Watching Big Tech’s recent earnings releases, I sensed a real inflection point. The market is shifting away from valuing companies on technological progress alone. How efficiently capital deployed is being returned — in other words, ROI — is becoming the key criterion.
I think Alphabet and Meta’s diverging stock performance came down to exactly this ‘visibility of returns.’ Alphabet showed relatively clearly that AI is driving cloud revenue, while Meta stoked the market’s ROI concerns by sharply raising its CapEx guidance for AI investment. The biggest challenge facing the AI industry right now is a cost structure of growth that’s structurally different from the old SaaS model.
First, it’s worth looking at how AI’s economics differ from SaaS.
In traditional SaaS, once the initial infrastructure was built, the marginal cost of scaling the service approached zero. AI is different. Usage growth generates ongoing token costs, and sustaining that requires massive, continuous infrastructure investment.
The cash burn rate from infrastructure investment that OpenAI’s CFO recently mentioned, and the structural cost-settlement arrangements with infrastructure partners like Microsoft, illustrate this burden clearly. Anthropic broadening its collaboration with multiple cloud providers likewise suggests it isn’t immune to this same cost structure.
Against this backdrop, hybrid AI can be a realistic alternative. On-device models like Gemma 4 in particular look less like a passing tech trend and more like a strategic lever that can actually change the cost structure.
A structure that runs on-device and cloud in parallel brings a few changes.
First is a redistribution of cost.
Companies can offload some of the compute burden onto user devices, easing their infrastructure costs.
Second is personalization and security.
Users can process sensitive data locally, strengthening privacy while also cutting token costs.
I’ve been feeling out this hybrid structure myself, running Claude alongside local models, and I’m genuinely enjoying it — there’s a lot of room to experiment.
Ultimately, the question companies need to work through is how to integrate the foundation model and the on-device environment into a single experience, and how that creates strong lock-in. As with Google’s efforts to organically connect cloud and local models, the crux of future AI competition is likely to hinge on how naturally that connection is designed, and how each company evolves it to fit its own strategy.
As the AI industry keeps advancing, I expect efforts to go well beyond simply using cloud AI and instead maximize internal infrastructure efficiency will only diversify further.
Everything above is my own observation and interpretation. I’d love to hear from anyone with a different take on the same topic.
#AIStrategy #BusinessInsight #OnDeviceAI #ROI #TechStrategy