The DeepSeek Sequel: When Openness Becomes Geopolitical

Twelve months after the shock, the real sequel to DeepSeek appeared on different ground: technical openness as an instrument of power, not as its opposite.

Twelve months after the initial earthquake, the real aftershock of DeepSeek showed up in a different field from the one analysts anticipated. The conversation started with chip efficiency. It ended with soft power. The Hangzhou startup that shook the markets early last year did not, in fact, cause a cost crisis for Silicon Valley. It produced a mutation in the grammar of open source, one that no free software manual had anticipated.

The figure that frames the conversation is blunt. Close to a third of global artificial intelligence usage now runs through open Chinese models. Building applications on that infrastructure is more than ninety percent cheaper than paying for OpenAI's API. When the cost gap reaches that size, we leave the territory of ideological preference and enter the territory of pure economic viability. For a team in Lagos, Kuala Lumpur or Santo Domingo, the decision about which model to adopt leaves the geopolitics seminars and lands in spreadsheets about whether the project survives.

Here the first distinction arises, one the international press rarely makes explicit. Open source means one thing, open weights mean another, and auditable architecture means a third. Most Chinese models that circulate as open source publish the trained parameters, not the training datasets, not the alignment procedures, not the safety filters. The user can run the model. Rebuilding it from scratch is another story. And in the Chinese case, alignment operates under an explicit regulatory mandate on topics sensitive to the Communist Party's political line. Whoever adopts these models imports, without intending to, an architecture of silences.

The paradox that deserves to be named plainly is this one. A country known for strict controls over technology and information became the world's leading exporter of open models. Technical openness stopped being a synonym for political openness, and American corporate closure stopped being a synonym for democratic protection. The categories inherited from free software no longer explain the terrain.

We are in a zone where openness works as an instrument of power, not as its opposite. South Korea, Taiwan and Australia have already banned their officials from using DeepSeek. The same piece of software that a developer celebrates for its availability operates, for a national security official, as a vector of exposure.

Sam Altman said he wanted to make sure democratic artificial intelligence won out over authoritarian artificial intelligence. The phrase is a textbook example of narrative entropy: the discursive disorder that results when actors with very specific material interests borrow vocabulary from political philosophy so their arguments sound universal. There is a legitimate version of that concern: authoritarian regimes will use artificial intelligence to reinforce social control, and that is factual.

There is also the instrumental version, where democratic ends up meaning mine and authoritarian ends up meaning the competitor's. When the political audience in Washington cannot tell one version from the other, we end up with regulation that protects market share under the language of protecting democratic values. That confusion impoverishes the debate and empties of meaning concepts we need to preserve for future discussions.

The pattern China is executing in artificial intelligence is recognizable. A good-enough product at an aggressive price, capture of adoption, productive scaling, improved margins, eventual dominance of the value chain. We saw it in textiles in the nineties, in solar panels in the first decade of the 2000s, in lithium batteries in the last decade. The difference with cognitive models matters. What gets externalized when you build the digital economy on a foreign model goes beyond computing costs.

Part of the judgment gets externalized: what counts as valid information, what counts as an appropriate response, what counts as permitted language, what counts as an acceptable interpretive frame. This is algorithmic sovereignty, and it is ceded silently, without public debate, one technical decision at a time.

For the Dominican Republic, the current scenario represents the greatest democratization of cognitive computing capacity in recent history. A small team in Santo Domingo, Santiago or La Vega can now build applications that eighteen months ago required six-figure contracts with American vendors. That openness is real, it is positive, and it deserves to be celebrated. At the same time, we have to separate out something else clearly. Switching from an American provider to a Chinese one reduces costs. It leaves our structural position intact. We remain importers of cognitive capacity, assemblers of solutions built from parts we do not manufacture, integrators of systems we did not design.

The country's drop in the Global Innovation Index, from rank 90 in 2020 to 97 in 2024, suggests that connectivity alone is no longer a competitive differentiator. What is missing is the cognitive layer. We have mobile penetration close to 94 percent, a digital economy that barely reaches 4.3 percent of GDP excluding telecommunications, and national investment in artificial intelligence of around ten million dollars in 2024.

Three staged routes can change that configuration:

The first is an explicit public policy on the use of foreign models in critical state infrastructure, with criteria of transparency, traceability and reasoned selection that replace today's case-by-case decisions with no common framework.

The second is incentives so that the technical talent that today figures among The invisibles —employed by foreign companies from the Dominican Republic without building local institutional capacity— can build companies that produce models, tools and applications for the regional market, where SMEs account for 98 percent of the business fabric.

The third is infrastructure to run open models —Chinese, American, European, whatever— on locally controlled compute, with clear policies for protecting citizens' data and tracing algorithmic decisions. A reasonable defensive investment, within reach of the available fiscal budget, and compatible with our position as a middle economy in the region.

The lesson of this anniversary is austere and useful. Open source, on its own, never produces sovereignty. What produces sovereignty is the set of decisions a country makes about what to do with the open source available. China built a coordinated strategy to use openness as an instrument while controlling the internal channels of adoption. The United States is losing global ground because its ideological discourse fails to translate into policies that keep an accessible technical presence.

Countries on the periphery gain space when we turn the availability of cheap models into capacity of our own, built with judgment. Availability comes free. Capacity costs coordinated effort. And the window to start that effort has shortened by twelve months.