Conference Presentation, Panel
Powering AI Infrastructure: Solving the 21st Century Manhattan Project | RAISE Summit 2026
RAISE SummitShaun O'Meara, Alex Saroyan, Vik Malyala, Ben Richardson, June Paik, Jeremie Eliahou Ontiveros, Jeremy, Sean Amara, Jun
- Anthropic projects 1.5 gigawatts of capacity by the end of 2025, targeting 10 gigawatts by 2027, a scale four other companies are expected to reach within the same two-year period.
- Inference workloads are forecast to comprise 75% of all GPU workload by the end of 2027, driving a long tail of new entrants seeking training capabilities alongside deployment needs.
- CoreWeave reported order book growth from $23 billion on March 31 of the previous year to $99 billion on March 31 of the current year, with expectations that nine out of ten largest model labs will utilize its infrastructure.
- CoreWeave intends to operate infrastructure at the speed of its fastest component and plans to expand international assets, scaling its workforce from one to 250 employees over two to three years.
- Smaller European cloud providers face limited power capacities, necessitating retrofitting for liquid cooling or the adoption of modular data centers to maintain operations.
- Netris utilizes digital twin technology to simulate clusters for immediate go-live status upon hardware arrival, aiming for deployment within days or weeks while ensuring zero mistakes for stability and security.
- Supermicro plans to validate solutions from the initial stage through to online status to optimize time-to-online, while maintaining multiple hardware and software choices to mitigate design, supply chain, and pricing risks.
- More than 50% of U.S. data center projects are expected to be canceled due to contract, permitting, or delay issues, while power delivery standards shift from 415 or 277 volts to 400 and 800 volt DC bus bars.
- Jun (Furious AI) anticipates enterprises will increasingly control their own compute infrastructure to manage data security and costs, as AI energy and CapEx costs approach unsustainable levels.
- Sean Amara expects enterprise clusters to grow from roughly five live to ten coming soon, with a market shift toward a balance of large-scale training and highly distributed inference deployments as inference moves into the enterprise sector.
- Enterprises are projected to struggle with defining AI policies and achieving ROI, often adopting strategies of starting small in sandboxes before moving to production.
- Alex Soroyan notes that as AI infrastructure complexity grows, the market will require specialized automation tools rather than in-house scripts to prevent mistakes.
- CoreWeave expects to respond to demand for non-Nvidia builds when specifically requested, though current market demand remains predominantly for NVIDIA stacks.
- Suppliers including TSMC, HBM, Hynix, and Supermicro are expected to collaborate to enable the production phase of server appliances, while enterprises seek alternatives to single-vendor reliance.