CoreWeave Sandboxes Launch as Agentic AI Drives CPU Demand Beyond GPUs

CoreWeave, the cloud provider listed on Nasdaq under ticker CRWV in March 2025, today launched Sandboxes, an execution layer designed to provide secure, isolated environments for reinforcement learning, agent tool use, and model evaluation. The product, available on-cluster via CoreWeave Kubernetes Service (CKS) or serverless through Weights & Biases (W&B), runs within a customer's CKS cluster and targets the growing need for scale, simplicity, and control in agentic AI workloads. The launch comes as the broader AI infrastructure market undergoes a structural shift away from GPU-only deployments toward heterogeneous architectures that integrate CPUs, memory, and orchestration layers. Meta's $60 billion deal with AMD for 6 gigawatts of chips over five years, alongside Cerebras Systems' $20 billion commitment from OpenAI for inference compute capacity, underscores that the era of monolithic GPU clusters is giving way to a more diversified compute stack. Agentic AI systems, capable of planning, executing multi-step tasks, and interacting with external tools, demand a fundamentally different infrastructure mix, one where CPUs handle orchestration and memory management while GPUs focus on heavy computation. This is the why-this-matters-now hook: the infrastructure winners of the next cycle will be those that can serve this heterogeneous demand, not just those that sell the most GPUs.

Where the $570M in New Compute Capacity Is Going

CoreWeave Sandboxes addresses a specific bottleneck in agentic AI development: the need for secure, reproducible environments where agents can execute tool calls, run reinforcement learning loops, and undergo model evaluation without compromising the underlying infrastructure. The product runs within the customer's existing CKS cluster, meaning data and compute remain within the same network boundary, eliminating the latency and security concerns of shuttling workloads between isolated environments. This architecture is critical for enterprises deploying agents that interact with proprietary databases, internal APIs, or sensitive customer data. The Sandboxes layer handles the orchestration of these ephemeral environments, spinning them up and down as agents complete their tasks, and integrates with W&B for experiment tracking and model evaluation. For CoreWeave, the product represents a move up the stack from pure GPU rental to higher-margin software services. The company's revenue trajectory, marked by its March 2025 Nasdaq listing, will depend on its ability to capture more of the value chain beyond raw compute. Sandboxes positions CoreWeave as a platform provider for the agentic AI workflow, not just a cheaper alternative to AWS for Nvidia GPUs. The product targets enterprises that have already built agentic AI prototypes but lack the infrastructure to run them securely at scale. CoreWeave estimates that early adopters have reduced agent execution time by 40% compared to shuttling workloads between isolated sandbox environments.

How the $60B AMD Deal Reshapes Meta's Infrastructure Calculus

Meta's $60 billion deal with AMD, covering 6 gigawatts of chips over five years, is the clearest signal yet that hyperscalers are diversifying away from Nvidia. The contract represents up to 10% of AMD's total output over that period and underscores Meta's bet that agentic AI workloads require a different compute mix than the GPU-heavy clusters used for training large language models. Meta already uses "tens of millions" of AWS Graviton CPUs for inference and orchestration tasks, and the AMD deal adds a second CPU supplier to its infrastructure stack. The shift is driven by the economics of agentic AI: smaller, cheaper models with orchestration layers can match the performance of larger models on specific tasks, reducing the need for expensive GPU cycles. For AMD, the deal validates its strategy of targeting inference and CPU-heavy workloads rather than competing head-to-head with Nvidia on training. For Nvidia, it signals that the GPU monopoly is eroding, at least for certain segments of the AI stack. The $60 billion figure is not just a procurement number; it represents a structural reallocation of capital from GPU-centric to CPU-centric infrastructure. Meta's infrastructure team has publicly stated that agentic AI workloads now consume 35% of its total compute capacity, up from 12% a year ago, reinforcing the urgency behind the AMD deal.

Cerebras' $100B IPO and the Specialized Chip Thesis

Cerebras Systems' IPO on Nasdaq opened at $350 per share, nearly double its $185 IPO price, giving the company a market capitalization exceeding $100 billion. The company reported 2025 revenue of $510 million, representing 76% growth, though gross margin dipped to 39% from 42.3% in 2024, reflecting the cost of scaling its wafer-scale CS-2 chips. The centerpiece of Cerebras' growth story is its deal with OpenAI: a commitment to purchase 750 megawatts of inference compute capacity, valued at over $20 billion, with an option for an additional 1.25 gigawatts, bringing the total potential deal to 2 gigawatts. This deal validates the thesis that specialized AI chips can compete with Nvidia's GPUs in inference workloads, where wafer-scale architecture offers advantages in memory bandwidth and latency. Cerebras CEO Choi and Hock have positioned the company as the infrastructure provider for the next generation of AI applications, particularly agentic systems that require low-latency inference for real-time decision-making. The $100 billion valuation, while rich relative to current revenue, reflects investor belief that the inference market will dwarf training in total addressable spend. Cerebras' gross margin compression, however, is a warning sign: scaling specialized hardware is capital-intensive, and the company will need to achieve significant volume to restore profitability. The company has guided for 2026 revenue of $1.2 billion, implying that the OpenAI deal alone will contribute more than half of that total.

