LLM Hosting Roundup: Nebius and CoreWeave financing, Meta Muse Ember and Google TPU 8 variants (13 Nov 2026)
Nebius and CoreWeave lean on customer prepayments to fund expansion, Fractile raises $720 million for inference chips, Meta launches low-cost Muse Ember, Google reshapes TPU 8 for inference and Artificial Analysis updates its endpoint reliability benchmark.
This week's scan shows AI hosting moving in two directions: neoclouds are financing rapid capacity expansion with customer commitments, while model and hardware developers are tuning inference for cost, latency and agentic workloads. The evidence also points to continued regional buildout and a growing focus on the software and networking layers around hosted models.
- Nebius outlines a customer-funded 2027 expansion: Nebius plans $40–$45 billion of 2027 capex and expects customer prepayments and asset-backed financing to cover up to 85% of it.
- CoreWeave leans on prepayments and contract-backed finance: CoreWeave posted $3.66 billion in quarterly revenue and is using prepayments, contract-backed loans and shorter deals to fund growth.
- Fractile raises $720 million for inference chips: British chip developer Fractile raised a Series C at a reported valuation of about $7 billion to ship its first chips and systems.
- Meta launches Muse Ember: Meta's smallest Muse model targets high-volume agentic execution at $0.10/$0.50 per million input and output tokens.
- Google reshapes TPU 8 variants for inference: TPU 8tx and TPU 8i Max are aimed at lower-latency inference, larger memory footprints and confidential workloads.
- AWS opens its Chile cloud region: AWS launched its Chile region for select customers, with AI services expected within a year.
- Rackspace signs capacity agreements: Rackspace will operate dedicated GB300 NVL72 infrastructure for Together AI and provide about 2MW of AMD capacity to Blacksmith.
- Endpoint Accuracy Index v1.1: Artificial Analysis found 24 of 103 serverless endpoints scored statistically below a self-hosted reference.
- Multimodal Knowledge Agent benchmark launches: A new Google DeepMind benchmark tests agentic knowledge work over complex visual files.
- NaiveAI targets fast, low-cost coding inference: Latent Space tested NaiveAI's open-weights model and inference engine, which peaked at about 2,100 tokens per second.
- AMD launches EPYC 9006 Venice: AMD's new server CPUs offer up to 256 cores per socket and are pitched at AI host and sandbox workloads.
- Meta shows the MetaRoCE adapter: Meta displayed an open RDMA network adapter designed to scale AI networks over commodity Ethernet.
- Gartner forecasts strong datacenter-systems growth: Gartner's forecast puts 2027 datacenter-systems spending growth at 55.4%.
- OpenAI expands Codex Cloud workflows: Codex Cloud can now run multiple tasks in isolated, published environments handed off from local Codex.
Nebius plans a customer-funded 2027 AI infrastructure expansion
TechCrunch reported that Nebius plans $40–$45 billion of 2027 capex to bring another gigawatt of capacity online. The company expects customer prepayments and asset-backed financing to cover up to 85% of that spending, while Microsoft-linked capacity at the Vineland, New Jersey site has been delayed into 2027. Data Center Dynamics separately reported $12 billion in Q3 contract value, $973.8 million in quarterly AI Cloud revenue and customer-funded capex coverage for the contracts signed in the quarter. [1] [2]
For hosting buyers, the announcement highlights both the scale of planned capacity and the delivery risk created when power, construction or third-party sites fall behind schedule.
