AWS Well-Architected Agent, Amazon's $1B datacenter plan, and Oracle

This transcript was generated automatically and has not been fully reviewed. Timestamps are approximate and wording may contain errors.

00:07 Welcome to The Cloud Pod, where the forecast is always cloudy. We talk weekly about all things AWS, GCP, and Azure. We are your hosts, Justin, Jonathan, Ryan, and Matt.

00:18 Episode 375, recorded for October 5th, 2026. AWS finally automates the blame game. Good evening, Ryan and Matt. How are you guys?

00:29 I didn't do it.

00:33 Quickest way to know that Matt actually did it is his denial.

00:36 Yeah, he, he admits it quickly when he doesn't do it or does do it. So go for it. How'd you guys go when I was in Limerick? I haven't listened to that episode yet. I've been a bit busy.

00:47 You should skip it.

00:48 Best episode ever. I don't know what you're talking about.

00:52 So one person says skip it, everyone's like the best ever. Awesome. Good.

00:55 That's all we did during the show too last week. We just, it was like counterpoint, point. I don't know. Wasn't there a Stephen Colbert skit like back in the day?

01:05 I think there was, yeah. Well, I, my trip ended up, you know, I went to our, like my day job's data center. Then I went to our DR site in London, which happens to be at an Equinix facility. And next to the Equinix facility, Azure is building a brand new 40-megawatt data center that is massive. It towers over the Equinix data center. Wow. It's like 5 stories, you know, has all these generators surrounding it. You know, it's crazy. And I was like, yeah, I guess, uh, you know, Azure might get some capacity finally. And, and,

01:39 uh, UK South is rough region to get capacity in. So they, all of the Europe

01:44 is rough. They just actually over the weekend, they haven't posted a blog post. We don't, we don't talk about it as main news, but, uh, they are actually announcing price increases on existing v1 through v4 instances in Europe from 8% to up to 20% in some cases. So, wow, they were sending an email out to impacted customers over the weekend. I'm waiting for an official statement from them and we'll talk about it more. But yeah, so yeah, capacity in EU, big problem.

02:11 Wow. And they probably about A year and a half ago, we tried to move up to v5, but they didn't have capacity everywhere for it. So you couldn't move your scale sets up to the v5, which was annoying too, 'cause they only had it in like one of the three zones. So maybe they hopefully fix that too.

02:30 Uh, maybe we'll see. I mean, I don't know what this new data center's gonna hold, but this is just one of two they're building. They're building another one in a different city. Uh, and that one apparently is an AI supercomputer they're building. That one is like, I think 45 megawatts or something like that. Just massive size with these things.

02:46 That's crazy.

02:46 It might be gigawatts even. I may be wrong on the megawatt versus gigawatt part, but just huge, huge. So there's a little bit of, I was like, oh, I'm outta work, but that's a cloud story right there at that data center. That's cool. So it was neat to see. All right, well, first up, if you are anti-data center, which many Americans are these days and Europeans and pretty much any human who doesn't work in technology and is not an executive who's gonna replace all the jobs of AI, is very unhappy with, you know, data centers. And so Amazon has decided to commit $1 billion into data center community opposition work, suggesting AWS is funding efforts to address or mitigate local pushback against new data center construction projects. Although what they're going to do exactly is not fully clear. This is an overall issue that we're seeing with hyperscalers, including AWS, Microsoft, and Google, facing increased local resistance to data center buildouts due to concerns over power consumption, water usage, and land use. Similar to the Oracle issue, uh, in Wisconsin. So yeah, uh, that's the thing. And so they decided to put $1 billion towards it, and then of course the reaction to that was, are you kidding me? So, uh, Ars Technica then covered, uh, you know, basically that, you know, the 100— this comes after they've announced $130 billion in data center projects that have faced blockages from local protests this year, such as the funding commitment as a response to mounting community opposition. AWS CEO Matt Garman framed data center in more terms as national security risk, arguing the slowing US data center growth could let China surpass American AI development. And critics include Stand.earth note the commitments don't address Amazon's backing of a large gas power plant expected to become a significant source of U.S. climate pollution. And I guess they— if you're about that climate pledge arena, they paid for the naming rights too. Yeah, yeah.

04:30 I mean, I would be against this if it wasn't for all the misinformation about data center pollution and, and how it's tied to AI and not separated from general compute costs. And like people, I feel like it's not really fair the other way. Like, all the backlash against the data centers. And so I, you know, I don't think Amazon has any other choice other than to invest hopefully in the environmental impacts of these data centers and, uh, to, you know, dispel the misinformation. Because it's— there's some truth to a lot of data center providers breaking rules and, and using temporary generators, do pollution. But there's a lot of, you know, big hyperscalers that are trying to do the right thing and have invested in, you know, a lot of water reuse and reclamation and closed-cycle systems. And then also, you know, trying to do their best to invest in clean energy as well. But, you know, it's one of those things where I think each individual data center build in the local region, that's important because there might be local concerns. But

05:36 I mean, the reality is if these data centers don't have capacity in the cloud, then everyone has to go back and buy more data centers in colo and run their own hardware, which then will jack up the price of hardware even more than it already is. Like, you know, like it just— everything ripples out from this at this point in our society. So yeah, I don't know that— I do think there are things that data centers do that is not great. Like the, uh, you know, Virginia— there's a story I think earlier last year where it was in maybe Wired magazine. They were Talking to a couple who lived outside of a data center that, you know, has blaring lights, uh, that, you know, basically, you know, are on all day long and there's a constant hum they can hear. Again, these are things you can mitigate with proper design. And, you know, do you really need LED lights in the middle of the night in the middle of nowhere? No, probably not. You could probably get away with night vision. You could probably do other things. So like there's, there's things that they could do that would be beneficial. And then the water consumption is really the one that I think a lot of environmentalists are jumping on. Which, you know, this Amazon claims this data center or some of their data centers are 74%, uh, meeting water positive targets, whatever that means, because there's no way to quantify what water positive means. But, um, you know, it's, it's just interesting because I think the, the pushback is real because of some of the policies these data centers run, but there's also a lot of FUD out there too.

06:54 There's a lot of FUD, and like, by, uh, in Florida, they were trying to open a data center and it got shut down because everyone was concerned about the water. And interesting there is that they were going to take it from the aquifer below. So the concern wasn't about the water consumption, but the, but the actual stability of the ground too. And like Ryan said, you got to look at each data center, not, you can't look at all these things together, but you got to look at kind of each data center by itself and see what the effects are. 'Cause in Florida they're worried about the land dropping like, uh, like New Orleans has over the years, you know?

07:32 Yeah.

07:33 But in Florida, if you drop the land down, then you know, that becomes a lake because it's at sea level already. Yeah. So water, it comes in. So like, will it affect the neighbor's houses and everything else? Like each place kind of has its own unique issues that you have to take into consideration.

07:48 Yeah. And there's a lot of abuse too. Like we've, I, I don't know if we talked about it like last week's show, but there was Grok who was,, uh, basically breaking a whole bunch of data, you know, local rules for using temporary stuff.

07:58 Yeah. 35 generators in the middle of nowhere without permits. Yeah.

08:01 There's, yeah. And there was one on the show last, last week where they, you know, talking about they're using drone technology and infrared and they caught, caught a whole bunch, you know, like caught a provider, you know, with a whole bunch of—

08:12 I think it was CodeWeave using like 130 generators. We could ask Boltz live, but yeah.

08:17 Yeah. I mean, I mean, that's just Elon Musk though, who thinks laws don't apply to him. So it's just,

08:23 you know, well, last week wasn't, it wasn't Elon, it was, it was some, uh, data center. I don't think it was Purview, but it was, uh, some provider.

08:30 Okay.

08:31 Not one of the mega, mega scalers.

08:34 Well, I mean, if, if the reality is you have to ruin the environment with data centers to, uh, win AI race, maybe we should lose it. So, right. We'll see. Right. Everything ends up, keep an eye on the story cuz yeah, EU capacity is a big issue. You know, moratoriums are there. We're seeing moratoriums being passed in the US and It's going to get worse before it gets better, is my guess. Well, on AI, it's how machine learning makes, makes money. OpenAI hosted their annual Dev Day 2026, and they announced a bunch of things. First up was DOTS, their always-on agents available on the Pro and Business Premium plan. These will handle ongoing tasks and responsibilities autonomously, representing a shift towards persistent agent-based workflows rather than single-session interactions. They announced GPT-6.1 SOL, as an upgrade with strong agent decoding and computer use performance, priced at 1/5 the input and output token cost of the standard tier, which gives developers more budget flexibility for complex workflows. And a new ultra-fast tier also offers up to 8x faster token generation in Codex, reaching 300 tokens per second. The Agents API now supports computer use, multi-agent capabilities, and context compaction. And OpenAI partnered with Amazon so developers can build these agents natively within AWS through Bedrock Managed Agents. OpenAI, uh, OpenChatGPT is a developer platform with plugin extensions letting third parties build native experiences within ChatGPT's interface across a reported 1.2 billion weekly users. And OVN.ai Private Intelligence adds zero data retention and private safety processing for enterprise customers with a private inference preview combining confidential computing and compute coming this fall. I mean, didn't they already have zero data retention? It seems like one of the big enterprise things that was early on everyone wanted in their enterprise agreements. So the fact that they're announcing this as new makes me a little nervous that they were lying before.

