Spending plans jump and big tech is told to bring its own power

Alphabet guided 2026 capital expenditure to US$175 billion to US$185 billion and Amazon to about US$200 billion, and OpenAI announced US$110 billion of new investment, part of it conditional, alongside 3 GW of dedicated NVIDIA inference capacity. Power stayed the constraint, with the US President calling on large tech companies to provide their own power, PJM proposing terms for co-located load and Malaysia saying it has been limiting approvals for new data centres not linked to AI.

What moved in February

OpenAI announces US$110 billion and names inference and training capacity separately (27 February)

OpenAI announced US$110 billion of new investment at a US$730 billion pre-money valuation: US$30 billion from SoftBank, US$30 billion from NVIDIA and US$50 billion from Amazon, of which US$35 billion follows later once conditions are met. Alongside the money, OpenAI's expanded NVIDIA partnership names 3 GW of dedicated inference capacity and 2 GW of training on Vera Rubin systems. Separately, OpenAI expanded its existing US$38 billion AWS agreement by US$100 billion over eight years, including a commitment to consume about 2 GW of Trainium capacity. More financial investors were expected to join the round, and the capacity figures are commitments to future use.

Alphabet guides 2026 spending to US$175 billion to US$185 billion and Amazon to about US$200 billion (4 to 5 February)

Alphabet guided 2026 capital expenditure to between US$175 billion and US$185 billion, against US$91.4 billion of property and equipment purchases in 2025, and put its Google Cloud backlog at US$240 billion. On the call Sundar Pichai said Alphabet has been supply constrained while ramping capacity. A day later Amazon said it expects to invest about US$200 billion in 2026, mostly in AWS. Andy Jassy told analysts Amazon added 3.9 GW of power in the previous 12 months, 1.2 GW of it in the fourth quarter. All of these forward figures are the companies' own projections.

Washington tells big tech to supply its own power, and PJM proposes terms for co-located load (23 to 27 February)

In his State of the Union address on 24 February, President Trump said he had negotiated a new "ratepayer protection pledge" and that the major tech companies have an obligation to provide for their own power needs, including by building their own plants. He named no companies and gave no details of the agreements.

PJM Interconnection filed two sets of proposals with the Federal Energy Regulatory Commission. The first, on 23 February in response to FERC's December 2025 order, would count on-site generation used to net down a retail load as behind-the-meter generation only up to a cumulative 50 MW of nameplate capacity, with a transition to 18 December 2028 and protection for arrangements under contracts in effect on 18 December 2025. It also proposed three new transmission services for co-located load. The second, on 27 February, would set up an expedited interconnection track for new generation, covering up to 10 large requests a year and targeting about ten months to an interconnection agreement. Both were awaiting FERC at month end.

Malaysia says it has been limiting new data centres not linked to AI (24 to 27 February)

Prime Minister Anwar Ibrahim told Parliament the government has limited approvals for new data centres not related to AI over the past one to two years because of pressure on electricity and water. Other reports of the same remarks put the period at nearly two years, and he said projections showed enough power supply for new applications over the next couple of years. W.Media noted that experts differed on how to read the restriction. Tenaga Nasional's results briefing on 27 February showed the scale involved: as at December 2025, 35 data centres were in operation on its system with total maximum demand of about 4.5 GW, against actual data-centre load of 850 MW that month. TNB signed 18 supply agreements totalling about 1.7 GW in 2025, and its data-centre electricity sales grew 148% year on year.

India makes its pitch, with a tax holiday and three large pledges (1 to 19 February)

The Union Budget on 1 February proposed a tax holiday until 2047 for foreign companies that provide cloud services globally using data-centre services from India, with Indian customers served through an Indian reseller, and a 15% safe harbour on cost where the Indian data-centre provider is a related entity. These were Budget proposals. Around the India AI Impact Summit, Adani committed US$100 billion to renewable-powered AI data centres by 2035, expanding from a 2 GW platform toward 5 GW with its 30 GW Khavda renewable project at the centre of the plan, of which more than 10 GW is operating. Reliance set out a plan of about US$110 billion over seven years, with more than 120 MW expected online in the second half of 2026, according to TechCrunch. OpenAI became the first customer of TCS's HyperVault data-centre business, starting at 100 MW and scaling to 1 GW over time, and the government said it would add 20,000 GPUs to the existing 38,000 in the coming weeks. All capacity and spending figures are the companies' and government's own plans.

HBM stays sold out, DRAM price forecasts jump and the H200 stays stuck (2 to 26 February)

NVIDIA reported data-centre revenue of US$62.3 billion for its quarter to 25 January and guided next-quarter revenue to US$78 billion, assuming no data-centre compute revenue from China. Its CFO said small amounts of H200 for Chinese customers had been approved by the US government but had produced no revenue, and Bloomberg Law reported that David Peters, the Commerce Department's assistant secretary for export enforcement, told a House committee on 24 February that NVIDIA had yet to sell any H200s to China. Memory tightened further. Samsung said on 12 February it had begun shipping HBM4 and expects its HBM sales to more than triple in 2026, Micron's CFO said on 11 February that its 2026 HBM supply remained sold out, as it had said previously, and TrendForce forecast first-quarter contract prices for conventional DRAM up 90% to 95% on the previous quarter, which it said would be a record quarterly rise.

