Amazon CEO Andy Jassy said AWS could become a trillion-dollar annual-revenue business within five to 10 years, with revenue growing above 40% in both 2027 and 2028. AWS reported 39% year-over-year quarterly growth and a $184 billion annualized revenue run rate. Amazon expects approximately $220 billion of 2026 cash capital expenditure, most directed toward AWS, and meaningfully higher spending in 2027.
The AWS Machine Learning Blog published a walkthrough of an enterprise Agentic AI Platform built on Amazon Bedrock and Amazon Bedrock AgentCore, covering a multi-account operating model, a central Platform control plane, repeatable team-owned Workstream cells, separate inference and MCP tool gateways, and an evaluation pipeline that gates agents before production promotion. AgentCore is described as providing modular runtime, gateway, identity, memory, policy, registry, and evaluation services, while a companion blueprint adds ownership, account boundaries, software delivery, observability, cost governance, and retirement as infrastructure as code. The post states the deployment and available features are based on the eu-west-1 (Ireland) Region and that the repository is a reference implementation developers can adjust.
AWS Machine Learning Blog provides a walkthrough for fine-tuning Qwen2.5-1.5B across eight NVIDIA L4 GPUs on two g6.12xlarge instances using Amazon EKS, KubeRay, Ray Train, PyTorch FSDP, and EFA. AWS says its Ray Train Deep Learning Container packages and patches the required CUDA, PyTorch, NCCL, EFA, and Ray components, while the sample uses one RayCluster manifest and requires no custom image build. AWS reports that enabling the EFA path doubled throughput over TCP with unchanged code, though the performance claim comes from the vendor.
AWS says its Amazon Nova Act-based QA Studio reference solution now organizes tests and test suites under applications, with ownership and health metrics including pass rate, executions, failures, and a flaky-test-adjusted score. It also adds team-based user management with five roles, application-level permissions, and bounded cross-team sharing. Teams can define tests in natural language, run them through AI-powered visual navigation, and use a CLI for batch regression testing in CI/CD pipelines.
German embodied AI startup MicroAGI publicly described a two-hop platform for delivering large multimodal robot-training datasets to S3-compatible object storage. The first hop stages source data in Hugging Face storage; the second uses Cloudflare Workers and Queues to transfer it to a customer's storage. MicroAGI said the process does not share customer credentials and can deliver petabyte-scale datasets. The company also reported throughput, cost, integrity-checking and dataset-browsing capabilities, but these figures have not been independently verified in the supplied report.
TrendForce expects TSMC's CoWoS-L to remain the mainstream advanced-packaging option for AI chips through 2028, citing its technology maturity and yield advantages. Intel's EMIB-T is pursuing AI-chip customers with a localized silicon-bridge approach, but competing routes must demonstrate stronger maturity and yield. The analysis also examines how larger AI models and more complex inference are driving demand for Chiplet, HBM and higher-bandwidth packaging.
AWS announced Seekable OCI (SOCI) indexing for Amazon SageMaker AI Training and processing jobs, allowing container images to be lazy-loaded instead of fully downloaded before execution. Users can select pre-indexed AWS Deep Learning Containers or index a custom image, though AWS says benefits vary with image size and startup behavior; its tests found GPU idle time during image pulls fell by up to 90%.
Property Finder discovered that 96.8% of its evaluation labels came directly from predictions by its incumbent vision model rather than independent human verification. In an AWS Machine Learning Blog case study, the team rebuilt expert-annotated ground truth and used evolutionary prompt optimization on Amazon Bedrock to select a model with higher accuracy and 86% lower cost.
HolonIQ’s Q3 2026 Education AI Watch Report says roughly 12 of every 100 global higher-education partnerships now involve AI and data, compared with about two five years ago. Skills training and capability development account for 40% of education AI initiatives, while degree and curriculum development represent 23%, according to the report as summarized by 36Kr. The underlying figures come from HolonIQ and were not independently verified by the supplied article.
An Edge AI Daily report said AWS ended the second quarter of 2026 with $496 billion in remaining performance obligations and $42.2 billion in quarterly revenue, up 37% year over year. Amazon CEO Andy Jassy said the company’s 2026 capital-spending guidance had been raised to $220 billion but still would not meet all demand, with a capacity gap expected in 2027 and demand described as “shocking” for 2028. The report also said Amazon added 3.9 GW of data-center power capacity in 2025 and plans to double total power capacity by the end of 2027.
