SoftBank's $50 Billion Malaysia Bet Lands on Turbines and Bonding Tools First — and It Bypasses the Credit Stress Now Choking US AI Capex

Investors should hold the two facts apart: what the Malaysian government said on October 24, and what SoftBank Group has actually signed. Anwar's own remarks identified only \"one of the world's largest investors\" and cited a video conference with the investor the previous Tuesday; SoftBank Group's disclosure trail contains the Sarawak framework, a third tranche of its OpenAI investment on October 1, the DigitalBridge acquisition on September 30 and a euro note early redemption on October 23 — and no Malaysian capital line.

That is not a quibble. A $50 billion pledge with no megawatts has no addressable supply-chain content yet, so the entire question for investors is whether the money arrives as one programme or as a drip of site-level approvals. The Sarawak framework's own language — a \"roadmap for the phased development\" — is the tell: phased means incremental, and incremental means order flow arrives in tranches tied to power connections, permits and chip allocations.

Scale does the sorting. On J.P. Morgan's estimate, the five largest US hyperscalers will spend $697 billion on capex in 2026, and S&P Global expects the top six — including Oracle and SpaceX — to spend more than $7 trillion on data centres and AI-related capex between 2025 and 2030. A $50 billion Malaysian programme spread over several years is roughly 7% of a single year of US hyperscaler capex, or well under 1% a year.

Against Malaysia, the same number is enormous. Johor's data-centre maximum demand is already about 3.8 GW, roughly 51% of peninsular Malaysia's total, and Wood Mackenzie expects that demand to keep compounding while Johor's installed generation base is only about 6.8 GW against current demand of 2.6 GW. On SoftBank Group's own French terms — 5 GW for up to €75 billion — $50 billion is roughly 3 GW of comparable scale, which would be most of Johor's pipeline delivered inside five years.

The comparison that matters for investors is not dollar-for-dollar but financing-for-financing. US hyperscaler capex is now roughly one-third debt-financed, and the marginal buyer of that debt is paying up: Goldman Sachs measures investment-grade new-issue concessions on a large hyperscaler deal widening to as high as about 20 basis points from 2-3 before the market turned \"indigested,\" with 17 of 23 data-centre joint-venture deals trading wide to their originated yields. Malaysia's $50 billion is not competing for that credit. It is political capital and industrial partners. Which is precisely why it extends the equipment cycle instead of cannibalising it.

Malaysia has already rewritten its own rules to ration AI capacity, and it has done so in a way that tells you exactly where the money goes. In February 2026 the government suspended approvals for new data centres that are not AI-linked, citing power and water, and from July 16, 2026 the Data Centre Task Force has required power and water assessments before approving. Regulators have insisted the regime is \"not a moratorium\" but a holistic one, and the task force logged more than 40 applications in 2026.

The physical constraint matches the policy one. Wood Mackenzie finds system-wide transmission and distribution utilisation near 30% while flagging the genuine bottlenecks: shortages at 132 kV main intake substations and limited nodal injection points. Data centres took a record 9.3% of overall Malaysian power consumption in the second week of August 2026 against a 7% year average, and could reach 31% of peninsular demand by 2035 as the last coal units retire by 2044. Money cannot be spent until a substation is energised.

The vendors of that equipment are already converting the first wave. Vertiv lifted Q2 2026 revenue to $3.27 billion, up 24.1% year on year, with gross margin widening to 37.7% from 34.0%. Eaton grew revenue 21.4% to $8.53 billion, though gross margin fell to 33.5% from 37.0% as it defended share. GE Vernova grew revenue 21.9% to $11.10 billion and runs a net cash balance sheet, which is what lets it take turbine and grid orders without new leverage.

Malaysia cannot fabricate leading-edge accelerators and is not trying to. It is trying to own the layer around them, and the evidence is concrete. Intel's US$7 billion Penang complex includes a 710,000 square-foot cleanroom with EMIB and Foveros technology, and its first phase is expected to begin advanced packaging operations later in 2026. Chipbond opened a Penang facility in February 2026 for wafer bumping and wafer-level chip-scale packaging, SPIL is developing a roughly RM6 billion Penang plant, and ASE continues to expand its long-standing Penang base.

Five Malaysian companies — SkyeChip, Inari Technology, FusionAP, Pentamaster Instrumentation and NSW Automation — have formed the Malaysia Advanced Packaging Consortium, targeting 7% of the global advanced packaging market by 2035, with a first Malaysian test chip built on HBM4 advanced packaging as the anchor project. This is the industrial logic of the whole wave: a country with abundant power, a credible packaging cluster and no leading-edge fab has every incentive to sell the box around the chip rather than the chip itself.

