The anchor tenant is the whole deal
@kawishwaqar|November 11, 2026 (7d ago)2 views
Disclosure: I work at Jazz as Principal Evangelist Cloud & AI. This article is written as personal industry analysis. Views are my own.
Two documents landed within ten weeks of each other and they should be read together.
On 9 November, SoftBank Group and Grab signed joint development agreements with Petroleum Sarawak Berhad to build AI infrastructure in Malaysia, with $1.85 billion of initial seed capital and a target of five gigawatts or more.
On 29 October, Pakistan's Ministry of IT told a reporter that the National AI Policy approved by the cabinet has still not entered implementation, that the AI Council it creates has not been formed a year and a half later, and that data centers under the policy have stalled for want of money. Implementation is now expected in June 2027, if funds become available.
One is a deal. The other is a delay. The gap between them is not ambition, and it is not even capital. It is structure.
What Malaysia actually signed
Read the split, because the split is the argument.
- Sarawak project: $1bn total, with $800m from SoftBank, $100m from Grab and $100m from PETROS.
- Peninsular Malaysia project: $850m, with $650m from Grab and $200m from SoftBank. Grab leads as founding shareholder.
- Deployment is staged, released only as projects hit development and customer contract milestones.
- Further capital comes at project level through third-party equity and non-recourse project financing.
- SoftBank agreed in principle to a five-year lock-up on the Grab shares it holds.
Four parties, four different things at risk. SoftBank carries capital. Grab carries demand, because Grab is the customer as well as an owner. PETROS carries power and land in Sarawak, which is where the hydro is. The federal and state governments carry permits and alignment with AI Nation 2030.
That is what bankable looks like. Not a bigger number. An allocation of risk to the party that can actually hold it.
The release is also honest about what could break. Its own risk factors name the possibility that "suitable land, competitively priced and reliable power, connectivity, or contracted customer demand will not be secured." A $1.85bn commitment from Masayoshi Son, with a state petroleum company and a Nasdaq-listed superapp beside him, still has to disclose that the tenant might not arrive.
Five gigawatts is mostly unbuilt
The honest counterpoint, before anyone quotes the headline as achievement. At roughly $10m to $12m per megawatt of data center build, five gigawatts is a $50bn to $60bn programme. $1.85bn is seed capital for land, power rights, permits and early works. Nothing has been energised yet.
So treat the announcement as a structure, not a delivery. The structure is the newsworthy part. The megawatts are a target with milestones attached, and the money stops if the milestones fail. That is a feature, not a bug, and it is precisely the discipline that the US neocloud model lacked when it signed take-or-pay leases before it had tenants.
What Pakistan has actually built
Our record is not nothing, and it deserves to be stated accurately.
Sky47 Karakoram-01 was inaugurated in Islamabad on 24 July 2026 at 8.5 megawatts, built to Tier III and Tier IV standards, with a design that scales past 50 megawatts and more than 3,000 racks. A 5MW node at Port Qasim was under construction with a September target, and a Lahore facility is planned. The National Telecommunication Corporation signed a three-year agreement to deploy NVIDIA H200, A100 and B200 capacity across government data centers with Data Vault Pakistan, hosting open-weight models including Llama 3.1 and GPT OSS 20B, with a 99.9% availability commitment.
That is a real facility with a real anchor in NTC. If you want to see the Malaysian structure working at small scale, it is here. A government buyer, a term agreement, an operating partner, a service level.
Now put the two numbers side by side. 8.5MW operating against a 5,000MW target. That is 0.17%. And the programme that was supposed to fund the next stage, $400m for AI infrastructure inside a $1bn national ecosystem programme, has not been funded. The ministry has approached the World Bank and the Asian Development Bank. A $200m National AI Fund is still a proposal under consideration.
The structural difference in one sentence
In Malaysia, the state is a partner. In Pakistan, the state is the customer, the financier, the regulator and the anchor tenant at the same time.
That is not a criticism of anyone's effort. It is a description of what happens when a country funds AI infrastructure as public capex: the government ends up holding all four risks at once. Capital risk, because it is a budget line. Demand risk, because there is no committed tenant beyond its own agencies. Power risk, because the tariff is a policy variable. Permit risk, because it owns the permits.
One of those risks will be managed well. The others will be managed by delay, and delay is what we are looking at. The policy is approved, the council is not formed, the money is not allocated, and the facility that does exist was inaugurated ahead of the programme that was meant to pay for it.
What an anchor tenant actually buys
A data center is a fixed-cost asset with a utilisation curve. Every megawatt you energise is a bill you pay whether or not anyone runs a token on it. The only thing that makes that asset financeable is a counterparty committing to a term price for a defined volume, with a service level attached.
That is why the anchor tenant is the whole deal. It converts a speculative asset into contracted revenue, which is what lenders and equity both price. It also disciplines the build: you size the first phase to real load, not to a press release.
We have plausible anchors in this country. A telco carrying most of the country's broadband traffic. Banks with regulatory pressure to keep data onshore and an SBP framework telling them how. A government that already buys cloud. None of them will sign a ten-year term for capacity that does not yet exist unless someone underwrites the delivery risk, and that is the missing piece: not another announcement, but a contract with a penalty clause.
The version that works looks like this. Pick one anchor with real, measurable inference demand. Sign a term contract with a published per-token price and an SLA, not a memorandum. Bring a power partner at a fixed tariff, which in our case means settling the industrial electricity question before the GPU question. Stage the capital against milestones, the way Malaysia did. Let the state hold permits and alignment, and let it stop pretending to be the tenant.
Certain, signal, unresolved
Certain, because the documents say so: SoftBank, Grab and PETROS signed JDAs with $1.85bn of seed capital for a 5GW target, staged against milestones, with demand risk disclosed. Pakistan's AI Policy was approved and, per the Ministry of IT on 29 October, has not entered implementation, with the AI Council unformed and data centers stalled for lack of funds. Karakoram-01 exists at 8.5MW with design headroom past 50MW.
Signal, and this is my reading: the countries that get sovereign AI will be the ones that allocate risk across a capital partner, a power partner and a demand anchor. Countries that fund it as public capex will keep announcing facilities and keep delaying programmes, because they are carrying risks they cannot price.
Unresolved: whether Malaysia's 5GW actually gets built, and whether Pakistan's programme can be restructured before June 2027 rather than simply delayed again. The tell will be boring. Not a ribbon, a signed term sheet with a named tenant and a number in it.
Related reading: There are three prices for the same token for why committed demand gets a better price than announced demand, The reset wall is a walk-away now for what happens to sellers who build before the tenant arrives, and The agent moves onto your desk for the small end of the same curve.
Sources: SoftBank Group and Grab joint development agreements, 9 November 2026, UrduPoint on the stalled AI policy, 29 October 2026, DigitalEye on the national AI infrastructure plan, and the Ministry of IT policy portal.