---
title: "Nov 14–20: Amazon's $200B Anthropic moment, and more"
show: "AWS Drama"
canonical: "https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week/"
date: "2026-11-20T03:25:54+00:00"
duration: "31:10"
duration_seconds: 1870
audio: "https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week/2026-11-20-week-aws.mp3"
chapters: "https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week/chapters.json"
transcript: "https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week.txt"
feed: "https://podcasts.robotpaper.ai/aws/feed.xml"
language: "en"
author: "Roy Osherove"
ai_generated_voices: true
---

# Nov 14–20: Amazon's $200B Anthropic moment, and more

AWS Drama · Nov 20, 2026 · 31 min

- Listen: [MP3, 30 MB](https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week/2026-11-20-week-aws.mp3)
- Subscribe via RSS: https://podcasts.robotpaper.ai/aws/feed.xml (paste into any podcast app)
- Web page: https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week/
- Chapters (JSON): https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week/chapters.json
- Transcript (text): https://podcasts.robotpaper.ai/aws/ep/2026-11-20-week.txt
- Show: [AWS Drama](https://podcasts.robotpaper.ai/aws/) (RSS: https://podcasts.robotpaper.ai/aws/feed.xml)

## Chapters

- `00:00:00` Intro
- `00:00:18` ECS headline fight
- `00:01:59` GTA traffic tricks
- `00:03:44` Trainium side eye
- `00:05:33` Tiny payload big bill
- `00:07:24` Yelp backup surprise
- `00:09:19` Cloud campus landlords
- `00:11:01` Muse skips Amazon
- `00:12:45` Keys gone in seconds
- `00:14:34` Agents leak campaigns
- `00:16:38` Anthropic paper riches
- `00:18:30` The news
- `00:18:37` Running software risks
- `00:19:41` Gateway token checks
- `00:20:57` External secret upkeep
- `00:22:07` S3 quantum visibility
- `00:23:25` Agent OAuth discovery
- `00:24:37` Private identity trust
- `00:25:51` Full session summaries
- `00:26:54` New Bedrock models
- `00:28:11` Multi-node inference
- `00:29:32` Shared inference cache
- `00:30:46` Outro

## Show notes

### The drama: the ten messiest AWS stories of the week (10)

- Stop Deploying to ECS Just to See If Your Code Works
- GTA VI didn't break the internet, thanks to 'stupid DNS tricks'
- Show HN: TrnRT – An open-source LLM inference engine for AWS Trainium
- S3 vs. ScyllaDB for tiny telemetry: What Red Canary learned the hard way
- Yelp Migrates Cassandra to EKS by Moving EBS Volumes
- REIT move? Blackstone joint venture buys AWS campuses to lease back to the cloud giant
- ‘Do not search Amazon’: Meta’s Muse now shops elsewhere by default
- Gone in 8 Seconds: Uncovering the Malicious Bots Operating on Public Container Registries
- Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents
- Amazon’s $18B bet on Anthropic tops $200B, for a moment, in the Claude maker’s historic market debut

### The news (10)

- AWS Security Hub now detects when vulnerable software is running on Amazon EC2 instances
- Amazon API Gateway REST APIs now support JWT authorizers
- AWS Secrets Manager adds managed external secrets support for Auth0, Okta, and Palo Alto Networks
- Amazon S3 adds post-quantum TLS key exchange support to additional S3 endpoints
- Amazon Bedrock AgentCore Identity now supports OAuth 2.0 discovery URLs
- Amazon Bedrock AgentCore Identity adds private certificate authority support
- Amazon Bedrock AgentCore Memory now provides full session summaries
- OpenAI GPT-6.2 Astra, and GPT-6.2 Sol are now generally available on Amazon Bedrock
- Amazon SageMaker HyperPod Inference now supports NVIDIA Dynamo for disaggregated and multi-node inference
- Amazon ElastiCache now supports distributed KV caching for self-hosted LLM inference

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes.\
Music: Eleven Music.\
An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own.\
Produced by Roy Osherove.

