Record hyperscaler backlogs meet slower power as regulators weigh who connects and who pays

Amazon, Alphabet and Microsoft reported record order books and rising capital spending in the week to 30 October, with Amazon saying customers consume capacity almost as quickly as it is added, while SK hynix and Samsung posted record memory profits and SK hynix contracted more than half of its 2027 and 2028 output. Grid regulators moved the other way, as PJM was due to refile with FERC over who pays for data-centre power, Australia's rule-maker delayed its large-load connection standards to November, and Singapore scheduled a decade of grid and import build to meet demand it says depends on when projects come onstream.

Our view, up front

We build power-anchored, behind-the-meter compute because power, as we have argued from the start, is the binding constraint on the AI build-out, ahead of compute demand and chip supply. This week pushed that reading further. The hyperscalers posted record order books while describing capacity that is consumed almost as fast as it is added, the memory makers sold out years of output in advance, and grid regulators in the United States, Australia and Singapore moved on how and on what terms the loads behind all of it connect. Demand for compute is not in doubt. The open question, and the one our model is built around, is when the power to run it is energised. What follows is the week as it happened, with every figure sourced and our fuller reading set out below.

Key numbers

• Amazon reported committed customer contracts, primarily AWS, of US$578 billion at 30 September, and said on its earnings call that customers are consuming capacity almost as quickly as it is added.
• Alphabet signalled a further significant increase in capital spending in 2027 after a third quarter in which Google Cloud revenue rose 103% to US$30.7 billion.
• SK hynix said long-term agreements now cover more than 50% of its memory production capacity for 2027 and 2028, and that it expects the memory shortage to persist through 2030.
• The US Federal Reserve held its policy rate at 3.75% to 4.00% on 28 October, on a 10 to 2 vote, with two members preferring an increase.
• Moody's raised its estimate of 2026 hyperscaler capital spending to about US$785 billion, from US$700 billion in March, while a basket of eight listed neoclouds has fallen more than 40% since late July.
• Singapore's Energy Market Authority granted a conditional approval to import 1.3 GW of low-carbon electricity, with first supply targeted only from 2035.

What moved this week

Hyperscalers report record order books as capital spending climbs (28 to 29 October)

Alphabet reported third-quarter revenue of US$127.6 billion, up 25%, with Google Cloud revenue up 103% to US$30.7 billion and quarterly capital spending of US$51.4 billion. It raised its full-year capital spending guidance to a range of US$195 billion to US$205 billion and signalled a further significant increase in 2027. Amazon reported that AWS grew about 39% and that its performance obligations, the committed customer contracts primarily related to AWS, reached US$578 billion at 30 September, and it kept its 2026 cash capital spending forecast at about US$220 billion. On its earnings call, Amazon said customers are consuming capacity almost as quickly as it is added, and that large blocks are reserved in advance with visibility into 2028 and parts of 2029. Microsoft reported revenue of US$92.28 billion, up 19%, with Azure up 48%, and capital spending up 103% to US$39.37 billion for the quarter. The quarter was the first to reflect an accounting change Microsoft set out on its July earnings call, extending the estimated useful life of its data-centre assets and buildings to 25 years from 15 and shifting part of its future lease commitments from finance to operating leases, both of which lower reported capital spending without reducing the underlying build.

Why it matters. The order books grew faster than the capacity standing behind them. Our reading is that a backlog is signed demand, and the binding question is when the power and the buildings behind it energise. Amazon's own description, capacity consumed almost as fast as it is added, states that gap from the seller's side, and Microsoft's longer asset life and lease reclassification change how the same build-out is reported without changing how quickly it can be powered.

Memory makers post records and lock up years of output (27 to 29 October)

SK hynix reported its first quarter above 100 trillion won of revenue, at 104.6 trillion won, with operating profit of 83.3 trillion won and an operating margin of 80%. It said long-term agreements signed with 14 customers now cover more than half of its memory production capacity for 2027 and 2028, that those contracts typically run five years, and that it expects memory demand to exceed supply through 2030. Samsung Electronics reported record group operating profit of 107.4 trillion won, of which its semiconductor division alone earned 106.5 trillion won, driven by sales of its latest high-bandwidth memory; its consumer division fell into a small operating loss as higher memory prices raised its costs. Samsung said strong demand for semiconductors, driven by rising investment in AI infrastructure, is expected to continue.

