Breaking Analysis with Dave Vellante

The AI buildout is making an enormous bet on what businesses will eventually be
willing to pay for intelligence. In our research this summer on Forecasting the
AI Bubble, we argued that a key indicator to watch is productive utilization by
enterprise customers. Specifically, we said if infrastructure supply catches up
before productive monetization materializes, the risk of a bubble pop increases
to dangerous levels. We forecast that supply constraints and healthy capital
funding appetites could provide the conditions for a “soft landing” and we
pegged 2029 as the high risk year. That brings us to the next obvious question:
what has to happen inside mainstream enterprises for their AI investments to
become economically self-funding? Michael Burry’s analysis of Anthropic
highlights the scale and interconnected nature of the commitments. Apollo’s
research by Torsten Slok raises questions about how broadly the financial
benefits are showing up today. And the history of the PC and internet eras
reminds us that productivity gains are not immediate - and that powerful
technology can coexist with what initially seem to be poor investment returns
that take time to appear.

In Breaking Analysis 322, we shared with you that our forecast of a $1T silicon business arrived 2 years earlier than we had forecast. The reason for our twenty-four month miss was the bottleneck had moved from GPUs to memory. Micron gave us a remarkable datapoint in its FY Q3: DRAM prices rose by more than sixty percent sequentially in the quarter, while the number of bits shipped barely grew. Today we address what that means. The key questions are: Is this temporary scarcity or a structural change in the economics of memory? We believe both forces are at work. This week we participated in a Webcast with Micron President and Chief Technology and Products Officer, Scott DeBoer. The roadmap he showed us underscores the importance of memory with respect to AI performance. New customer take or pay agreements change how risk is shared. At the same time, customers are being forced to adapt to higher costs. Qualitate’s proprietary research gives us a window into those decisions. Our question is whether these changes can make memory less cyclical, and what evidence would tell us that the familiar cycle is returning.

ServiceNow is making one of the more consequential and visible bets in enterprise AI. The company says AI Control Tower can discover and govern any agent, work across models and clouds, track cost and value, and shut down behavior that violates policy.
That sales pitch plays directly to ServiceNow's greatest asset – its installed base. ITSM, CMDB, risk workflows, security operations and years of operating context give the company a distribution and data advantage most AI governance vendors do not have.
Our research suggests that moat is real. But proprietary customer data shows a more nuanced picture. Control Tower is already strong as a governance workflow and system of action. It can inventory AI assets, connect risks to owners, route approvals and remediation, and preserve an audit trail.
What the field has not broadly validated is the claim that Control Tower is a complete technical governance and runtime enforcement plane across the heterogeneous enterprise.

In this week’s episode, we explore the AI control gap and the starting point for addressing AI safety, namely alignment. We question prevailing narratives and propose a public policy regime that has more teeth than what has been put forth thus far.

Ahead of CoreWeave’s Fully Connected conference, we have made a notable investment in proprietary customer research with Qualitate. Rather than simply repeat the earnings call, we went to the people evaluating, buying and running the infrastructure. This analysis draws on thirteen in-depth interviews and more than seven hours of interview time. We then compared that evidence with CoreWeave’s financial disclosures and the public statements of CEO Michael Intrator and CFO Nitin Agrawal.
Our research indicates that GPU scarcity opens the door, but performance, cost and the operating experience give customers reasons to stay. Inference is growing alongside training. Notably, training is not declining at the expense of inference. Inference is growing on a very steep curve and training workloads continue to grow as well. At the same time, the hyperscalers remain deeply embedded in the application estate, and not every successful CoreWeave eval turns into a signed customer.
The question is whether CoreWeave is converting an availability advantage into a durable AI cloud. We believe the customer evidence strengthens that case. It also shows exactly where the case still needs work.