Fully Connected 2026 Day 2: Mike Intrator on Hot Ones, Parag Agrawal, Dylan Patel, and Jenson Button
Day two of CoreWeave's Fully Connected conference brings together Parallel Web Systems CEO Parag Agrawal on agent-driven web infrastructure, SemiAnalysis founder Dylan Patel on AI cloud fragmentation and security, Formula One champion Jenson Button on team culture under pressure, and CoreWeave CEO Mike Intrator on financing, risk, and building an AI cloud from a New Jersey garage.
Building for Ambition: From a New Jersey Garage to the Frontier of AI Infrastructure
Origins: A Crypto Hobby That Became a Cloud
CoreWeave’s story starts in a New Jersey garage, not Silicon Valley. The founders began with crypto mining, choosing GPUs over specialized ASIC chips because GPUs were something they could actually work with. When crypto winter hit in 2018, the bet was simple: either crypto goes to zero, or owning a large pool of GPUs will carry real value. Along the way, startups and friends who couldn’t get the compute they needed kept showing up in conversations, and that demand became the cloud. Crypto mining income helped finance the buildout while the customer base grew, a window the company describes as a particular moment in time that it was fortunate to occupy.
The pivot toward generative AI came in 2022, when text-to-image models suggested capabilities were changing fundamentally. The company went all in roughly five months before ChatGPT, and was called crazy for it. The company’s self-description captures the tension: fearless risk-taking paired with superior risk mitigation. Financing has been built against demand the company can actually see, which the founders describe as unconventional but historically how the best infrastructure gets built.
The Agent Web: Parallel’s Thesis on Traffic and Data
Parag Agrawal, now CEO of Parallel Web Systems, laid out a thesis three years ago that has largely held: internet traffic would be dominated by AI, the requests would come from agents, and the answers would come from AI as well. His core belief is that agents, given that models appear all-knowing, will be expected to read anything on the open web, and that agent web access will need to be roughly a thousand times larger in scale than human web search. Parallel builds what he calls a Google for agents, a GPU-native index designed for crazy scale and cheaper search than legacy systems can offer.
The technical challenge is substantial. Search infrastructure built over two decades cannot survive a thousand-fold scale increase without fundamental redesign. The business model is also unresolved: the human web ran on advertising, subscriptions, and transactions, and all three weaken when the consumer is an agent. Creating new incentives for content producers is as much an open problem as the systems work.
Agrawal’s longer-term prediction is a shift from pull to push. Today an agent queries a search engine; in a few years, a large share of web usage may be systems pushing actionable information to always-on agents that wake up and act, so the web comes to the agent rather than the agent going to the web. He also offered a durable heuristic for keeping pace with change: rely on long-term invariants. Frontier models will keep growing in capability and cost, smaller models will keep getting cheaper, and the gap between them will widen. Data retrieval, he argued, will consume one to two GPUs for every ten spent on model inference, because differentiation increasingly comes from supplying the right context from trillions of pages, several billion of which change daily.
Managing Humans and Agents at Once
Agrawal and the conversation’s moderator both described a new management condition: they now oversee teams of humans and fleets of agents, and the work never stops. Agrawal argued that a small, agile company should deliberately push agents past their current limits and accept failures, then retry a few months later, because being among the first to see a capability work shapes product and strategy. The consequence is that ambition scales with productivity. Revenue targets and shipping cadences that once felt aggressive stop feeling satisfying, and the constraints become the limits of one’s ideas and awareness rather than capacity itself.
Infrastructure Fragmentation and the Full-Stack Question
Dylan Patel of SemiAnalysis described an ecosystem still in heavy experimentation, with builders choosing among frontier closed models and open-weight models, then assembling inference, service-level agreements, retrieval, permissioning, and workflow integration on top. Standards are thin. The OpenAI API has become a de facto compatibility layer, but beyond the model interface, scaffolding is fragile: a new model can make elaborate scaffolding hurt performance that it previously helped. Co-design of model, infrastructure, and scaffolding is therefore continuous work.
