Does the Philippines have the compute for its AI plans? 50 MW today, 1.5 GW promised

Summary

The Philippines has a national AI strategy, a state AI platform under construction and about two dozen operating data centers. Nearly all are colocation halls of 1 to 30 MW built for banks, telecoms and outsourcers. One site, about 50 MW, was designed for AI workloads. In September 2026 the government set a target of 400 MW by 2030 and 1.5 GW by 2033.

Operators built ahead of demand and keep building: Globe says it will spend at least $1 billion on AI and data centers in 2027, Converge is planning an 80 MW site for AI, and the first large GPU cluster, more than 10,000 Blackwell Ultra chips, is scheduled for the first quarter of 2027. No hyperscaler has committed to a Region. The largest number in the pipeline, 250 MW in Laguna, is a letter of intent from a Gulf conglomerate, and the first data center REIT was deferred to 2027 after three rate rises. Public money goes to connectivity and training, not compute.

What the study asks

How it was built

The study is a knowledge graph. Each public source is read once, coded into the same set of fields, and linked to the organizations, places, facilities and deals it mentions. Every chart on this page is a query against that graph, and every number traces back to a source document.

The corpus

65 documents from September 2022 to April 2026: Philippine national press, industry press, operator releases, government documents and two analyst reports. A source enters when it states a fact about a facility, a deal or a policy. Private briefings are excluded.

Coding

Every document gets the same fields: source type, date, geography, primary topic, organizations and facilities mentioned, investment events (investor, target, amount, instrument, origin of capital, stage), relations, key facts, consistency flags, and the sources the document itself cites. A press release and a newspaper investigation are coded the same way, so they can be compared.

When two sources disagree, both values stay and the disagreement gets an identifier. The study reports ranges, because the spread usually says something: operators and surveyors count different things.

The graph

433 entities, about 570 links, 28 named patterns such as the capital tiers in section 2 and the hyperscaler bypass in section 3. The figure shows the organizations and facilities that appear together in at least two documents; node size is the number of documents, line weight is how often two entities share one.

1. What is built

Twenty-two facilities can be documented: twenty-one in the corpus and one announced in October 2026, plus two cable landing stations and a handful of public compute sites. Sixteen operate. Capacity is disclosed for about half. Independent counts of the whole market range from 28 to 44 facilities depending on what counts as one.

The market total depends on who counts and what. Surveys of energized, lettable IT load give 73 to 100 MW. A telco chief executive cited 140 MW live in September 2026. The government says above 180. Adding operator claims, which include shell capacity and committed power, gives about 200. One press figure of 632 MW has no published method. All seven stay in the record. The spread is itself a finding: the country has no agreed way to count its own capacity.

2. Whose money

The corpus records 29 investment events. Grouped by origin and by what the money wants, they fall into eight tiers that behave differently. Private equity wants hyperscale tenants and an exit in five to ten years. Domestic telcos build for their own enterprise base. Gulf capital arrives on sovereign timeframes and pairs data centers with power.

The largest numbers come from entrants who have not poured concrete. The public contribution to AI compute, the NAICRI platform at about $46 million, is one hundred to one thousand times smaller than the private foreign commitments. Public money goes to connectivity (the World Bank loan to DICT) and training, not to compute.

The eight tiers

Capital that prefers Singapore, Malaysia and Indonesia (tier 1, and the BlackRock–GIP–MGX consortium) treats the Philippines as adjacent. Capital arriving now (tiers 7 and 8) treats the power constraint as the opportunity.

3. The hyperscaler question

Every major cloud serves Philippine customers from Singapore. AWS runs a Local Zone in Manila for latency-sensitive workloads; Azure and Google Cloud have no Philippine presence. In July 2026 AWS proposed up to $5 billion over 15 years, including a possible Region. It remains a proposal. Two unnamed US hyperscalers were doing due diligence on Luzon sites of about 200 MW each in September 2026; the DICT secretary who said so resigned a month later.

STT GDC's Fairview campus and VITRO's planned Cavite site are both designed for a hyperscale tenant that has not been named. Capacity built ahead of the anchor customer, on the bet that a Region follows, is the hyperscaler bypass pattern. The bet paid in Malaysia and Indonesia; in the Philippines it is still open.

4. Who sets the agenda

DICT is mentioned in more documents than any operator. Labor is the fourth largest topic, which is unusual for a data center market and is the subject of a separate study.

5. Where

Metro Manila dominates the record. Laguna is second, ahead of any single Manila district, because the three largest AI-oriented projects (VITRO Sta. Rosa, DAMAC, NARRA1) sit there. Davao appears because of the cable.

6. Power

The grid generated about 62 percent of its electricity from coal in 2023, the highest share in the region. A fleet growing from tens to hundreds of megawatts adds concentrated load unless each site brings generation. Two responses appear in the record: dedicated switching stations built with Meralco for Fairview and Sta. Rosa, and the Masdar–DAMAC pairing, a renewable platform and a data center platform announced by the same government in the same week. If the pairing holds, the power constraint becomes a reason for Gulf capital to enter. If it does not, the 1.5 GW target has no grid to plug into.

Since the corpus closed

Developments between the end of the coded corpus and publication, checked against primary sources on November 3, 2026.

Findings

Questions this page answers

What to watch in 2027

The weekly monitor tracks these six. When one moves, the table changes and the changelog records it. Dates for the region's conferences are on the events calendar.

Sources and citation