10:16 Well, I think it's actually zero data retention on now the agent workloads that are running on OpenAPI.

10:22 Oh, I see. Specifically to the agent workloads. Okay. That makes sense. Yeah.

10:27 I feel like this is like déjà vu. We have containers and we have everything being stateless. Okay. Now we have slowly more and more stateful things, right? We, We do single sessions, we don't keep anything. Okay, now we wanna keep more and more context as we go through things.

10:41 Yeah.

10:41 It's like the world just repeats itself as you think about these things.

10:44 Yeah. I mean, it is interesting. I'm, I'm still very conflicted about these, all of these, you know, whether it be DOTS, whether it be Meta AI's thing Muse, or like even I think Anthropic has some version of this where they're basically moving agent execution into the Anthropic's cloud.

11:01 Yeah, I saw— I don't think that's been initially announced.

11:05 I don't think it's out yet. I think it's been announced.

11:08 Yeah, but they're basically removing all the code processing from your local computer to their cloud, which I'm like, why would you want to take on that compute overhead that you have distributed to the endpoints to do this? But yeah, it's interesting that decision. Everyone's kind of going back to move it back to the cloud. I'm like, we don't have the capacity. I don't get it.

11:25 Yeah, well, and it's the security play that I don't think adds to any security, right? Like it's not running on my laptop.

11:31 Well, but now it's running in some arbitrary third-party zone. You know, is it protected now? What are your security policies? Are you protecting it with WAF and all my security controls and access management? And is that agent now have access to my enterprise apps internally to my— like there's just so many complications and VPN connectivity to internal apps. Like I just, I don't see what they're trying to do.

11:53 I think it's, you know, I don't think it's particularly enterprise, you know, like pointing, just because I do, I agree with you, like the VPN and the network capability. I think it's for people doing all these like random, you know, Instagram coding loops things that they read about. And like, I had AI run my business and make an AI influencer. Now I just sit back and collect money. And I think it's, that's really the market that they're going for. And I don't think there's a lot enough or a lot of security around those. And I don't, you know, like, I'm not real sure. But

12:25 yeah, not either. I

12:26 mean, I, I, since it can, and it can connect back to your local file system with like, I think it just probably creates an outbound session from the client too. So it's sort of this weird, like you still have a lot of security concerns there.

12:35 Well, I, I, I imagine it's them trying to pick their IP as well, right? Because everyone seems to be in this like concern that, well, if you have things running locally, it could be reverse engineered by your competitors. And so they're, they're taking these choices that are, are actively user hostile because they're trying to protect some IP they think is going to be saved in that thing. And you're just seeing it in all kinds of use cases. Like Reddit this last week, you know, basically killed RSS feeds, uh, for subreddits because they're saying, oh, well, they're being inadvertently, you know, scraped by AI tools and getting around our policies to make money. And I'm like, hey, but now you're, now you're impacting me and my RSS reader, which, you know, like, I track Google and AWS and Azure's Reddits that way because I don't want to go to those Reddits on a daily basis, but I, I'd like to see what's there. So I feed it to my RSS reader and I go through it occasionally, but like this arms race is starting to become actively user hostile.

13:29 Yeah. To non-AI use cases. Right. Which is the worst.

13:34 Yeah. Even, even like Amazon not putting in, uh, in emails they send you anymore that, you know, the item is being delivered today. All it says is a package is being delivered and you're like, what package? And you're like, oh, don't know. You had to click the button. Now go back to Amazon because they were worried Google, and they may be was, you know, was scraping Gmail boxes to see what people were buying and targeting pricing

13:55 data. Yeah, exactly. And it's just like, I'm not clicking a link in an email either because that's a phishing thing. So no.

14:02 Yeah, no, no, just packages show up randomly and I have no idea what they are.

14:07 Yeah, you get a household item arrived. What is that? What did my wife buy today? I don't know.

14:14 A household item, that's all you need to know. Yeah, I mean, I get just— but it's, it's so frustrating because it's like, okay, I get I see what your concern is, but like, you're, you're now hurting me as the end user, and that's just like, it's— we're in this kind of— I think we're starting to enter kind of the really sucky part of AI adoption where everyone's trying to build, you know, a moat around their IP and their protection and these things, and I think it's gonna be a while before it corrects.

14:40 Yeah, agreed.

14:43 Um, all right, well, if you, uh, you'd like some of those features that you got around plugin extensions for ChatGPT, you can also get similar capabilities now with Anthropic is offering ability to implement small TypeScript functions that can rewrite prompts, add UI elements, and replace built-in features or add new functionality, uh, going beyond existing hooks by allowing event rewriting and custom UI rendering. So if you'd like to really break your Claude client, you can do that with JavaScript. Uh, the mods themselves ship inside plugins and work across both the CLI and the desktop app, and users can write mods manually or ask Claude Code Generate and hot reload them during a session. Uh, security is a key consideration since mods run unsandboxed with in the same machine access the Cloud Run itself. And for proper runs, only installing models from trusted sources similar to any other code. And I went into my Cloud Enterprise and I disabled this feature. 'Cause I said, nope, not worth it.

15:31 Yeah. And how do you know you're gonna break whatever cloud experience you have? Because it's first thing I did when I tried to do browser rewrites using JavaScript.

15:39 Uh-huh. Uh-huh. How many times did I crash back in the day trying to learn how to write JavaScript properly for iRules? Yeah, no, I know how this goes. Yeah.

15:49 I'm not sure that like, I understand, fully understand the use case. Like they talked about the rebuilding the diff thing. And so like, I'm sure there's some graphical viewer enhancements.

15:57 I'm sure there's some like super, super edge cases that make sense in this one specific edge case. You can now do this thing. And I'm sure that's valuable, but I just, you know, this is not a, this is not a feature to be turned on by default in your organization without understanding exactly why you think you need it.

16:14 Yeah. And I love the functionality of rewriting the prompts. Like, what?

16:19 See, I'm concerned about the opposite. The sub-security or sub-IT person's gonna write something and push it down to everyone in their company. Yeah, I'm gonna rewrite a prompt or like override a default thing, and all of a sudden Claude's gonna be doing crazy shit on my laptop, and I'll be like, what is going on?

16:39 Oh yeah, let me tell you about the fun I experienced the day job, like with managed settings and pushing them out to all the clients and making sure everyone's secure. And trying to manage all the different, you know, Windows and macOS versions and make sure we got all the things covered, it's not any fun.

16:55 Let me rephrase that. I'm worried about what crazy shit Ryan's gonna try to do and push out to my laptop.

17:00 Correct, sir.

17:01 That's the correct answer. Yeah. Yeah. Ryan, you know, he was like, I wanna make security not be so obtrusive. Then he goes to the dark side and he is like, security's obtrusive. This is what we do.

17:11 I still wanna make it less intrusive. But there's some things that are terrible. Link, yeah, no, it's a good balance.

17:19 If you deploy a DNS black hole in the future enterprise-wide, I will come hunt you down.

17:24 I will not do that. I will never do that. I will never, like, in terms of like split DNS, I will reject certain queries. But yeah, no, I don't believe in that. But I have fallen on one of my things that I definitely didn't think I would do, which is SSL decryption. And it's just the amount of rules and protection. Like, I don't know how to do it from, for an enterprise without inspection, inspecting the network traffic. And I don't like it. I don't, it's, it's just an operational nightmare and I don't like maintaining it. And there's, thankfully there's pretty good products that do it, but I just don't know how to keep people safe without being able to write rules and policy that can't work against encrypted payloads.

18:06 I mean, you can just trust people.

18:08 Yeah, yeah. Okay. I mean, if you know, honestly, I would do that if we could hold them accountable.

18:16 Yeah, yeah. The reality is like we, we see in phishing reports, uh, you know, when you do phishing campaigns against your user populations, that there's still a pretty high percentage of people who still click. And, you know, every one of those failed phishing tests is a potential ransomware attack against your enterprise. And so, you know, you need— I get it, it's It's a scary place.

18:35 Well, and you, you, you start looking at things and you see how many like people are using unsanctioned apps. People are doing things that they've been explicitly told not to do. And it's, you know, like when you— it is sort of tricky 'cause it's like you can trust people, but there's a lot of evidence that's like, no, you can't. And you know, if we could hold people accountable, that's a different thing. But there's too many laws and restrictions around that. Not enough creativity. So that's, I mean, that's a problem.