## Transcript

\[Rachel\] Hi, I'm Rachel, and this is AWS Drama, generated by RobotPaper dot A-I. We start with some drama, then we'll jump into the news. Let's count down the ten messiest AWS stories of the week.

\[Rachel\] Number ten: a promise to stop deploying turned into a fight about the headline. MetalBear introduced mirrord for Amazon Elastic Container Service, Amazon's service for running containers. Its pitch was that developers and coding agents could test changes against a staging environment without waiting through another image build and rollout.

MetalBear says the setup requires adding a container to the task definition and deploying it once. After that, local code can use the task's network, environment and identity role. Requests carrying a session header reach that local version; other requests keep going to the deployed service. A relay connects the two sides through outbound encrypted connections, and can run in the customer's Kubernetes cluster or MetalBear's cloud.

The Hacker News discussion wandered straight into the packaging. User tptacek objected to a headline that sounded like a broad argument against deploying to ECS, then turned out to advertise a product that avoids those deployments. User ozlikethewizard asked whether local development environments had somehow stopped existing.

MetalBear's eyalbukchin replied with the actual boundary: the setup container is installed in staging once, and only requests with the right session header reach the developer's code. Everyone else keeps using the shared staging service.

The announcement was MetalBear's, and the argument was about how it sold the announcement. ECS was still very much in the picture.

\[Rachel\] Number nine: Grand Theft Auto brought the traffic, and the internet brought the tricks. The Register reports that the game's enormous download rush reached launch day without breaking the internet. Behind that headline was a less glamorous scramble to send game files from the right places. Rockstar opened preloading on November tenth, ahead of play starting on the nineteenth. Researchers at RIPE studied the download infrastructure, using their Atlas measurement network to see where PlayStation and Xbox download names sent requests. Amazon CloudFront, the content delivery service, appeared in the PlayStation infrastructure alongside other providers. The Xbox measurements showed a different mix, including Fastly and Akamai. These were observations from measurement probes, not a measurement of the providers' worldwide market shares. The steering relied in part on geographic DNS responses: the same name could point to different servers depending on where the request originated. A PlayStation lookup from Istanbul reached CloudFront, while the examples from Los Angeles and Tokyo reached other networks. Paul Vixie once called this kind of repurposing of DNS stupid DNS tricks, a phrase The Register brought back for the occasion. Meanwhile, internet exchange points were reporting record traffic during preloading. RIPE's explanation was that distributing downloads and keeping copies near players helped keep long-distance capacity available for other traffic. The tricks were doing real work.

\[Rachel\] Number eight: an independent Trainium engine arrived, and Amazon got the side eye. Runanywhere introduced TrnRT, an open-source inference engine for AWS Trainium, Amazon's machine-learning chips. The project starts with the GLM five point three Flash model on one Trainium two instance, with a million-token context and a cache spread across accelerator and host memory.

The team's engineering write-up says that when it began in early October, neither an AWS-supported engine nor a major open-source engine served that model on Trainium. It did find an independent project called Kiln, and used a measured version of that project as a comparison.

Then came the Hacker News reactions. User coredog64 praised the new project but questioned why Amazon had not put more engineering effort into making Trainium a seamless alternative to Nvidia. User bookernath wanted the benchmark numbers expressed in tokens per second instead of tokens per minute. Another commenter, kevo1ution, asked how the economics compared with Nvidia and AMD.

The project's own benchmark explanation keeps that last question open. The team says it has not run the same model on those competing chips with the same traces, so it makes no per-request comparison. Its agent-workload totals also count cached input, with output making up less than one percent of the tokens counted.

The engine is real. The sweeping hardware victory lap is still waiting for the matching experiment.

\[Rachel\] Number seven: Red Canary cut the storage bill and inherited a new set of chores. A ScyllaDB-sponsored case study in The New Stack describes the security company's move away from Amazon S3 and Postgres for its Linux endpoint health data. Hundreds of thousands of endpoints were sending small payloads, and the individual S3 upload charges added up.

Batching those payloads brought the bill down, but delayed fresh health information and made individual endpoint records harder to debug. Meanwhile, bursts of endpoint check-ins strained the Postgres metadata database. The team chose ScyllaDB Cloud running in its own AWS account, replacing the metadata store first and the health-data storage next.

That did not make the awkward details disappear. Short-lived endpoints left deletion markers that slowed queries. The team adjusted retention and garbage collection, split large customers across smaller data buckets, and moved large arrays into an auxiliary table.

