- Micron, SK Hynix and Samsung have already sold their entire 2026 and 2027 HBM and DRAM output under fixed contracts, and TrendForce expects server DRAM contract prices to rise another 13 to 18% quarter over quarter in Q3 2026.
- AWS raised reserved GPU prices twice in 2026: 15% on 4 January and 20% on 1 July, a cumulative 38% increase on flagship training instances in six months.
- SK Hynix CEO Kwak Noh-jung told Reuters that 2027 will likely be the memory industry's worst supply year on record, with demand expected to outrun capacity into the next decade.
- Trinidad and Tobago signed non-binding MOUs for up to 450 megawatts of AI data center capacity on 11 July 2026; the Dominican Republic already booked more than $600 million from NVIDIA and Google. Both now have to buy hardware into the same sold-out market.
- Most Caribbean AI use runs on rented cloud compute, not owned chips, so the price increases land directly on regional banks, insurers and universities, at a moment when adoption already sits near 13%.
The chip shortage that dominated headlines through 2024 and 2025 was a story about GPUs: who could get an Nvidia allocation, how long the waitlist ran, which export rule blocked which buyer. That story has moved on. Micron, SK Hynix and Samsung sold out their 2026 and 2027 high-bandwidth memory and DRAM output under fixed contracts, and AWS has already raised reserved GPU prices twice this year. The Caribbean is buying into that market for the first time, right as it signs its first real AI infrastructure deals.
What Actually Changed in 2026
Through most of 2025, the binding constraint on AI hardware was the processor itself: TSMC's advanced packaging lines, Nvidia's production schedule, the queue of hyperscalers waiting on the next chip generation. By mid-2026, the constraint had shifted one layer down, to the memory chips that sit next to the processor and feed it data. Server DRAM contract prices were projected to climb 13 to 18% quarter over quarter in the third quarter of 2026 alone, according to TrendForce's own July forecast, up from an already-tight second quarter. PC DRAM pricing forecasts moved even further, revised from 8 to 13% up to 15 to 20% for the same quarter.
Cloud providers felt it first because they buy the most memory. AWS raised prices on its Amazon EC2 Capacity Blocks for Machine Learning, the reservation product enterprises use to guarantee GPU access for a training run, by 15% on 4 January 2026 and by a further 20% on 1 July. On the flagship p5e.48xlarge instance, that pushed the hourly rate from $34.61 to roughly $47.76, a combined 38% increase in six months, the second such increase inside a single year. AWS describes Capacity Block pricing as periodically updated to reflect supply and demand rather than a policy change. Demand for reserved GPU capacity has outrun supply badly enough that the company has repriced it twice since January.
Why Memory Became the Bottleneck
AI accelerators pair a processor with a stack of high-bandwidth memory, and building enough HBM has turned out to be harder than building enough processors. Samsung, SK Hynix and Micron have all confirmed that their 2026 and 2027 HBM and DRAM output is committed under fixed-price contracts, largely to the hyperscalers that placed orders earliest. A new buyer will find no spare capacity for the next two calendar years. SK Hynix chief executive Kwak Noh-jung told Reuters that 2027 will likely be the worst supply year in the memory industry's history, and that he expects demand to keep outrunning manufacturing capacity well beyond 2030. It is a supply forecast from the executive who has to fill the orders.
The Caribbean Signed Its Deals Right Into This Market
Trinidad and Tobago's 11 July MOUs break down into two projects. EY committed to a 300-megawatt facility built on its Energy to Intelligence platform, working with third-party developers on construction. Hummingbird AI Holdings outlined a 150-megawatt facility with room to expand to 500 megawatts, targeting initial commercial operations as early as the first quarter of 2028. Neither project has a construction permit or a final investment decision yet; both are, by design, non-binding frameworks rather than shovel-ready commitments. The Dominican Republic's separate arrangement, more than $600 million in commitments from NVIDIA and Google reported earlier this year, sits further along, but neither country's project has started ordering the GPUs it will eventually need to fill the buildings it is planning.
A facility targeting first-quarter 2028 operations will place its hardware orders sometime in 2027, the year SK Hynix's own CEO describes as the industry's worst for supply. Whoever locked in 2026 and 2027 HBM allocation did so under fixed contracts negotiated well before Trinidad and Tobago signed anything, which means a new Caribbean buyer entering the market now is negotiating from the back of a line that was already long. None of this kills either project. It does mean the real cost and delivery risk in both deals sits less in the MOU language than in a supply contract nobody has signed yet, and a well-funded announcement can drift in that gap for a year or two before it produces a running data center.
