Prediction is a dangerous game in technology, and in AI it's particularly treacherous — the field moves at a pace that makes year-ahead forecasting more like educated guessing than analysis. I've been wrong about AI timelines before, and I'll be wrong about some of what I write here. But reasoned engagement with where the evidence points is more useful than agnosticism, so here's my honest assessment of what 2026 holds for generative AI.

I'll organize this around what I'm most confident about, what I think is probable, and where I hold genuine uncertainty — rather than presenting a single confident narrative that obscures real disagreement within the research community.

High Confidence: These Are Happening

Reasoning models become standard, not premium. OpenAI's o1 and o3 series, which use extended test-time compute to "think through" problems step by step before responding, demonstrated in 2025 that trading inference speed for reasoning quality is a productive trade-off for many use cases. By the end of 2026, I expect reasoning model capabilities to be available across the full price range — not just at frontier model pricing. The architecture insights are now broadly understood, and competition will drive these capabilities toward commodity. This has significant implications for complex reasoning tasks (mathematics, code generation, multi-step planning) that will become substantially more tractable.

AI agents move from demo to standard enterprise feature. The major enterprise software vendors — Salesforce, Microsoft (with Copilot), ServiceNow, SAP — are all shipping agentic features that allow their software to take actions rather than just answer questions. By year end, "AI agents" will be a standard feature of enterprise software platforms, not a differentiating premium. The battleground will shift to which agents are most reliable, most deeply integrated, and best governed. The standalone agentic AI startup layer will face increasing competition from platform vendors.

Inference costs fall another 5–10x. The trend in inference cost reduction has been remarkably consistent. Hardware improvements (Nvidia's Blackwell architecture generation, increasingly competitive custom silicon from AMD, Google, and Amazon), software optimizations (better quantization, speculative decoding, KV-cache efficiency), and competition among providers will continue driving costs down. By end of 2026, running GPT-4 class inference will cost roughly what GPT-3.5 costs today. This makes economically viable a new tier of AI applications that don't work at current pricing.

Context window sizes cross 1 million tokens as standard. Gemini 1.5 Pro's 1 million token context and the subsequent research from multiple labs have established that very long context is technically feasible and commercially valuable. The engineering challenges of making long context efficient (sparse attention mechanisms, memory-efficient transformers) are being solved. By end of 2026, 1 million token contexts will be available across most frontier model providers, enabling application categories — full codebase analysis, entire document corpus synthesis, long-form research — that are currently constrained by context limits.

EU AI Act compliance drives systematic change in enterprise AI governance. With high-risk AI system obligations taking effect in August 2026, European enterprises and companies selling into EU markets will have completed the work of building formal AI governance infrastructure: risk classification, technical documentation, conformity assessments, and monitoring systems. This compliance investment will produce better-governed AI more broadly, as the discipline required for regulatory compliance tends to spill over into general practice.

Probable: Strong Evidence Pointing This Direction

A multimodal model demonstrates state-of-the-art performance in video understanding for enterprise use cases. Video understanding has lagged behind image and text understanding in commercial quality, but the gap is narrowing fast. Google's Gemini architecture is furthest ahead on long-context video, and the competitive pressure from OpenAI and Anthropic's multimodal roadmaps will accelerate this. By mid-2026, I expect practical video understanding quality that enables real production use cases in healthcare (surgical video analysis), manufacturing (assembly quality control), and enterprise knowledge management (meeting and call analysis).

On-device AI becomes a genuine competitive differentiator for consumer and edge applications. Apple's focus on private AI computation with on-device models (demonstrated through Apple Intelligence on current hardware), Qualcomm's Snapdragon X Elite AI performance, and the rapid maturation of model compression and quantization techniques are collectively making capable AI computation on-device realistic. Applications that can guarantee user data never leaves the device will have a meaningful trust and privacy advantage over cloud-only alternatives. I expect significant product innovation in this space, particularly in health (continuous monitoring with local inference), productivity (on-device personal assistant that knows your data), and enterprise (edge inference for latency-sensitive and data-sensitive applications).

A significant AI safety incident reshapes industry practice. I don't make this prediction with certainty, and I am not predicting a catastrophic or existential incident. But the deployment of increasingly capable and autonomous AI systems at scale creates the conditions for consequential failures: an agentic AI causing significant financial or operational damage through a misunderstood instruction, a clinical AI recommendation contributing to a patient harm event that receives major media attention, or an AI fraud attack at unprecedented scale. The realistic scenario is an incident serious enough to catalyze concrete regulatory action and industry practice changes — not a Hollywood scenario. The industry would benefit from proactively preventing such incidents rather than responding to them.

