บทถอดความฉบับเต็มของ Rethinking the AI Data Stack: From Storage to the AI Operating System

these RAG pipelines and so forth that 32 00:01:56,290 --> 00:02:00,280 we'll talk a little about, have to be very, very current, meaning within 33 00:02:00,280 --> 00:02:03,670 seconds of, you know, new documents being added to a realm, being able 34 00:02:03,670 --> 00:02:05,720 to utilize those in these workflows. 35 00:02:05,769 --> 00:02:09,440 Very different requirements as we move into this inference era. 36 00:02:09,589 --> 00:02:13,330 Maybe tens of thousands of agents will be accessing this stuff, and it really 37 00:02:13,330 --> 00:02:17,170 changes the whole dynamic of what's required for computing, and that's really 38 00:02:17,170 --> 00:02:19,170 what we are bringing to the forefront. 39 00:02:19,400 --> 00:02:22,749 Existing legacy architectures weren't designed for streaming inference 40 00:02:22,749 --> 00:02:26,690 pipelines, multi-tenant AI workloads, or zero trust data access at scale. 41 00:02:26,839 --> 00:02:28,959 And this is where a lot of agencies get stuck. 42 00:02:29,159 --> 00:02:32,650 They don't fail on GPUs or models, they fail on data movement, 43 00:02:32,709 --> 00:02:34,419 latency, and security boundaries. 44 00:02:34,419 --> 00:02:38,309 It's a very different conversation than what capacity needs do you have or what 45 00:02:38,309 --> 00:02:39,520 fees and feeds and those kind of things. 46 00:02:39,520 --> 00:02:42,830 So we're really getting away from that quite a bit, especially talking about 47 00:02:42,830 --> 00:02:46,910 the AI workflows as everyone's trying to figure out how to do that, um, and 48 00:02:46,910 --> 00:02:48,720 we can help to simplify a lot of those. 49 00:02:48,999 --> 00:02:51,209 So I want to touch on the DAY's architecture in case you're, 50 00:02:51,209 --> 00:02:52,629 uh, not familiar with this. 51 00:02:52,679 --> 00:02:54,590 As I said, it's, it's a unique architecture. 52 00:02:54,590 --> 00:02:55,330 It's very different. 53 00:02:55,489 --> 00:02:59,699 Some of the limitations that even a scale-out architecture built on HA 54 00:02:59,699 --> 00:03:03,440 pairs has were kind of glaring and, and that's where the VAST system 55 00:03:03,620 --> 00:03:05,849 filled all of those gaps, if you will. 56 00:03:05,919 --> 00:03:08,499 This architecture is very different in the, in the obvious part of being 57 00:03:08,499 --> 00:03:11,249 shared everything, and what that really means is that if you think of 58 00:03:11,249 --> 00:03:15,160 a controller-based implementation, controllers manage disks, right? 59 00:03:15,160 --> 00:03:16,830 Or, or Flash in, in our case. 60 00:03:16,870 --> 00:03:19,590 Um, and so we don't actually have that same relationship. 61 00:03:19,830 --> 00:03:23,599 So what does disaggregated and share everything actually mean? 62 00:03:23,840 --> 00:03:26,610 VAST DAY's architecture fundamentally changes how 63 00:03:26,610 --> 00:03:28,300 performance and scale are delivered. 64 00:03:28,654 --> 00:03:32,795 In plain technical terms, that means C nodes handle protocols and compute, 65 00:03:33,014 --> 00:03:38,794 those being NFS, S3, SMB, Kafka, NVMe, and TCP, while D nodes store 66 00:03:38,794 --> 00:03:43,195 state and data with no controller ownership and no HA pair bottlenecks. 67 00:03:43,455 --> 00:03:47,835 This also includes NVMe over Fabrics connecting everything at low latency. 68 00:03:48,005 --> 00:03:52,124 And storage class memory absorbs write pressures, and QLC flash is used 69 00:03:52,124 --> 00:03:54,864 efficiently and gently for long-term data. 70 00:03:55,175 --> 00:03:59,114 Having an understanding of storage and integration experience really matters in 71 00:03:59,114 --> 00:04:02,744 this conversation because mapping mission workloads to performance profiles while 72 00:04:02,744 --> 00:04:07,084 avoiding over-provisioning is an extremely important part of this implementation. 73 00:04:07,174 --> 00:04:10,694 You do have to consider independent scale of performance versus capacity, 74 00:04:10,804 --> 00:04:14,124 simpler life scale management, and fewer architectural choke points 75 00:04:14,124 --> 00:04:15,865 from long-term mission systems. 76 00:04:16,045 --> 00:04:19,365 But, uh, we're all kind of in it together and, and kind of creating this whole 77 00:04:19,365 --> 00:04:22,084 vision for kind of simplifying AI. 78 00:04:22,214 --> 00:04:26,194 And just from a little history, the company was started in 2016. 79 00:04:26,334 --> 00:04:29,404 Um, the whole idea was this revolutionary new architecture 80 00:04:29,404 --> 00:04:31,885 for storage and for storing data. 81 00:04:32,025 --> 00:04:34,375 Um, it was really all about the data, which is why we're 82 00:04:34,375 --> 00:04:36,615 Vast Data and not Vast Storage. 