Five AI industry leaders warn of current model limits at Milken Conference

On May 6th at the Milken Global Conference in Beverly Hills, five people operating at different points in the AI supply chain sat down with TechCrunch for candid conversation. The result is not an optimistic one: chip shortages, orbital data centers, and doubts about whether the dominant technical architecture is the right one. When the people building the system start questioning whether the foundations can hold, it's worth listening.

The conversation, covered in this TechCrunch report, is not your typical conference panel where everyone smiles and projects growth curves. It's a collective diagnosis of real tensions affecting any organization relying on AI infrastructure, from startups using APIs to enterprises deploying their own models.

Chip shortages remain the real bottleneck

Demand for GPUs and specialized accelerators has not found significant relief. Participants pointed out that wait times to access computing capacity remain unpredictable, directly affecting training cycles and product planning. It's not just a price problem: it's a physical availability problem that no amount of money solves in the short term if the silicon doesn't exist yet.

This has immediate practical consequences. Companies depending on training windows to keep their models current are forced into compromised decisions: train less, train worse, or pay spot prices that distort any business model. For those integrating Claude or other models via API, the impact is less visible but equally real: it translates into latency changes, price adjustments, and occasionally access restrictions.

Orbital data centers: solution or distraction?

One of the most striking moments in the conversation was the mention of orbital data centers as a potential response to terrestrial infrastructure saturation. The idea isn't new in speculative terms, but that it appears in a conversation among industry executives as a mid-term option says something about the pressure level on conventional infrastructure.

The logic is straightforward: space removes some cooling and energy access constraints, though it introduces others around latency and maintenance that aren't trivial. It's worth not treating it as pure science fiction, but also not as an imminent solution.

The question nobody wants to ask aloud

The most uncomfortable point in the debate was questioning the underlying architecture. Several participants raised the possibility that the current paradigm—enormous models trained on massive amounts of data using backpropagation in transformer networks—has structural limits that aren't resolved by adding more compute or more data.

This doesn't mean current models will stop working tomorrow. Claude Opus 4.7, with its one million token context window, represents the state of the art that anyone can use today. But the question of whether this path leads where the industry promised it would is legitimate, and the fact that people with skin in the game are asking it carries different weight than it would in an academic paper.

Who should care about this

This diagnosis matters at several levels:

• Engineering teams planning their own infrastructure or evaluating providers: instability in the compute supply chain must enter their risk analysis.
• Product leads building on external model APIs: price and availability changes aren't temporary anomalies, they're part of the structural environment.
• Executives making AI investment decisions: the narrative of unlimited increasing returns has cracks acknowledged by the system's own architects.
• Developers building integrations and MCP servers: the layer of tools and agents we build on top of models inherits the fragility of what lies beneath.

---

At ClaudeWave, we've been watching how the AI conversation swings between triumphalism and catastrophism for some time. What happened at Milken is more interesting than either extreme: people who know the system from the inside recognizing there are loose bolts, without pretending they have them all tightened. That, at least, is an honest starting point.