Open-weight model capabilities reach within 10% of frontier closed models on standard benchmarks. Meta's Llama and Mistral's releases have been advancing capability rapidly, and the trajectory suggests that by end of 2026, the best open-weight models will be genuinely competitive with frontier closed models on the major benchmarks (with the possible exception of the most advanced reasoning tasks). This will intensify the debate about open model governance and will present closed-model providers with significant pricing pressure.

The AI search war produces a restructuring of web economics. Perplexity, Google's AI Overviews, Bing's AI integration, and the broader shift toward AI-mediated information access are eroding the traffic-based economics that have funded web content creation for two decades. By end of 2026, the impact on publisher revenue will be significant enough to force new business model experiments (content licensing deals, AI-specific subscription models, premium behind-AI-paywalls content) and will be a live policy debate. The New York Times's lawsuit against OpenAI is a harbinger, not an outlier.

Genuine Uncertainty: Where I'm Unsure

The AGI timeline question. Anthropic, OpenAI, and Google DeepMind leadership have all suggested that human-level AI on many tasks may be 3–7 years away. Sam Altman has described OpenAI as potentially building the last technology humans need to build. I hold significant uncertainty about these timelines. The progress has been real and rapid, but the history of AI prediction includes many confident claims about imminent breakthroughs that didn't materialize on schedule. I think the current trajectory could hit human-level performance on narrow cognitive tasks within a few years while leaving broader general intelligence elusive for longer. But I am genuinely uncertain, and I think anyone who claims certainty in either direction is overconfident.

Whether scaling continues to deliver. The implicit assumption in most AI investment and roadmap thinking is that scaling — more compute, more data, more parameters — will continue to yield capability improvements. The evidence through 2025 generally supports this, but there are real questions about where the current paradigm hits a ceiling. Efficient frontier methods, mixture-of-experts architectures, and test-time compute improvements may extend the scaling hypothesis's useful life, but a regime change (where more compute stops producing predictable capability gains) would significantly alter the economics and strategy of frontier model development.

China's AI competitive position. The combination of U.S. export controls on advanced semiconductors and the extraordinary research output from Chinese AI labs creates a complex picture. On one hand, compute constraints from chip export restrictions should slow frontier model development in China. On the other hand, Chinese research on efficiency (getting more from less compute), the enormous domestic market and data advantages, and the strong talent base make a simple "China falls behind" narrative too simple. Qwen, DeepSeek, and other Chinese frontier models have continued to impress. The competitive dynamics here will matter for geopolitics as much as for technology, and the 2026 picture remains genuinely uncertain.

The role of synthetic data. A major constraint on continued scaling is the availability of high-quality training data — the internet is finite, and concerns about "running out of data" are real. The hypothesis that synthetic data (AI-generated training data used to train better AI) can substitute for human-generated data and extend the scaling trajectory is being actively tested. OpenAI's o1 training leveraged synthetic reasoning traces. Whether this approach scales to produce continued capability improvements without the quality degradation that naive synthetic data generation produces is one of the most important open technical questions.

The Meta-Trend: AI as Infrastructure

Stepping back from specific predictions, the most important development of 2026 may be conceptual: AI moving from "the thing we're adding to our product" to invisible infrastructure that underlies everything. Just as compute, storage, and networking became infrastructure rather than features, AI capability is becoming a foundation layer that everything else is built on.

When this transition is complete — and we're in the middle of it now — the relevant competition is no longer about which product has an AI feature. It's about which organizations have built the proprietary data infrastructure, the human expertise, and the governance systems to deploy AI effectively across their operations. The technology itself will be broadly available; the competitive advantage will be in the application.

This is a more mature and more interesting moment than the initial excitement about any particular model or product. We're building the infrastructure of the next era of knowledge work, and the decisions being made now — about governance, equity, safety, and the human role in AI-augmented work — will shape that infrastructure for a long time. I think that's worth taking seriously, with both excitement and care.

The organizations and researchers who contribute most to making AI genuinely useful, genuinely safe, and genuinely accessible to more people will look back on 2026 as one of the pivotal years in a genuinely pivotal moment. I intend to be among them.