83 00:04:36,664 --> 00:04:40,475 Um, storage is a thing, but managing the data and making data 84 00:04:40,475 --> 00:04:42,375 useful is what we're all about. 85 00:04:42,534 --> 00:04:45,884 The idea behind Vast was to create this new architecture. 86 00:04:45,884 --> 00:04:49,384 We call it DASE, the Disaggregated And Shared Everything. 87 00:04:49,685 --> 00:04:52,055 And that was, you know, kind of counterintuitive to every other 88 00:04:52,055 --> 00:04:55,734 storage company, um, because they were all about shared nothing 89 00:04:55,734 --> 00:04:57,335 implementations and still are today. 90 00:04:57,585 --> 00:05:01,475 While talking to Vast, we also saw a shift from the, quote, "storage 91 00:05:01,475 --> 00:05:03,905 platform to the AI operating system." 92 00:05:04,005 --> 00:05:07,174 Because Vast has said that they're not trying to win on storage anymore. 93 00:05:07,295 --> 00:05:10,515 It's all about collapsing multiple AI infrastructure layers into 94 00:05:10,515 --> 00:05:12,225 a single operational platform. 95 00:05:12,395 --> 00:05:16,734 And their latest platform expansion includes unified data stores at the file 96 00:05:16,765 --> 00:05:21,335 object and structured data levels, SQL and vector databasing to allow for

massive 97 00:05:21,335 --> 00:05:25,975 scale, Kafka native data streams, and data engines that leverage functions, 98 00:05:25,975 --> 00:05:29,475 triggers, and event-driven automation, while agent engines and insight 99 00:05:29,475 --> 00:05:34,284 engines use NVIDIA NIMs, GPU-assisted embeddings, and inference workflows. 100 00:05:34,579 --> 00:05:38,070 So what that means for your environment is it eliminates brittle multi-vendor 101 00:05:38,070 --> 00:05:42,849 pipelines, reduces security accreditation scope, and enables real-time AI reactions 102 00:05:42,920 --> 00:05:46,960 to data retrieval And most importantly, between all of these Sync Engine, Insight 103 00:05:46,960 --> 00:05:50,509 Engine, Agent Engine, having the same security, so having them in this, you 104 00:05:50,509 --> 00:05:55,349 know, single data platform allows us to manage who can use what agents, which 105 00:05:55,349 --> 00:05:59,470 users can go out and do inferencing or, uh, embeddings and things like that, 106 00:05:59,720 --> 00:06:03,249 and make sure that those, uh, Insight Engines and Agent Engine components 107 00:06:03,410 --> 00:06:07,680 can get access only to data that the user is allowed to have access to. 108 00:06:07,680 --> 00:06:09,349 And so that's one of the clear differentiators. 109 00:06:09,349 --> 00:06:13,140 But I also want to highlight some non-hypothetical use cases that 110 00:06:13,140 --> 00:06:16,250 we discussed with VAST and the expansion of its new platform. 111 00:06:16,559 --> 00:06:18,620 Those include cyber network security. 112 00:06:18,800 --> 00:06:24,360 The new VAST platform includes AFRL packet capture and NetFlows, instant pivot from 113 00:06:24,360 --> 00:06:28,789 metadata to full packet, reduced Splunk licensing footprints, and faster zero-day 114 00:06:28,789 --> 00:06:33,700 investigations Secondly, we talked about citizen and analyst AI assistants. 115 00:06:33,819 --> 00:06:38,550 We discuss a recent IRS chatbot example and continuous documentation updates. 116 00:06:38,899 --> 00:06:42,939 Thirdly, we discuss multi-tenant agencies and how the VAST platform is a 117 00:06:42,939 --> 00:06:47,470 shared platform with logical isolation, cost amortization, and role-based 118 00:06:47,470 --> 00:06:49,099 and attribute-based boundaries. 119 00:06:49,249 --> 00:06:52,919 In this example, I'm gonna use some S3 buckets, and I have set 120 00:06:52,919 --> 00:06:56,089 up some triggers so that something gets dropped into the S3 bucket. 121 00:06:56,239 --> 00:07:00,409 I can see that as a trigger, and this is where we use our Kafka engine internally 122 00:07:00,409 --> 00:07:05,009 to impact that and, and, and basically execute a function based on that. 123 00:07:05,129 --> 00:07:08,489 Um, with the Insight engine, function could be to call out 124 00:07:08,489 --> 00:07:12,259 to a, a, a GPU and do, you know, vector embeddings or other things. 125 00:07:12,259 --> 00:07:15,590 So the idea is that within the VAST Data