TMTPost, citing IDC data, reported that Alibaba Cloud held 45.1% of China’s automotive public-cloud infrastructure market in the first half of 2026. Volcano Engine ranked second in infrastructure; the report estimated its share at more than about 14.9% using aggregate public data rather than a directly disclosed IDC figure. In the smaller AI cockpit solutions market, Volcano Engine led with 45%, while Alibaba ranked second. The report said the companies are competing for autonomous-driving training resources and in-vehicle AI deployment projects.
Amazon CEO Andy Jassy said AWS capacity for 2026 is sold out and 2027 is largely sold out, while a good portion of 2028 capacity is already committed. The Information reports that AWS still reserves capacity for on-demand computing, inference services, Bedrock and EC2, so the statements mainly concern large capacity bookings rather than every available resource. Jassy attributed the strong bookings to continuing demand from AI laboratories and enterprise customers.
SemiAnalysis reports that Anthropic is running production inference and reinforcement-learning rollouts on AWS Trainium2 and Trainium3, while AWS Neuron is moving Trainium toward native PyTorch training with eager mode, torch.compile and standard APIs. Its engineers found the Trainium3 direction promising after testing the TorchTitan and TorchNeuron stack, but encountered unspecified bugs, long compilation times and a platform that remains far less mature than NVIDIA's CUDA ecosystem. AWS says upcoming TorchNeuron releases will use precompiled NKI kernels and a redesigned compile stack to target faster time to first step, though that target has not yet been demonstrated.
The SGLang Team, in collaboration with AWS, announced a beta plugin for running LLM inference serving and reinforcement learning rollouts on AWS Trainium2 and Trainium3. Built on TorchNeuron, the beta initially covers Qwen3, Qwen3-MoE-30B-A3B, DeepSeek-V2-Lite, GLM-4, and GLM-5.2/5.3, with parallelism, Radix Cache, BF16, FP8, and an asynchronous scheduler. In an early end-to-end test, SGLang generated Qwen3-MoE-30B-A3B rollouts on Trainium3 while Miles trained a GRPO policy with Megatron on one H200 node; performance work remains ongoing.
According to the release notes, Codex CLI 0.164.0 adds ten numbered /presets slots for reusable prompts and slash commands, inline /model and /permissions arguments, and fullscreen image previews on terminals supporting Kitty or Sixel graphics. It also documents enterprise MCP authentication options, credential masking for supported remote Bash commands and Zsh snapshot sessions, and an Amazon Bedrock setup wizard enabled by default. Bug fixes address compaction failures, session resumption, Windows proxy authentication, configuration overrides, queued prompts, and macOS sandbox checks.
AWS announced an upgrade to the Amazon Polly Generative engine, which it says delivers more expressive, more accurate speech output with low latency. The release adds 14 new voices across 9 locales, including the first generally available Generative voices in Hindi and Finnish, and AWS says the upgrade is now generally available in production. AWS describes it as a drop-in change: voice names and API calls are unchanged, existing applications keep working, and pricing stays the same, while standard and neural voices are unaffected.
AWS announced the general availability of training sessions within Amazon SageMaker AI serverless model customization, letting developers write and coordinate their own post-training loop in Python — including custom loss functions or reward signals — while SageMaker AI handles GPU allocation, distributed execution, fault recovery and checkpoint storage. The launch supports gpt-oss-20b, gemma-4-31B-it, Qwen3.6-27B and Qwen3.8-27B, and AWS released a cookbook with ready-to-run SFT and RL recipes. AWS reports that a rank-32 LoRA adapter for GPT-OSS 20B trained for 50 steps on GSM8K rose from 0.46 to 0.60 training reward and from 40% to 50.5% held-out accuracy, figures the vendor supplied without independent replication.