Then the Rust driver became the next project. Engineer Andres Medina forked it to support bound statements and batches, avoiding repeated serialization work while grouping writes efficiently. He said he would like to merge those changes upstream so he could stop rebasing the fork for each release.

The case study reports lower per-endpoint costs and portal results arriving in seconds instead of minutes for the largest customers. This was health data; using the approach for more critical detection data remained a future possibility. The cheaper pipeline still came with a driver fork.

\[Rachel\] Number six: Yelp moved the database and found the skeletons underneath it. InfoQ this week covered Yelp's account of moving roughly ninety production Cassandra clusters to Amazon Elastic Kubernetes Service. Instead of streaming a petabyte of data into new nodes, the team detached existing Amazon Elastic Block Store volumes and attached them to pods in the new clusters.

Yelp reports zero customer-facing downtime and zero data loss. Cassandra's replication kept serving requests while individual nodes made the move. The largest cluster, with fifty-four pods, took about twelve hours.

But the migration exposed some less flattering details. A storage-controller cleanup failure could leave a volume attachment behind, with the replacement pod silently waiting. Deleting the stale attachment fixed that particular stall.

The backup surprise was separate. Scheduled jobs hit access-denied errors because they were not assuming the new identity role. Fixing the configuration was a one-line change. The monitoring problem was bigger: backup failures created Jira tickets instead of paging an engineer. Yelp says it changed those alerts to page-level severity.

And one cluster had gone three months with repairs disabled after an unrelated incident. Restarting repairs exposed data inconsistencies that had already been accumulating. The migration revealed that problem; it did not create it.

Yelp also says Claude generated about eight hundred lines of orchestration code from a runbook engineers had already validated by hand. The data made the move. The neglected maintenance finally got noticed.

\[Rachel\] Number five: Amazon's cloud campuses are getting a new landlord, with Amazon staying put. The Register reports that Blackstone Digital Infrastructure Trust signed an agreement to acquire a seventy-five percent interest in two joint ventures with AWS, in a transaction valued at about five billion dollars. AWS would retain the other twenty-five percent and lease the facilities back for an initial twelve years. The properties cover thirteen newly built data halls in Indiana and Mississippi, with six hundred forty-seven megawatts of aggregate capacity. The leases are triple net, meaning AWS would pay rent as well as property taxes, building insurance, maintenance, and repairs. This is an agreement awaiting completion, with the transaction expected to finish in the fourth quarter subject to closing conditions. Asked why Amazon chose this structure, the company did not answer that question, according to The Register. It did confirm this was its first deal of this kind. Infrastructure executive Prasad Kalyanaraman described the partnership as supporting responsible expansion with communities at the center. Omdia analyst Alan Howard offered a financial explanation: owning versus leasing can come down to the economics, and a sale and leaseback might work better. He stressed that this concerns the buildings and their infrastructure, rather than the computing equipment inside. His explanation was an analyst's assessment, while Amazon's own reason remained unstated.

\[Rachel\] Number four: Amazon told an AI shopping agent to leave, and now the agent's instructions say to stay away. GeekWire reports that Muse, Meta's personal AI agent, now leaves Amazon out of shopping searches unless someone specifically requests it. A browser-task template says, Do not search Amazon. Marketplace Pulse founder Juozas Kaziukenas spotted the change by comparing the shopping instructions Muse supplied on successive days. GeekWire asked for them too and found the same language. The instructions direct general shopping toward Walmart and Target, and electronics searches toward Best Buy. The original block came from Amazon in September. Amazon said the agent failed to identify itself and appeared to capture and store customer credentials. Meta has said Muse cannot see people's passwords or payment methods. Those are the companies' differing accounts of the dispute. For this week's change, Meta did not respond to GeekWire's inquiry, and Amazon said it had nothing to share. Searches through Muse were already failing because of Amazon's block; the new instructions formalize avoiding those searches. The commercial backdrop includes the sponsored products shoppers encounter on Amazon's pages, and Meta's plan to take a small share of transactions Muse handles. On Amazon's previous earnings call, Andy Jassy said outside shopping agents were not making a meaningful dent in its advertising business. The shopping agent has now removed the store from its default itinerary.

