When the Best Open Model Came Out of Hangzhou

Note from the President

The first months of every year ask the same two questions of an association like ours. What did we miss last year, and what should we be doing about it now. Lisette Boucher's piece this month addresses both, on the back of the DeepSeek R1 release on the twentieth and the broader story of where capital, talent, and capability are flowing in the AI industry as the new year opens.

I will add one item from this office. The CAIRA Founders Cohort started its first session on January 9 with twelve Caribbean entrepreneurs. Half are building businesses that did not exist as concepts when this Association was founded six months ago. The other half are AI-rewiring businesses that were not viable at our last gathering. I encourage every CAIRA member to follow what they are doing through this year. Their progress is the best measurement instrument we have for the regional opportunity.

January is also Cervical Cancer Awareness Month. Cervical cancer is one of the most preventable cancers in the world, and our region still loses too many women to it. If the women in your life have not been screened in the last three years, this is a good month to make the appointment with them.

Adrian Dunkley Founder and President, Caribbean AI Association

---

Feature

When the Best Open Model Came Out of Hangzhou

By Lisette Boucher

On the twentieth of this month a Chinese research lab called DeepSeek released a reasoning model named R1, with weights, paper, and a permissive licence. Within seventy two hours it was the most discussed AI model in the world. Within five days it had wiped roughly six hundred billion US dollars off the market capitalization of US technology stocks, briefly become the most downloaded application on the US iOS store, and forced the largest US AI firms into a defensive crouch they have not been in since the launch of ChatGPT in November 2022.

If you read this newsletter in part because you want a window on where the industry is going and what it means for opportunity in our region, the DeepSeek event is the most important thing that has happened since I started writing for CAIRA. Let me say why.

The cost structure of frontier AI has just been publicly challenged. DeepSeek's published technical reports place the training cost of their reasoning model in the low millions of US dollars, against tens or hundreds of millions for the equivalent US frontier model. Some of the cost claim is contested. Some of it survives serious scrutiny. The point for our purposes is not whether the number is six million or sixty. The point is that the floor of what is possible has been demonstrated to be much lower than the industry consensus assumed, and that demonstration has been documented openly enough that other labs around the world will now reach for the same playbook.

The open weights question is settled, at least for one cycle. The argument over whether the highest-performing models would remain closed and proprietary, or whether competitive open models would catch up, was a live debate in the US AI industry through 2024. As of last week, the answer is provisional but clear. The best open model is competitive with the best closed model on the tasks people actually use these systems for, and is available to download to any sufficiently equipped machine in the world. That changes what it costs for a Caribbean firm to build on top of frontier AI, in three specific ways.

First, the cost of inference, the per-query cost of running the model in production, drops substantially when you can run an open model on infrastructure you control or rent on commodity terms, rather than paying API rates set by a small number of US providers. For a Caribbean fintech that wants to embed AI into a banking workflow, this is the difference between a feature that breaks unit economics and one that improves them.

Second, the data sovereignty question becomes tractable. Several CAIRA member firms in financial services and health technology have told me, on background, that the regulatory blocker to AI adoption in their industries is not capability, it is the inability to send sensitive customer data to a US provider's servers. Open models that can be deployed inside a Caribbean data centre, or on premise, are a regulatory unlock that did not exist twelve months ago.

Third, the differentiation thesis shifts. If the model is a commodity, then the value migrates to what surrounds the model. To the proprietary data you have. To the workflow integration. To the user interface. To the customer relationship. These are the places where a well-run Caribbean software business has a fair chance of competing with a Silicon Valley equivalent. They are not the places where capital alone wins.

The talent picture is changing too. DeepSeek's published author list, and the broader Chinese AI lab output of the last six months, makes it clear that frontier AI research is no longer concentrated in a handful of US institutions. Beijing, Hangzhou, Shanghai, Paris, Toronto, Tel Aviv, Bangalore, and Seoul are all on the map now. Caribbean students considering graduate study in machine learning should expand their list. The two best places to study reasoning model research in 2025 are not in California. They are in Hangzhou and Paris.

What does this mean for our region in concrete terms.

For Caribbean founders. The AI-native business you wanted to build six months ago is now meaningfully cheaper to build, both in compute cost and in deployment complexity. If you held back because the unit economics did not pencil, run them again this month. Two of the cohort members in CAIRA's Founders programme are quietly re-architecting their products this week to swap a US API for an open model deployed on a regional data centre, and the numbers are coming out very differently.

For Caribbean enterprises. The case for waiting on AI adoption because the technology is too expensive and the providers too few is much weaker today than it was on December 31. If your IT director has been telling you the procurement decision is hard, ask for a fresh memo. The option set has changed.

For Caribbean policy makers. The deployment of open weights AI inside our region's data centres raises legal questions that our existing frameworks do not yet handle well. Who is liable for the outputs of a model your firm has fine-tuned and is hosting yourself. What does cross-border data flow mean when the model is downloaded and runs locally. What does compute reporting look like when there is no longer a single foreign vendor on the other end. These are the questions for 2025, and we are not late to them, but we cannot afford to be slow either.

For Caribbean investors. The case for a regional AI infrastructure fund is materially stronger today than it was last quarter. Not a fund for chip-scale investment, our region cannot economically justify that, but a fund for the data centre capacity, the model deployment platforms, and the application layer companies that will run on top of them. The Caribbean Development Bank, the Inter-American Investment Corporation, and selected family offices in the region should be looking hard at this thesis.

I have been writing about industry structure for twenty years, in the Caribbean and outside it, and I have learned to be suspicious of moments that get described as turning points. Most of them are not. This one might be. It might also be a flash, and the closed model labs in California might assemble a response by the second quarter that puts the world back where it was. Both possibilities should be priced into your planning.

What I will tell you with confidence is this. The next eighteen months in AI will be more competitive, more open, more multi-polar, and less predictable than the eighteen months that preceded them. The Caribbean has spent its history being the place where global flows came together. AI is a flow. Read the wind, set the sail, and do not wait for the captain in the next harbour to tell you when to go.

---

Lisette Boucher is an industry analyst covering technology and capital markets across the Caribbean and Latin America. She is the founding partner of Cane Row Partners and serves on the CAIRA Industry Council.

Originally published in The Monthly Intelligence Report, January 2025.