← All news

LLM Hosting Roundup: GPT-6.2 Astra and Sol, Gemini 4 Neon and Cerebras adds Microsoft (20 Nov 2026)

NewsBy Lawrence Ladomery

OpenAI launches GPT-6.2 Astra and Sol with day-one availability on Amazon Bedrock and Microsoft Foundry, Google ships Gemini 4 Neon, Cerebras signs Microsoft, NVIDIA posts record revenue and tightens neocloud financing, and Lambda brings co-packaged optics into production.

This week's coverage points to faster model release cycles, broader managed access to frontier models and continued investment in inference infrastructure. Cloud and hardware providers are also pushing efficiency and geographic diversity as demand for AI serving grows.

  • GPT-6.2 Astra and GPT-6.2 Sol launch: OpenAI's new models list $8/$40 and $2/$10 per million input/output tokens and are generally available on Amazon Bedrock and Microsoft Foundry.
  • Gemini 4 Neon arrives: Google DeepMind's fourth model in seven weeks scores 61 on the Artificial Analysis Intelligence Index at $3/$15 per million tokens.
  • Cerebras adds Microsoft: Microsoft becomes the third hyperscaler to deploy Cerebras chips, while CS-5 systems are prepared for 2027 production.
  • NVIDIA reports record revenue: Fiscal third-quarter revenue reached $115.8 billion, with $107.6 billion from datacenter, as Vera Rubin systems ramp.
  • NVIDIA clarifies neocloud financing: Loan guarantees apply only once customers secure financing and meet construction and deployment conditions.
  • Co-packaged optics reaches production: Lambda and NVIDIA have deployed a co-packaged-optics cluster with more than 10,000 GPUs.
  • Huawei Ascend cluster targets token export: A Shantou data centre built on Ascend 950DT accelerators serves overseas users.
  • Cloudflare builds small serving models: Cloudflare is launching its Clef and Clairic System One models alongside Workers AI.

GPT-6.2 Astra and GPT-6.2 Sol expand managed model access

Latent Space reported that OpenAI launched GPT-6.2 Astra and GPT-6.2 Sol, with rollout to ChatGPT Work and Codex and developer access through the API. It listed API pricing of $8/$40 per million input/output tokens for Astra and $2/$10 for Sol; both models have a 1M-token context window and a 128K output limit. [1]

AWS and Microsoft separately announced general availability in Amazon Bedrock and Microsoft Foundry, giving teams additional managed serving options for both models from launch. [2] [3]

Gemini 4 Neon brings a new price and performance point

Artificial Analysis reported that Google DeepMind released Gemini 4 Neon, its fourth model in seven weeks. Its evaluation scored Neon 61 on the Artificial Analysis Intelligence Index and reported $3/$15 per million input/output token pricing, a 1M-token context window and public availability through Google's API and Gemini Enterprise Agent Platform. [4] [5]

The analysis also measured lower cost and time per task than Gemini 4 Carbon at the cited comparison settings, which makes Neon a new option to weigh for teams choosing between Gemini tiers.

Cerebras adds Microsoft while preparing its next inference system

The Next Platform reported that Cerebras is preparing its CS-5 system for production in summer 2027 while expanding cloud and manufacturing capacity for inference contracts. [6]

Data Centre Dynamics reported that Microsoft became the third hyperscaler to sign on to deploy and offer Cerebras chips, after AWS and Oracle, although deal details were not disclosed. DCD also described capacity constraints and said larger portions of planned capacity were expected to deploy in mid-2027. [7]

NVIDIA's datacenter scale continues to set the infrastructure backdrop

The Next Platform reported that NVIDIA's fiscal third-quarter revenue reached $115.8 billion, including $107.6 billion from its datacenter business, as Vera Rubin systems ramped. It described $47.47 billion in sales for NVIDIA's Datacenter AI Clouds, Industrial, and Enterprise division, which includes neoclouds and other inference-focused customers, and reported guidance of about $130 billion in fourth-quarter sales. [8]

NVIDIA's results release separately stated that Vera Rubin NVL72 had shipped. [9]

NVIDIA clarifies selective financing for neoclouds

TechCrunch reported that NVIDIA clarified its loan guarantees for AI data centres take effect only after customers secure financing and meet construction and deployment conditions. Interim CEO Colette Kress said NVIDIA would continue supporting AI customers but would be more selective about future investments, commitments and guarantees. [10]

For hosting buyers, the change matters because vendor-backed financing has supported part of the neocloud capacity buildout.

Co-packaged optics reaches a large production cluster

ServeTheHome reported that Lambda and NVIDIA had deployed a production co-packaged-optics cluster with more than 10,000 GPUs. The cluster uses NVIDIA Quantum-X co-packaged-optics switching. [11]

The article said moving optics onto the switch package can reduce the power used by high-speed links, which matters when facility power limits the number of GPUs that can be deployed.

Huawei Ascend cluster targets overseas token demand

Data Centre Dynamics reported that a China Telecom, Huawei and Tianyima collaboration launched an AI data centre in Shantou using Huawei Ascend 950DT accelerators for a token-export business serving foreign users. The infrastructure had begun operating the previous week. [12]

DCD said the project represents approximately $150 million of investment, while the site's deployed capacity and chip count were not disclosed.

Cloudflare builds smaller serving models alongside Workers AI

Latent Space reported that Cloudflare is building and launching System One models called Clef and Clairic alongside its Workers AI infrastructure. The article described Clef as a decision model and Clairic as an extraction model, and said Cloudflare serves open-weight models while exploring further post-training and potential LLM work. [13]

The company's stated emphasis is on smaller, faster models for tasks where a full LLM may be unnecessary.

Conclusion

Model access is broadening across managed platforms while price, latency and task-level efficiency remain central differentiators. At the infrastructure layer, hyperscaler partnerships, power-aware networking and region-specific accelerator deployments continue to shape where inference capacity can be delivered.

News Sources

  1. Latent Space - GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse
  2. AWS - OpenAI GPT-6.2 Astra and GPT-6.2 Sol are now generally available on Amazon Bedrock
  3. Microsoft - GPT-6.2 Astra and GPT-6.2 Sol now generally available in Microsoft Foundry
  4. Artificial Analysis - Google continues to accelerate its release cadence with Gemini 4 Neon
  5. Google - Introducing Gemini 4 Neon
  6. The Next Platform - Cerebras Readies CS-5 Systems For 2027 Launch As It Bags Another Hyperscaler
  7. Data Centre Dynamics - Cerebras secures Microsoft as a customer
  8. The Next Platform - Nvidia Breaks Revenue And Profit Records, And Will Keep Doing It
  9. NVIDIA - NVIDIA Announces Financial Results for Third Quarter Fiscal 2027
  10. TechCrunch - Nvidia clarifies when it will — and won't — keep financing neoclouds
  11. ServeTheHome - NVIDIA and Lambda Disclose First Production Co-Packaged Optics Cluster
  12. Data Centre Dynamics - Data center powered by Huawei Ascend chips launches in China
  13. Latent Space - Cloudflare Joins the Jevolution and Labels Itself a Neolab

LLM hosts mentioned