Every industry has a moment when its favourite assumption gets repriced. For AI, that moment arrived when DeepSeek showed that a near-frontier model could be trained and served for a fraction of what the market assumed it cost.
The shock was not that a Chinese lab built a good model. It was the combination: strong results, radically lower prices, and weights anyone could download. Each element alone was manageable; together they attacked the business model.
The cost story that broke the spell
The AI boom rested on a simple belief: intelligence is expensive, so the winners will be whoever spends the most on chips and data centres. Valuations, power deals and construction plans all leaned on that belief.
DeepSeek’s published figures suggested its models reached competitive quality with far less compute than expected. Even allowing for debate about what those figures included, the direction was unmistakable: clever engineering could substitute for brute force.
Investors reacted to the implication, not the spreadsheet. If models get cheaper to build, the moat around the biggest spenders gets shallower, and the selling price of intelligence falls toward the cost of producing it.
Open weights turned a launch into a flood
A closed model launch is an event. An open-weight launch is a release into the wild: within days, the model is mirrored, fine-tuned, quantised and embedded in products its creators never imagined.
That is what happened with DeepSeek. Developers who would never sign an enterprise contract could suddenly run a serious model on their own infrastructure, including on local-AI hardware like the machines covered in our RTX Spark laptop report.
Once thousands of teams build on a free model, its ecosystem becomes self-reinforcing. Tutorials, tools and integrations accumulate, and switching away gets harder even for users who never paid a cent.
Cheap changed what buyers expect
DeepSeek also reset price expectations. Its API pricing undercut comparable services so sharply that “good enough and nearly free” became a real option for routine tasks like summarising, drafting and classification.
That pressure travels upward. Premium providers now have to justify their prices with reliability, safety, support and genuinely better results, rather than with the mere fact of being a frontier lab.
Trackers estimate DeepSeek properties draw roughly 319 million visits a month in 2026, with about 130 million monthly users in China. Numbers on that scale mean the pricing pressure is not a niche experiment.
How rivals answered. The first response was price. Within months of DeepSeek’s breakout, API prices across the industry slid, free tiers grew more generous, and “efficiency” replaced raw scale as the word every launch event reached for.
The second response was openness itself. Labs that had treated weights as crown jewels began releasing smaller open models of their own, calculating that an ecosystem they partly own beats a standard set entirely by someone else.
The third response was narrative. Big spenders reframed the argument around reliability, safety certification, enterprise support and the very frontier, the places where cheap and open are not automatically enough.
Hardware felt the shock differently. If each unit of intelligence needs less compute, demand shifts rather than disappears: more users, more agents, more countries building their own stacks, all still needing chips.
The market’s deeper adjustment was psychological. Before DeepSeek, “they spent more” passed as analysis. Afterwards, every spending announcement had to survive a new question from investors: what does this buy that clever engineering cannot?
What the shock did not change
Frontier research is still expensive, and the very best models still come from the biggest training runs. DeepSeek narrowed the gap; it did not repeal the economics of scale.
It is worth remembering how complete the surprise was. Analysts who covered the sector for a living had modelled many risks (regulation, power shortages, talent wars) but not a small lab collapsing the assumed cost curve in public.
The aftershocks are still working through budgets. Every cloud provider, chip buyer and foundation-model startup now writes its plans in a world where a DeepSeek-style efficiency jump is a standing possibility, not a black swan.
What changed permanently is the burden of proof. Spending is no longer its own argument. Every lab now has to show what its extra billions actually buy, in quality users can feel.
0 Comments