An AWS Machine Learning Blog post states that Amazon Quick is a HIPAA-eligible service, with that eligibility applying across its web, desktop and mobile clients subject to the shared responsibility model and customer configuration. The post says only the web version provisioned through the AWS Management Console can be brought under a signed BAA via AWS Artifact, and it separates HIPAA eligibility from HIPAA compliance. It also describes governance features such as ACL-aware retrieval, chat and feedback logging and controls over what each chat agent may do, alongside a fictional health-plan walkthrough covering up-front, irreversible choices such as customer managed KMS keys and document-level ACLs on knowledge bases.
An AWS Machine Learning Blog post walks through streaming each sentence an AI agent generates to a hosted Kokoro-82M text-to-speech model on Amazon SageMaker AI, so synthesis starts on sentence one instead of waiting for the finished response. The post frames the problem as a latency gap: a one-sentence response might generate in 600 milliseconds while a full response takes five seconds, during which the customer hears nothing. Kokoro-82M's 82 million parameters run in a custom container on an ml.g5.xlarge real-time endpoint with two routes, a POST streaming route and a bidirectional WebSocket route; AWS also compares returning audio to a browser over Server-Sent Events, where base64 encoding inflated payloads by close to the expected 33 percent and MP3 frames did not align with delivery chunks. AWS reports time to first audio fell by more than half in isolation and by 2.4 to 3.5 times in a deployed agent, and notes the handlers run a strictly sequential loop rather than concurrent synthesis.
An AWS Machine Learning Blog post describes deploying Kokoro-82M, an 82-million-parameter non-autoregressive text-to-speech model, as a custom container on an Amazon SageMaker AI real-time endpoint via AWS CDK stacks. Instead of guessing auto scaling target values, the authors measure two numbers against the live endpoint: Q_max, the highest concurrency one instance sustains before breaching a 3,000 ms latency SLO, and T_total, the end-to-end scale-out time. The post reports Q_max of 50 concurrent requests on ml.g5.xlarge (concurrency 60 breached the SLO) and a measured T_total of 260.2 seconds, planned against as 320.2 seconds with bounded policy detection lag, and derives scale-out and scale-in thresholds and cooldowns arithmetically from those values.
Rescale used AWS's AI-Driven Development Lifecycle (AI-DLC) framework to build an intelligent triage agent that classifies automated end-to-end test failures and routes them to the right team, according to an AWS Machine Learning Blog post. The agent was built with the Strands SDK, containerized on Amazon EKS inside Rescale's existing Argo Workflows pipeline, and retrieves log evidence through MCP, after the team moved log retrieval from a third-party service to Amazon CloudWatch. AWS says Rescale went from problem definition to staging deployment in five days and to production within weeks; the post also says humans define scope, validate work and approve production changes while AI proposes and executes.
A commentary republished by TMTPost and written by Naojiti asks who actually funds Anthropic, citing Reuters figures that its 2025 revenue approached $4.6 billion, roughly 12 times the previous year. More than 80% of that came from usage-based API and token billing rather than subscriptions, and about 47% of sales flowed through Amazon and Google cloud platforms, the piece says, while 2025 operating losses exceeded $8 billion, including $7.3 billion in compute and infrastructure spending. The article also flags that talk of a $2 trillion valuation and a record IPO sits among unverifiable rumours, and argues that the loudest Chinese-language discussion — account resale and access workarounds — is not Anthropic's core business, with the US market supplying nearly two-thirds of sales.
TrendForce forecasts that server CPU shipments will increase 31% in 2026 and more than 40% in 2027 as AI inference and agentic AI workloads raise CPU-to-accelerator ratios. The research also expects Arm-based server CPUs to reach 23% market share in 2026 and 30% in 2027, while Intel faces delays affecting new product shipments. These figures are forward-looking estimates reported by TechNews, TMTPost and IT之家, not realized shipment results.
An AWS Machine Learning Blog walkthrough describes how enterprises can connect on-premises data sources to Amazon Quick without traffic crossing the public internet. It assumes an existing AWS Direct Connect or Site-to-Site VPN link to the VPC, then adds PrivateLink-powered VPC endpoints plus security groups and endpoint policies for defense in depth. The post says the setup keeps Quick traffic on the AWS backbone and is aimed at organizations facing HIPAA, FedRAMP or PCI-DSS requirements that sensitive data stay inside controlled network boundaries.
