On 27 January 2025, Nvidia lost 17 percent of its value in a single trading day. Around 593 billion dollars in market capitalisation evaporated, the largest one-day loss any company has ever posted on a US exchange. No product recall lay behind it, no weak quarter. A Chinese lab had given a model away for free.
That tells you most of what you need to know about the year. Everyone expected the models to improve. Far fewer had the collapse in what it costs to use them anywhere on their map.
January: the model that cost nothing
DeepSeek published R1 on 20 January with open weights under an MIT licence. It performed roughly on par with OpenAI's o1, and anyone could download it and run it themselves. Until that point the market had priced in an assumption: serious AI requires billions in compute. On 27 January that assumption was re-examined in the space of an afternoon.
The panic itself passed within a fortnight. The consequence did not. For the rest of the year, per-token pricing fell at essentially every vendor, and open models became a genuine option for anyone who preferred running something on their own hardware to shipping customer data into an American cloud.
March: the standards war that never happened
On 26 March, Sam Altman said OpenAI would support the Model Context Protocol, the open standard for connecting AI models to real systems that Anthropic had released the year before. A competitor adopted a rival's protocol in public, without a fight. The industry is rarely that grown-up.
That this also mattered to a six-person company was not obvious from the headlines. MCP is the reason an assistant can now read your calendar, pull an invoice out of your accounting system and draft a reply in one motion. With three competing standards, every such integration would have had to be built three times, which in practice means it would never have been built for small firms at all.
August: the regulation, then GPT-5
On 2 August, the EU AI Act's rules for providers of general-purpose AI models started to apply. Five days later OpenAI shipped GPT-5, with a router that decides on its own when the model should think longer and when it should simply answer.
August's obligations landed on the people who build the models. The cleaning firm in Tyresö using them was untouched, and models already on the market were given until 2 August 2027 to comply. The date that mattered in practice arrived later: the AI Office only gained its full enforcement powers on 2 August 2026.
Sweden, by the numbers
Statistics Sweden published its measurement on 6 November 2025. The share of Swedish companies with at least ten employees using some form of AI went from 25.2 percent in 2024 to 35.0 percent in 2025. Just under ten percentage points in twelve months.
The breakdown by size, from the press release of 11 December, is the figure that actually says something:
- 10–49 employees: 30.8 percent
- 50–249 employees: 49.6 percent
- 250 employees and above: 71.9 percent
The gap between the first and the last line comes down to staffing. A large company has an IT department that can be told to evaluate something, given time, and asked to come back with a proposal. A company of twelve has nobody whose job that is. It usually ends with someone trying ChatGPT one evening last autumn, finding it fairly impressive, and going back to answering email by hand.
One concrete thing you can do before building anything at all: look at what people nearby are actually asking for help with. On Önskemål, private individuals and businesses post their jobs in plain words, and that list is more honest market research than most reports you can buy.
What the year left behind
In 2023, getting started with AI cost money and expertise. By the end of 2025 it cost an afternoon and a card. The barrier everyone had discussed for three years simply disappeared, quietly, somewhere between DeepSeek in January and Christmas.
A different problem remains, and it is the harder one. Knowing what to automate first. Leaving alone the thing that worked perfectly well as it was. We have previously worked through what automating customer replies actually costs and where the line runs between a chatbot and an auto-reply — both questions are really the same one.
The answer to it decides whether that 30.8 percent reads 50 the next time Statistics Sweden counts.
The story continues in our review of the first quarter of 2026, when AI moved from answering in a chat window to working directly in your files.
Sources: Statistics Sweden, Use of ICT in enterprises 2025 (2025-11-06) and Larger companies use AI more (2025-12-11); European Commission on the AI Act's rules for general-purpose AI models (2025-08-02); OpenAI (2025-08-07).