The story of AI in 2025 is not about who shipped the smartest model. It is about which forces have crossed from speculation into measurable, durable structure. Five have. Everything else is noise that will look obvious in retrospect.
This is a signal-backed survey of those five forces, drawn from the live Steek index and the underlying signal stream. We do not rank AI trends by how often they appear on social media. We rank them by accumulated, typed evidence of change.
The five forces
- Reasoning models. Long-thinking, RL-trained inference replaces brute-force completion as the frontier UX.
- Agentic workflows and computer use. Models begin operating real software on a user's behalf.
- Inference cost collapse. Per-token cost falls 80–95% per year while quality keeps rising.
- Open-weights frontier. The gap between open and closed models compresses to under 6 months.
- Sovereign AI. Nation-states fund domestic compute and models as strategic infrastructure.
1. Reasoning is the new pre-training
For five years, every frontier release was an answer to "how big a model can we afford to pre-train?" In 2025 the question changed to "how long can we afford to let a model think?" OpenAI's o1 in late 2024 was the ignition. DeepSeek R1 in January 2025 was the moment the technique escaped a single lab. By Q2, every flagship API exposed a reasoning_effort parameter as a first-class control surface.
The implication is structural. Pre-training scaled along one axis (parameter count) with very predictable cost curves. Reasoning scales along two — model quality and inference budget — and the second axis is decided by the user, per request, in production. Pricing pages, latency budgets, and product UX all rebuild around it. The full mechanics are documented in our velocity model.
2. Agents stop being a demo, start being a product surface
"Agents" was the most over-promised word of 2023. In 2025 it became the most under-priced one. Three signal classes converged:
- Tool-use protocol consolidation. Anthropic's Model Context Protocol was adopted across vendors. Read the trend page.
- Computer-use capability. Vision-grounded agents began driving real screens. See the timeline.
- Per-task pricing. Vendors started selling outcomes ("close this ticket") instead of seats.
Together, these turned every SaaS category whose value is sequential clicks into contested territory. CRMs, ITSM, RPA, ETL, scheduling, support — anything that is mostly orchestration is now a candidate for an agent layer that does the orchestration on the user's behalf.
3. Intelligence per dollar collapses again
From March 2023 to early 2025, the marginal cost of a token of equivalent quality fell roughly an order of magnitude every 12 months. It is not slowing. Three forces compound: distillation of frontier capabilities into smaller models, accelerator competition (Groq, Cerebras, AMD MI300, custom silicon at every hyperscaler), and improved serving stacks (speculative decoding, KV-cache reuse, batching).
When the unit cost of intelligence falls 10× per year, the strategic implication is simple: any product priced per seat is exposed to a per-task competitor that uses the savings to give the work away. The full dataset is on the inference-cost-collapse trend page.
4. The open-weights gap closes to months
DeepSeek R1 was the moment the open-weights story became impossible to dismiss. Frontier-quality reasoning shipped under a permissive license, available for any organization to download, fine-tune, and deploy. By mid-2025 the gap between the best open-weight and best closed model is measured in months rather than generations.
That does not destroy closed labs — distribution, data, and product remain durable moats — but it caps how much rent any closed lab can extract per token. Trend page.
5. Sovereign AI restructures the supply chain
NVIDIA naming "sovereign AI" as a revenue category in early 2024 was the start. Through 2024 and 2025, France, UAE, Saudi Arabia, Japan, India, and the UK announced multi-billion-dollar sovereign-AI funds. The bet is no longer about whether AI is a strategic technology — every government has accepted that — but about who controls the training compute and the model weights.
The downstream signals are visible in NVIDIA's geographic revenue mix, in domestic model launches from regional champions, and in export-control rule updates. See the trend page.
What is not a 2025 trend
A short list of stories that received disproportionate attention but show weak signal density:
- "AGI is here." No benchmark, hiring, or pricing signal supports the framing. Capability is rising; "general" remains undefined operationally.
- "AI bubble bursting." Capital deployment and enterprise revenue both still accelerating. The bubble vocabulary is a sentiment trade, not a signal trade.
- "Crypto-AI fusion." Persistent narrative, scarce evidence. Few tokenized AI projects show production utility beyond speculation.
Reading the rest of the year
The five forces above feed each other. Reasoning makes agents work. Agents create demand for cheaper inference. Cheap inference makes open weights viable. Open weights enable sovereign deployment. Sovereign capital funds the next round of training compute. The compounding is the story.
To watch it in real time, the right surfaces are the live trends index and the underlying signal stream. The methodology behind both is documented in The AI Signals Report.