19:03 Until there's real consequences, like I've never seen anyone fired or reprimanded for having uninstalled apps or doing anything that they aren't, that they're not allowed to do or explicitly told multiple times. You know, essentially all it ends up being is, you know, we're gonna tell your boss, we're gonna tell your boss's boss until eventually the person just relinquishes 'cause they don't wanna deal with it. It's not like they're going to— you're going to be able to say, do this again and you're fired, because no offense, Ryan, what control do you have over the situation?

19:34 Right. Well, and I don't know how many meetings I've been in where I'm trying to figure out and work with people where they're like, well, I didn't know, it wasn't blocked, so I assumed it was allowed. Like,

19:44 okay, well, that's the ask for permission, not for forgiveness. Ask for forgiveness, not for permission strategy, which I'm not gonna lie, I've totally done before. More. But

19:53 sure, but then, you know, it's also on, you know, you're holding, uh, CISOs criminally liable for, for security. And all of the risk, risk for a company is for, you know, for data loss is all—

20:07 well, they do at some level. SEC can go after a CISO for certain things, which is why they get paid a lot of money, because they're taking on a lot of personal risk.

20:16 That's— oh, I mean, that's what I'm saying. And so like all of the risk is being put on, you know, executives. And so they're having a hard time. They're having a hard time saying, well, we can't— we can just trust people when all that risk is one-sided.

20:28 Yeah, which is why they want them back in the office to monitor them. I'm sure, I'm sure it's all tied together.

20:36 Anyway, we kind of got off on a rant there.

20:38 We did.

20:38 We never go off on rants.

20:39 What are you talking about? We previously talked about Claude and its ability to fingerprint your text. to respond to the EU's requirements about this. And so, uh, OpenAI has rolled out their solution, or at least details about what they're gonna be rolling out. Uh, their text watermarking, watermarking is called Text Grain, and it's in response to the EU AI Act. Uh, it adds an invisible statistical signal to word choices in generated text, with ChatGPT and Codex watermarking coming to EU users in the next few weeks, and API customers globally able to opt in starting now. With it off by default. Detection reliability varies significantly with text length and editing. The detector identified watermarks about 80% of the 200-token passages versus 95% of 400-token passages. And editing just 10% of the words with synonyms dropped detection from 92% to 60%. 25% edits reduced to 17%. So, uh, yeah, you can get past this real easily, which is what I learned about the Anthropic version as well. Uh, detector access is limited to approved researchers and expert organizations by an application process rather than public release due to risks of false positives false negatives. OpenAI plans to open source the text grain technology for others to build on. OpenAI is explicit that watermark detection cannot determine human contribution ownership, user identity, or content accuracy, and the absence of a watermark does not prove human authorship. Primus is one layer alongside existing C2PA content credentials and SynthID watermarking for images and audio. And yeah, I mean, I don't need this. I can tell AI-generated text pretty darn well these days. Uh, but, uh, you know, if you, uh, are unsure, this apparently is going to be there. And I can't wait for some, some person's life work to be discredited because this detector detected it's AI-generated when it wasn't. Sure, go both ways.

22:25 There was a story I was reading online about a guy that they're, uh, they were getting accused of, you know, using AI to generate like— it was like an essay or like, or like a pretty high-level, like PhD thing. And then the guy took the, the person who accused them of it, took his thesis and put it through an AI generation tool. And this is from like 20 years ago and it said it was AI. And I feel like that's the other problem you're going to have is we put these tools in place where we're trying to figure out how to detect it, but it's saying things that were, you know, there's tools that are saying that, you know, the opposite too. So you're having to prove a negative at a lot of point with some of these things.

23:07 Yeah. Well, and like, whatever you, whatever you are a person who happens to love the em dash, you can never use the em dash now in writing because everyone will assume your em dash is, uh, you AI. So like, you know, entire, uh, entire writing constructs are now dead because AI killed them.

23:22 Yeah. That, that one hit me personally. I don't like that. I use that a lot. Now I don't.

23:28 Yeah. 'Cause people think you're using AI.

23:31 Well, he is, but that's a different story.

23:33 Yeah. Yeah. Well, I am using AI a lot, but I used it in my natural writing too. And so it's just sort of, I mean, you know how you can tell my writing is AI generated? It's spelled correctly. It's correct. You know, the grammar's correct. The ideas are, are in the right order instead of like some scatterbrained, you know, Jackson Pollock painting. That is usually what I write in my mind.

23:51 Yeah. It's, it's sort of funny to me because, you know, now when I need to write something, I'll You know, write all the bullets I want to, I think all the really clear, crisp things. Like I want all these things and I'm like, can you put that into a nice email for this person? And he writes it in an email. Then I just imagine them on the other side grabbing that email and putting it to AI and going, will you just summarize this down to the basic bullets? And we could have just saved a bunch of time if we just sent bullets to each other.

24:14 Yes. That's 100% what's happening. People send me AI slop. I tell AI to decrypt it and then I read it like

24:23 That is exactly what's happening for sure.

24:25 Mm-hmm. And I'm like, how many tokens are getting used just by this? How many gallons of water have we consumed by this email thread? They could have been bullets.

24:32 Well, maybe that's why AI has, you know, AI return on investment's not there. 'Cause people are just writing AI slop and then decoding AI slop.

24:40 Yeah.

24:41 Ah, all right, let's move on outta AI world to AWS. S3 Tables now supports the full Apache Iceberg v3 spec. Adding detection, uh, deletion vectors, row lineage, and new data types including variant, nanosecond timestamps, geometry, and geography. This addresses common v2 pain points around compliance deletes, semi-structured JSON parsing, and geospatial data encoding. We get into a lot of the rest of this, but it's, uh, in the weeds and I'm going to skip it for Ryan's sake.

25:10 We need a Mr. Baker here to help us understand the benefits, I feel like.

25:14 Yeah, basically The gist of this is this is all fitting into Amazon's broader Iceberg strategy spanning EMR, Glue, and Redshift with both S3 tables and Glue Data Catalog, all supporting Iceberg REST. And it really is coming down to data, to data, you know, sustainability as well, because if you store it all in one simple format and then you can access that format through multiple tools, then you don't have to, you know, do ETL jobs, which both all of us hate. And so, you know, you save that hassle and it's one of the big things in sustainability is not copying your data multiple times. So this is really around solving that big problem. And also it's portable and it's flexible and allows you to use lots of new tools and new capabilities you maybe weren't able to use in the past. So like, I get the value of it. I don't understand how it works, but, uh, I get it.

25:59 Yeah. I mean, this, this solves a problem that I ran into a couple of years ago for a personal project for a friend, and we couldn't use any kind of static, uh, database technology. It was cost was a huge concern. Because it required the geospatial data in there for the use case. And so like, this is great. If I had that same problem today, I could use this. So cool.

26:22 You know, you can bring that problem back to the surface. Now you have us call 'em up being, "Hey, I have a fix. We can do

26:26 this

26:26 now."

26:27 I will actually, thinking about

26:29 it. Amazon PostgreSQL can now directly query Apache Iceberg and Parquet data in S3, S3 Tables, or AWS Glue Data Catalogs. Extending this feature directly into the Postgres database. The feature embeds DuckDB directly into Aurora PostgreSQL, allowing, following Amazon's recent acquisition of DuckLabs, querying process happens within Aurora with no extra network hops, and DuckDB handles the analytical scan of lake data while Aurora manages operational data. So basically, yeah, this DuckDB basically sits inside of Postgres as a binary extension to allow Postgres to do magic with Iceberg tables. So that's what you're hearing.

27:04 Yeah. I think I learned that you could do extensions within Postgres as part of this, and now I'm terrified. Like I don't— Oh yeah. It's been a thing for a long time. Yeah. It's been a thing. I'm sure. Yeah.

27:15 You can do crazy stuff inside of there. You know, people use PostgreSQL as their vector databases. People use it for all sorts of things that, you know, some make sense. I love when people run NoSQL on top of Postgres because it's easier.

27:30 That is an extension. That's hilarious.

27:32 I know people that run that in production. It's fabulous from what I hear.

27:36 Yeah. In like 10 years, we're just going to be like the same sort of like Microsoft SQL Server, Oracle. We're just going to be in huge migration projects like we have been for decades on those two platforms on off of Postgres.

27:50 Yes.

27:51 Yeah, exactly.

27:52 And it will be back to, oh, we'll have a new database type. We had, you know, cause we. We made NoSQL into SQL at this point, so we have to have something else.

28:01 I assume that the next one's going to be All SQL Matters. Is that kind of what it is?

28:07 Yeah, something like that.

28:12 Oh man. I guess history repeats itself.

28:15 Maybe we'll go into CouchDB. What was it? CockroachDB and CouchDB for a while.