CoreWeave uses prepayments and contract-backed finance as AI cloud demand lifts revenue
TechCrunch reported that CoreWeave generated $3.66 billion in third-quarter revenue, up 168% year over year, alongside a $132.5 billion backlog, a $709 million net loss and $846 million in quarterly interest expense. The company expects about $9 billion in customer prepayments and deposits over three months, and is using contract-backed loans and shorter-term deals where higher prices justify the risk. [3]
Data Center Dynamics reported that CoreWeave brought more than 1.7GW of power online and is also deploying its managed CoreWeave Omni stack inside customer data centres. The Next Platform added that the company raised prices across several GPU offerings. [4] [5]
Fractile raises $720 million for inference-chip development
TechCrunch reported that British AI-chip developer Fractile raised a $720 million Series C at a valuation reported as about $7 billion. Fractile is pursuing a full-stack approach, including a machine-learning compiler, and said the funding will support shipping its first chips and systems next year. [6]
Data Center Dynamics reported that the company's in-memory-compute design is intended to improve power efficiency and inference performance, while noting that the UK Sovereign AI Fund and British Business Bank participated without disclosing their investment sizes. [7]
Meta launches Muse Ember for high-volume agentic execution
Meta's AI Developer Blog introduced Muse Ember in Muse Code and the Meta Model API for repeated coding subtasks, tool calls and document analysis across text, image, audio and video inputs. Meta listed $0.10 input and $0.50 output per million tokens for prompts up to 200k input tokens, with higher pricing above that threshold. [8]
Artificial Analysis measured Muse Ember at 36 on its Intelligence Index and estimated about $0.13 per task at maximum reasoning effort. The analysis also reported stronger results on some agentic and coding evaluations than similarly priced models, alongside weaker knowledge accuracy and a high answer-decline rate. [9]
Google reshapes TPU 8 variants for inference and confidential workloads
The Next Platform reported that Google is preparing TPU 8tx and TPU 8i Max variants for 2027 installation. TPU 8tx replaces the larger training-oriented torus with a smaller Boardfly-connected system for lower-latency inference, including complex agentic workloads, while TPU 8i Max combines two accelerator complexes to provide a larger memory footprint for context and KV-cache workloads. [10]
ServeTheHome reported that TPU 8tx pairs the accelerator with AMD Turin hosts for confidential computing and that TPU 8i Max reaches 17.2 TB/s of HBM bandwidth. Google had not published pricing for the new systems in the available coverage. [11]
AWS opens its Chile cloud region to select customers
Data Center Dynamics reported that AWS launched its Chile region for select customers, with general availability expected by the end of 2026. AWS plans to invest $4 billion in the region, which starts with compute, storage, networking and database services; AI offerings are expected within a year of launch. [12]
The region gives customers in Chile a local option for data residency, regulatory requirements and lower-latency services, while its AI-hosting availability remains a future phase rather than a launch-day capability.
Rackspace signs dedicated inference capacity agreement with Together AI and compute deal with Blacksmith
Data Center Dynamics reported that Rackspace will provide and operate dedicated Nvidia GB300 NVL72 infrastructure for Together AI, with deployment planned for 2027. The companies also signed a non-binding letter of intent for an enterprise AI go-to-market approach; separately, Rackspace agreed to provide Blacksmith about 2MW of AMD Epyc-powered capacity in Northern Virginia. [13]
The report said the Together AI facility will include power, liquid cooling, networking and storage, but did not disclose its capacity or site. Rackspace said financing for the GPU purchase was still under discussion and not yet committed.
Endpoint Accuracy Index v1.1 measures provider trade-offs in hosted inference
Artificial Analysis launched version 1.1 of its Endpoint Accuracy Index, covering 103 serverless endpoints across six open-weight models. Twenty-four endpoints scored statistically below the self-hosted reference; the analysis attributed differences to output caps, reduced reasoning, tool-call handling and request or stream limits. [14]
The reference deployments ran on vLLM or SGLang on NVIDIA B200, B300 or GB200 systems. The results were measured between 8 October and 4 November 2026, so they are point-in-time comparisons rather than a permanent ranking of providers.
Artificial Analysis and Google DeepMind launch the Multimodal Knowledge Agent benchmark
Artificial Analysis reported a new benchmark from Google DeepMind for agentic knowledge work over complex visual files. It covers 200 private tasks across 10 industries, six output types and roughly 70 weighted criteria per task, with evaluation run in the Stirrup agent harness. [15]
Claude Fable 5.2 led the reported leaderboard at 57% Climb WSR, followed by Gemini 4 Carbon at 54% and Claude Opus 5.5 at 53%. Cost per task varied nearly 30-fold within ten points of the top score, making the benchmark relevant to hosted-model selection as well as raw capability comparisons.
NaiveAI targets fast, low-cost coding inference
Latent Space tested NaiveAI's open-weights Naive-N0.5-Flash model and its NaiveRT inference engine against DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. The article reported a peak of about 2,100 tokens per second on a single eight-GPU Blackwell node and target pricing of $0.10 input, $0.40 output and $0.01 cache-read per million tokens. [16]
The report cautioned that the peak is an upper bound from a single-stream speculative-decoding setup and that the sample was small. In the test, median response time was 3.1 seconds for Naive-N0.5-Flash with thinking enabled, versus 5.3 seconds for GLM 5.3 Flash and 10.1 seconds for DeepSeek V4.1 Flash 1017 in the compared configurations.
