What Is vLLM? The Open Source AI Model Server, Explained
vLLM is free, open source software that serves open AI models to many users at once. What PagedAttention does, how it differs from Ollama, and who should use it.

The compute, data centers and software infrastructure that AI runs on. Recent coverage includes GPU cluster orchestration with Slurm, Ray and Kubernetes, embedding serving engines, Microsoft's cancelled Nscale GPU deal and Anthropic's Nova Scotia data center lease.
vLLM is free, open source software that serves open AI models to many users at once. What PagedAttention does, how it differs from Ollama, and who should use it.

Foxconn Q3 2026 net profit rose 34% to a record NT$77.5B on AI servers. It sees high double-digit AI rack shipment growth in 2027 but tight chip supply.

Nebius Q3 2026 revenue rose 571% to $980.8M, but 2026 revenue will land at the low end and its ARR goal slips to February. 2027 capex: $40B to $45B.

Naver signed a 2 trillion won (about $1.5B) AI Factory computing deal with a US AI inference company: about 20 MW, B300 and Vera Rubin GPUs, from 2027.

Meta says infrastructure head Santosh Janardhan leaves early 2027 and ex-AWS executive Dave Brown succeeds him. Meta expects $135B to $145B 2026 capex.

Bloomberg reports CoreWeave plans an investment-grade loan of at least $4B at a 7% to 7.5% yield, about half what its junk bonds pay. How it works.

Google's MaxText v0.2.4 splits long training sequences across Cloud TPU chips with four context-parallel strategies, reaching 14M tokens in Google's tests.

CoreWeave posted $3.66B in Q3 revenue, a $132.5B backlog and higher 2026 guidance, but also a $709M net loss and $846M in net interest expense.

Ex-DOE founders launch Mavria with $130M raised and a $1B valuation. Its first product, Terrain, is a data center built in a factory in Houston.

Hankyung reports a planned 300MW AI data center in Gumi, South Korea, fell through after lease talks with Anthropic ended over site issues.

Google opened Pathways, its system for training large models such as Gemini, to Google Cloud users. How it runs JAX jobs on TPUs and how it differs from Ray.

SemiAnalysis toured Envision's Galaxy campus in Inner Mongolia: up to 800MW planned, 120MW delivered within a year, own wind power and an 800V DC pilot.

The FT reports Anthropic agreed to pay about $3.4 billion for 1.7 GW of OnePWR gas generators and the capability to install them at its data centers.

Nvidia GPUs will run prefill and Rebellions' R200 NPUs will run decode in SK Telecom racks linked by NVLink Fusion. R200 systems are expected in 2028.

CoreWeave's credit default swaps sit above 1,000 basis points, a level analysts call distress. The debt figures, its reply and what to watch on November 10.

Google will invest US$15 billion in Malaysia in 2027 and 2028 for data center campuses, more than 1 GW of clean energy capacity and water projects, it said on November 4.

The FT reports Microsoft rejected an option to more than double its Nscale lease in Texas, weeks after leaving the Essex project. Reuters could not verify it.

DCD reports most anti-data center measures passed on November 3, with at least 14 Ohio bans or limits. Maine, Illinois, Kansas, California and Florida voted too.

Fal announced $1.65B across a Series E and F. Bloomberg reported the $1.3B Series F values it at $16B, double April's $8B.

The FT reports Oracle canceled a roughly $7 billion lease that would have given Tencent about 100,000 chips, a day after it reported Nvidia told cloud partners to sell at least 51% to non-Chinese firms.
