Revisiting Our Microsoft Thesis After the Stargate Norway Takeover

The last note flagged three concerns. Azure AI revenue growth was decelerating from the 29% print in the September quarter. Capex had crossed $80 billion annualised, with the free cash flow line visibly compressing. Copilot attach rates were behind the bull case run-rate, though still ahead of any reasonable scepticism.

We landed constructive at roughly $420, with a fair value band between $480 and $520. The rationale was that the compute commitment, while uncomfortable, was the correct competitive response to a market where the cost of being under-built materially exceeds the cost of being over-built by 20%. That logic still holds. What has shifted is the mix of how the compute is being funded.

The Stargate Norway takeover and the Nscale rental arrangement mean Microsoft is pursuing a hybrid ownership model for its next leg of AI compute. That is the first meaningful capex optimisation story the company has told since the AI capex narrative began in late 2023. Historically, when hyperscalers shift from pure-ownership compute to mixed-ownership compute, the ROIC line stabilises within four quarters and the multiple re-rates inside six.

The Stargate partnership with OpenAI was always a hybrid financial structure. OpenAI was the named operator for some sites, with Microsoft and SoftBank providing the capital. The Norway site being absorbed into the Microsoft footprint means the ownership has consolidated on one side of the partnership. The operational and accounting implications are meaningful.

For Microsoft, absorbing the site simplifies the capital accounting and brings the depreciation schedule on-balance-sheet. That sounds like a negative but it is actually neutral to modestly positive, because the site was always going to be paid for one way or another, and owning the asset gives Microsoft long-term operational flexibility. For OpenAI, it reduces the standalone infrastructure commitment at a moment when OpenAI is trying to manage its own cash burn. The deal is a rational reallocation of who holds what in the partnership.

The Nscale rental deal is the more telling signal. Microsoft is renting 30,000 Nvidia chips for a Norway deployment instead of buying them outright. That is the first public admission that Microsoft is willing to carry part of its AI compute footprint on an opex basis rather than a capex basis. For a company that has been buying compute like every dollar was the last one, that is a structural change in the capital allocation philosophy.

This is the kind of move that looks invisible in any single quarter and shows up with force over four to six quarters, as the capex line decelerates faster than revenue does. The accounting optics also matter. Shifting a portion of AI compute into opex reduces the depreciation headwind that has been the quiet drag on earnings growth through the back half of 2025.

Rental compute is a lower-return, lower-risk structure than owned compute. Return on invested capital on an owned data centre, once fully utilised, runs in the high teens to low twenties. Return on rental compute is lower, because the rental cost includes the provider's margin, but it is positive day one and does not require the $8 to $12 billion of capex to light up a new site.

For Microsoft, the mix shift matters because the marginal dollar of compute does not need to deliver the full owned-asset return. It just needs to deliver incremental revenue above the rental cost. That materially lowers the hurdle rate for incremental AI capacity, which means Microsoft can service more demand without committing to the full multi-year capex cycle. This is exactly the kind of move a company makes when it is confident in the demand curve but wants to preserve optionality on the long-term compute architecture.

Management repurchased $18.5 billion in shares during fiscal 2025 at an average price below $430, a 7% discount to today's level. That is capital allocation competence. Adding the rental-compute lever on top of the buyback discipline is the mark of a team that is thinking about the dollar-for-dollar return on every incremental move.

Historically, large-cap tech names that pivot from pure capex growth to mixed opex-capex structures trade at a premium multiple to peers within 12 to 18 months of the pivot. The pattern was visible with Google in 2017, Amazon in 2019, and Meta in early 2024. The catalyst in each case was the first quarter where the capex line decelerated while revenue held its growth rate. Microsoft is not there yet, but the Nscale deal points to that quarter arriving sooner than consensus thinks.

Amazon's capex line has been the loudest in the hyperscaler group for 18 months, and AWS has been the beneficiary. Google has mostly matched the spend and underperformed in perception. Microsoft has run the tightest margin profile of the three, and this week's moves should widen that gap.

The rental model from Nscale is a template that Amazon and Google will inevitably explore, but Microsoft has struck first with a public deployment of meaningful size. That creates a 12 to 18 month period where Microsoft carries a structurally different capex-to-revenue ratio than its closest competitors, which is the period that matters most for the next valuation rerate.

By comparison, Oracle has gone the other direction with the Stargate investment, committing to almost pure capex with less flexibility. That positions Oracle as the highest-risk, highest-upside play on AI compute, but Microsoft as the more balanced exposure. The relative trade has started to reflect this. On a price-to-free-cash-flow basis, Microsoft trades at 32x, Oracle at 28x, and Amazon at 35x. The spread is tighter than it has been in three years.

Our revised model has Microsoft delivering roughly $315 billion of revenue in fiscal 2026, a 12% lift on the prior year. Azure is 33% of the top line and growing in the mid-20s, with the AI-attached portion running faster. Capex lands at roughly $88 billion, a lower growth rate than the 2024-25 jump, reflecting the rental-compute lever.

Operating margin holds near 44%. Earnings per share lands at roughly $14.10 against consensus of $13.75. At the current $430 share price, the forward multiple sits at 30.5x. Strip out the net cash and the core business multiple is closer to 29x, against a five-year average of 32x. That is a 10% discount to the trailing multiple during a period when the business is growing faster and allocating capital better.

The buyback programme remains the under-discussed leg of the story. Microsoft has roughly $60 billion of remaining authorisation and is buying at current levels. Management repurchased $4.2 billion in the last quarter at an average price of $442, a 3% discount to today. Consistent buyback execution at these levels is a direct transfer of value to remaining shareholders. Dividend coverage remains comfortable at 26% of free cash flow, well below the 35% ceiling management has indicated for the payout.