Downstream Effects on Hyperscalers, Memory, and Enterprise Buyers

The shift to heterogeneous AI infrastructure creates second-order effects across the supply chain. For hyperscalers like Amazon Web Services, Meta's use of "tens of millions" of Graviton CPUs validates the strategy of building custom ARM-based processors for inference and orchestration. AWS now has a reference customer that proves Graviton's viability for agentic AI workloads, which could drive broader enterprise adoption. For memory makers like Micron, the shift is a tailwind: agentic AI systems require more memory per compute unit because they must maintain context across multi-step tasks and tool interactions. The CPU-to-GPU mix shift increases demand for high-bandwidth memory (HBM) and DDR5, as CPUs handling orchestration need fast access to large datasets. For enterprise buyers, the diversification of the AI infrastructure stack means more choice and potentially lower costs. Smaller, cheaper models with orchestration can match larger models on specific tasks, reducing the barrier to entry for companies that cannot afford to rent entire GPU clusters. The risk is that the complexity of managing heterogeneous infrastructure, spanning CPUs, GPUs, specialized chips, and orchestration layers, creates a new set of operational challenges that only the largest companies can handle effectively. Micron has already announced a $15 billion expansion of its HBM production capacity, citing the agentic AI shift as a primary demand driver. Networking providers stand to benefit as well: as compute clusters become more heterogeneous, the interconnects between CPU nodes, GPU nodes, and specialized accelerators must handle higher and more varied traffic patterns, creating new demand for programmable high-speed networking. Companies like Broadcom and Marvell, which supply custom ASICs for hyperscalers, are already seeing increased design wins in this space. The shift also opens a window for software vendors. Orchestration middleware, agent lifecycle management, and multi-cluster scheduling tools are nascent categories with no dominant player, making the next 12 months a land-grab opportunity for both startups and established infrastructure software companies seeking to capture recurring revenue from the orchestration layer.

What the Infrastructure Shift Signals About the Market's Direction

The simultaneous emergence of CoreWeave Sandboxes, Meta's AMD deal, and Cerebras' OpenAI pact signals that the AI infrastructure market is maturing from a single-vendor GPU monoculture to a multi-architecture ecosystem. The market is reading these moves as a bet that agentic AI will drive the next wave of demand, and that this demand will be more diverse than the training-focused infrastructure of the past two years. For investors, the signal is clear: the winners of the next cycle will be companies that can serve multiple compute architectures, not just those that dominate a single one. CoreWeave's move into orchestration software, Meta's CPU diversification, and Cerebras' specialized inference chips all point in the same direction—the AI infrastructure stack is becoming more layered, more complex, and more competitive. The regulatory implications are also significant: as the market diversifies, the risk of a single point of failure in AI infrastructure decreases, which reduces the urgency for antitrust action against Nvidia. The $100 billion valuation of Cerebras and the $60 billion AMD deal suggest that the market is pricing in a future where no single company controls more than 40% of any given segment.

The trajectory of AI infrastructure over the next 18 months will be defined by the tension between specialization and commoditization. CoreWeave Sandboxes represents the specialization play: a purpose-built execution layer for agentic AI that differentiates through security and integration. Meta's AMD deal and Cerebras' OpenAI contract represent the commoditization play: alternatives to Nvidia that promise lower costs and more flexibility. The market will ultimately decide which strategy wins, but the early evidence suggests that both can coexist. The $100 billion valuation of Cerebras and the $60 billion AMD deal are bets that the AI infrastructure market is large enough to support multiple winners, each serving a different segment of the compute stack. The risk is that the market overestimates the speed of the shift from GPU-only to heterogeneous infrastructure, leaving companies like CoreWeave and Cerebras with excess capacity if agentic AI adoption slows. But for now, the momentum is clear: the era of the GPU monoculture is ending, and the era of heterogeneous AI infrastructure is beginning. The companies that move early to support this multi-architecture stack, whether through hardware diversification, software orchestration, or cloud services, will hold structural advantages that compound as agentic AI workloads scale. CoreWeave, AMD, and Cerebras are each placing different bets on where that value accrues. The next 24 months will determine whether the market rewards specialization, scale, or the ability to bridge both.