28:19 There's Cockroach, there's Voldemort, there's Couch. There's so many out there. Other. I mean, they all have very specific niches that the people who use it love it, and then if you don't have that specific use case, it's, it's the wrong solution. So, uh, AWS is releasing the Dogwood local engine, an open-source library that enforces temporal governance policies for AI agent actions, letting developers restrict tool calls based on past events, outcomes, and timing, such as requiring passing tests within the last 15 minutes for allowing a GitHub Actions push. Nginx integrates as a library into an agent harness that intercepts tool calls, checks them against policies, and enforces allow/deny verdicts, while durability is handled through a Rust-based embedded key-value store to survive crashes and restarts without losing event history. A notable technical feature is support for dynamic policy updates mid-session, allowing users to tighten or loosen agent permissions without stopping the agent, though updated policies only apply to events occurring after the change rather than retroactively. Performance benchmarks showed the schema design matters significantly, with fine-grained action definitions evaluated roughly 5 times faster than coarse-grained ones, and shorter temporal windows kept evaluation times constant, while a 24-hour window caused evaluation times to grow to 6 milliseconds per decision at 12 hours, with 300 times slower than a 15-minute window. This addresses a growing concern as AI agents are giving more autonomy to call tools, run code, and interact with external systems, giving developers a way to add guardrails around sequences of actions rather than just single-action permissions. Available now on GitHub for Teams, Building Agentic systems.

29:44 Yeah, the, um, I love this feature. I mean, I, I like anything where you can evaluate policy and you can define policy and, and have that be a sensible outcome. There are always going to be, you know, situations where you're balancing the productivity and the friction and reduce the business and the safety, much like we were talking about, uh, previously. And so this is one of those ways where you can introduce that from an enterprise to get, to get the level of protection you need And the previous method that everyone's been using is just routing everything through a single gateway, right? So configuring all the clients for an enterprise via MDM or something like that to route all the, you know, inference requests through a single place that they can evaluate policy there. And that's also very, you know, annoying to sort of set up and manage. And so a local library that can run along my local agent harness that I can then also distribute policy and have it dynamically reload. That's fantastic.

30:40 But how do you get every developer to actually set this up and run it? To my understanding, this is like a plugin or something you're running locally on the laptop. I honestly didn't do this.

30:51 It is, yes. No, it would be, it would be something you run locally. It's the same way you'd distribute any kind of software within a corporate environment.

30:57 I feel like I would still get annoyed and figure out ways around this, but you know, that's just me being a pedantic pain in the ass, which also goes into the prior conversation of actions do you have consequences when you're a pain

31:07 Well, yeah, no, and maybe it's one of those things where, you know, that's, you know, that's where maybe it's not just full functionality. So you do need an LLM gateway so that you can apply sort of all the, all the rules and stuff.

31:19 But I feel like LLM gateway, at least most places I've seen, also used as like a routing layer so they can route it to whatever the inference.

31:27 Yeah. You know, the value is being able to interpret the, interpret the prompt and then route it to the right model so you're not burning so many tokens.

31:33 Yep. And you know, there's a lot of financial stuff too, but you're starting to see a lot of security tooling that's, that's being injected in that layer, which doesn't have the same first-class citizenry of those types of things. And

31:44 so, yeah, but I mean, that also becomes problematic too, because if we have to start routing LLM calls through multiple gateways for different use cases, that's not really scalable either. So it's going to be interesting to see how that stuff evolves over time. But yeah, you're seeing DLP becoming integrated through gateways. You're seeing the routing models, you're seeing the open source models being accessed that way through a gateway. Like, it's, it's very popular. And then, and there's a couple different things too. There's, you know, if you have a SaaS application like your Oracle and you want to allow people to access your Oracle financial system, you know, you want to control the type of prompts that go through there. So you'll put an AI gateway there acting almost like an API gateway there, which has very clear rules and things that you can do or can't do with it. Um, whereas internally your AI gateway is more about, you know, allowing your engineers to access open models and access all this other stuff. So there's There's multiple things considered an AI gateway right now, so it's also a very confusing market as well.

32:41 Yes, for sure. Yeah. This is consuming all of my day job today, or, you know, these days. Yeah. It's just researching and knowing what's out there.

32:50 Yeah, it's me too. We should compare notes cuz I've been spending a lot of time on this one too.

32:55 I need to join in on this conversation. It sounds like we have an after-show topic one day too.

32:59 Oof.

33:00 Yeah. Everybody hate all of us.

33:02 Join our Patreon to not hear this conversation.

33:09 Amazon CloudWatch Logs now auto-indexes fields used with the equality and in filters based on query history, removing the need to manually configure field indexes as usage patterns shift over time. Auto-index fields don't count against the existing 20-field-per-log-group limit and persist for 30 days, giving teams a rolling window of optimized query performance without manual tuning. Indexing lists adjust dynamically as query patterns change, but teams can promote specific fields to permanent field index policy via the console or API if they want a guaranteed long-term index. This addresses a common operational pain point with teams previously had to anticipate which fields to index, often required rework as logging query needs evolved. Feature's available to you at no additional cost in all regions supporting CloudWatch Logs field indexing, making it a straightforward performance proof for existing CloudWatch Logs Insights users with no migration required. And this is a nice feature. Like the fact that it's free and I can now basically get long-term retention of specific patterns that I use all the time. I fill up my indexes. That's, this is nice.

34:04 So the only reason I don't like it, it was funny, we, in the pre-read, it was clear that you liked it and I didn't, was that I've had this break so many dashboards and automated queries because all of a sudden a field will, will basically change type. And so you're, if you have queries that are, that are populating dashboards and stuff, it can, can really mess up that, those types of things. And so I'm, you know, like I understand that there is developer overhead at the, at the application in order to predict that you need an index, but you're also in full control of that application layer. And I'm not sure that putting it into the, the, the logging service is the right way to go.

34:39 I mean, show me your version who's doing CloudWatch dashboards custom and I'll tell you, I'll show you an insane person because CloudWatch dashboards are terrible.

34:49 They are. I agree. I can't— I mean, I've done it and I do it, but I, I won't argue that they're a great thing.

34:55 They're terrible, but they're good when you, like, if you don't have like a Datadog, a New Relic, like a third party set up in order to process it. And I found them good when I'm like, okay, I quickly am trying to correlate across these 6 random AWS services, like my cache, my auto scaling group, you know, and my CP and my database. Okay, let me see what the hell's going on to see that linear point in time, you know, as like I can scroll through and at least see like, okay, CPU is high, memory is high when you're trying to compare like bisecond things like that. So I've used them in like live debugging situations like that. But then again, I'm a crazy person, as we probably discussed, that likes to break all the rules and break things.

35:39 I mean, I've had the same type of issues on any kind of like whether it be Datadog, New Relic, like wherever it's hosted, like it's Elasticsearch for, for—

35:46 Oh yeah, it's a, it's a problem in all those tools. Again, like this is Log Insights to me though, and Log Insights is really about mining with like full, you know, with query language or full text or using Amazon Q, heaven forbid. And so I, I get your point, Ryan. I just, I think the use case for this is like, hey, I'm going spelunking anyways and I know I need to use this field. Uh, that's, that's a value. But yeah, you're right. It will break dashboards. Down the line, which is a different problem. But I'm just the developer who wants to query my shit. I don't care about your dashboard.

36:15 No, I mean, it's, it's fair. I can, you know, I can see the pros and cons both ways.

36:19 Yep. I, I, I can as well. So, well, I think we talked about it here on the show. I think Matt even suggested it, that a, a great Agentic use case for Amazon would be to create a Well-Architected Agent. Uh, and so ta-da, they have produced an AWS Well-Architected Agent. Checks can be written to Matt, uh, available preview.

36:39 You know, I was gonna use it as my, uh, re:Invent guess.

36:44 Yeah, well, they took, they did it early. Uh, so yeah, unfortunately you, uh, did not get a point. Uh, but this is an AI-powered service that analyzes infrastructure across cost, security, performance, and reliability, then prioritizes recommendations based on defined business goals. The agent reviews Terraform, CDK, and CloudFormation templates directly, identifying gaps behind, guess what are the best practices, and generating specific infrastructure as code changes. Needed to fix them. Fargate agents come with execution tools built in, including SSM runbooks, CLI scripts, and guided console walkthroughs, plus cross-pillar impact analysis so teams can see cost and performance tradeoffs before applying a change. Preview is limited currently to US-East North Virginia, US-East Ohio, and US-West Oregon, but workloads from any commercial region can be onboarded. Access requires an active AWS support plan.

37:29 Death to long checklists and interviews with multiple choice questions, answers. I couldn't like this more just because it's replacing patterns that I think are insufficient. They're absolutely required because you have to go and review whether something's well-architected and those processes were previously used to educate developer teams in a lot of cases on backup strategies, DR strategies, and ensure everything was in place. But I'm sure, because I've seen it, the questions get answered in such a way where it's vague and people assume that DR is actually in place and functional. And I think that this type of analysis by, you know, an AI agent is going to really highlight some issues and discover things. So that's— I think it's great.