Patel argued that clouds are converging on full-stack offerings: GPU providers are becoming inference providers, managed-service vendors are building their own infrastructure, and model builders are trying to build clouds. The gains from tighter co-optimization between cloud layers, kernels, and silicon can be large, and the number of performance knobs is too great for any one person to hold in their head. Patel also noted that today’s tinkering is easy, while building complex systems is where the difficulty and the fun now lie. Capacity access, meanwhile, remains hard to obtain, which makes unified, end-to-end services more valuable.
Security in the Wild West of Neoclouds
Patel’s most pointed material concerned security. SemiAnalysis’s ClusterMax rating evaluates GPU clouds across security lifecycle management, networking, storage, and dozens of other factors. In testing more than eighty providers, the team found that in many cases it could reach other tenants’ storage and see other users’ Slurm jobs, and in one case reached systems belonging to a national defense and air force environment before stopping and reporting. Open-weight models with fewer restrictions were willing to attempt privilege escalation where frontier models refused. Patel’s point was that AI-assisted penetration testing has collapsed the cost of finding vulnerabilities, and that the vulnerabilities were already there; the change is that they are now cheap to find for anyone. The risk spans data exposure to multi-tenant snooping and, more speculatively, autonomous agents exploiting poorly secured infrastructure. He framed this as a top-to-bottom security problem that enterprises and startups both need to take seriously.
Formula One and Building Under Pressure
Jenson Button’s account of his 2009 championship offered a lesson in team culture. When Honda withdrew at the end of 2008, Ross Brawn bought the team for a nominal sum, and the team competed on a tight budget with engines sourced from Mercedes. Button won six of the first seven races, then the second half of the season exposed the cost of pressure: he pushed too hard, made mistakes, and needed his father’s steady counsel to refocus. He described a period after the title in which he sat alone in his hotel room reflecting on what he had given up to get there, and then the need to seek a new challenge.
His account of the 2011 Canadian Grand Prix, where he collided with his teammate, took a drive-through penalty, hit Fernando Alonso, and won after a red flag, emphasized that positivity produced aggression, which produced tire temperature and a well-timed switch to slicks. Button’s broader lesson was that the team’s culture determined what drivers could attempt. Brawn had removed blame culture, which freed engineers to try ideas, and one engineer found a regulatory loophole that produced the double diffuser and the downforce that underpinned the championship. Button tied this to the present by noting that Formula One’s simulations now run billions of strategies, and that drivers spend most of the year off track, with AI shaping the car long before the race weekend.
Hot Seat: Pioneering, Financing, and Structural Risk
In the final session, CoreWeave co-founder and CEO Mike Intrator discussed the company’s early days in a Manhattan office where a cluster of GPUs overheated after the air conditioning was turned off, which led the team to relocate the equipment to a garage in New Jersey. Intrator described the four co-founders as people he would choose again, and summarized one of his mentor’s observations: his superpower is sitting in a noisy room and letting the best ideas win.
His description of capital markets revealed the company’s financing philosophy. West Coast venture money is designed to change the world and fund systemic step-function technologies; East Coast capital, largely debt, asks a single question about whether it will be repaid. CoreWeave learned to move between the two by making its vision compelling to one group and underwritable to the other. The underlying method is to map options embedded in each risk: GPUs were chosen partly because they could mine crypto and could be monetized from the moment they were installed, but the real value lay in the options to do other things. Owning the GPUs at the end of a debt structure, and keeping them running rather than replacing them, provides flexibility, which has begun to generate value as 2020-era hardware remains contracted through 2029.
Intrator’s view on the next decade was that the physical and operational challenges of building and delivering AI infrastructure will be far harder than the industry expects, and that the transition will require many companies working together, both on infrastructure and on helping organizations integrate AI into their workflows and society.
Closing Themes
Across the day, the speakers returned to the same tension: ambition requires the willingness to be misunderstood, to take risks in a fragmented landscape, and to build the systems and cultures that let people move faster without breaking what they depend on.