand for compute that is more CPU-intensive and memory-bandwidth-constrained than traditional deep learning workloads. The OpenAI-Apple dispute over iPhone AI integration, now approaching litigation, adds further uncertainty: any shift in Apple's AI partnerships redirects substantial inference demand toward independent infrastructure providers like CoreWeave and Cerebras.

What the Moves Signal About Market Direction

The CoreWeave Sandboxes launch and Cerebras' $100 billion market cap are not isolated events. They are signals about where the AI infrastructure market is heading. The industry is moving from a single-vendor GPU monoculture to a multi-architecture, multi-layer stack where orchestration, memory, and CPU compute are as important as GPU flops. CoreWeave's product strategy reflects this: by building an execution layer that works across hardware types and deployment models, the company is positioning itself as the infrastructure layer for the agentic AI era, not just a GPU rental business. Cerebras' valuation, meanwhile, tells investors that there is room for multiple chip architectures in AI inference, especially as workloads diversify from large language models to multi-step agentic reasoning. The OpenAI commitment to purchase 750 megawatts of Cerebras compute, with potential expansion to 2 gigawatts, is the strongest signal yet that inference will be the dominant compute demand driver over the next three to five years. For regulators and policymakers, the concentration of AI infrastructure in a few hands (Nvidia, AWS, Microsoft, Google) is becoming a concern, and the emergence of viable alternatives like Cerebras and CPU-based inference could reduce systemic risk.

The next twelve months will test whether these bets pay off. CoreWeave must prove that Sandboxes can attract enough enterprise customers to justify the development cost, while Cerebras must deliver on its OpenAI commitment and show that its wafer-scale architecture can scale to meet growing inference demand. The broader market will watch whether Meta's CPU pivot and AMD's $60 billion deal signal a permanent shift away from GPU dominance or a temporary diversification. If agentic AI workloads continue to grow at their current trajectory, the winners will be the companies that provide the orchestration, memory, and CPU infrastructure that these systems require, not just the fastest GPU. The era of GPU monoculture is ending, and the era of heterogeneous AI infrastructure is beginning.