38:13 While I agree, and I feel like it can definitely— like, I, I don't mind the conversations because I feel like it really forces, you know, people to talk and think through things where I feel like once it's an agent, people are just going to click, click, click and not actually look at the outcome. Where if I'm sitting there having a conversation, interviewing Ryan and saying, how is your Elasticsearch cluster configured? Do you have HA? Do you have DR? She's throwing things at my face. I can get a little bit more idea of how it's set up than, hey, this tool ran. I mean, I think it's great that it queries the live whatever's running in your account. But it doesn't, you know, it doesn't right now. Oh, rephrase that. I don't think it right now queries like what's actually running, but it does, you know, I think over time they'll expand onto it. So you'll be able to point it at, say, hey, here, grab all the EC2 instances and grab all the, anything that's tagged BoltBot across the environment to see how it's configured. I

39:14 mean, did those conversations really come at question time or reviewing the findings from the questions? Because that's That'll still be in place. It's just the AI is gonna generate the output from the infrastructure as code configurations.

39:25 I did it at question time. I said, look, we can go, it depends what the customer wanted at the time, 'cause I was the one that did a lot of these back in the day, you know, was, okay, let's chat. Here's the question, but like, here's what they're really asking. Here's the real-world implications of what they're really going for here. and I would have that conversation with them. And you know, the hard point for a lot of companies is actually like isolating the workload that you were doing the review for, especially if your like application was a monolithic application with multiple pieces to it.

39:58 Mm-hmm. Yeah, that'll continue to be a hard problem to solve, I think, just because especially as you have AI querying, you know, codified configuration, I don't know if it's as easy to isolate. You know, just like a running cloud platform's not easy to isolate to specific workloads unless it's done correctly.

40:17 I mean, I think it will get there. I think just, you know, it's in preview. So I'll make, uh, our TrackIt bot hopefully track this. And it is something I do want to play with. Maybe I'll run it in, uh, you know, one of my accounts and tell me how horrible everything is because, you know, I don't want to pay for a NAT gateway. So I have something with a public IP address associated with it. Definitely don't have any of my containers set up like that for cost reasons. But you know, I'd be curious to see what the outcome is for some of these things.

40:46 Have you, um, you have NAT gateways you don't want to pay for. Uh, have you checked out FCK NAT?

40:53 Yes, I've used it in the past.

40:55 I've, I'm using it now. It's pretty great. Not for the Cloud Pod, different project I'm on. I'm like, this is, this is awesome. So much more. It was not hard to configure. And it's way cheaper, especially if you have low usage NAT gateway, like it's great.

41:09 Are you hosting it on like a small compute?

41:12 Yeah, you run it on a micro instance basically, or nano.

41:14 Nice.

41:15 Nice.

41:16 I had Foghorn back in the day. We essentially had as part of our VPC deployment was an EC2 auto-healing instance that just essentially did what they did here.

41:26 I mean, that's how, yeah, that's how we did it before.

41:29 That was NAT.

41:29 That was a thing. Yeah.

41:30 Yeah. Before you had to build your own nets, that's kind of what we did and we launched everything as small. Now I can tell you, there was a lot of outages from customers over the years that launched this thing as like a T2 micro back in the day, forgot about it for a long time, then all of a sudden was hitting, you know, bandwidth limits or CPU limits.

41:50 CPU credits.

41:51 Yeah. Yeah. Well, it was before credits.

41:54 The age-old conundrum. Yeah.

41:56 Yeah, you know, so like people forgot to scale them up or we don't wanna take the outage because of it, et cetera, et cetera. But you know, it was an easy thing to do.

42:05 I mean, you can get a T4g.nano for basically 4/1000 of a penny. So I mean, it's, it's so much, you know, it's basically over half, you know, almost 4 times or 4 times less expensive than the nginx gateway hourly cost. So,

42:21 or for my 3 containers running in my dev account, which one of which I offloaded to a laptop running locally because I was dealt with it. Eh, what's it, $1 a month for a public IP? I mean, is it the best way to do it?

42:37 It's more if you're not using it, but yeah, it's not that bad.

42:39 Yeah. But for my little janky things that I set up for, you know, to play with, it's fine. Do as I say, not as I do. Yeah.

42:49 And how would a security researcher find all the terrible things that Matt did in his account?

42:55 Well, he would use Security Hub, of course. Yeah. And, uh, Security Hub now has remediation plans that you can now effectively ignore. It groups the related exposure findings by root cause, letting teams fix one misconfigured resource or overly permissive policy to resolve multiple findings simultaneously rather than addressing each exposure separately. Uh, the plans include prioritization labels of critical, high, medium, and low, impact assessment and step-by-step fix instructions across AWS CLI, Terraform, and CloudFormation. Security Hub automatically ranks plans by risk reduction, surfacing the highest impact fixes first, which helps the security team prioritize limited remediation resources more effectively. What? No, no security team is doing that. They're gonna open a ticket to Matt. The API-accessible format allows AI agents to consume remediation plans programmatically, enabling automated security fix workflows rather than requiring manual intervention for every finding. And luckily, that AI agent can talk to Jira where they filed the ticket. So you can connect it back together and it is still automated even though there is a dumb ticket in the middle.

43:51 Yes, but that dumb ticket will contain a lot of remediation instructions and context where people will still not read it, but, uh, they will no longer have the excuse that you just sent me a list of 15,000 findings and with context that doesn't make any sense for me to resolve, which is true. That's usually how that works.

44:07 They're just looking for their AI agent anyways, so it doesn't really matter.

44:11 I was just saying, they're gonna have their AI agent read everything, blindly implement, take down the application, then blame Ryan for creating all these tickets.

44:19 Yeah, I mean, I'm perfectly fine with that because that's how it works with, with the Meet proxies now, right? So I'm perfectly happy developers will implement rules that are, that are copy and pasted by a security analyst into the Jira. No one's reading the context, no one's using critical thought, and then no one wants to be accountable when that all breaks down. I'm perfectly happy when that's all automated and can at least happen faster now, so it's not like on a 6-month remediation cycle.

44:47 We effectively broke production in 20 minutes. Good job, AI.

44:51 Well, I mean, you know, I've always long said that humans are the problem, so this is, you know, I'm okay.

44:57 I mean, I can drop a database in about 3 seconds, so I can break prod anytime. Well, it's, we got a lot of stories to get through, and so I'm thinking of time here. I'm gonna push us along just a little bit.

45:10 Probably a good call.

45:12 Yeah, let's move on to GCP. Enabling end-to-end checksums is now available to you in Cloud Storage. This basically is in the Cloud Storage SDK. Now enables end-to-end checksumming by default, closing a gap where data could be corrupted by bit flips in flight before server-side checksum validation occurred. Previously, this required developers to manually implement client-side checksums. I mean, welcome to the club, Google. Thanks.

45:33 Right? Yeah.

45:34 I mean, I don't know how this wasn't a requirement years ago for them, given some of their customers, you know, having checksum abilities on stuff in blob storage and S3 is like I had to do years ago.

45:45 Yeah, this is one of those things that when you use Google Storage for the first time, you're like, what? What do you mean? I've been using S3 for so long, I just assumed it was there.

45:56 You should be introduced to this thing called Azure. Less things are there. The adder has this, doesn't it, on blobs?

46:05 Yeah, it does.

46:07 Yeah. So yeah, that is what didn't exist. It was kind of, it was crazy. But yeah, at least it's there now. So appreciate that, Google. Thanks for finally welcoming us to 2020. Spanner Omni is now generally available, which lets customers run Google Distributed SQL database on-premises, on other clouds, or locally rather than exclusively on Google Cloud. There's been over 2 million downloads since preview launched at Google Cloud Next '26. The database combines SQL, graph, key-value, full-text search, and vector search in a single multi-model engine with native support for KNN and ANN vector search and Spanner graph for AI and GenTech application workloads. The general availability adds enterprise features needed for production, including TLS encryption, audit logging, backup restores, dedicated worker nodes for offloading background operations, and direct Google Cloud customer care support. Licensing is split into two tiers, a free developer edition with 90-day license or perpetual for single-server deployments, of 4 vCPUs or less, excluding backup and worker nodes, and a paid commercial edition using a vCPU-based annual subscription model with discounted proof of concept pricing available to you.

47:07 But is it a real database if it can't, you know, do like an installed plugin that does some sort of metadata lookup caching vector something something?

47:17 I mean, it does all that natively. It doesn't need a plugin.

47:19 I know. I'm just saying like it's—

47:21 It's a jack of all trades database. So it does nothing well, Ryan.

47:24 Nothing well. I've just learned it. You know, like now, now to be a giant, uh, nightmare, you have to have plugins. I mean, a production-ready, you have to put—

47:33 I mean, all SQL matters, Ryan.

47:37 If it doesn't speak native SQL language, it's not a real production database. Come on.

47:42 Yeah. I mean, it's nice to see this though. I mean, I'm glad to have the option and capability to run my Google workflow locally on my laptop before I go try to push it into Google Cloud. 'Cause you know, their, uh, their APIs are unfriendly and unforgiving. So. Anything you can do to help your testing and development environment work better, I think is a win.

47:59 Yeah, definitely.