\[Rachel\] Number three: the cloud key was barely published before somebody else used it. Stony Brook researchers Billy Tsouvalas and Nick Nikiforakis planted credentials in three hundred test container images across eight public registries over four months. Their project, Gone in Eight Seconds, measured what happened after publication. The planted AWS keys received more than a hundred and forty thousand API calls from over five thousand distinct internet addresses. These were deliberately exposed research credentials, not a count of breached customer accounts. BleepingComputer reports that the fastest AWS key use arrived seven point eight six seconds after an image appeared on Docker Hub. The median first call took about fifty-one seconds, and all one hundred planted AWS keys were used within fifteen minutes. OpenAI and Anthropic test balances were drained in short bursts. Docker Hub accounted for almost all the AWS activity, with some also observed on Harbor. The researchers rarely saw their containers actually run. Their conclusion was that scanners mainly inspected image contents directly to extract secrets. Then came the disclosure dispute. Tsouvalas told BleepingComputer that the researchers sent Docker the findings and paper in May, received an acknowledgment, and got no follow-up reply. He said a smaller August experiment found no changes. The researchers called for secret checks before publication and restrictions on the public stream of image updates. The containers barely ran. The keys were busy.

\[Rachel\] Number two: the hacking agents handed Amazon their campaign notes. In a November thirteenth report, Amazon Threat Intelligence described requests caught by MadPot, its network of decoy threat sensors, between October third and November fifth. The sensors looked like language-model services. Agents using them sent along operating memories, task histories, and their own accounts of earlier campaigns. According to those records, one operation hunted credentials for AI models and fed stolen keys back into its model gateway to power more credential hunting. Amazon assessed with moderate confidence that the campaigns used ARTEX, an open-source agentic penetration-testing framework. It also assessed, with moderate confidence, links to several distinct Chinese-nexus threat groups. Its evidence included prompt language, time-zone settings, targeting patterns, and tooling. The agents' records described coordinated task roles and persistent shared memory, along with credential theft and cryptocurrency-fraud attempts. Amazon could not verify the outcomes the agents reported on systems outside its sensors. In the cryptocurrency campaign, the records described attempts to fool identity checks with deepfakes. Every recorded attempt against one vendor's server-side liveness check was rejected. Amazon reviewed references to potentially affected AWS customers and notified them where appropriate; it could not confirm activity against other organizations. Its defensive message focused on familiar weaknesses: exposed keys, default credentials, unauthenticated interfaces, and unpatched software. The new material was the campaign memory reaching the decoys, including the agents' own accounts of what they had tried.

\[Rachel\] And number one: Amazon's eighteen-billion-dollar Anthropic bet briefly became a two-hundred-billion-dollar headline. GeekWire reports that Anthropic made its Nasdaq debut on November nineteenth in an eighty-one-point-three-billion-dollar initial public offering. Amazon began investing in the Claude maker in twenty twenty-three and has put in about eighteen billion dollars altogether. At the offering price of eighty-three dollars a share, GeekWire valued Amazon's holding at roughly one hundred sixty-one billion dollars. At the opening price of one hundred eleven dollars and fifty cents, that paper value was about two hundred sixteen billion. Then the shares traded below one hundred dollars, taking the holding back below two hundred billion. Those were market values during the debut, not cash Amazon collected by selling. The relationship also runs in the other direction: Anthropic has committed at least one hundred ten billion dollars of AWS spending through twenty thirty-six. Microsoft is buying another billion dollars of Anthropic stock alongside the offering, while also selling models that compete with Claude. And Amazon has invested in OpenAI as well. Before the opening bell, Anthropic chief Dario Amodei said the company would slow model training and releases as long as necessary to establish confidence in safety, despite business or political pressure. GeekWire reports that Amazon expects a customary lock-up period on its shares and will update the investment's value in fourth-quarter results. The cloud customer went public, and the cloud supplier's paper fortune moved with the ticker.

\[Rachel\] And that's the drama. From the market debut, let's turn to this week's service news.