The AWS Solutions Architecture team published the first post in a three-part series describing Deep Insight, a production multi-agent system that takes a CSV file and a natural-language prompt and produces a professional .docx report with citations. The system runs a hierarchy of specialized agents — Coordinator, Planner, Supervisor, Coder, Validator, Reporter and Tracker — built on the Strands Agents SDK, Amazon Bedrock, Amazon Bedrock AgentCore, an Application Load Balancer and AWS Fargate, with Claude models assigned by role (Haiku, Opus and Sonnet). The authors say Deep Insight is open source and that a human-in-the-loop Plan Reviewer, which auto-approves after a 300-second countdown in the web version, addresses the problem that agents make context-based judgments rather than following fixed instructions. All performance and adoption claims come from the vendor authors and were not independently measured.
An AWS Machine Learning Blog post walks through the Adaptive Agentic Intelligence Framework (AAIF), built with Strands Agents, which uses a 1B-parameter llama3.2 orchestrator running locally through Ollama to classify incoming requests. Routine tasks such as device telemetry retrieval and camera-feed analysis stay on local small models, while complex analysis is sent to cloud frontier models through Amazon Bedrock. The post says the pattern builds on a research position from Belcak et al. at NVIDIA arguing small language models are sufficiently powerful and more economical for many agentic invocations, and uses an IoT monitoring scenario as its reference implementation.
An AWS Machine Learning Blog post walks through training Qwen3 on a single trn2.48xlarge Trainium instance using torchtitan, saying TorchNeuron registers Trainium as a native PyTorch device so torchtitan can use it through standard PyTorch interfaces. The post describes three configurations that share the same torchrun launch pattern and config structure: Hugging Face Transformers v5 as a modeling backend running Hub models through AutoModelForCausalLM with published safetensors, torchtitan's native Qwen3 selected by changing config.model_spec, and an NKI mixture-of-experts path using kernels from the torchtitan-neuron recipes attached via torchtitan's override registry. AWS says the setups support both pre-training from random initialization and fine-tuning published weights, and pins versions including torch 2.14.0, torch_neuronx 2.14.3, transformers 5.16.0 and torchtitan-neuron release-2.33.
NVIDIA announced an expanded open-model portfolio spanning language, retrieval, structured business data, robotics, autonomous driving, protein design and quantum computing. The lineup includes Nemotron 3.5 Super VL and Bolt, Nemotron Retriever, Kumo, Cosmos 3.5, Alpamayo 2 Edge, BioNeMo RFD4-Proteina and Ising, with availability through Hugging Face, GitHub, NVIDIA NIM and partner platforms. NVIDIA says these models can be customized with local data and deployed on infrastructure controlled by users.
NVIDIA announced the OpenShell Reference Architecture at GTC Berlin, a full-stack design for building, containing, observing and enforcing safe AI agents on enterprise and edge infrastructure, and said it is now available. It combines OpenShell secure runtime software with Vera Rubin systems for inference, Vera CPUs for agent harnesses and tool use, and BlueField-4 DPUs running NVIDIA Sentry for in-silicon enforcement, extending to RTX Spark, DGX Spark and DGX Station at the edge. NVIDIA says Vera Rubin NVL72 delivers 30x higher throughput per megawatt on agentic workloads and that Sentry stops an agent leaving its boundary in milliseconds while raising agent capacity by up to 20%; those figures are vendor claims not independently measured. Anthropic, AWS, Cisco, CrowdStrike, Dell, Google Cloud, HPE, IBM, Meta, Microsoft, OpenAI, Palo Alto Networks, SAP and ServiceNow are among companies NVIDIA lists as working with the architecture, and Sentry is planned for availability early next year, with DOCA libraries available now.
Anthropic publicly filed an S-1 registration statement with the SEC and plans to list Class A shares on Nasdaq under the ticker ANT. The filing reports first-half 2026 revenue of $16.286 billion, usage-based revenue of $13.064 billion, and a first quarterly operating profit of $974 million in the second quarter. A secondary analysis estimated compute commitments at about $1.01 trillion on an upper-bound basis, including contracts still under negotiation and potential additions. TechCrunch reported, citing an S-1 leaked to The Wall Street Journal, that Anthropic has had difficulty securing sufficient capacity from AWS, Google, and Microsoft, which increasingly view the Claude maker as a competitor. Anthropic did not immediately respond to TechCrunch’s request for comment.
