Hello world, it’s Friday, September 25th.
Oracle’s force majeure notice on its New Mexico data center campus is rattling power equipment suppliers and raising questions about big infrastructure commitments, even as Goldman Sachs projects the five largest US hyperscalers will push AI CapEx past $1.2 trillion by 2027. Akamai is moving fast to capture that spend, signing an $11.6 billion cloud infrastructure deal with Anthropic, and Google sent a fridge-size TPU satellite into orbit aboard SpaceX Transporter-18.
Let’s get into it. — Austin & Vik
Be sure to check out the Semi Doped podcast on YouTube or your favorite podcast player!
Tell us how to make Semi Doped better! Give us your feedback! ❤️
Oracle Force Majeure Notice on New Mexico Campus Rattles Power Equipment Suppliers
Oracle sent a force majeure notice to Blue Owl Capital over the Project Jupiter data center campus being built in New Mexico, citing a series of setbacks that have stalled construction on the Stargate-affiliated site. The move lets Oracle seek cover from contractual obligations tied to the development timeline. Bloom Energy and GE Vernova both saw their shares fall on the news, as investors repriced the revenue visibility each company had built around steady hyperscale construction schedules. Barron’s flagged GE Vernova specifically as exposed, given its power equipment backlog depends on data center builds proceeding on time. Force majeure claims of this kind can suspend or void supplier delivery commitments, leaving infrastructure vendors holding contracted capacity with no clear installation date.
Vik: Austin Lyons made this really keen observation on the Semi Doped podcast, a week before this happened. He was totally right. It’s the local governments who will set the pace of AI. Not Silicon Valley.
Akamai Commits $11.6 Billion to Anthropic, Crossing $500 Billion Compute Threshold
Akamai’s multi-year agreement with Anthropic covers $11.6 billion in cloud infrastructure, with Akamai separately committing $5.5 billion in capital expenditure to support the build-out. The deal isn’t structured as a standard cloud contract: forkast.news describes it as an equity position in the model layer, recasting Akamai from a CDN operator into a direct stakeholder in frontier AI. Anthropic’s cumulative compute spending crossed $500 billion in under a year, with Akamai’s commitment pushing it past that threshold. That pace, for a company still privately held, is striking on its own terms. Reuters reports Anthropic is separately seeking 50.1% voting control for co-founders ahead of a planned IPO.
Vik: This is for their CPUs. Conspiracy theory: Anthropic will release a personal agent like Muse.
Goldman Sees Hyperscaler AI CapEx Hit $1.2 Trillion as Nscale Closes Pre-IPO Round
Goldman Sachs strategists project spending on AI infrastructure by the five largest US hyperscalers will rise more than 50% in 2027 to $1.2 trillion, sustaining semiconductor and power-equipment demand well beyond the current buildout cycle. Against that backdrop, European AI cloud provider Nscale closed a $3.36 billion pre-IPO convertible note round led by Third Point, giving the company fresh capital to expand its GPU-dense data center platform ahead of a public listing. The convertible structure lets Nscale defer equity dilution while locking in growth funding now, at a moment when hyperscaler commitments give lenders unusual confidence in forward AI compute demand. Reuters Breakingviews called the implied valuation a stretch, but the capital is closed.
Google Launches Fridge-Size TPU Satellite Aboard SpaceX Transporter-18
Google’s Project Suncatcher sends its first hardware into orbit next week aboard SpaceX Transporter-18, carrying a fridge-size unit it calls a minimum viable product. The satellite packs four TPUs and 1,000 watts of solar power, though overheating constrains each compute run to just 15 minutes despite the spacecraft being rated for a full year of orbital operations. That thermal ceiling is itself a core finding the mission is designed to produce: Google needs real vacuum-and-radiation data before it can engineer around the problem. The payload rides a SpaceX rideshare rather than a dedicated launch, keeping early validation costs in check. No other major hyperscaler has flown an AI compute payload of this kind, leaving Google alone in collecting radiation-hardened TPU performance data from low Earth orbit.
Vik: Its cute they sent up a 4-TPU scale-up domain to test in space. I always maintain that launch costs are the biggest impediment to datacenters in space. Watch for drop in launch costs as a leading indicator for space compute.