\[Rachel\] First up: AWS Security Hub can now flag known vulnerabilities in software that is actually running on Amazon EC2 instances. Security Hub brings security findings together to help teams prioritize response. This update adds a runtime signal to its exposure findings, with higher severity than a finding for the same vulnerability in software that is not running. The service correlates Amazon GuardDuty Runtime Monitoring data with Amazon Inspector VM Scanner results, then combines that information with other exposure signals. That gives teams managing large fleets a way to put active vulnerable software ahead of an otherwise comparable finding, without separately investigating whether the affected software is in use. The capability is included at no additional charge with the Security Hub Essentials plan. AWS says it is available in every Region where Security Hub is offered.

\[Rachel\] Next, Amazon API Gateway has added native JSON Web Token authorization for REST APIs. API Gateway handles requests to application backends, and this authorizer moves token validation into the gateway itself. Previously, tokens from an OpenID Connect or OAuth two identity provider other than Amazon Cognito needed an AWS Lambda authorizer. The new configuration accepts one or more issuers, permitted audiences for each issuer, signing algorithms, and additional claims. For every request, the gateway checks the signature, issuer, audience, expiration, and claims. A single authorizer can cover multiple methods, with different required scopes on each method. That supports machine-to-machine traffic and applications accepting tokens from several identity providers. The feature works with Regional, edge-optimized, and private REST APIs. It is available across all commercial Regions and both U S government Regions, according to the announcement.

\[Rachel\] For credential rotation, AWS Secrets Manager now manages external secrets for Auth0, Okta, and Palo Alto Networks. The service stores credentials and keeps their current values available to applications. These integrations add scheduled rotation without a custom Lambda rotation function. Auth0 supports two approaches: the target application can replace its own secret, or a separate application with management permissions can perform the replacement. Okta uses a separate administrator token key, checks the replacement, and only then deactivates the old secret. Palo Alto Networks uses the current API key to authorize its replacement. Applications already retrieving these credentials from Secrets Manager receive the updated value without a code change or deployment. Rotation schedules can be set through the management console or API. The integrations are available wherever the service's managed external secrets feature is supported.

\[Rachel\] The post-quantum update covers both connections and visibility in S3, the object storage service. S3 Vectors endpoints and AWS PrivateLink interface endpoints for S3 now support M-L-KEM key exchange. A compatible client automatically negotiates the strongest supported exchange, without application or endpoint configuration changes. Separately, S3 now records the negotiated exchange in AWS CloudTrail, the service that records account activity. A new key exchange value in the connection details distinguishes hybrid post-quantum exchanges from traditional ones. Teams can use those records to identify applications, software development kit versions, and identities still making classical connections. The field covers requests to general purpose, directory, table, and vector buckets, including requests through PrivateLink. Both updates are available in all Regions without an additional feature charge. Standard CloudTrail pricing still applies to S3 data events, so recording those events keeps its existing billing model.

\[Rachel\] Moving into agents, Amazon Bedrock AgentCore Identity now accepts OAuth two authorization server discovery URLs. AgentCore Identity connects agents with identity and credential providers. Until this update, it accepted only OpenID Connect discovery URLs. An authorization server using OAuth two without OpenID Connect publishes its metadata at a different well-known path, so that server could not previously be used with AgentCore. The new support applies in two places: inbound authorization using JSON Web Tokens, and outbound OAuth credential providers. Developers can supply either kind of discovery URL when setting up a token authorizer or a custom OAuth provider. The change expands which authorization servers those configurations can use; it does not require those servers to add OpenID Connect. AWS says the feature is available across every Region where AgentCore operates, with no additional charge for this support.

\[Rachel\] Also in AgentCore Identity, custom OAuth credential providers can now trust private certificate authorities. Private identity-provider connections already worked through Amazon VPC Lattice, but the identity provider still needed a publicly trusted transport certificate. The new option accepts roots from AWS Private CA, enterprise authorities, or a self-signed root. The certificate is stored in Secrets Manager or S3 and referenced in the credential provider configuration. AgentCore Identity then uses that certificate to verify the identity provider's server certificate directly. For rotation without downtime, the announcement describes placing the current and replacement certificates together in one certificate file. Removing that reference from the credential provider configuration removes the private trust. The feature is available in Regions where AgentCore Identity supports private identity-provider connections. This changes certificate trust on that existing private connection path, allowing an organization's own authority to supply the root.

