AWS Infrastructure Build, September 2026: Sydney 235 MW Approval and Qualcomm Chip Deal Frame Amazon's Expanding AI Push

A planning approval for a reported 235 MW AWS-linked data centre in Sydney and a multi-generation chip partnership with Qualcomm valued at up to $60 billion in press reports arrive as Amazon's capital expenditure hit $131.82 billion in FY2025 — the largest absolute figure among tracked hyperscalers.

Planning authorities in New South Wales reportedly approved a 235 MW data centre development linked to AWS in Sydney, adding the latest coordinate to what has become a rapidly expanding map of Amazon's global infrastructure commitments. The approval extends AWS's Asia-Pacific footprint at a moment when its capital programme is running at a pace few anticipated even two years ago. Amazon shares stood at $258.45, a gain of 1.9% on the session.

The most strategically significant development surrounding AWS in September, however, is the partnership with Qualcomm. Multiple outlets describe a multi-generation agreement under which Qualcomm will supply custom AI inference accelerators and 1.6 terabit-per-second optical interconnects for AWS data centres, with contract value cited across several reports at up to $60 billion. The arrangement is notable on two levels. It signals a deliberate effort to diversify silicon supply beyond Nvidia — even as AWS separately committed to deploying two million additional Nvidia GPUs by 2028 in an announcement made with Nvidia in August. AWS and Qualcomm characterised the collaboration as bidirectional: AWS runs Qualcomm accelerators for inference workloads while Qualcomm uses AWS Bedrock to help design future chip generations. That mutual embedding is unusual and points toward a deep architectural partnership rather than a straightforward procurement transaction.

The financial scale behind these moves is striking. Amazon's capital expenditure reached $131.82 billion in FY2025 — a 58.8% increase from $83.00 billion in FY2024 — equivalent to 18% of $716.92 billion in revenue and the largest absolute capital investment figure among the hyperscalers Slicast tracks. Amazon subsequently raised its 2026 AI spending target to $220 billion, according to August reports. AWS's chief executive told investors that accelerator supply is largely spoken for through 2028, a framing that positions the Qualcomm deal as much a supply-security measure as a technology bet. In the most recent quarter AWS reported revenue growth strong enough that Wall Street analysts described it as easing concerns about whether the infrastructure investment could be monetised at pace.

AWS's geographic expansion is running on multiple axes simultaneously. Alongside the Sydney approval, the company has committed $5.3 billion and 50 MW of initial capacity to a Saudi AI Zone with partner Humain, targeting a cloud region launch by December 2026 — a timeline confirmed by About Amazon as recently as August. A 420 terabit-per-second submarine cable from Japan to Washington State is under construction to support AI-scale data flows across the Pacific. AWS and Microsoft have also announced a 100 Gbps private multicloud interconnect between their competing platforms, a pragmatic acknowledgement that enterprise AI workloads increasingly span providers. Against that momentum, one report described AWS directing clients to permanently abandon resources in Middle Eastern facilities following Iranian drone strikes that damaged operations in Bahrain and the UAE, with no recovery timeline offered. The company also withdrew from a planned campus near a nuclear facility in Maryland, citing infrastructure priorities.

The opportunities are real and the pipeline is large, but several risks deserve clear-eyed attention. Geopolitical exposure is now material: the reported Middle East withdrawal, if accurate, represents a reputational and contractual liability that geographic diversification elsewhere cannot straightforwardly offset. TSMC's 3nm and 2nm nodes and CoWoS advanced packaging remain supply-constrained, with AWS competing against MediaTek, Google TPU teams and others for limited wafer starts — a structural bottleneck that no capital expenditure commitment can resolve unilaterally. Permitting friction is real too: the Maryland withdrawal followed community and political resistance, and AWS is separately exploring power options at a 4.5 GW natural gas site in Pennsylvania, pointing to the difficulty of securing grid-connected land at scale. Internally, Amazon has begun enforcing tighter CPU utilisation limits on engineers as agentic AI demand pressures existing EC2 capacity, a sign that efficiency constraints are sharpening even as the build-out accelerates. Three signals are worth tracking: whether the Saudi cloud region launches on its December 2026 schedule; how rapidly Qualcomm inference silicon reaches production deployment relative to the Nvidia 2M-GPU roadmap; and whether Amazon's 2026 capital expenditure tracks the $220 billion guidance or encounters the permitting-driven slippage that has affected the broader hyperscaler land acquisition cycle.