AMD launches EPYC 9006 Venice processors for AI hosts and agentic workloads
ServeTheHome reported the launch of AMD's EPYC 9006 "Venice" server CPUs, with up to 256 cores per socket, PCIe Gen6 and higher memory bandwidth. The flagship SP7 stack includes up to 8TB of memory capacity and up to 600W peak power, according to the review's specification table. [17]
The Next Platform connected the processors to AI host and sandbox workloads, reporting AMD's claim of 1.9x the throughput of prior Turin processors and higher agents-per-watt in its comparisons. Those figures are vendor comparisons and should be read as such. [18]
Meta shows the MetaRoCE adapter for scalable open Ethernet AI networks
ServeTheHome reported that Meta displayed a MetaRoCE network adapter at OCP Summit 2026. The open RDMA transport approach includes per-path reliability, congestion control, packet-loss tolerance and multi-path streams over commodity Ethernet; Meta's stated ambition is to scale the protocol toward million-GPU training. [19]
For hosted inference and training operators, the development is relevant because network behaviour can become a limiting factor as accelerator clusters grow, although the report did not provide production availability or performance benchmarks for the adapter.
Gartner forecast points to strong 2027 datacenter-systems growth
The Next Platform analysed Gartner's November forecast, which put 2027 datacenter-systems spending growth at 55.4%, after 73.8% growth in 2026. It said spending on datacenter systems plus IaaS could reach $1.73 trillion in 2027, compared with $1.68 trillion for enterprise software. [20]
The article qualified the outlook with political opposition, electricity constraints and supply-demand price pressure. It is a forecast and analysis, not a commitment from cloud providers or a guarantee of available hosting capacity.
OpenAI expands Codex Cloud workflows with isolated environments
Latent Space reported on OpenAI's AI Engineer Code Summit workshop, where the company described Codex Cloud as a background executor that local Codex can hand work to. The revised workflow lets developers publish an environment and run multiple tasks in isolated instances, with access to selected GitHub repositories. [21]
The update is relevant to hosted coding-agent infrastructure because it treats the cloud as a persistent execution layer rather than only a separate interface. The report described the workflow and demonstrations, not new API pricing or a general-purpose model-hosting offer.
Conclusion
Customer-backed financing is allowing neoclouds to keep expanding, but the week's coverage repeatedly qualifies that growth with delivery, power, debt and utilisation risks. At the model and systems layer, lower-cost execution, inference-specific accelerators, local regions and more rigorous endpoint benchmarks are all becoming part of the hosting decision.
News Sources
- TechCrunch - Nebius tests the neocloud playbook with $40 billion in planned spending
- Data Center Dynamics - Nebius signs $12bn in deals for Q3, all deals reportedly cover its capex
- TechCrunch - CoreWeave turns to customer prepayments, shorter contracts to fund AI growth
- Data Center Dynamics - CoreWeave CEO addresses "volatility in the capital markets" as company posts $3.66bn in revenue
- The Next Platform - CoreWeave Is Accelerating To Escape Velocity Thanks To GPU Shortages
- TechCrunch - With $720M round, Britain's Fractile gets a leg up in the race for AI chips
- Data Center Dynamics - UK chip startup Fractile valued at $7bn following $720m Series C funding round
- Meta AI Developer Blog - Start building with Muse Ember, our smallest, most efficient model
- Artificial Analysis - Meta's Muse Ember is the cheapest model scoring 36 on the Intelligence Index
- The Next Platform - Google Rejiggers TPU Accelerators For Broader AI Inference Work
- ServeTheHome - Google TPU 8tx Pairs AMD EPYC Turin Hosts with Confidential Computing while TPU 8i Max Doubles Memory Bandwidth
- Data Center Dynamics - AWS launches Chile cloud region for "select customers"
- Data Center Dynamics - Rackspace signs capacity agreements with Together AI and Blacksmith
- Artificial Analysis - Endpoint Accuracy Index v1.1: Tighter long context intervals and expanded model coverage
- Artificial Analysis - Multimodal Knowledge Agent benchmark: testing knowledge work capabilities with complex visual files
- Latent Space - NaiveAI Promises 2,100 Tokens per Second. How Well Does It Code?
- ServeTheHome - AMD EPYC 9006 Results Are Here and AMD Delivers a Knockout Part Deux
- The Next Platform - AMD Tunes Up CPUs For Agentic AI With "Venice" Epyc Processors
- ServeTheHome - This is the Meta MetaRoCE Network Adapter
- The Next Platform - Datacenter Systems Spending To Go Through The Roof In 2027
- Latent Space - AI Engineer Code Summit: OpenAI on New Ways of Working With Codex
Related model families
| Family | Developer | Weights | Licence | Models |
|---|---|---|---|---|
| DeepSeek | DeepSeek | Open | MIT | 6 |
