Microsoft Azure Nuclear and Power Deals, September 2026: Three Mile Island Restart and Chevron PPA Alongside $115.95B in Annual Capex

Microsoft's FY2026 capital expenditure hit $115.95 billion — up 79.6% year-on-year and 35% of revenue — as the company secured long-duration power supply through Three Mile Island's undamaged reactor and a 20-year deal with Chevron, making energy procurement co-equal with compute acquisition as Azure's binding constraint on AI growth.

Microsoft's power infrastructure strategy has entered a new phase. According to reports, the company's purchase agreement is helping bring Three Mile Island's undamaged reactor back online — a facility idle since 2019 — while a freshly disclosed 20-year power purchase agreement with Chevron secures electricity for a Texas data center campus. The near-simultaneous moves by Meta, which reportedly signed a 20-year agreement backing an Illinois nuclear plant, and Amazon, which contracted for up to 1.92 gigawatts from Pennsylvania through 2042, confirm that long-duration power contracting has become table stakes for hyperscalers competing in AI infrastructure. For Microsoft, these energy commitments are not peripheral to the compute story; they are the foundation on which the rest of the Azure buildout rests.

The capital allocation data explains why energy security is now this urgent. Capital expenditure at Microsoft reached $115.95 billion in fiscal year 2026, a 79.6% increase over the $64.55 billion reported in FY2025, and consumed 35% of the company's $331.84 billion in revenue for the same year — ranking second in intensity among tracked hyperscalers, behind Oracle, and second in absolute terms, behind Amazon, according to Slicast's compilation of SEC XBRL filings. In the single highest reported quarter, ending September 2025, Microsoft deployed $19.39 billion in property, plant, and equipment payments alone. Against that backdrop, Microsoft's decision to disclose Azure's revenue and profitability as a standalone segment for the first time fills a transparency gap that had previously forced outside observers to rely on inference: the cloud unit is now generating enough discrete revenue to benchmark directly against the capital it consumes.

On the compute side, the past two weeks have seen Microsoft receive its first production NVIDIA Vera Rubin systems, deploying the next generation of accelerators ahead of most peers. Reports indicate the company signed a deal with Nebius worth up to approximately $19 billion for exclusive Azure GPU capacity. Lambda, a San Jose-based GPU cloud provider, raised $1 billion in debt financing to acquire NVIDIA chips that feed Microsoft's infrastructure — one of several third-party financing arrangements through which the supply chain is being extended to meet demand, particularly notable given that NVIDIA simultaneously notified customers of a 15% AI server price increase. IREN's Horizon 1 facility was accepted by Microsoft in late August, triggering the initial drawdown of a reported $9.7 billion contract. Microsoft and AWS separately announced a 100 Gbps private interconnect service in early September, a pragmatic acknowledgment of enterprise multicloud demand that both companies had previously been measured about endorsing.

Microsoft's Maia 300 accelerator introduces a parallel optionality track. The chip is being benchmarked against NVIDIA's dominant position in AI inference workloads, following the established pattern of Google's TPUs, Amazon's Trainium, and Meta's MTIA — large platforms building custom silicon to reduce long-run dependency on a single vendor. Maia 300's competitive standing in production inference settings has not yet been established publicly, and its commercial impact on Microsoft's unit economics remains to be demonstrated. The announcement nonetheless signals a deliberate investment in architectural flexibility, made more relevant by NVIDIA's latest pricing action.

The balance sheet position invites scrutiny. Morgan Stanley flagged a widening gap between AI capital expenditure and realized revenue, a concern that contributed to a 3% slide in Microsoft's share price in late August; the stock was most recently trading at $499.70, down 2%. Analyst commentary has pointed to a margin squeeze embedded in the infrastructure strategy, where revenue growth is real but the capital required to sustain it compresses near-term returns. A Microsoft-affiliated data center complex in Vineland, New Jersey — a facility valued at $19.4 billion — faces community opposition over alleged unpermitted gas turbines, a 1.5-million-gallon LNG tank, and other construction irregularities, illustrating that permitting and social license risks are now material operational variables in the buildout, not tail risks.

Three signals will define whether the infrastructure investment translates into durable margin expansion. First, Azure's standalone revenue trajectory over the next two to three quarters will test whether new capacity is being absorbed at sufficient utilization rates; the first discrete disclosures will set the baseline against which subsequent performance is measured. Second, the permitting and grid-interconnection timelines for large power projects — the Three Mile Island nuclear restart and the Chevron Texas build-out among them — carry execution risk that could delay planned capacity additions. Third, Maia 300's performance against NVIDIA inference benchmarks will determine whether Microsoft's custom-silicon program can meaningfully reduce the per-unit cost of serving AI workloads, or whether the 15% NVIDIA price increase simply flows through to the income statement. The energy and compute bets are made at a scale that forecloses easy course correction; the pace of monetization is the variable that markets are still pricing.