Here is the part of the story that can turn a sovereign-scale commitment into a shell. Malaysia has required a Strategic Trade Permit, with 30 days' advance notice, for the export, transshipment and transit of high-performance AI chips of US origin. Separately, Nvidia has more than halved the number of Asian customers authorised to buy its AI chips after introducing a compliance white list that covers Singapore, Malaysia and Japan.

The two rules point in opposite directions. The white list throttles who can lawfully receive accelerators; Malaysia's own permit regime exists because Chinese AI firms have been routing advanced Nvidia compute through Southeast Asian data centres, which is precisely what regulators are trying to close. A Malaysian AI factory can therefore be fully financed, fully built and still unable to fill its racks, and that is why the packaging layer — substrate, tools, servers, networking, cooling — gets paid before the silicon does.

The bear case for a second-region build-out is that it drains capital, power and equipment from a US cycle that is already financing itself with debt at deteriorating terms. The evidence on the US side is genuinely uncomfortable. Goldman Sachs estimates roughly 33% of AI capex is debt-financed in 2026, hyperscalers have issued $194 billion of their own debt so far this year against $108 billion for the whole of 2025, and AI-related issuance has run close to $500 billion. Meta is the cleanest warning: Q3 2026 capex of $33.9 billion against $32.8 billion of operating cash flow produced negative free cash flow of $1.16 billion, funded by $23.9 billion of long-term debt issued in the prior quarter. Microsoft's capital spending reached $35.8 billion in the June 2026 quarter, more than double the $17.1 billion of a year earlier, taking trailing twelve-month capex to $115.9 billion against $182.9 billion of operating cash flow.

And yet the Malaysia wave does not take anything away from that cycle. It is funded outside it. Malaysia's version is sovereign-anchored, partner-funded and — crucially — aimed at power and hardware rather than at hyperscaler balance sheets, so the marginal Malaysian dollar competes for turbine slots, switchgear and substrate capacity, not for the same credit. The equipment suppliers win either way; the losers, if any, are the second-region operators who lever up against unfilled racks. GDS Holdings is the clean example of that risk: Q2 2026 revenue of RMB3.09 billion, up 6.5%, with committed area up 18.2% year on year, 2026 capex guided at RMB9-10 billion, and about RMB31.2 billion of net debt after long-term borrowings of RMB36.9 billion against RMB14.9 billion of cash.

The second region does not compete with the US cycle for capital, but it does share one balance sheet, and that is where the risk concentrates. SoftBank Group's own capital intensity has gone from 9.2% of revenue in FY2024 to 11.8% in FY2025 and 22.2% in FY2026, and 34.1% on a trailing basis, while free cash flow to the firm has swung to negative ¥2.72 trillion. Leverage tells the same story: net debt to EBITDA fell from 6.8x in FY2024 to 4.8x in FY2025 and 2.6x in FY2026, then moved back to 3.6x on trailing numbers as the build-out began.

That funding is mark-to-market in nature. SoftBank's net asset value rose 80% in the June 2026 quarter, mainly on a 134% increase in its Arm position, and its FY2026 net income of ¥4.98 trillion against ¥1.13 trillion the year before is a valuation outcome, not an operating one. So the honest reading of the Malaysia headline is this: a company whose funding capacity is a function of AI equity valuations is committing capital to a second AI build-out. If the Arm and OpenAI marks that fund it compress, the Malaysian programme is the most deferrable item on the list.

Facts: the prime minister of Malaysia said on October 24, 2026 that a global investor had chosen Malaysian locations for $50 billion of AI, new technology and data-centre investment; Bloomberg reported the investor was SoftBank Group; SoftBank Group has confirmed only an October 8 Sarawak framework agreement with Grab and PETROS that discloses no capacity, cost or schedule; Malaysia needs 9 GW of new gas-fired capacity by 2032; and exports of US-origin high-performance AI chips through Malaysia require a Strategic Trade Permit.

Inference: because the binding constraints in Malaysia are power connections and chip permissions rather than capital, the first dollars move to turbines, switchgear, cooling and advanced packaging — Vertiv, Eaton, GE Vernova, ASMPT and Samsung Electro-Mechanics — while NVIDIA silicon arrives last and most rationed. Because the programme is financed by political and industrial capital rather than by hyperscaler credit, it adds to the equipment cycle instead of cannibalising it, leaving the genuine credit exposure with the second-region operators such as GDS Holdings rather than with the hyperscalers. And because SoftBank Group funds both France and Malaysia out of a mark-to-market equity position, the programme's real vulnerability is the AI equity cycle it is meant to be insulated from.