newsfilter
← Interviews
20VC with Harry Stebbings1h 13m

Why Neo-Clouds are Great Businesses & The State of Compute with Kush Bavaria, Co-Founder @ Ornn

Kush Bavaria, Harry Stebbings

Interview, Fireside Chat, Podcast
  • Humans are anticipated to transition into biological vessels for AI within the next few years to teach the systems emotions, a shift accompanied by a "bold" 2030 prediction for human longevity extending to 150 or 200 years, though the speaker acknowledges this specific timeline is unlikely to be fully realized.
  • The GPU compute market is currently undersupplied with lead times for purchasing chips from vendors like CoreWeave extending to Q4 2027 or Q1 2028, while acquisitions from OEMs involve a 20 to 22-week delivery period plus a two to three-week installation phase, resulting in a total build time of approximately half a year.
  • Chip procurement timelines are forecast to lengthen from the current 18 to 20 weeks to 22 to 26 weeks soon, and forward pricing curves are expected to reverse from previous assumptions, with prices rising to $20 or $30 per GPU hour compared to current rates of seven or eight.
  • The compute market is projected to surpass the oil market's 100 to 150-year dominance, potentially reaching a trillion-dollar opportunity, with the speaker's company aiming to become a "Moody's for data center debt" within a few years and potentially building physical data centers to compete with entities like CoreWeave and Modal.
  • While infrastructure investors are expected to benefit as long as AI remains the primary tenant and GPU demand persists, a bear case predicts that if large enterprises (such as McDonald's or the S&P 500) find AI offers no business change, it could lead to payment defaults and assets valued only for their metal content.
  • Enterprise demand risks include a potential ROI crisis where cracks in spending trickle down to providers, causing GPU hour prices to fall, contrasting with the current consensus that inference demand will set forward prices due to its non-speculative, revenue-generating nature.
  • GPU hardware, including older A100 and H100 models, is expected to remain in use for three-year periods or longer by migrating to specialized inference markets rather than facing rapid obsolescence.
  • The inference market is predicted to segment into fast, slow, and specialized categories (e.g., biology, video, physics), evolving into a landscape dominated by one or two horizontal providers alongside approximately a hundred specialized vertical companies.
  • Companies with cash are advised to buy chips immediately to avoid escalating costs, while founders with excess capacity are expected to operate as "neoclouds" charging high rates, though this model faces risks of lending to failing neolabs without adequate hedging.
  • The competitive landscape for neolabs is expected to be defined by a combination of talent and compute, with the latter potentially serving as a stronger moat, while customer support and specific legal verticals like "rogo" and "harvey" are considered saturated and closed to new entrants.
  • Market dynamics are driven by fear-based momentum investing and FOMO among VCs seeking the next major AI firm, creating a divergence where AI capabilities in the "San Francisco bubble" advance exponentially while enterprise adoption follows a linear growth curve.
  • The company plans to adopt a "family business" structure intended to last for generations, prioritizing long-term stability over short-term exit strategies despite the speaker's uncertainty regarding the feasibility of the 2030 longevity predictions.