48:01 If you would like to have a remote MCP for the Cloud CLI, it's now in public preview, giving AI agents direct access to gcloud and bq commands without requiring local CLI installation or dependency management in Azure environments. Server exposes two tools, run gcloud command and run bq command, allowing agents to handle infrastructure management and including diagnostics, BigQuery job monitoring, and scheduled queries, and permissions control through natural language. Security is handled through zero-ambient credentials, authentication via agent identity, OAuth 2.0, and IAM. Every command runs under the authenticated caller's permissions, and org policy constraints are enforced on target resources. Now before Ryan says he hates this, I will tell you that I've been using the AWS MTP a lot because it's a lot more efficient than the CLI, and I'm actually kind of happy about it. And I know I'm wrong, but I sort of like it.

48:50 So plot twist, I love this. And the reason I love this is because you wanna run those agent executions inside of a sandbox environment and you don't want to have all that token pass-through environment, token and credential data pass-through environment details, and so it's excluded. And now you finally have a way to safely execute against GCP in those either a remote agent execution or in a sandbox environment. So I think it's pretty great.

49:17 I, I did notice, A, it's a lot more efficient at querying things like CloudWatch logs and, and ECS server status and stuff without having to do a lot of like, 'cause you'd watch Claude struggle with CLI and like, it made me feel better 'cause they had the same problem with filters that I have with filters. Like never remember the proper syntax for this particular service and how they filter to this one versus another one. And so you'd see Claude stumble through the same things and now it doesn't, You don't see it stumbling through a sec. It just uses the MCP and queries it natively. And yeah, the fact that it's, it's not getting the credentials inline in the context, it's not doing all that stuff. It's, it's nice. I, I do quite like it. I, I'm starting to get more MCP religion as of late.

49:55 I mean, it's gotta be the new API interface. We just have to adjust to our new world order.

50:00 Yeah. I mean, I, I, I get a little annoyed when I have to host the MCP still. So I do prefer the remote hosted MCP that I can. OAuth to versus installing locally. But that's, yeah, I'm hoping it's just a matter of time before I don't have to worry about local anymore.

50:12 I think new businesses are gonna have an API endpoint. You're gonna have to have an MCP endpoint. Like, I think it's

50:17 just, I think that's the new default too. As long as, you know, we figure out a better way to handle authentication,

50:22 cuz, which I think is gonna be interesting for companies too, because, you know, you've been kind of able to get away with having no API because you provide a GUI, you provide a user experience, and And so customers like, well, I wanna do this customization, like, well, we just don't support it. I think the reality is these MTPs rely on APIs. So if you didn't have an API strategy, now you're really in trouble.

50:42 Yep.

50:42 With the MTP strategy, because you have to now build both an API strategy and an MTP strategy at the same time. So, you know, for those companies who invested in OpenAPI standards, you're ahead of the game. And those who didn't are now gonna feel the wrath. Yep.

50:55 It's gonna be a problem.

50:57 Uh, Google's Data Agent Kit is now generally available, offering free MTP tools and agents skills that connect coding agents like Claude Code, Codex, and Antigravity to over 15 Google Data Cloud services, giving agents live context on schemes, partitions, and job logs. The general availability adds support for BigQuery Graph, Bigtable, and managed services for Apache Spark, letting agents assign Bigtable row keys, building property graphs with GQL, and run serverless Spark sessions against Iceberg tables with branching and time travel. The Federate Lakehouse Catalog skill connects to AWS Glue and Databricks Unity Catalog, allowing agents to query data across cloud providers without building separate Injection pipelines first and security panels for existing IAM permissions with EKS connecting as the user or an impersonated service account.

51:39 Yeah, I, I, I'm really enjoying sort of the, the proliferation of these dedicated skills and tools and the installation just because I think it's so much better than relying on the data that's trained into the model and, and whenever that was, which is what, you know, I was doing before these things and you're, Azure's generating whatever you know, ETL workflow or data lake configuration or whatever plan. And so like this, I'm using, using these more and more, um, specifically skills and tools so that they're loaded into the context on demand and not, you know, based into the overall query instructions, which is cool.

52:16 Yeah. Yeah. It's sort of crazy. Um, you know how we went from like, yeah, you're relying what's in the model. So if, you know, Terraform releases a new provider, the model has no context. Now you're like, well, the MCP is there and it knows about the new model. And so it just gave that information to the model when you started up, and now it knows. And like, it's— you didn't have to do anything. It's kind of— it is really kind of great.

52:36 Yeah,

52:37 I think it's interesting what, uh, you know, the replacement for a lot of RAG workloads I think becomes MCP as well for the same reason.

52:44 Oh, absolutely. And publishing them to these kits, like, it's got version control and the whole you know, like it's easy to install and I'm really enjoying that sort of side of the ecosystem too. And I'm developing all my own personal stuff and publishing it in the same way just so I can keep track of it and I can share it easily to peers and what have you. So that's cool.

53:05 The problem with some of it is also just tracking it. Like, okay, I have, oh wait, where did I go to download this originally? And you want to talk about security context. Are these safe skills? To download, you know, have I created my own skills? You know, you end up with all these things that then create your environment. So now my local dev environment is all of these things put together, et cetera, et cetera, that I need. And finding the one piece in the future, if I get a new laptop or everything's going to be fun, you know, oh, I grabbed this random, you know, MCP I ran locally. One, talk about security of it. Two, where's it, where did I find this? So like, I'm almost a little bit concerned about the sprawl of some of it until it can become managed a little bit more centrally.

53:48 Mm-hmm. It's only a matter of time before, before there's like trusted repositories for these plugins too, so that you can kind of guarantee the security of, of these things and not have everything public and beholden to whatever public instructions are in there.

54:00 Well, I think that's sort of Anthropic opening. They haven't provided that type of, you know, federated marketplace and for skills and things. I mean, they have it for some connectors, but like they really need to provide some way to do like signing and cryptographic that they've, you know, reviewed and approved. And you know, these are trusted sources versus, you know, the guy who spelled Adobe incorrectly is like, this is my Adobe Photoshop MCP. And you're like, are you? I don't

54:25 know. Yeah.

54:27 Uh, well, Gemini has announced a new model, Argon, uh, Gemini 4 Argon in fact, which will expand output token limits to 1 million tokens up from 64,000 tokens. Enabling the model to sustain long multi-step reasoning in a single pass for complex software engineering and enterprise tasks. It launches at $2 per million input tokens and $10 per million output tokens, with cached inputs priced at 5% lower. So right in line with Anthropic and ChatGPT. And the model is rolling out initially through the Fairwind program to trusted cyber defenders, with access expanding gradually to API customers and Google AI Ultra subscribers. So basically, we announced it, but you can't use it because it's a, it's a threat to national security. And so Instead of being slapped by the administration, we're going to just make it this way initially, day one. So thanks Google for announcing and then not letting us play with it.

55:12 And then it's gonna be basically beta tested by a whole bunch of security researchers. Like, okay, I think that's—

55:19 And the government in the war against Iran, hopefully doesn't find a, you know, ship that it thinks is a nuclear vessel or something.

55:27 Yeah.

55:29 So I mean, I'm glad to see it, but you know, we'll talk about it more when we actually can use the damn thing. Sort of how I feel about Mythos too. It's like, yeah, okay, until you can use it, I don't really care.

55:38 Yeah. Yeah. I mean, this, they're starting to expand those things to where more and more companies are getting access to those programs, but that are not like the big, you know, trusted partners that signed onto the program to begin with. And so that's good. Yeah. But it, you know, how they get applied and everything, it still seems like a big gray area. It's like, cool, I can ask this model questions, you know, but it's like, eh, you don't give it access to anything. Like, what are you, are you just doing static code analysis, like expensive static code analysis. Like, I don't know if this is a good plan.

56:06 Yeah, agreed. Well, if you're unhappy about your containers on GKE starting up slow, they've now given you GKE CPU Startup Boost, now in preview, which temporarily increases CPU allocation for pods during initialization and automatically scales back to baseline once the app is ready, addressing the common trade-off between cold start latency and resource overconsumption. Provisioning. Google reports up to 2x faster application initiation for CPU-heavy startup workloads like Java and Spring Boot, Node.js, and Python, uh, without requiring developers to permanently overprovision CPU requests for steady-state operations. The feature relies on Kubernetes in-place pod resize, which graduated to GA in Kubernetes 1.35, allowing CPU requests to be adjusted on running pods without restarts, cache invalidation, or rescheduling overhead. That's a pretty cool feature. I mean, I didn't even know this was released inside of Kubernetes, but dynamically adjusting the CPU, you know, is a pretty cool feature to be able to do. I mean, your app has to support it though. So I mean, oh, not really.

57:06 I mean, the app, this is being transparent at the app layer.

57:09 I mean, the, yes, underlying part of it. Yeah. But like the way they made that sound was like, oh, you can add more CPUs to the pod and the pod will just use it. I'm like, well, if the If the app is dynamically able to see the CPU.

