Cerebras IPO doubles to $350, CoreWeave launches Sandboxes for agentic AI

Cerebras Systems opened at $350 per share on its first day of trading, nearly double its $185 IPO price, pushing the company’s market capitalization above $100 billion within hours. The explosive debut, which values the AI chipmaker at roughly 200 times its 2025 revenue of $510 million (up 76% year over year), signals that public markets still have an insatiable appetite for AI infrastructure plays. The listing comes as CoreWeave, the cloud provider that went public in March 2025 under the ticker CRWV, launched Sandboxes, a product that provides secure, isolated environments for reinforcement learning, agent tool use, and model evaluation, available on customer infrastructure or serverless via Weights & Biases. Together, these two events underscore a market that is simultaneously frothy and fragmenting: the AI infrastructure boom is real, but senior investors now warn it has three years or less to run, and the hardware mix is shifting from GPU-only to heterogeneous architectures that include CPUs, memory, and orchestration. Why this matters now: the next wave of agentic AI will demand a fundamentally different infrastructure stack, and the companies that control that stack from chip design to cloud orchestration will capture the next cycle of value.

The $20 billion OpenAI deal behind the valuation

Cerebras’ market cap surge to over $100 billion is anchored in a single, massive customer commitment. OpenAI has signed a deal to purchase 750 megawatts of Cerebras inference compute capacity, valued at over $20 billion, and the two companies are co-designing future models. That revenue visibility, essentially a locked-in pipeline that dwarfs Cerebras’ current $510 million revenue base, drove the IPO price to double on day one. The implied valuation multiple of roughly 200x trailing revenue is extreme even by AI standards, but investors are pricing in a future where Cerebras captures a meaningful slice of the inference market that Nvidia currently dominates. Cerebras CEO Choi has positioned the CS-2 wafer-scale chip as a superior alternative for inference workloads, particularly for large language models that require low latency and high throughput. The OpenAI partnership validates that thesis in the market’s eyes, but it also creates concentration risk: if OpenAI shifts its inference load elsewhere, Cerebras’ revenue base collapses. The $20 billion deal is structured as a capacity commitment, not a guaranteed revenue floor, meaning Cerebras must deliver the compute infrastructure and OpenAI must actually use it. For now, the market is betting that OpenAI’s model roadmap will require Cerebras’ unique architecture, but the history of AI hardware is littered with single-customer dependencies that turned into liabilities.

How the money flows through CoreWeave’s P&L

CoreWeave’s Sandboxes product represents a direct monetization of the agentic AI trend that is reshaping cloud demand. Sandboxes provide secure, isolated environments for reinforcement learning and agent evaluation, exactly the workloads that are growing fastest as enterprises move from chatbot-style AI to autonomous agents that take actions in the world. The product is available on a customer’s own CoreWeave infrastructure or serverless via Weights & Biases, which means CoreWeave captures revenue both from the underlying compute and from the orchestration layer. This is a higher-margin business than raw GPU rental: Sandboxes include pre-configured tool-use environments and evaluation harnesses, reducing the engineering time customers would otherwise spend building their own. For CoreWeave, which competes with Amazon Web Services and other hyperscalers, Sandboxes is a wedge into the enterprise market where agentic AI is still early but growing rapidly. The company’s Kubernetes Service (CKS) provides the underlying container orchestration, and Sandboxes sits on top as a managed service. The revenue model is consumption-based: customers pay for the compute hours their agents use, plus a premium for the Sandbox environment itself. If agentic AI follows the adoption curve that large language models did in 2023–2024, CoreWeave’s Sandboxes revenue will scale from near-zero to hundreds of millions within two years. The key metric to watch is agent runtime hours, which CoreWeave will likely disclose in future earnings calls.

