Grid access comes with more conditions

In January, access to grid power for large data-centre loads was increasingly tied to conditions, led by PJM's move toward curtailing new load that does not bring its own generation, while Malaysia's penalty on data centres that use less than 85% of their declared demand drew fresh attention. Capital kept coming regardless, with Meta guiding 2026 capital expenditure to as much as US$135 billion, DayOne raising more than US$2 billion and Digital Edge committing US$4.5 billion in Greater Jakarta.

What moved in January

The largest US grid moves toward curtailing new data-centre load in emergencies unless it brings generation (14 to 16 January)

The board of PJM Interconnection, which runs the largest US grid, directed its staff on 16 January to design a "connect and manage" framework, due by the end of 2026. Under it, new large-load growth that is not matched by new generation could be curtailed before pre-emergency demand response when supply runs short, with states and transmission owners deciding which loads. The board proposed defining large loads as additions of 50 MW or more at one connection point, encouraged loads to bring their own new generation on a voluntary basis, and called for an expedited interconnection track for that generation, due by August 2026. PJM's most recent capacity auction, for 2027/28, cleared 5.6% short of its target reserve margin.

The White House's National Energy Dominance Council and governors of PJM states signed a statement of principles on 15 January, announced at an event the next day, calling for a reliability backstop auction by September 2026 with 15-year price certainty. The non-binding principles ask PJM to charge its costs first to utilities serving new data centres that have neither secured their own capacity nor agreed to be curtailable. On 14 January the Federal Energy Regulatory Commission accepted Southwest Power Pool's process for large loads, including a service that caps a paired generator's grid injection at the forecast hourly load of the site it serves.

Microsoft's spending stays high and Meta guides higher, with demand still ahead of supply (12 to 28 January)

Microsoft reported December-quarter capital expenditure of US$37.5 billion, with cash paid for property and equipment of US$29.9 billion. It said it added nearly one gigawatt of capacity in the quarter and that customer demand continues to exceed supply, while guiding capital expenditure to fall from the December quarter to the March quarter. Its commercial remaining performance obligation, the contracted revenue it has not yet booked, rose to US$625 billion, about 45% of it from OpenAI.

Meta reported 2025 capital expenditure of US$72.22 billion and guided 2026 to between US$115 billion and US$135 billion, which is Meta's own projection. On 12 January it created Meta Compute, a top-level unit for its data-centre build, with Mark Zuckerberg saying the company plans to build tens of gigawatts this decade.

Washington opens H200 sales to China, and Beijing hesitates (13 to 28 January)

The US Bureau of Industry and Security moved licence reviews for chips such as the NVIDIA H200 and AMD MI325X from a presumption of denial to case-by-case, in a rule filed on 13 January and published and effective on 15 January. To qualify, exporters must show that shipments to China and Macau stay at or below 50% of what they ship to US customers, and every shipment must pass independent third-party testing in the United States. On 14 January a 25% Section 232 tariff was imposed on the same class of chips, with imports for US data centres exempt according to KPMG's reading of the proclamation.

China's response was mixed. Reuters reported on 14 January that customs agents had been told the H200 was not permitted into China, then on 28 January that ByteDance, Alibaba and Tencent had been approved to buy more than 400,000 H200s between them, with conditions still being set. Both reports rest on unnamed sources, and Chinese ministries had not commented when they were published.

NVIDIA says Rubin is in full production, and buyers contract for inference separately (5 to 26 January)

At CES on 5 January, NVIDIA said its Rubin platform is in full production, with partner products due in the second half of 2026. It claims Rubin cuts the cost per token of mixture-of-experts inference by up to ten times against Blackwell and trains those models with a quarter of the GPUs. On 14 January OpenAI agreed to buy 750 MW of compute from Cerebras, delivered from this year through 2028 and aimed at low-latency inference. TechCrunch put the deal's value above US$10 billion, citing an unnamed source. Microsoft launched Maia 200, its own inference chip, on 26 January, claiming 30% better performance per dollar than the latest hardware in its fleet. The performance figures are the vendors' own claims, and we have not seen independent tests of them.

DayOne raises more than US$2 billion (5 January)

DayOne announced a Series C of more than US$2.0 billion, led by Coatue with Indonesia's sovereign fund INA taking part, priced at a 100% premium to its previous round. GDS then said on 13 January that DayOne would buy back US$385 million of GDS's DayOne shares at the Series C price. Later in the month W.Media, citing Bloomberg's unnamed sources, reported that DayOne was considering a US listing that could value it at US$20 billion.

Power and policy

Malaysia's terms for reserved power were back in the news in mid-January. In statements published by The Edge on 29 December and reported again by W.Media on 15 January, the trade ministry said that under Tenaga Nasional's electricity supply agreement a data centre must use 85% of its declared annual demand in years one to four, or pay RM8.50 for each kilowatt of shortfall, calculated monthly. The ministry also said the Data Centre Task Force will be the sole licensing authority and can move reserved capacity from projects that are slow to ramp up to AI projects. On 5 January a Johor state official said a special technical committee of state and federal agencies has been set up to assess each data-centre project, and that meeting Johor's water usage efficiency (WUE) threshold of 1.8 does not guarantee approval where local water is short. Power contracting continued: TM Nxera signed a 280 MW electricity supply agreement with Tenaga Nasional for its Johor campus, reported on 14 January.