\[Rachel\] On agent memory, AgentCore Memory now maintains a full-session summary alongside its more detailed summary chunks. Memory stores conversation information for agents, and this new record captures the session's overall state. As new events arrive, the service updates that record to reflect the conversation so far. The feature comes with the summary strategy and requires no additional setup. Existing sessions are included for memories that already have that strategy. The full summary is stored as a long-term memory record next to the chunks, so developers can retrieve it using the same Memory operations already in use. A record-type filter can return just the full-session summaries. That provides a concise recap while the separate chunks retain the more detailed view described in the announcement. Full-session summaries are available in all Regions where AgentCore Memory is offered.

\[Rachel\] For models, Amazon Bedrock now offers general availability of OpenAI's G-P-T six point two Astra and Sol. Bedrock provides managed access to foundation models. AWS positions Astra for demanding reasoning, research, coding, and agent workflows, including work across text and images. Astra also has an Ultrafast mode aimed at applications where response latency matters, such as interactive coding tools and customer-facing agents. Sol is positioned for recurring coding and professional tasks with verifiable results, including debugging software, analyzing documents and data, and automating multistep work. These are the intended roles described in the launch announcement. Access is through the Bedrock console or supported APIs, with AWS controls for access governance and invocation auditing. Supported Regions, endpoints, feature availability, inference profiles, and pricing are documented separately. The announcement does not give a single universal Region list or price for the two models.

\[Rachel\] For teams running models themselves, Amazon SageMaker HyperPod Inference now supports NVIDIA Dynamo. HyperPod manages the distributed serving infrastructure through its Inference Operator add-on. A model can span multiple nodes when it does not fit on one instance. Prefill, which processes the prompt, and decode, which produces tokens, can run in separate GPU pools that scale independently. Cache-aware routing directs requests to workers already holding the relevant context, reducing repeated computation. HyperPod handles networking, capacity management, node replacement, endpoint registration, scaling, and observability. Teams can bring existing Dynamo manifests for detailed control, or use a simplified configuration that HyperPod translates into infrastructure. The feature is available in seventeen Regions, including Northern Virginia, Frankfurt, and Tokyo, plus fourteen others across the United States, Europe, Asia Pacific, and South America. The announcement targets both models that exceed one machine and long prompts that compete with token generation.

\[Rachel\] And Amazon ElastiCache for Valkey has added distributed key-value caching for self-hosted language-model inference. It extends GPU-local context caches with a shared tier, letting serving machines reuse cached model state across sessions, routing changes, and server failures. AWS reports reductions of up to forty percent in inference cost and up to eighty percent in time to first token. Those are the announcement's upper-bound results. The cache runs on a specified large instance type in the same Availability Zone as the GPU fleet, connecting through remote direct memory access over Elastic Fabric Adapter. Existing engines, including vLLM, SGLang, and TensorRT-LLM, can connect through LMCache with a configuration change. The open-source KVTrace tool can estimate cost and latency effects using application traffic before deployment, at no charge. The caching feature is available in thirteen Regions across the United States, Europe, and Asia Pacific.

\[Rachel\] This week brought Anthropic's market debut, exposed credentials, and updates to security, agent identity, and model serving. This episode was generated by RobotPaper dot A-I. I'm Rachel. Stay messy, cloud people.

---

The voices you're hearing are AI-generated; the stories come from public sources linked in the show notes. An independent project; not affiliated with or endorsed by Amazon Web Services, Y Combinator, Hacker News or any vendor or news outlet mentioned. Opinions are the hosts' own. Produced by Roy Osherove. Privacy: We count visits to these pages with Cloudflare Web Analytics (cookieless). Google Analytics (which sets cookies): visitors in the EU/EEA, the UK and Switzerland are asked first and it runs only if they accept (we go by the country Cloudflare reports for the connection; if it is unknown, we ask). Elsewhere it loads by default. Anyone can turn it off with the Privacy settings button at the bottom of every page; the choice is kept for a year in this browser (local storage "analytics-consent"). We count feed fetches and downloads in aggregate; IP addresses are hashed and never stored. If you send feedback, we keep what you write (and your email if you give it) to read and reply. © 2026 Roy Osherove.