platform itself, we 126 00:07:15,590 --> 00:07:19,070 can now create triggers and functions, minimal coding required. 127 00:07:19,140 --> 00:07:21,480 Uh, most of this stuff is written in Python, but could be 128 00:07:21,539 --> 00:07:22,570 written pretty much in anything. 129 00:07:22,809 --> 00:07:27,210 All we need to do is take that code and put a Docker wrapper around it, and now 130 00:07:27,210 --> 00:07:31,030 we have a Docker container that we can execute as a trigger or, or, or function. 131 00:07:31,079 --> 00:07:35,190 Once you create those elements, then you can actually go in and create pipelines. 132 00:07:35,440 --> 00:07:38,110 And that leads us into a conversation about zero trust, 133 00:07:38,110 --> 00:07:39,930 ABAC, and security realities. 134 00:07:40,169 --> 00:07:43,510 This particular part of the discussion emphasizes where AI 135 00:07:43,510 --> 00:07:47,180 platforms fail in government today when security is bolted on later. 136 00:07:47,260 --> 00:07:50,949 And this is why VAST's new architecture bakes it in from the beginning, because 137 00:07:50,949 --> 00:07:55,729 it includes native RBAC and ABAC, active directory integrations, metadata-driven 138 00:07:55,729 --> 00:08:00,010 access controls, zero trust principles embedded at the data layer with the same 139 00:08:00,010 --> 00:08:04,769 policies distributed across AI agents, databases, object storage, and pipelines. 140 00:08:05,019 --> 00:08:08,040 As far as the data store and the database, that's kind of where stuff's stored. 141 00:08:08,199 --> 00:08:10,470 I didn't talk too much about the data space, but that gives us the 142 00:08:10,470 --> 00:08:12,269 ability to have this global namespace. 143 00:08:12,269 --> 00:08:15,640 So I have maybe something in Amazon, something in Azure, something on-prem. 144 00:08:15,729 --> 00:08:19,559 Um, we can share data across those different realms, have that as a single 145 00:08:19,559 --> 00:08:23,610 global namespace and, and again, manage the security across that, move data 146 00:08:23,610 --> 00:08:27,890 only when necessary, give access to data across wide area networks, so. 147 00:08:28,099 --> 00:08:33,599 We actually handle, uh, about 99% of the zero trust without any external softwares. 148 00:08:33,669 --> 00:08:36,049 It's all built into the VAST system, so there's a lot of 149 00:08:36,049 --> 00:08:37,320 security components built in. 150 00:08:37,320 --> 00:08:41,459 But ABAC is, uh, is a core component of, of our security infrastructure and 151 00:08:41,459 --> 00:08:45,319 how we manage that and allow capturing of, of lots of different metadata 152 00:08:45,319 --> 00:08:49,259 and giving it specific attributes based on, uh, on need there, so. 153 00:08:49,679 --> 00:08:53,699 In our view, this is where VAST really adds value in validating these controls 154 00:08:53,699 --> 00:08:58,490 against RMF, DoD zero trust scores, and agency-specific security overlays. 155 00:08:58,600 --> 00:09:01,750 So in the end, the AI accelerates approvals instead of blocking them. 156 00:09:02,229 --> 00:09:06,039 So what's the bottom line up front of VAST's new AI operating system? 157 00:09:06,500 --> 00:09:10,949 It's all about data architecture, security, and operational simplicity. 158 00:09:11,259 --> 00:09:15,149 Because federal AI isn't about flashy models, and platforms like VAST 159 00:09:15,149 --> 00:09:19,490 show what's possible when storage, databases, streaming, and AI pipelines 160 00:09:19,539 --> 00:09:24,190 are treated as one system instead of six It means that AI success in 161 00:09:24,190 --> 00:09:26,239 government is a data problem first. 162 00:09:26,549 --> 00:09:29,419 Daze architectures remove scale and performance bottlenecks, 163 00:09:29,569 --> 00:09:32,920 unified platforms reduce complexity and accreditation friction, and 164 00:09:32,920 --> 00:09:36,880 real-time AI requires event-driven data pipelines, while ABAC and zero 165 00:09:36,880 --> 00:09:38,909 trust must exist at the data layer. 166 00:09:39,169 --> 00:09:42,999 And in the end, success in this arena still depends on integrators, integrators 167 00:09:42,999 --> 00:09:46,019 who understand federal networks, zero trust, and mission constraints 168 00:09:46,019 --> 00:09:47,620 well enough to put this tech to work. 169 00:09:47,730 --> 00:09:51,470 So if you're exploring AI platforms, data modernization, or zero trust 170 00:09:51,470 --> 00:09:55,400 alignment and want to understand what actually works in your environment, reach 171 00:09:55,400 --> 00:09:59,799 out so we can help translate fast new architecture into real mission outcomes. 172 00:10:00,330 --> 00:10:07,970 Be sure to reach out to ATPGov today at www.atpgov.com or email info@atpgov.com or 173 00:10:07,970 --> 00:10:10,290 check us out on social media on LinkedIn. 174 00:10:10,530 --> 00:10:13,700 Thanks for listening, and be sure to subscribe to the Bottom Line Up Front 175 00:10:13,700 --> 00:10:17,409 wherever you get your podcasts, and stay tuned for more distilled insights from the 176 00:10:17,409 --> 00:10:19,810 front lines of tech and national security. 177 00:10:19,950 --> 00:10:24,098 So until next time, stay secure, stay mission ready.