AI Infrastructure —
Share
What changed in the AI trade, and why it mattered.
Anthropic signs $11 billion infrastructure deal with Akamai while Oracle hits a snag in New Mexico and faces questions on its giant Project Jupiter buildout
A hyperscale AI lab committing $11 billion to Akamai signals continued acceleration in compute demand and diversifies where that capacity gets built, while Oracle's New Mexico setback and Project Jupiter jitters show execution risk is creeping into even the largest cloud capacity buildouts investors are counting on.
Oracle reaffirmed its 2.4 GW fuel cell power deal despite the Project Jupiter concerns, and the stock is on track for its best month since April
Power availability is the binding constraint on new AI data center capacity, so Oracle standing behind a multi-gigawatt fuel cell commitment even amid buildout doubts is a signal that it intends to keep securing power at scale.
Broadcom fell 2% as its $230 billion AI revenue target is seen doubling the delivery challenge it must execute against
Broadcom is a core supplier of custom AI silicon and networking to hyperscalers, so a stock move tied directly to skepticism about hitting an outsized target is a read on whether the market believes current AI capex commitments can actually convert into delivered, monetized hardware.
ASML said it sold nothing in Europe in 2026, underscoring that its business is now entirely dependent on demand from Asia and the US
ASML's lithography tools are the bottleneck asset for advanced chip manufacturing, so a complete absence of European sales this year highlights how concentrated global leading-edge capacity investment has become outside Europe.
Microsoft committed $10 billion to Gulf states through 2030, adding another major hyperscaler capital pledge outside the US
This extends the pattern of hyperscalers locking in long-dated, geographically diversified infrastructure commitments, which underpins demand for compute, networking and power buildout suppliers over the rest of the decade.
ByteDance obtained access to more than 2,000 Nvidia B200 chips through a Norway-based data center, and Musk said the Colossus 2 facility will ramp Nvidia AI chip deployment by year-end
Both items point to continued high-end Nvidia chip demand finding routes to deployment, including workarounds for Chinese buyers via third-country data centers, which matters for how export controls are actually functioning in practice.
Tower Semiconductor unveiled a $4 billion Japanese investment plan centered on a new optical chip hub, and Woodway is building a 22-mile gas pipeline for an undisclosed US data center
Both are capacity-side commitments, one in specialized optical semiconductor manufacturing and one in dedicated gas infrastructure for data center power, showing supply chain investment continuing to build out ahead of demand across geographies.
AI infrastructure names broadly strengthened over the past week, with semicap and networking names moving from roughly half or three-fifths in an uptrend to the large majority, while power and data center infrastructure names stayed split and neoclouds remained weak
The advance was broad rather than confined to a handful of names, with memory, servers and compute holding firm at strong levels throughout, though Broadcom's move against the trend in compute and Siemens Energy's slip in power show the strength was not universal.
Compute · Semiconductor equipment · Memory · Networking · Servers · Data-centre infrastructure · Power · Neoclouds· AVGO, ENR.DE, CRDO
Published 25/09/2026, 08:12 UTC · claude-sonnet-5