AI Infrastructure & Compute
The physical and financial layer under AI: data centers, chips and GPU supply chains, energy and power constraints, neoclouds and inference providers, AI capex and its financing.
Modal targets underserved mid-sized AI training workloads
Modal is expanding beyond inference to serve researchers and biotech firms with smaller training runs. The company anticipates non-NVIDIA accelerators will become viable alternatives by 2027.
OpenAI grants $80 million to Savannah community for data center project
OpenAI is providing an $80 million grant to the Savannah, Georgia community to build trust for a new data center. The community retains full decision-making power over how the funds are deployed over five years.
Chinese firms order six million Nvidia GPUs despite export limits
Alibaba, ByteDance, and Tencent have ordered over six million Nvidia RTX Pro 5500 cards to run AI applications, a volume that exceeds Nvidia's global manufacturing capacity for 2027. The shortage highlights the strain on supply chains as US restrictions block access to more powerful chips.
Celestica shares surge thirtyfold on AI data center demand
Celestica's stock has risen from approximately $10 to over $330 since early 2023, driven by strong revenue growth in AI infrastructure. Analysts note that the current valuation prices in significant future growth, making the company vulnerable to any slowdown in AI capital expenditure.
SF Compute builds physical settlement layer for gpu trading
SF Compute is operating a full cloud stack to facilitate physical settlement of compute markets, arguing that this independence is essential for market integrity. The company claims that major hardware and cloud providers are structurally unable to provide this service due to conflicts of interest.
German project escate achieves up to 89% energy savings in ai
The ESCATE project concluded with the development of neuromorphic hardware and spiking neuron algorithms that reduce energy consumption in AI data centers. Professor Wolfgang Maas stated that the technology is being transferred to successor projects for use in drones and military initiatives.
California shifts data center costs to operators as US projects face delays
Governor Gavin Newsom signed seven laws requiring data center operators to cover infrastructure costs and undergo full environmental reviews. This state-level tightening contrasts with federal efforts to remove construction barriers, even as 47 large US projects face delays or cancellations.
CoreWeave bond yields hit 15% amid rising default fears
CoreWeave's bond yields surged to 15% in October 2026, signaling a 60% to 66% probability of default over the next six years. The move reflects growing investor anxiety over the company's heavy reliance on debt-financed data centers without a diversified core business.
Gartner predicts 80% of AI cloud providers will fail by 2030
Gartner forecasts that most specialized AI cloud providers will not survive independently by 2030 due to high costs and rapid hardware depreciation. Market stress is already evident as CoreWeave bonds trade at distress levels and its stock falls sharply.
Oracle invokes force majeure in New Mexico data center deal
Oracle is using a force majeure clause to lower interest payments on the Project Jupiter facility after local authorities denied key permits. The dispute highlights stress in AI financing as the CFTC delays the launch of GPU compute futures.
Nvidia rumored to acquire Reflection for $30 billion
Nvidia is reportedly in talks to buy open-source model company Reflection for $30 billion, following its $12 billion acquisition of Poolside. The move aims to secure the software layer of the AI stack and reduce reliance on independent model developers.
ASML sees record orders extending into 2030
ASML reported unusually strong demand from customers like TSMC and Intel, with orders placed for 2029 and partially for 2030. Production capacity remains constrained to roughly 30% annual growth, limiting the company's ability to meet this surge.
Microsoft to deploy Maya AI chips at scale in 2027
Microsoft plans to use its proprietary Maya chips for the majority of its AI inference workloads starting in 2027. The move aims to reduce reliance on external graphics processors while improving energy efficiency for products like Copilot.
DX study shows AI tool spend up 28x with flat innovation
Spending on AI tools has increased 28-fold over the past year, yet overall innovation metrics remain flat according to a new DX report. The findings suggest that new friction points in code review are offsetting the velocity gains from faster code generation.
Data and RL environments become critical bottlenecks for model progress
Nathan argues that high-quality data and reinforcement learning environments are the primary levers for capability gains in non-verifiable domains. He predicts a multi-year transition as labs move to in-house data creation to address current supply chain flaws.
Dyna Robotics builds full-stack hardware to overcome supply chain bottlenecks
Dyna Robotics is developing its own actuators and inference stack to minimize latency and address US manufacturing gaps. Jason Ma argues that hardware constraints, not model performance, are the primary bottleneck for scaling complex robotic tasks.
Microsoft unveils MAIA 300 chip for multi-gigawatt inference
Microsoft announced the MAIA 300, claiming 40% more tokens per watt than leading GPUs. The launch pushed Microsoft's market cap above $4 trillion while rival chipmakers' stocks declined.
Ukrainian strikes disable three Yandex data centers in Russia
Drone attacks knocked out facilities training Yandex GPT, disrupting services across Russia. The incident highlights the physical vulnerability of AI compute infrastructure.
Amazon negotiates $7 billion acquisition of Descartes AI
Amazon is in talks to buy Descartes AI at a $7 billion valuation, a 75% premium over its last round. The deal centers on Descartes' Optimization System, which claims to cut inference costs significantly.
Simple Closure brokers over 100 data deals from dead startups
The firm has facilitated over 100 transactions since April, returning more than $10 million to closed startups. Google reportedly outbid Merkur for Spirit Airlines data in a $10 million deal.
Everpure expands AI infrastructure strategy with OneTouch acquisition
Everpure acquired OneTouch to add data intelligence capabilities, aiming to reduce unnecessary token usage in LLMs. The company targets enterprises, neoclouds, and hyperscalers with a new memory hierarchy.
Enterprise AI KPIs should prioritize task latency and accuracy
Everpure's Rob Lee advises CTOs to focus on task completion time and accuracy rather than raw GPU utilization. He argues that optimizing for speed naturally improves utilization in bursty enterprise workloads.