UpTrajectory Review

TheStreet's interview with Liquid AI CEO Ramin Hasani lands at a moment when the AI trade is visibly wobbling. CoreWeave's debt load and the sheer scale of data-center spending spooked investors last week, triggering a sell-off that has since partially reversed. Against that backdrop, Hasani is making a structural argument: the value in AI is migrating away from the model layer — where dozens of U.S. and Chinese companies are producing functionally similar large language models — and toward the infrastructure that lets AI run on physical devices. His company, spun out of MIT, builds models designed to operate across varied hardware rather than live exclusively in cloud data centers. The piece itself is thin, cutting off mid-thought, but the thesis it introduces is worth taking seriously on its own terms.

For a small-business operator, the practical significance is not about whether Liquid AI specifically wins. It is about where leverage accrues as the AI stack matures. If Hasani is right that the model layer is commoditizing, then the businesses best positioned are not the ones building chatbots but the ones embedding AI into products, workflows, and physical environments — and the vendors supplying the tools that make that embedding possible. A local manufacturer, a fleet operator, or a retail business evaluating AI integrations should care less about which brand-name model powers a feature and more about whether the solution runs reliably on the hardware they already own, at a cost that does not require a cloud contract the size of a mortgage.

The genuinely contested point here is whether the model layer really lacks differentiation. Hasani claims it does, but that is a convenient position for a company whose pitch is hardware-agnostic infrastructure. OpenAI, Anthropic, and Google would argue their models are meaningfully distinct in reasoning quality, safety, and capability — and enterprise buyers currently pay premiums that suggest they agree. What is under-reported in the excerpt is the counterargument: that model quality differences still drive real product outcomes, and that edge deployment involves tradeoffs in accuracy and update cadence that cloud models avoid. We are sympathetic to Hasani's broader point about physical-world AI being underpriced by markets obsessed with chatbot metrics, but skeptical that differentiation has fully evaporated at the model layer.

The second-order effects cut in several directions. If AI inference shifts toward cars, phones, and embedded devices, the winners include chipmakers like Qualcomm — whose summit served as the interview venue — and the manufacturers who can integrate AI without ballooning unit costs. The losers, or at least the pressured, are pure-play cloud AI providers whose economics depend on centralized compute demand continuing to climb indefinitely. For small businesses, cheaper and more localized AI could mean smarter point-of-sale systems, predictive maintenance on equipment, or voice-driven customer tools that work offline. But it also means a more fragmented vendor landscape, where integration costs and compatibility headaches replace subscription fees as the hidden expense.

What to watch: whether major device makers begin shipping AI features that demonstrably work without cloud connectivity, and whether enterprise buyers start treating model choice as a commodity decision the way they treat cloud storage. For operators, the actionable move now is to audit any AI tool you are paying for and ask whether the intelligence runs on your hardware or someone else's — and what that dependency costs you in money, latency, and data control. If Hasani's thesis holds, the next round of AI value will not be visible in a chat window. It will be sitting in your vehicle, your equipment, and your store.

“If you just look at the model layer, you will see no differentiation.” — TheStreet

Takeaway: Audit whether your AI tools run on your hardware or a vendor's cloud — the next wave of AI value is shifting to embedded, physical-world deployment.