What problem it solves
Developers using AI coding CLIs need a way to delegate work among multiple agents and coordinate their execution. CAO addresses this by running provider CLIs in isolated terminal sessions and giving a supervisor tools for coordinating workers.
Key features
Multi-agent delegation — A supervisor delegates work to specialist agents, with parallel and sequential execution suppor…
What problem it solves
AI-generated code can violate the rules its developers intended, especially when those rules are described in natural language and the resulting code is not reviewed closely. Bend addresses this by representing important rules as laws and requiring mathematical proofs when the code is edited.
Key features
Proof-enforced application laws — Applications can declare rules in `L…
What problem it solves
Agents can understand diagram content but often produce generic layouts that do not match a site's visual identity. Diagram Design addresses this by applying a defined editorial style to agent-generated diagrams and producing publishable, self-contained files without a manual design workflow.
Key features
Self-contained HTML and SVG output — Each diagram is written as one HT…
What problem it solves
Organizations need a distributed storage layer that can serve objects, files, tables, and vectors while keeping administration and horizontal expansion straightforward. SeaweedFS addresses this by providing these capabilities on shared storage and allowing capacity to grow by adding volume servers.
Key features
Unified distributed storage — A single cluster exposes S3 object…
What problem it solves
Modern systems need controlled access to database credentials, API keys, and other secrets across platform-specific infrastructure. Teams also need secure storage, key rotation, encryption, revocation, and detailed audit logs without building a custom solution.
Key features
Secure secret storage — Vault stores arbitrary key/value pairs and encrypts data before writing it to …
What problem it solves
VPN traffic can expose protocol characteristics that are examined during network inspection. The project addresses this by encrypting UDP tunnels, applying configurable traffic masks, and rotating session keys while providing bounded packet handling and live traffic-quality checks.
Key features
Encrypted and forward-secret sessions — The InitialV2 exchange confirms a fresh s…
What problem it solves
Leaked credentials can be difficult to identify and assess because they may be stored across repositories, chats, wikis, logs, filesystems, and other systems. Finding a credential pattern alone does not establish whether it is active, so the project combines discovery, classification, validation, and analysis.
Key features
Broad secret discovery — Scan sources including Git …
What problem it solves
Creating intelligent AI agents that can collaborate on complex tasks can require code and substantial integration work. Docker Agent addresses this by letting users define agents in YAML, assign tools, and run them through the Docker CLI.
Key features
Multi-agent orchestration — Create teams of specialized agents that delegate tasks automatically.
Source: README
Extensible t…
What problem it solves
AI coding agents need AWS-specific tools, knowledge, and guardrails to work effectively with AWS services. Teams using different coding agents also need a consistent way to configure AWS API access, guidance, and reusable skills.
Key features
AWS MCP Server — The managed MCP Server gives agents access to AWS APIs, isolated script execution, and current documentation through …
What problem it solves
Organizations need to build and maintain custom applications such as dashboards, admin panels, customer 360 tools, IT automations, and service management systems. This platform streamlines those application-development workflows while supporting databases and APIs.
Key features
Low-code custom application development — Build internal applications through a low-code platform …
What problem it solves
Ad-hoc AI coding can lose context as projects grow. AI-DLC keeps requirements, decisions, implementation, tests, and operational work connected through an audited lifecycle.
Key features
Harness-neutral workflow core — A single core runs across multiple coding-assistant harnesses, with thin harness-specific integrations.
Source: README
Structured delivery lifecycle — The wor…
What problem it solves
Teams need to capture, search, and analyze multiple observability signals without managing fragmented tools or complex infrastructure. OpenObserve addresses this with a unified platform, a single-binary deployment option, and familiar SQL and PromQL query languages.
Key features
Unified observability — Logs, metrics, traces, RUM, dashboards, alerts, incidents, pipelines, and…
What problem it solves
Developing against AWS services often requires a cloud account, authentication token, or paid feature gates. Floci provides AWS-shaped services locally so existing AWS SDKs, CLI commands, infrastructure-as-code workflows, and test suites can run without those dependencies.
Key features
Local AWS services without an account — Run AWS-compatible services locally without a clou…