ASML Sold Nothing to European Fabs as Dutch PM Eyes China Deal
ASML’s European revenue share has dropped to zero, with the company saying it has sold “absolutely nothing” in the region this year. That stark absence of domestic orders arrives as Dutch Prime Minister Rob Jetten confirmed he has discussed the MATCH Act with President Trump, legislation that would further restrict ASML’s China sales. Jetten told Bloomberg Television he is “pretty sure we come to some agreement that is workable for each and every one,” offering cautious optimism about the negotiations. Europe produces none of ASML’s revenue while its own government works to tighten the screws on the one market still buying heavily.
Vik: That’s because there are no chips happening in Europe. Absolutely nothing too surprising.
Muse Gives Every User a Free Isolated Linux VM in the Cloud
Meta’s Muse AI assistant now provisions each user with a dedicated Secure VM running a full Ubuntu Linux image, giving users root-level visibility into the runtime cell filesystem. A Sentinel process running outside the runtime cell oversees sensitive actions, while passwords and other secrets are stored separately from the cell. Users can install software, compile code, browse the web, and download their personal agent data via Settings > Data controls > Download your agent data.
Sector Watch
Compute & Edge
Microchip Technology closes acquisition of edge-AI chip firm Hailo, accelerating consolidation in the edge inference silicon market. (Jon Peddie Research)
Alibaba Zhenwu V900 AI accelerator with 216 GB memory joins Huawei Ascend 960 in competing for Nvidia’s vacated China AI compute share. (digitimes.com)
Mobilint delays mass production of its second-generation Regulus NPU by six months, pushing the schedule back as competition in the Korean edge-AI chip segment intensifies. (The Elec)
Foundry & Packaging
Korea Herald maps the post-CoWoS advanced-packaging race, with multiple Korean firms targeting next-generation AI chiplet integration platforms. (The Korea Herald)
SkyWater Technology reaffirms its role as an open-access quantum foundry following IonQ’s acquisition, pledging continued third-party customer access. (EE Times)
Tata Electronics boardroom battle prompts India government to seek assurances on the stability of the company’s $14 billion chip investment plans. (bloomberg.com)
Power
Eaton signs agreement to acquire COL Group, expanding power-distribution manufacturing capacity for data-center and utility markets across EMEA. (Business Wire)
ROHM SiC MOSFETs are selected for BMW’s Neue Klasse electric powertrain, marking a design win in a flagship EV platform. (Charged EVs)
Amazon signs a deal with Generac to secure backup power supply, putting standby generation capacity into the AI infrastructure spotlight. (Data Center Frontier)
Memory
SK hynix CEO Kwak Noh-Jung publicly reaffirms commitment to sustaining HBM and next-generation memory leadership as AI demand intensifies. (Gagadget.com)
Rambus memory-interface IP stocks attract renewed investor attention as AI-driven demand for high-bandwidth memory architectures broadens. (Barron’s)
YMTC secures its first-round patent win against Micron in German court over NAND technology, marking a significant IP offensive milestone. (YMTC)
Optics & Networking
InfraTech Cables selects Poland for a new European optical fiber cable manufacturing and R&D campus targeting subsea and terrestrial demand. (Business Wire)
Vocus launches a 4,000 km ducted fiber route connecting Brisbane and Darwin as its second project under Australia’s ADIP initiative. (Light Reading)
Iridium stockholders approve the pending acquisition by Rocket Lab with 99.6% of votes cast in favor, clearing a key deal milestone. (Light Reading)
Data Centers
Japan’s FSA steps up scrutiny of AI data-center financing by the country’s largest banks and life insurers as leverage exposure grows. (Bloomberg Tech)
Nscale UK supercomputer launch is delayed by power supply problems, underscoring the grid-access bottleneck facing new AI compute builds. (The Guardian)
xAI founder Elon Musk targets doubling the Nvidia GPU count at Colossus 2 by year-end, escalating the Memphis data center buildout. (Trending Now Infrastructure)
Policy & Trade
Texas Attorney General opens probes into data center projects over water usage, adding a new regulatory front to the state’s infrastructure scrutiny. (Bloomberg)
Foxconn-led Taiwan consortium announces plans to build a major AI and EV industrial hub in Poland, diversifying manufacturing geography. (Taiwan News)
India government seeks boardroom stability assurances from Tata Electronics as a leadership dispute clouds the country’s flagship chip investment. (bloomberg.com)
Tell us how to make Semi Doped better! Give us your feedback! ❤️
The interesting part is how quickly the bottleneck is moving from chips to execution. Capex can keep rising while power, permitting and financing determine which projects actually get built.