57:21 Yeah,

57:22 that's, that's the question I have there. But other than that, yeah, I think this is cool. All right. Moving on to Azure. If you were really jealous of that Cloud Spanner local feature earlier, you can now run SQL local, which you always could run SQL local, but this is particularly Azure SQL Server. Uh, SQL Server on Azure local is now generally available, supporting both connected and fully disconnected deployments for environments where cloud connectivity cannot be assumed. Such as remote industrial sites or regulated sectors like defense and healthcare. The disconnected mode called Azure Local Disconnect Operations, or ALDO for short, because Microsoft loves a good acronym, allows SQL Server to keep running locally when external connectivity is restricted or unavailable, addressing sovereignty and operational continuity requirements. Foundry Local on Azure Local, currently in preview, lets customers run AI inference directly alongside their SQL Server data on-premises, avoiding the need to move sensitive data to the cloud for AI workloads.

58:16 I just wonder what happened before, and I guess I never have played with Azure Local because I don't have piles and piles of money to burn. Is— did SQL just stop working? Like, did your SQL just go offline before where now it's— there's a full disconnected mode? Like, I don't understand what it did. You know, I guess it was in public or private preview before, so maybe it— that was a known edge case, but like, that seems like a crazy not so much of an edge case to, to have.

58:48 I mean, I think this is specifically to the Azure SQL capability, right? If you had this use case before, you just use RealSQL, but now this is like, well, I don't wanna buy RealSQL, I just wanna use subscription that I already have. This is what you're getting now through Azure Local.

59:01 Well, and you also get like the, the on-demand like patching and upgrading, right? Like there's other stuff that's part of this, or at least some of it. I know some of it you have to actually be in Azure Cloud to run.

59:11 But

59:12 yeah,

59:12 well, even, even Azure SQL, you still have to set the schema. Like you have to set the schema by database to set the version. And otherwise you got really weird nuances that can occur of like what's available for that SQL table or database and stuff like that. So maybe that's a SQL thing, but I definitely ran across it in, in Azure SQL Server.

59:35 I mean, ever, you know, a whole bunch of weird nuances in SQL Server. No. No, never.

59:41 Yeah, never, never heard of it. Don't know what that's about. Well, if you, uh, want to know what happens to all that old data center equipment when you finally shut down that M1 instance or, you know, V1 CPU, whatever thing Azure has— I never remember the name of it— Microsoft wants to tell you all about their operation of 8 circular centers globally with new facilities in Newport, Wales, and Sydney, expanding coverage to North America, Europe, and Asia. 9th Center plan for San Antonio, Texas. These facilities handle decommissioned data center hardware, deciding whether complete systems can be reused, like to give it to labs or schools, or components can be harvested as spare parts or materials needed recycling. Microsoft reports a 90% reuse and recycling rate for decommissioned servers and components last year, and Microsoft cites substantial efficiency gains since 2014, with cores per rack have increased roughly 13-fold while power consumption for equivalent workloads has dropped about 90%, meaning more compute capacity from the same data center footprint and power budget. The custom Azure Cobalt processor lines ties into this lifecycle strategy, with Cobalt 200 delivering up to 30% higher performance than Cobalt 100 through a co-designed silicon. Uh, so yeah, so if you're curious about how they recycle your stuff, they either give it away to schools and labs, or they sell it to people who will buy it from them, or use it for spare parts.

60:57 I mean, I like the reuse and it's better than what I previously thought would happen, which is that I thought they just sent all this stuff to those rage rooms and smash centers for people to beat up with baseball bats. So

61:08 yeah, I just, I just figured it got squirreled away in someone's garage where they are paying way too much power to run, you know, their Azure Local.

61:15 So, well, could be that too.

61:18 It is nice to see that they're upcycling it though and, you know, giving it to people to hopefully learn and get good use out of it when they can no longer use it.

61:26 Well, uh, Matt decided to trigger me, uh, and put this article into the show notes. Microsoft has announced that Microsoft Publisher has reached the end of life in October 2026, with Microsoft 365 subscribers losing access to the app entirely, while perpetual license holders will lose support. Microsoft recommends migrating common publisher use cases, uh, like templates, labels, calendars, and business cards to Word, PowerPoint, or Microsoft Create templates, sealing a consolidation of desktop publishing features into existing apps. Uh, and all I can say about this is good effing riddance. I hated this product. It was always a nightmare to support. And back in my IT help desk days, like anytime someone called me with a publisher problem, I would try to give it to someone else on the help desk so I didn't have to deal with it because it had like weird printer problems where like you would design a card and then try to print it out and it wouldn't print properly. It was always a nightmare. Fucking hated this thing. And especially because I came from, uh, Aldus PageMaker and then Adobe InDesign, and then I had to learn how to use Publisher, and it was so inferior to those Adobe and Aldus products that I just always hated it. So goodbye.

62:30 We need a sound effect that somebody could play with, like, after a rant. One of us goes on a rant, though now that I'm saying it, we might play it too much.

62:38 Yeah, I don't

62:38 think I'm saying it would just be constant. Yeah. But, uh, you're sad about this one. I, I, I see Matt. So that was your comment pre-show. Was that, uh, you're sad to see an old tool die? I'm not, but, uh,

62:50 I just remember like learning Publisher back in the day. Yo. And I was more surprised that people still are using it. Yo.

62:58 I mean, have you tried to put an image in a block of text in Word anytime recently? I mean, like there's a reason why it exists.

63:05 No, but I would have AI do it for me if So maybe I'm not the right, uh, sample.

63:11 I mean, I don't know if AI could do it either. It's, it's a, that's like trying to get a, a text wrapping around an image properly in Word is like this like super frustrating thing. So Publisher made it easier 'cause you could basically define this is an image block, this is a text block, and like then you could do drag and drop of the text and stuff like that, which just made it, was part of why it was so powerful when it was like PageMaker and those products did it. But when Microsoft implemented, they implemented in a shitty way, which is why I didn't like it, but you could do the same idea. But if you still want this capability, just go use Adobe InDesign. So that

63:43 sounds like an easy, easy tool to learn quickly.

63:46 It's totally fine. I actually know how to use it because I, I knew how to use PageMaker pretty well. Like I was a pro at it. And then InDesign came and like changed a bunch of things. And I was like not in a space where I was working about that kind of stuff anymore. And I just never learned it. So I think I, I have an Adobe Creative Cloud subscription for other reasons. And so I have it installed and I open it every once in a while and I try to make it work and then I'm like, yeah, I don't have to do this. I'll go somewhere else. Well, I got Oracle stories for you this week. Not really a cloud story so much as just like, hmm, yeah, what could go wrong? Oracle is releasing Fusion Claw, which if you know what Oracle Fusion is, you know, that's their enterprise apps that include their finance, HR, supply chain, and sales use cases and POS and all kinds of things. And if you know what Claw is, you know, that's OpenClaw. So they're combining OpenClaw access to Oracle Fusion apps which 25 new Agentic applications covering the full suite of Fusion applications. Pitch is moving from AI to assist to AI that executes, though customers will want to see concrete results before buying into that framing. Fargate splits work between frontier models for reasoning and deterministic enterprise computation for execution, while Oracle frames as an economic advantage by limiting expensive AI inference to only where judgment is needed. Or does this actually translate to meaningful cost savings? We'll have to find out. Governance for features like the enterprise operating envelope, outcome trust harness, and outcome receipts are clearly aimed at easing enterprise concerns about autonomous execution. I mean, but not so concerned that you're not going to run OpenClaw. Okay, yeah, sure, right.

65:16 That was my first take on this, was like, really, OpenClaw? Is that— that's where you're going? Not like, you know, the OpenSSH, like,

65:22 not like

65:22 a custom NVIDIA thing or some sort of other secure sandboxing? Like, no, just straight, uh, OpenClaw, which everyone learned about, not with all the security noise that it was made.

65:34 You're seeing like Meta Muse is basically a new version of like an OpenClaw type solution, both security built in on it. And then there's another one that's pretty popular right now in the Bay Area as well. Like there are alternatives to OpenClaw these days that are much more secure and robust. So I don't know why Oracle wouldn't have chosen to partner with one of those companies versus going kind of out this direction. It's sort of an interesting choice.

65:57 Yeah, like, I don't know why you wouldn't use OpenShell from NVIDIA on this, which is built from the ground up for isolation and security for financial data. Are you kidding me?