Power and policy

The IEA's Electricity 2026 report, published on 6 February, forecasts global electricity demand growing an average 3.6% a year from 2026 to 2030. It projects US demand growing nearly 2% a year, with about half of the increase from data centres, and says more than 2,500 GW of projects, spanning renewables, storage and large loads such as data centres, are stalled in grid connection queues worldwide.

In Japan, TEPCO restarted Kashiwazaki-Kariwa unit 6, a 1,356 MWe reactor, on 9 February after a delayed start and began sending power to the Tokyo area on 16 February for the first time in about 14 years. It paused the unit on 20 February for turbine checks and restarted it on 24 February, with commercial operation planned after a final check on 18 March, subject to regulatory approval.

About 50 people protested at a data-centre construction site in Gelang Patah, Johor, on 7 February over dust and water supply, in what W.Media, citing Bloomberg, said was described as the first such protest in Malaysia. In the Philippines, ICT Secretary Henry Aguda said on 24 February that the plan is 18 GW of data-centre capacity within ten years, against about 200 MW today, which is a government target.

Capital and deals

KKR and Singtel agreed on 4 February to buy the remaining 82% of ST Telemedia Global Data Centres for S$6.6 billion, an implied enterprise value of about S$13.8 billion. KKR will hold 75% and Singtel 25%, with closing expected by early in the second half of 2026. STT GDC has 2.3 GW of design capacity across 12 markets. In Malaysia, AIMS completed the purchase of about 10 acres in Cyberjaya for a planned AI data centre of up to 200 MW, subject to final engineering confirmation, an investment of about RM4 billion due for completion in 2027. AirTrunk was reported by the Australian Financial Review to be close to an A$1.8 billion debt package for its 150 MW Johor campus. Bloomberg reported on 13 February that DayOne had picked banks for a US listing that may raise about US$5 billion, and Indonesia's sovereign fund Danantara said it would allocate Rp21 trillion to a data-centre platform with global operators. In Singapore, Nxera opened its 58 MW DC Tuas with more than 90% of capacity committed before opening.

AI labs and cloud providers raised and committed large sums. Anthropic raised US$30 billion at a US$380 billion post-money valuation on 12 February, in a round led by GIC and Coatue with MGX among five co-leads. Oracle said it expects to raise US$45 billion to US$50 billion in 2026 to build capacity for contracted cloud demand, about half through equity, and issued US$25 billion of bonds on 4 February. Meta agreed to deploy up to 6 GW of AMD GPUs, with the first gigawatt shipping from the second half of 2026, a week after NVIDIA announced a deal for millions of its GPUs. Nscale arranged a US$1.4 billion GPU-backed loan for clusters in Norway, Portugal, Iceland and the UK.

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.

Contracted capacity ran well ahead of capacity that is live or billing across February's results. CoreWeave had more than 850 MW of active power and about 3.1 GW of power contracted from its suppliers. NEXTDC had 416.6 MW contracted by customers against 119.8 MW billed. Tenaga Nasional's 35 operating data centres carry about 4.5 GW of maximum demand against 850 MW of actual load in December. The three measures differ (active capacity, billed capacity and metered load), but each shows a large gap between what has been contracted and what is in service today.

The US President has now said publicly that large tech companies should provide their own power, but the detail matters. PJM's proposal would let on-site generation net down a retail load only where the generation is 50 MW or less, with legacy treatment for contracts in place on 18 December 2025 and a transition to December 2028 for other existing set-ups. Co-located loads above that would choose among network service and three new transmission services. If FERC accepts it, supplying your own power inside PJM becomes a defined, regulated arrangement with its own terms.

February also brought a study programme for siting inference where spare grid capacity exists. On 3 February EPRI, NVIDIA, Prologis and InfraPartners announced a collaboration to assess micro data centres of 5 to 20 MW at or near utility substations with spare grid capacity, aiming for at least five US pilot sites in development by the end of 2026. OpenAI's NVIDIA expansion also names inference capacity separately from training, though it says nothing about where that capacity will sit.

What to watch in March

• FERC's response to PJM's co-located load and expedited interconnection filings.
• Whether the White House sets out terms for the ratepayer protection pledge, and which companies sign it.
• Whether any H200 shipments to China take place.
• Whether Malaysia publishes criteria for its limit on non-AI data centres.
• Kashiwazaki-Kariwa unit 6 moving to commercial operation after its 18 March check.
• The Hong Kong government's tender result for the Sandy Ridge data facility site, which its 25 February Budget said would be announced shortly.

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 February 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.