Why it matters. This is the week's clearest counter-evidence to a power-first reading, because the scarcity being monetised first is memory, not grid supply. Our reading is that a memory maker can sell out years ahead because its product ships on a truck, while a data centre's power cannot. These contracts show that demand for compute is already there and priced; they say nothing about when the sites meant to house it will be energised.

PJM prepares to refile its reliability backstop at FERC over who pays for data-centre power (29 October)

PJM, the operator of the largest US grid, was due to file a revised Section 205 plan at the Federal Energy Regulatory Commission by 29 October, after FERC accepted but suspended its one-time Reliability Backstop Procurement for five months over how the costs would be allocated. The backstop is meant to procure about 6.8 GW of new supply to cover a shortfall of 6,831.3 MW identified in the 2028/2029 capacity auction, with a price cap of US$555 per MW-day and a cost that one trade estimate puts at about US$1.38 billion a year. In a statement of position on 7 October the US Department of Energy said large electric loads, not households or other business ratepayers, should fund the generation and infrastructure the backstop is designed to secure.

Why it matters. The dispute is about who pays for the power that large new loads require. Our reading is that the direction is the same one Pennsylvania's proposal took earlier this month, placing the cost of the gap between contracted and available power on the loads that create it, which means the price of connecting a data centre now depends on a cost-allocation ruling as much as on the contract itself.

Australia delays its rules for connecting large loads (29 October)

The Australian Energy Market Commission extended the date for its final determination on new access standards for large loads, part of its work on improving access standards for the national electricity market, on 29 October, citing issues of material complexity. It will now publish the determination on 12 November. The draft determination would create a tiering framework for large-load connections and add access standards for large inverter-based loads, covering how they ride through disturbances and how instability is managed.

Why it matters. The rules that decide whether data centres and other large loads can connect in Australia's main market slipped by two weeks on complexity. Our reading is that this sits alongside last week's events in the same market, where a distributor's suspended line works took contracted and partly billing capacity out of an operator's guidance, so the terms of connection are moving while the capacity waits.

Singapore schedules a decade of grid build ahead of data-centre demand (26 to 27 October)

At the opening of Singapore International Energy Week, the Energy Market Authority set out a five-year infrastructure plan. It said peak electricity demand was 8.6 GW as at August 2026 and could reach between 9.9 GW and 13 GW by 2031, mainly from semiconductor manufacturing and data centres, and that the pace would depend on when those developments come onstream. The plan commits up to S$5.3 billion to expand the grid between 2025 and 2029, eleven times the spending of the previous five years, with seven new transmission substations by 2031 and a second interconnection with Peninsular Malaysia that could carry up to 2 GW by 2030. Separately, the authority granted Sembcorp Utilities a conditional approval to import 1.3 GW of low-carbon electricity, drawing on offshore wind from Vietnam and solar power from Malaysia, with supply targeted only from 2035 and still subject to jurisdictional approvals, power purchase agreements and financing. On a panel on 27 October, Tenaga Nasional's chief distribution network officer said the utility pledges to connect data centres to the grid within a year once they have cleared Malaysia's ministry-chaired approval committee.

Why it matters. A regulator laid out a decade of generation, grid and import build to meet demand it says depends on when projects energise. Our reading is that the Sembcorp approval is the clearest signed-versus-energised marker of the week, consent granted roughly a decade ahead of first power, and that Tenaga's one-year pledge puts a clock on the connection step specifically, which is where the delay usually sits.

