Cutting through the noise: Tech signals worth tracking as AI advances

Are foundation models reaching a plateau? Foundation models—large AI systems trained on massive data sets—face a critical question: Will they keep improving exponentially, or will their capabilities plateau? New models are still improving, but some metrics show they’ve not delivered the dramatic performance leaps seen in earlier generations.¹ Plus, bigger models drive up energy consumption and computing costs. New scaling approaches, like techniques that allow models more time to process complex problems,² could shift us away from “bigger model = better performance.” This means current models can improve by optimizing prompting and implementation strategies. How well businesses deploy, fine-tune, and integrate AI into redesigned processes will likely matter more than having the latest foundation model.

New data > synthetic data > old data. As foundation models train on similar publicly available data sets, data itself stops being a competitive advantage. Old data loses value as the world changes. Synthetic data—AI-generated content used to train other AI—helps fill gaps, with predictions that 80% of data used by AI tools will be synthetic by 2028, up from 20% in 2024.³ However, it has a performance ceiling, typically achieving 90% to 95% of real data quality.⁴ Worse, AI trained primarily on AI-generated content can create model collapse, a degenerative process where models lose information about rare patterns, confuse concepts, and eventually produce bland, repetitive outputs.⁵ Those with access to fresh information—real-time user interactions, proprietary business data, breaking research—have the advantage. Translation: Companies controlling the interaction layer (search engines, social platforms, AI assistants, smart devices) win.

Neuromorphic chips supercharge computing. Neuromorphic chips are brain-inspired processors that are more energy-efficient than traditional graphics processing units (GPUs) for certain AI tasks. GPUs have separate areas for memory and processing, while neuromorphic chips combine both in the same place. Neuromorphic chips are event-driven—they only process information when something happens—while GPUs run constantly at full speed. This means neuromorphic chips can use 80 to 100 times less energy for AI tasks involving sporadic signals, like analyzing sensor data or processing information in autonomous vehicles, though GPUs remain superior for continuous, high-throughput computation.⁶ As AI moves from data centers to billions of edge devices (see next signal), the energy efficiency advantage becomes critical. Widespread adoption of neuromorphic computing is expected by 2030.⁷

The rise of edge AI and on-device processing. Instead of sending data to distant cloud servers, edge AI runs directly on devices—your phone, smartwatch, security camera, or industrial robot. Why it matters: latency (autonomous vehicles can’t wait for server responses), privacy (data never leaves your device), exploding costs (cloud bills reaching tens of millions monthly), and internet dependency. Edge AI’s potential is reflected in the market for generative AI–capable smartphones, which grew nearly 364% year over year in 2024 to 234.2 million units sold annually, and heading toward 912 million by 2028.⁸ Real-world applications include smart cameras doing real-time recognition locally, industrial sensors predicting equipment failures, and health wearables monitoring vitals without broadcasting medical data. This is a fundamental shift already in motion.

Will AI-native personal devices and wearables become mainstream? Companies are experimenting with AI-native wearable devices beyond smartphones—pendants that record and transcribe conversations, smart glasses with real-time translation, and screenless pins with voice interaction. The global wearable technology market is projected to reach US$265.4 billion by 2026, and tech giants are investing heavily in next-generation form factors.⁹ However, market adoption remains deeply uncertain, and the landscape is littered with failed glasses, pins, and other wearable or pocket-sized form factors.¹⁰ Questions persist about whether consumers actually want separate AI devices or prefer AI integrated into the phones and earbuds they already use. The winning form factor—if one emerges at all—remains undetermined, with success dependent on addressing privacy concerns and delivering functionality that justifies carrying an additional device.

Biometric authentication as next-level cybersecurity. Because AI can replicate voices, forge documents, and mimic behavioral patterns, biometric authentication is becoming critical for verifying physical presence and identity. As deepfakes and AI-powered fraud grow more sophisticated, organizations are rapidly adopting biometric systems: In one study of chief information security officers, 92% of those surveyed said they have already implemented, are implementing, or plan to implement passwordless authentication.¹¹ However, biometrics won’t be the sole solution. Compromised biometric data cannot be changed like a password, and privacy concerns remain significant. The future points toward hybrid approaches where biometrics serve as the primary but not exclusive verification method.

The AI agent privacy trade-off. Truly capable personal AI assistants require unprecedented access to personal data—and that access is already being granted. To book restaurants, manage schedules, or filter emails effectively, personal AI agents need years of message history, calendar entries, browsing data, stored passwords, credit card information, and intimate personal preferences.¹² But the trade-off is stark: Once personal data is incorporated into AI models, the right to erasure becomes nearly impossible.¹³ And of course, security concerns are significant. The public response is already splitting: Some eagerly grant permissions for capability while others resist. But the consent paradox remains. Users must grant extensive permissions to make AI assistants useful, but most don’t fully understand the scope or permanence of what they’re agreeing to share.

GEO overtakes SEO. Users are increasingly turning to AI chatbots over traditional search engines. The race is on to appear in AI-generated answers—a shift from search engine optimization (SEO) to generative engine optimization (GEO). AI-generated answers already dominate search results across major search engines, reducing click-through rates to conventional websites by more than a third.¹⁴ AI platforms now drive 6.5% of organic traffic, projected to hit 14.5% within a year.¹⁵ GEO differs fundamentally from SEO, prioritizing semantic richness over keywords, author expertise over backlinks, and being cited in AI responses over page views.¹⁶ Just as paid search defined the 2000s and social media advertising dominated the 2010s, AI-generated responses are becoming the most critical marketing channel of the 2020s.

Some of these signals may mature into dominant forces. Others may fade. But all of them reflect the same underlying reality: The pace of technological change has fundamentally shifted. But the speed of adaptation matters more than the certainty of prediction. The organizations that thrive won’t be those that predict which signals become trends; they’ll be those that build the capacity to sense, evaluate, and respond quickly to what emerges. Those that wait for clarity will find themselves adapting to changes their competitors are already leveraging.