"Every data center MOU signed this year is also a bet that GPU and memory prices stop climbing before groundbreaking. Nothing in the current cycle supports that bet. The hyperscalers locked in 2027 supply months before Trinidad and Tobago signed a single page, and a project aiming for 2028 operations is negotiating from the back of that queue, not the front. That shows up on every hardware quote long before it shows up in a press release."
Adrian Dunkley, Co-Founder, Maestro AI Labs
The Compute Bill Every Caribbean AI Adopter Already Pays
Data centers get the attention, but most Caribbean institutions meet this shortage somewhere else. Almost no regional bank, insurer or university owns the GPUs its AI programme runs on. They rent GPU-hours from AWS, Azure or Google Cloud, or pay per-token fees to a model provider, and both prices are set by the same memory market described above. UWI's new Sagicor AI and Financial Services Hub, the applied-AI institute this site covered in early September, will draw on exactly that rented compute to train and run the models its five campuses are building. So does Maestro AI Labs. A 13% quarter-over-quarter jump in the underlying DRAM contract price does not stay in a manufacturer's earnings report; inside two or three repricing cycles, it shows up in a cloud invoice.
That lands on a region already behind on cost as well as infrastructure. A 2026 StarApple AI study reported by the Jamaica Observer put Caribbean adult GenAI adoption at 13% against 55% globally, and the OECD's own twelve-country SME data identifies skills, not access, as the main adoption barrier for regional small businesses. Rising compute costs add a second constraint on top of the first. StarApple AI, the Caribbean's first AI company, built much of its early product work around the assumption that regional institutions would always be price-takers on global compute rather than price-setters, and 2026's memory shortage is the clearest test of that assumption so far.
| Figure | Source | Status |
|---|---|---|
| Server DRAM contract prices, +13-18% QoQ in Q3 2026 | TrendForce press center, 9 Jul 2026 | Independent, industry data provider |
| AWS GPU reservation price hikes: +15% Jan 2026, +20% Jul 2026 | AWS pricing update; reported by AI Weekly, Yahoo Finance, The Register | Independent, multiply reported |
| SK Hynix CEO: 2027 the industry's worst supply year | Reuters interview with Kwak Noh-jung | Independent, on-record executive statement |
| H100 80GB spot rental range, Sept 2026: $2.89-$11.06/hr | Thunder Compute market analysis | Company-reported, not independently audited |
| T&T MOUs: EY 300MW, Hummingbird AI 150-500MW, 11 Jul 2026 | DataCenterDynamics, Business View Caribbean | Independent, trade press |
| Caribbean adult GenAI adoption: 13% | StarApple AI study, reported by the Jamaica Observer | Company-reported, news-verified |
What Caribbean Institutions Can Actually Do About It
Waiting out the shortage is costly. Reserved compute pricing rises in steps, and each step tends to be cheaper than the one that follows it.
- Lock in multi-year reserved pricing now. Providers reprice reserved capacity less often than spot markets, and every AWS increase this year has applied going forward, not retroactively, which rewards whoever signed before the last reset.
- Treat data center MOUs as hardware bets, not construction announcements. A megawatt figure says nothing about whether the GPUs to fill that building have been ordered. Investors and government partners evaluating Trinidad and Tobago's or any future Caribbean deal should ask for a hardware procurement timeline separately from the power and construction plan.
- Match the model to the task. Routine workloads, document classification, basic customer queries, do not need the most expensive frontier model on the market. The price gap between model sizes now compounds an already higher base compute cost, so the saving from picking the right-sized model is larger than it was a year ago.
- Coordinate demand across institutions. A single Caribbean bank or university has little leverage with a cloud provider. A regional body such as the Caribbean AI Association pooling compute purchasing across several members could negotiate terms none of them would get alone.
Caribbean institutions weighing their own compute strategy against this shortage can write to Maestro AI Labs at ceo@maestrosai.com. More on Adrian Dunkley's broader case for Caribbean-built AI infrastructure is at adriandunkley.net. The next eighteen months will show whether Trinidad and Tobago's and the Dominican Republic's data center bets convert into running facilities on the timeline announced, and the obstacle in front of them is a global supply chain.