The Fed holds as markets reprice AI spending against returns (27 to 28 October)

The Federal Open Market Committee held the target range for the federal funds rate at 3.75% to 4.00% on 28 October, on a 10 to 2 vote, with Lisa Cook and Beth Hammack dissenting in favour of a quarter-point increase. The decision came after equity markets had sold off on concern about the scale of AI capital spending against its returns, with chip and neocloud shares among the hardest hit going into the mega-cap results. In a 27 October analysis, Moody's Ratings was reported to expect hyperscaler capital spending of about US$785 billion in 2026, up from US$700 billion in March, and to warn that higher capital intensity and debt could lead to a reassessment of creditworthiness if profit growth fails to appear. A basket of eight listed neoclouds, which rent out AI computing power, has lost more than 40% since late July.

Why it matters. Rates held at a restrictive level, and the dissents pointed towards a rise, so there was no easing to lower the cost of a debt-funded build-out. Our reading is that two different pressures are now visible at once: the physical chain of consent, grid connection and energised power, and a separate returns question about whether the capital spending will earn its cost.

Our reading of the week

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

Our view is that the AI build-out is priced on signed capacity, while the constraint runs further up a chain: public and regulatory consent, then a grid connection, then energised power, and only then compute. This week the signed side set records and the supply side showed its limits. On signed-versus-energised, Amazon put its committed backlog at US$578 billion and said capacity is consumed almost as fast as it is added, SK hynix contracted more than half of its 2027 and 2028 memory output years in advance, and Singapore's regulator approved a 1.3 GW import that will not deliver power until about 2035. On the grid and consent links, PJM moved to return to FERC over who funds the power that large loads need, the US Department of Energy said those loads should pay for it, Australia's rule-maker pushed back the standards that govern whether large loads can connect, and Singapore scheduled a decade of grid and import build against demand it says depends on when projects come onstream.

The case against this reading is that compute supply, not power, is the binding constraint, and this week's records in memory support it: SK hynix and Samsung posted record profits on high-bandwidth memory they have largely sold out, and reports that TSMC may raise prices for its advanced packaging point to the same tightness in chip supply. Our reading is that a chip shortage decides which buyers are served first without bringing any site's power forward, and the hyperscalers' own backlogs show that demand is not the open question.

For investors, our first reading is that the clearest signal this week came from the sellers, not only from the stressed neoclouds. Amazon describing capacity consumed as fast as it is added, and SK hynix locking up half of two years of output, describe a build-out running flat out against its own supply chain, in which power is the slowest input to add. Our second reading is that the cost of consent and grid connection is being moved directly onto large loads, through PJM's cost allocation and the Department of Energy's position, while the rules that govern connection are still in motion, as Australia's delay shows. The week's capital signal, a Fed on hold and a market repricing the spending itself, raises the cost of getting that gap wrong.

Capital and valuation read-through

Money kept flowing to the suppliers of compute and to owners of capacity that already has power, while the companies carrying the build-out's debt stayed under pressure. SK hynix reported net cash of about 98.7 trillion won and set out plans to return about 100 trillion won to shareholders this year, and Samsung's semiconductor division earned almost the whole group's operating profit. On the other side, the listed neoclouds that rent out capacity they are still building and financing remained well below their summer levels, and the Fed's decision to hold kept the cost of that debt high ahead of the 3 November US midterm elections. In our view the income is accruing to those who supply the compute and those who already hold power, and not yet to those still waiting to energise.

What we are watching

• The Australian Energy Market Commission's final determination on access standards for large loads, now due 12 November.
• PJM's revised reliability backstop filing at FERC and the Commission's response.
• The data-centre-related ballot measures across several US states on 3 November, and whether high-yield issuance for data centres resumes after the vote.
• Nscale's planned New York listing, whose price range was still unset through the week.
• Jemena's restart of the M3 Melbourne line works, which it does not expect before February 2027.
• Any further step by Finland's environmental regulator after Google's response on the Muhos and Kajaani sites.
• Third-quarter results still to come from other large operators and chipmakers through November.

About this brief

This brief 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 24 October 2026 to the US market close on Friday 30 October 2026, and is written as at 2 November 2026. 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.