In Taiwan, TSMC's chairman named the island's power supply as his main concern for expansion on the company's 15 January results call, according to a translated transcript. The economics ministry said the same day that supply will stay adequate through 2032. TSMC guided its 2026 capital budget to between US$52 billion and US$56 billion.

Order books for generation equipment kept growing. GE Vernova reported on 28 January that its gas power equipment backlog and slot reservations grew from 62 GW to 83 GW in the fourth quarter. On 4 January an AEP subsidiary agreed to buy about US$2.65 billion of solid oxide fuel cells for a generation facility near Cheyenne, Wyoming, with a 20-year offtake for all of its output to one customer; the offtake is subject to conditions AEP expects to be met by the second quarter of 2026. Meta signed nuclear agreements with Vistra, TerraPower and Oklo on 9 January that it says support up to 6.6 GW of new and existing capacity by 2035.

Capital and deals

On 28 January Digital Edge announced a US$4.5 billion campus in the GIIC Industrial Estate in Bekasi, Greater Jakarta, planned for 500 MW of IT capacity, with the first building due in the fourth quarter of 2026. Digital Realty agreed on 19 January to buy CSF Advisers, owner of TelcoHub 1, a 1.5 MW interconnection hub in Cyberjaya, plus land for up to 14 MW, marking its planned entry into Malaysia; the deals are expected to close in the first half of 2026. Aware Super announced a US$300 million investment in the joint venture that owns Vantage Data Centers APAC on 20 January.

India produced the month's largest project pledges, in memoranda of understanding signed around Davos. AM Group agreed a 1 GW compute hub in Greater Noida, powered by wind, solar and pumped storage, at about US$25 billion, with first capacity planned for 2028. Lodha added Rs 1 lakh crore to its commitment, taking planned investment to Rs 1.3 lakh crore for a 2.5 GW data-centre park near Mumbai, and Tillman Global Holdings signed for 1 GW in Maharashtra. India's Economic Survey, published on 29 January, cites industry estimates of nearly 4 GW of national capacity by 2030, up from about 1.3 GW in June 2025. The memoranda are statements of intent, and only AM Group's gives a completion date.

In the US, xAI closed a US$20 billion Series E on 6 January with NVIDIA and Cisco as strategic investors and QIA and MGX among the backers. NVIDIA invested US$2 billion in CoreWeave on 26 January and said it would help CoreWeave secure land and power toward more than 5 GW by 2030.

What it means for distributed inference and power-anchored compute

This section gives our own reading of the month's events. Vailinor builds power-anchored, behind-the-meter compute, so we have an interest in the question.

Access to grid power for large loads was increasingly tied to conditions. PJM is designing a framework under which, in an emergency, new load without matching generation could be curtailed before pre-emergency demand response is called. Southwest Power Pool now has a route for loads paired with their own plant, and Malaysia's supply terms penalise data centres that use less than 85% of their declared demand in their first four years. Johor treats water as a second test. On those terms, a project that arrives with its own supply, or a contract for it, may face fewer obstacles than one that asks the grid to find capacity later, though the new routes are optional, limited in scope and, in PJM's case, not yet in place.

Several developers' January moves fit that pattern. Keppel secured rights to lease a site near Morwell, Victoria, with up to 720 MW of gross power capacity in phases, and is paying for early access so it can obtain approvals and contract power and water before its funds take long leases. TM Nxera signed a supply agreement for 280 MW, and AM Group's planned hub is to run on round-the-clock wind, solar and pumped storage. KCTV5 reported that the at least 800 MW needed by Nebius's project in Independence, Missouri will come from a new, privately financed plant developed by Independence Power Partners on the site of a retired power station.

Inference is also being bought as a product in its own right, with OpenAI contracting dedicated inference capacity and Microsoft building its own inference chip. If performance per watt improves alongside the cost gains NVIDIA and Microsoft claim, each megawatt will produce more output, which matters most where power is the binding limit. In the sources we reviewed, we found no notable January announcement of inference spread across small sites located at their power source, so that model has no direct evidence from this month.

What to watch in February

• The conditions China attaches to H200 purchases, and whether shipments begin.
• Samsung's first HBM4 shipments, which it said would begin in February.
• How Malaysia's Data Centre Task Force uses its licensing role and its power to reallocate reserved capacity.
• PJM's first steps on the connect-and-manage framework, the expedited track for generation and the backstop auction.
• Whether other US grid operators follow Southwest Power Pool's approach for loads paired with their own generation.

About this report

This report was prepared by Vailinor Analysts to help people better understand the AI and energy spaces, and the places where the two meet: power for data centres, the chips that set how much power they draw, and the money and policy behind both. It covers January 2026 only and is written as at the end of that month. Every figure comes from the public sources listed above, as reported by the companies, regulators and publications named, and forecasts are attributed to whoever made them. Nothing in it is investment advice.