OpenAI 임원진이 RAISE Summit에서 속도와 비용을 AI 구축의 새로운 경쟁 우위로 강조했다.

At the RAISE Summit 2026 on July 9, 2026, OpenAI and Cerebras (CBRS) outlined a broad push to make frontier artificial intelligence faster, more usable, and more deeply embedded in business workflows. The message was upbeat, but it also pointed to hard limits: infrastructure remains costly and complex, and the race to scale AI is now as much about data centers and supply chains as it is about model quality.

OpenAI executive Sachin Amodei and Cerebras Chief Executive Andrew Feldman framed their newly signed compute agreement as a direct response to a changing market. They argued that AI models reached a level of quality in 2025 sufficient for broad enterprise adoption, making speed the next key differentiator. Feldman called the arrangement “one of the largest deals in Silicon Valley history,” valuing more than $20 billion in compute over several years. He noted the agreement was designed to support workloads that are “made vastly better by being faster.” Amodei added that latency has become a central product feature, drawing a parallel to how search engines prioritized speed during the 2010s. Both executives emphasized that faster systems drive higher usage, better user experiences, and broader application to complex tasks. Feldman stressed that waiting is especially painful when software is part of a daily workflow, arguing that even a few seconds can feel too long.

The compute agreement, which closed on December 24, does not include a full economic breakdown, but major expansion figures were disclosed. Cerebras announced a 200-megawatt European data center buildout across Lyon, France, Norway, and Finland. Some capacity is expected to be operational by the end of 2024, with full completion targeted for the end of 2025. Feldman highlighted the company’s rapid scaling, noting that Cerebras had $25 billion less in sales 12 months prior to the announcement. Financially, Cerebras shares trade at $181.72, yielding a market capitalization of $41.17 billion. The stock has declined 41.58% year-to-date, sitting closer to its 52-week low of $160.81 than its high of $386.34. Despite recent volatility, the company reported 87.8% revenue growth over the last twelve months, reaching $603.88 million in sales. According to InvestingPro data, Cerebras trades at a high revenue valuation multiple with a price-to-earnings ratio of 152.31, and the stock appears overvalued relative to the platform’s Fair Value estimate. An InvestingPro Tip notes the company is "trading at a high revenue valuation multiple," placing it on the Most Overvalued list. Investors seeking deeper analysis can access Cerebras’ comprehensive Pro Research Report, which distills complex financial data into actionable intelligence for over 1,400 US equities. Additionally, InvestingPro offers 11 exclusive tips for Cerebras, alongside advanced screening tools and Fair Value analysis across thousands of stocks. The company’s overall financial health score stands at 1.55, labeled "WEAK" by InvestingPro metrics, with analysts forecasting a shift to negative earnings per share of -$1.04 for fiscal 2026.

On the product front, OpenAI revealed that its Phi-6 model will run on Cerebras infrastructure at 750 tokens per second. Amodei described that speed as “unheard-of,” likely an order of magnitude faster than competing systems, and confirmed Phi-6 is currently the only frontier model operating at that pace. He added that the experience should feel “magical” to users. OpenAI also accelerated its release cadence, now shipping new frontier models monthly—a pace Amodei attributed largely to internal reliance on Codex. Codex has become OpenAI’s default user interface and primary productivity tool, deployed across engineering, legal, finance, go-to-market, and human resources. Many engineers now interact directly with browsers through Codex due to its computer-use capabilities, and Amodei noted the company starts “basically” everything with it.

Internally, AI is transitioning from a support function to a core operating layer. The HR department has built AI agents to handle restructuring tasks, including reorganizing human teams. Amodei stated that OpenAI measures success through visible business outcomes, such as faster model releases and improved capability, rather than raw token usage. Despite the optimism, both speakers stressed that scale introduces new bottlenecks. The infrastructure stack is growing increasingly heterogeneous, spanning CPUs, GPUs, networking, storage, and memory. Data center capacity itself has emerged as a primary constraint, making software optimization critical for extracting greater output from existing hardware. The industry focus is shifting from speed-to-market toward efficiency and utilization. Feldman observed that every segment of the supply chain—chips, networks, storage systems, and physical facilities—represents both a growth opportunity and a limitation. This reality has made the current AI infrastructure race significantly more complex and capital-intensive than earlier cycles.

The European expansion aligns with a broader geopolitical trend: governments are increasingly treating AI infrastructure as a critical national resource. Executives noted that nations are seeking independent access to advanced AI capabilities, pushing for distributed, locally accessible intelligence rather than reliance on a few centralized frontier systems. While adoption patterns appear consistent across the U.S., Europe, and Asia, Feldman and Amodei emphasized that the goal is to make intelligence available globally, not exclusively at the frontier. They suggested that sovereignty concerns will inevitably drive more countries to build local capacity and reduce dependence on distant infrastructure.

Looking ahead, both executives expect the pace of change to accelerate further. Amodei remarked that 12 months is “an eternity in AI,” predicting continued capability gains and faster release cycles over the coming year. He identified the primary challenge as shifting from model development to enterprise deployment. Organizations will need to adapt their structures around AI agents, potentially reshaping team sizes, reporting lines, and job responsibilities. The most significant innovations, he suggested, will emerge from enterprises experimenting with agents within their own sectors, particularly finance and healthcare. During the Q&A, the speakers repeatedly returned to the idea that speed fundamentally changes user behavior. Faster tools encourage more frequent interaction and enable teams to tackle harder problems. Drawing another parallel to search engine evolution, they argued that reducing latency improves engagement and drives adoption. As Amodei concluded, business metrics—not token counts—should ultimately guide the evaluation of AI’s impact.