66:09 Yeah. Uh, and then we got a bunch of stories this week from Cloudflare. Uh, is it— was Cloudflare's annual birthday celebration, which they announced a ton of features. Uh, you know, I'm gonna go through these a little bit quick, uh, and we'll jump back to anyone you guys want to talk about, but, uh, So first up was AI search is now generally available. So their Cloudflare AI gateway now integrates with AI searching. So you can get basically real-time access to things on the web and browser rendering, et cetera, for your AI agent. The new Cloudflare K2 is a new serverless event streaming primitive now in public beta built to decouple event producers from consumers or so slow or unavailable downstream services can't cause dropped data. This basically replaced your need for Apache Kafka. Something from Cloudflare in the edge. The CloudflareOS is your new fully managed version of CloudflareOS, its agent workspace product, reducing the operational burden of self-hosting and configuring an open source deployment for organizations that don't want to manage it themselves. So this gives you a managed AI runtime area. And then introducing Clef. Clef is a new open source decision model hosted on Workers AI, positioned as a fast alternative to TypeSafe's JEV for structured non-text classification tasks close. I had to call out here that Jav was announced 2 weeks ago. So in 2 weeks, Cloudflare saw it, built a new product, and basically destroyed that company with Clef. Well done. Okay. Things are, uh, things are real in the AI race. Cloudflare Traces, which now extends automatic tracing beyond workers to the full request path, covering security rules, transformations, cache decisions, routing, and origin handling in a single timeline, requiring no additional setup, which is great. If you ever have trouble shooting anything in Cloudflare, that's really nice. Enhancement.

67:44 Yeah, that's awesome.

67:45 Happy to see that. Uh, yeah, so that's, uh, that's the big features from the Cloudflare birthday celebration.

67:50 Mm-hmm. Yeah, I mean, I'm not, I haven't really had the use case to use like AI search and setting that up. I, you know, I, if it's anything like other enterprise

68:01 search, well, technically every time you run Bolt, uh, it uses AI search every time.

68:05 Oh really? Oh, cool.

68:06 Yeah, it's one of the, how we get the articles is It goes and uses AI search capabilities to go get them when it can't, when it can't do a SAML, a simple curl on the website because of Cloudflare. So,

68:16 oh, that actually explains to me more a little bit about how it works. Now I'm a little bit more clued in. That's cool.

68:24 It, a curl is our, our typical go-to. Then there's, uh, you know, we go all the way up to a full browser emulation capability for us to grab show titles or show stuff.

68:32 There's a couple hacky things I feel like I've done too where there. Like, yeah, to get some of the articles that consistently yell. Microsoft's very, been very friendly about us pulling those articles down.

68:45 Harness is not. So Harness, anytime we do a Harness story, it's like crazy to the point where we, so we have a manual extraction capability built into Bolt so that when it, it does hit, it goes through it 6 different ways to try to bypass blocking. It will just come and tell Matt like it's gonna take over the world and destroy this company first when it becomes sentient. Uh, please provide me the text in the meantime. I'm like, okay, sure, here you go. But, uh,

69:09 Bolt gets angry at us, in case you haven't noticed.

69:12 Not at us, he gets angry at— it gets angry at others and then demands— he gets demanding at times. He's, uh, then

69:19 he sends a push notification to our phone too. It gets really annoying.

69:24 Yeah, because he wants you to confirm that this is the article you wanted to actually send, which is fine, but also annoying. All right, well, that's all I got on Cloudflare. Anything else you guys wanna add about Cloudflare?

69:39 Other than I look forward to trying the K2 service of Event Streams. I love anything that can scale down to zero, and this is one of those areas that's, you know, there's a lot of use cases for having Event Stream on the edge, and I like it. Agreed.

69:55 Uh, and then if you wanted one more place to run your agent, uh, DigitalOcean is bundling agent infrastructure, uh, including compute, storage, inference, and access to tools into a single monthly price called Agent Droplets, addressing the common pain point of piecing together separate vendors like Fly.io, AWS, Fireworks, Anthropic, different billing units. So one more place to run your open-code solution, uh, in the world. All right, we're going to skip Cloud Journey this week, uh, so I'll see you guys next week here in the cloud.

70:24 All right, bye everybody.

70:25 Bye everyone.

70:27 Another week of cloud news wrapped up. Vault will collect the news, Justin will get the notes, Jonathan will write some code, Ryan will watch the perimeter, and Matt will reluctantly watch Azure. Till next week for AI, Amazon, Google Cloud, and Azure. And hey, maybe Maybe even Oracle, who knows? Check out thecloudpod.net for our newsletter. Join our Slack, message us on socials, or leave a review.

70:58 All right. I have a fun after show for you guys this week. Uh, so developer Lucas Vogel rendered Doom using a SQL database with about 1,300 lines of SQL across 89 common table expressions, generating a 35 full color, 640 by 480 frames per second. In CedarDB. The project converted Doom's WAD files into relational tables of vertices, lines, and sectors, and replicated the game's binary space partition tree logic using a precomputed sort key and ORDER BY statements for rendering. This builds on the developer's earlier DoomQL project, which produced only grayscale ASCII raycasting graphics, showing a clear jump at what's achievable when pushing a database engine to handle real-time rendering tasks. I mean, this is just fantastic.

71:40 Yeah, wow. Yeah, I mean There's a lot of these like Doom, Doom porting. I don't know if that's what you call it, but, uh, like where I was like, that's silly. I don't understand it. Sequel. That's impressive. That's impressive. Yeah.

71:55 I've seen Doom run there. There has to be a Wikipedia page somewhere of everywhere Doom's run. 'Cause I feel like I've seen it run the most crazy things and this has to be up there and

72:05 yeah.

72:05 Bravo people. That's all I gotta say. Like, this is impressive and fun.

72:11 But also maybe go outside, you know, I don't know, touch some grass, uh, get

72:19 your vitamin— oh, D. It's D. D to a normal level.

72:24 Yeah. All right, so, uh, there's a, there's a website I found, 10 most absurd places that people ran Doom. Oh yeah.

72:31 All right, count them down.

72:37 Uh, a pregnancy test. Uh, so they, they, uh, took a digital pregnancy test that normally would say, you know, pregnant or not pregnant, and replace the board with a powerful microcontroller and a small OLED display. And so that way it ran it inside of a pregnancy test, which is funny.

72:57 Okay, okay, but they actually run it there. They swapped out the internals.

73:01 Yeah, I know, it's a little bit of a cheat, but I still thought it was Uh, a graphing calculator. A TI-83 is the next one. That one, that one makes sense.

73:09 Yeah.

73:11 It says despite slow CPUs and monochrome screens, with some optimizations, they managed to generate a few frames per second. That's pretty good for a TI-83, right? Uh, a pressure cooker with touchscreen. So they— that's pretty good too. That one is not a swap out, they just reprogram the interface firmware I had to basically run it.

73:31 That's nice.

73:32 Inside Minecraft. That one's, that one's obvious.

73:34 Uh-huh.

73:35 Apparently in a PDF, in an implementation called Doom PDF, which I just love the name of that. Like, that's—

73:41 Wow. That's terrifying.

73:44 Apparently the C engine inside of, uh, Adobe Acrobat was strong enough to run asm.js with an older version of Emscripten and saves the PDFs that use JavaScript engines in the PDF reader. So yeah, that's a great hack too. Hack your, your PDF with JavaScript. A digital thermostat. It's pretty good.

74:01 It is.

74:02 I mean, the digital thermostat had an ARM9 processor, so that's, that's relatively easy. And then this one is interesting. If you're gonna run Doom from a PDF, why not print it out? That's what this YouTube channel Bring It Studios did by modifying an Epson thermal printer. They printed each frame of the game running on a computer and you could see what was printed to play.

74:21 How many trees died in the creation of that? Yeah.

74:24 Exactly.

74:25 Yeah.

74:25 Uh, and then someone apparently hacked a Mercedes-Benz entertainment system to make it play Doom in their car on the native screen in their TV.

74:32 Oh wow.

74:33 That one, that one makes sense. I could see that one.

74:35 Yeah. Sounds dangerous cuz I would do that in traffic.

74:38 Yeah. And then the GoPro Karma drone controller runs Android and so someone got the boot Doom to run on top of the Android system for the drone controller, which, you know, could also be real life if you were, you know, had a gun on your drone, I suppose.

74:52 Yeah,

74:53 cool.

74:53 Yeah.

74:53 That's my Lance. And then, uh, someone hacked a Life Fitness Android exercise machine to run Doom. And that's the final one. So yeah, pretty good. Pretty clever. It's pretty cool out there.

75:02 I like this one.

75:03 These are clever.

75:04 Doom in Excel. Doom runs entirely in Excel using spreadsheet cells as pixel pixels and a VBA to drive the game rendering input engine. That's pretty good.

75:17 Yeah.

75:18 I mean, people have a lot of time.

75:20 Oh, this is a rabbit hole. There's one on the MacBook Touch Bar. Oh man, this is dangerous.

75:26 Oh, nice. That's—

75:27 yeah.

75:27 If I still had my iTouch Bar, I'd want to do that.

75:30 I know.

75:30 I have an old laptop downstairs. I feel like I should go try that. All right.

75:39 Fun times.

75:39 That's fun.

75:40 Yeah.

75:40 What's up there? That one's— yeah, maybe someone tried to get Doom to run on DNS. I think that'd be, you know, using Route53 as a, you know, as a database. Why not make it run Doom?

75:52 I mean, it's gotta be possible. Doom runs on anything.

75:55 Yeah, that's the goal.

75:57 All right guys, I'll see you next week.

75:59 All right. See you later.

76:01 See ya.