The competitive reshuffle: Nvidia, AMD, and the CPU comeback

The Cerebras IPO and CoreWeave’s product launch are symptoms of a broader shift in AI infrastructure: the era of GPU-only dominance is ending. Agentic AI workloads require a different compute mix, with more CPUs for orchestration and planning, more memory for state management, and more networking for coordination between agents. Meta is already using “tens of millions” of AWS Graviton CPUs for AI workloads, and AMD announced a $60 billion deal with Meta for six gigawatts of chips over five years. That deal alone dwarfs Cerebras’ entire revenue base and signals that hyperscalers are diversifying away from Nvidia. Nvidia remains the dominant GPU supplier, but its stranglehold on AI compute is loosening as customers realize that smaller, coordinated models can achieve similar results to a single massive model. The implication for Cerebras is that its wafer-scale architecture must prove it can handle the heterogeneous workloads that agentic AI demands, not just pure inference. For CoreWeave, the shift is an opportunity: its cloud is designed to support any hardware, and Sandboxes abstracts away the underlying chip choice. The company can position itself as the neutral layer between Nvidia, AMD, Cerebras, and CPU providers like AWS Graviton. The risk is that hyperscalers like AWS and Microsoft build their own agentic AI platforms, cutting out independent cloud providers. Morgan Stanley analyst Shawn Kim has noted that the AI infrastructure market is consolidating around a few key players, and CoreWeave must move fast to lock in enterprise customers before the hyperscalers do.

Downstream effects on memory, networking, and data center construction

The shift to agentic AI has second-order effects that ripple through the entire hardware supply chain. Agentic workloads require more memory per compute unit because agents must maintain state across multiple tool calls and reasoning steps. Micron and other memory makers are the direct beneficiaries: higher CPU-to-GPU ratios mean more DRAM and HBM per server, and the orchestration layer requires fast, low-latency memory to coordinate agent actions. The CNBC report on orchestration highlighted that the AI bull market’s obsession is shifting away from Nvidia to memory chip makers, and the numbers bear this out. Data center construction is also being reshaped: the 750 MW commitment from OpenAI to Cerebras will require new data center capacity, likely in locations with cheap power and low latency to OpenAI’s existing infrastructure. But the broader capex boom is nearing its peak. Senior infrastructure and private capital investors told the Financial Times that the AI spending spree has three years or less to run, which means the window for building new data centers is closing. Companies like CoreWeave that are investing heavily in capacity must be careful not to overbuild. The Sandboxes product, by enabling more efficient use of existing compute through better orchestration, could actually reduce the total amount of hardware needed for agentic AI workloads, a paradox that investors will need to grapple with. If agents become more efficient, the hardware demand per agent drops, even as the number of agents explodes.

Policy and strategy signals from the Cerebras IPO and CoreWeave launch

The Cerebras IPO is a strategic signal from the market that AI infrastructure is no longer a single-player game. The success of the listing, despite Nvidia’s dominance and the concentration risk of the OpenAI deal, tells other AI chip startups that there is a path to public markets if they can secure a marquee customer. This will likely accelerate the IPO timelines of companies like Groq, SambaNova, and others that have been waiting for a market window. It also puts pressure on Nvidia to diversify its own product line beyond GPUs, which it is already doing with its Grace CPU and networking acquisitions. For CoreWeave, the Sandboxes launch is a strategic bet that agentic AI will be the next growth wave, and that the company can capture the orchestration layer before hyperscalers do. The partnership with Weights & Biases is particularly telling: it gives CoreWeave access to the ML experiment management ecosystem that many AI teams already use, lowering the switching cost for enterprises to try Sandboxes. The broader policy implication is that the US government, which has been concerned about AI chip export controls, now has a more diverse set of domestic AI chip suppliers to consider. Cerebras’ wafer-scale technology is manufactured in Taiwan, but the company has announced plans to build a US fab. If the IPO proceeds are used to accelerate that plan, Cerebras could become a strategic asset for US chip independence. The next 12 months will determine whether the $100 billion valuation was a peak or a floor.

The next 12 to 18 months will test whether the infrastructure market can sustain the valuations that Cerebras and CoreWeave have commanded. The $20 billion OpenAI deal provides Cerebras with a revenue floor, but the company must prove it can win additional customers beyond its anchor tenant. CoreWeave’s Sandboxes product is well-timed for the agentic AI wave, but the company faces an uphill battle against hyperscalers that can bundle orchestration with their own compute and storage. The broader risk is that the AI capex boom peaks within three years, as senior infrastructure investors have warned, which would compress valuations across the sector. For investors, the key signals to watch are enterprise adoption rates of agentic AI, the pace of data center construction, and the quarterly revenue disclosures from both Cerebras and CoreWeave. If agentic AI drives a new wave of demand that offsets the plateau in training infrastructure, the current valuations will prove conservative. If not, the market will correct sharply. The companies that survive will be those that own the orchestration layer, the software that decides which chip runs which workload, because that is where the switching costs are highest and the margins are stickiest. Cerebras and CoreWeave are both betting that they can own that layer, one from the chip side and one from the cloud side. The next earnings season will provide the first real data point on which bet is winning.