Trinidad Signed an $800 Megawatt AI Bet. Six Days Later, a Chinese Lab Made the Math Riskier.

Two Signings, Six Days Apart

On 11 July 2026, Trinidad and Tobago's government signed memorandums of understanding with Ernst & Young, for a 300-megawatt data center, and with Florida-based Hummingbird AI Holdings, for an initial 150 megawatts with room to expand to 500. Combined, the proposals reach up to 800 megawatts. The government put the expected investment at US$5 billion and projected more than 5,000 jobs. Hummingbird has targeted the first quarter of 2028 for initial commercial operation.

Against Trinidad and Tobago's own grid, that is not a rounding error. PowerGen, Trinidad Generation Unlimited, and Trinity Power together run roughly 2,289 megawatts of installed national generating capacity. The MOUs signed on 11 July ask that grid to make room for demand equal to more than a third of everything currently installed, on an island where residential and commercial customers already share the same generation pool the data centers would draw from.

Six days after the signing, on 17 July, Chinese lab Moonshot AI released Kimi K3: a 2.8 trillion parameter model priced at $15 per million output tokens, a third of Anthropic Fable 5's $50. According to Moonshot's own benchmarks, K3 performed competitively with Fable 5 and ahead of Anthropic's Opus 4.8 and OpenAI's GPT 5.6 Sol. An independent Arena.AI benchmark ranked it the best model currently available, and it became the first open-source model to top the Code Arena programming benchmark, ahead of proprietary labs.

What a Cheaper Model Did to the Chip Trade

The market did not wait for a research paper. Taiwan Semiconductor fell 7% that Friday. SoftBank fell 9%. Chinese rival Z.ai plunged nearly 30% in Hong Kong trading. Nvidia dropped 1.2% and briefly lost its position as the world's most valuable company to Apple. Meta slid more than 2.4%, and the Nasdaq 100 was down a full point by early afternoon in New York. JPMorgan's note on the move called it a DeepSeek 2.0 moment, a direct reference back to the original shock a Chinese open-weight model delivered to US chip valuations. Per Fortune's reporting on the release, Anthropic chief executive Dario Amodei and others had not expected a Chinese lab to ship a competitive frontier model for another six months. Kimi K3 compressed that timeline to zero.

The selloff had more than one cause that week, and it is fair to say so. Weak Netflix and TSMC earnings, an escalation involving Iran, and a broader risk-off mood in equities all landed in the same seventy-two hours. Kimi K3 was one trigger among several, not the sole one. What it changed specifically, and what the broader macro noise did not, is the price the market puts on the assumption that frontier AI output requires large, expensive, scarce GPU clusters. K3 argued, in public, with a benchmark score and a price list, that it does not require quite as many of them as the capex plans assume.

The Arithmetic Under Every Megawatt MOU

Every Caribbean data center MOU signed this year is underwritten by the same premise driving the $7 trillion in global data center spending Goldman Sachs tracks through 2030: that AI compute demand will keep outrunning supply, so the capacity gets built first and the electricity bill gets passed along after. Goldman's own estimate has that buildout raising US electricity costs 6% between 2026 and 2027, with another 3% on top by 2028. PJM, the largest grid operator in the United States, expects consumer electricity costs to rise $6.3 billion over the next three years because of it. Virginia, the state carrying the heaviest data center load, saw residential electricity prices rise 13% in a single year.

Asa Watten, a co-author at the Electric Power Research Institute, put the risk in one sentence to Fortune: "If the grid builds capacity, expecting a lot of demand from data centers, and that doesn't show up, that could increase prices in the future." That is the mechanism Trinidad and Tobago's utility regulators, and every government negotiating a similar MOU across the region, now have to price explicitly. A grid that expands for demand that arrives late, arrives smaller, or gets served more cheaply by a more efficient model somewhere else, still has to recover the cost of the capacity it built.

Efficiency Did Not Make the Hardware Disappear

The counter-argument matters as much as the shock did. Moonshot AI's own deployment guidance still recommends clusters of at least 64 high-end AI chips to run Kimi K3 in production. The model activates only a portion of its 2.8 trillion parameters per request, an architectural efficiency that trims the compute bill, but the underlying hardware requirement has not gone away. Data centers were, on average, a deflationary force on electricity prices for a decade: research covering 2015 to 2024 found that every doubling of data center capacity coincided with a 3.5% national and 6% statewide fall in average retail electricity prices, because fixed grid costs spread across more paying customers. That history argues against reading Kimi K3 as the end of the Caribbean's infrastructure thesis.

What it argues for is a more specific kind of diligence than "the MOU was signed." A megawatt commitment priced on the assumption that GPU-hungry training runs will keep multiplying is a different asset than one priced on steady, contracted inference demand from named tenants. The first kind gets repriced every time a lab like Moonshot ships something cheaper. The second kind does not, because the workload was never speculative in the first place.

What Diligence Has to Check Now

Caribbean governments and the capital following them into these deals are not choosing between "build" and "don't build." They are choosing what kind of commitment to sign. The table below is the screen Maestro AI Labs runs against any Caribbean AI infrastructure MOU that crosses our desk after 17 July.

Question	Why it matters after Kimi K3
Fixed or indexed PPA?	A power purchase agreement priced on fixed capacity regardless of actual utilization leaves the utility, and by extension the ratepayer, carrying the cost if a cheaper model reduces the workload the facility was built for.
Named anchor tenant?	An MOU backed by a contracted workload from an identified customer is a different asset than one built against a general hyperscaler pipeline that has not committed to specific capacity.
Rate base exposure?	Virginia's 13% residential price rise in one year shows what happens when local ratepayers absorb the gap between built capacity and realized demand. The same exposure applies to a small island grid at a smaller scale.
Training or inference load?	Training-scale GPU demand is exactly what an efficient model like Kimi K3 is designed to undercut. Contracted inference serving for identified customers is far less exposed to that repricing.
Delivery timeline risk?	Hummingbird AI Holdings' own target of Q1 2028 for first commercial operation leaves eighteen months for compute pricing to keep moving before a single watt is billed.

None of this argues for cancelling what is already signed. Trinidad and Tobago's MOUs create real jobs and real capital inflow regardless of how the compute-pricing argument resolves. It argues for underwriting the next MOU, and re-underwriting the ones already on the table, against a compute market that just proved it can move faster than a data center gets built.

Where This Leaves the Region's Bet

StarApple AI, the first AI company built in the Caribbean, and its founder Adrian Dunkley, the region's most consistently cited AI voice, have tracked this exact gap for two years: the distance between an announcement a government can point to and a facility that generates revenue on schedule. Kimi K3 did not create that gap. It gave it a price, in the form of a chip selloff that landed six days after the region's biggest AI infrastructure signing to date.

The practical move for anyone with capital in this space is not to wait out the volatility. It is to ask each Caribbean MOU on the table the same five questions in the table above, before the next signing ceremony, not after the next selloff.