Written by Kate Kupriienko
At the start of 2025, we made a set of predictions about how technology — especially AI, infrastructure, and computing architectures — would evolve over the year. Now that 2025 is nearly complete, it’s time to do something many forecasts avoid: audit ourselves in public.
Some predictions landed squarely on target. Others were directionally right but off on timing or magnitude. And a few assumptions didn’t hold up as expected.
Below is a clear, side-by-side look at what we predicted vs. what really happened in 2025, followed by a deeper analysis.
Overview: 2025 predictions vs. reality
🎯 What we got right: the "utility" revolution
Generative AI quality
We predicted that in 2025 AI would become more accurate and sophisticated, but 2025 gave us something better: AI that reasons. In previous years, AI was like a fast-talker — it simply predicted the next word in a sentence. Then we discovered Chain of Thought (CoT) — the ability for AI to "show its work" by writing out its reasoning steps. However, in 2025, the breakthrough was how we teach it to do that. New training methods like GRPO (Group Relative Policy Optimization), famously used to create the DeepSeek-R1 model, allow AI to teach itself through a massive process of trial and error without requiring expensive manually curated thinking steps training datasets.
Instead of just imitating human text, the model now explores multiple ways to solve a problem, compares them, and learns which logic actually works. This "self-evolving" reasoning has pushed AI into solving PhD-level math and professional coding challenges. The "hallucination wall" is still there, but 2025 proved that we don’t just need a bigger brain; we need a more efficient way to train it to think.
AR/VR adoption
Regarding hardware, our prediction that "AR and VR successfully broke out of their gaming niches" and integrated into daily life was equally accurate. Our forecast for "AR-powered navigation apps" and virtual shopping became a more popular way consumers interact with their surroundings. In 2025, AR head-up displays (HUDs) transitioned from luxury experiments to mainstream safety features in the mid-to-high-end car market, and "virtual try-ons" have become a retail requirement for global brands. This move from "immersion" to "augmentation" fundamentally changed how people shop and travel, overlaying digital information onto the real world just as we suggested.
AI-generated content
In the world of media, we saw a surge in AI-Generated Content. As we forecasted, social feeds became "inundated with AI-generated content tailored to your interests." Already data-focused social platforms like Meta and TikTok have moved beyond recommendation algorithms to real-time generative assembly of video content. By late 2025, your "For You" page is no longer just picking videos for you — it is often generating narration, captions, and edits specifically to match your psychological profile in real-time.
⚠️ The reality checks: right direction, wrong magnitude
In these areas, our "compass" was pointing the right way, but we underestimated the obstacles.
Sustainability
While we correctly predicted a growing emphasis on sustainability and the rise of energy-efficient, renewable-powered data centers, 2025 delivered a harsh reality check. We forecasted an AI revolution, but it was bottlenecked by physics and geography. This 'AI Factory' wall proved that in 2025, securing physical utilities is now just as difficult — and just as valuable — as writing the code itself.
Edge computing
We predicted that edge computing would become "increasingly accessible to small businesses," and the 2025 data confirms we are partially right. According to Gartner, AI PCs will represent 31% of the total global PC market by the end of 2025, with worldwide shipments projected to total 77.8 million units. While the cloud remains the indispensable engine for high-end reasoning and heavy-duty model training, the "AI PC" has successfully carved out its role as a local efficiency partner.
By running Small Language Models (SLMs) locally, some SMEs are no longer sending every single query to expensive cloud APIs. They are using local hardware for routine tasks — like drafting emails or summarizing internal documents — while "bursting" to the cloud for complex data analysis or deep reasoning.
This hybrid model allows small businesses to avoid a constant "token tax" on basic workflows, significantly lowering recurring operational costs without sacrificing the raw power of top-tier cloud models when they truly need them.
Education & healthcare
We predicted 24/7 AI Tutors and personalized medicine. We were Half-Right. The win: AI "Ambient Scribes" revolutionized healthcare by giving doctors back their valuable time. A landmark 2025 analysis in NEJM Catalyst proved that TPMG’s generative AI scribes rescued physicians from 15,791 hours of documentation across 2.5 million encounters — the equivalent of 1,794 8-hour workdays — while simultaneously enhancing doctor satisfaction and restoring face-to-face patient connection. The miss: AI Tutors remained "Experimental." The "clunkiness" — specifically the risk of AI giving wrong answers while communicating with 100% confidence (hallucinations) — prevented the mass classroom takeover we expected.
Workforce & automation
In January 2025, we spoke of "enhancing productivity without replacing jobs." We were over-optimistic. While AI did enhance productivity, 2025 saw 60,000+ layoffs at major firms like Amazon, Microsoft, and UPS. Critics and experts identified a trend of "AI-washing" — where companies used AI as a convenient narrative to cut pandemic-era "bloat" and cover for a cooling economy.
Regulations
We originally expected regulation to "work on guidelines" and evolve slowly alongside the tech. We were Partially Right in January. We correctly identified that governments would focus on "responsible AI," but we underestimated the legal force they would use. In 2025, AI policy moved from a "set of suggestions" to a "license to operate" that varies depending on which border your data is crossing.
In the U.S., the government issued a December 2025 Executive Order to strike down state laws in favor of a unified National Security policy. Simultaneously, the EU AI Act began enforcing its first major bans and transparency rules, while China made AI watermarking a legal requirement for all digital platforms. Regulation didn't just 'lag' — it became a high-stakes competition between nations to see who could set the 'global standard' first.
What we learned for 2026
- Infrastructure as the new gating factor: We spent years focusing on model parameters, but 2025 proved that the real "gating factor" for AI is the physical world. The software revolution hit a construction wall where power (megawatts) and cooling (gallons) became as precious as the code itself.
- The shift toward task-driven automation: In 2025, tools like AI Scribes won because they solved specific workflow problems, proving that utility is more valuable than raw intelligence. We are seeing a fundamental shift in the AI roadmap: the goal is no longer to build a more sophisticated "chatbot," but to create autonomous agents that can handle end-to-end business processes without constant human prompting.
- The governance standard: With the end of the "Digital Wild West" and the rise of federal and global mandates, AI regulation has moved from a legal hurdle to a prerequisite for trust.
What does this mean for the next 12 months? We are currently finalizing our deep dive into the shifts that will define the coming year. From the rise of truly autonomous workflows, to agentic AI and AI agents, and to the restructuring of the global tech stack, our 2026 Tech Predictions will be published on our blog in early January.
At Proxet, we don’t just watch these trends; we help you navigate them. Whether you are looking to bypass the "token tax" or move from experimental "chatbots" to autonomous agents that drive actual ROI, our team is ready to help you build what’s next.