WEKA CEO discusses shift to AI inference, augmented memory, and sovereign cloud demand
Summary
WEKA CEO Liran Zvibel explains the industry's pivot from AI training to inference, where storage must manage KV cache and context windows. He claims WEKA's Augmented Memory Grid can increase token throughput five-fold by optimizing cache hit rates. The company offers both software and proprietary hardware (WEKAPod) to address supply chain constraints. Zvibel identifies sovereign AI and neo-clouds as primary growth drivers due to data residency needs. He predicts AI storage will eventually merge with memory and vector databases.
Key takeaways
- 1 WEKA's Augmented Memory Grid links NeuralMesh to inference engines to optimize token throughput.
Unidentified speaker (speaker_1) · 7:16
Episode sections
- 0:18 – 3:15
Introduction and Company Overview
Derek Mogull introduces Liran Zvibel, CEO of WEKA. Zvibel outlines WEKA's 2013 founding for scalable file systems and its pivot to GPU-heavy AI workloads.
- 3:15 – 5:04
Shift from Training to Inference and Agents
Zvibel describes the market shift from training to inference, where storage value lies in efficiency and context delivery for real-world applications.
- 5:04 – 7:00
KV Cache and Memory-Storage Hierarchy
Discussion on multi-turn inference requiring KV cache storage that exceeds GPU memory, necessitating low-latency flash storage as augmented memory.
- 7:00 – 11:31
Augmented Memory Grid (AMG) and Economic Benefits
Zvibel explains how linking NeuralMesh to inference engines optimizes throughput, claiming cache hit rate improvements yield significant economic benefits.
- 11:31 – 17:07
Product Portfolio: NeuralMesh, Axon, and WEKAPod
Details on WEKA's software-centric approach, including NeuralMesh, Axon for third-party hardware, and WEKAPod for proprietary appliance control.
- 17:07 – 23:16
Flash vs. HDD and Tiering Strategies
Zvibel outlines WEKAPod flavors (Nitro/Prime) and asserts active data will remain on flash while HDDs handle archival.
- 23:16 – 27:30
Sovereign AI and Neo-Clouds
Identification of sovereign AI and neo-clouds as key demand drivers due to data residency concerns and newer hardware availability.
- 27:30 – 33:56
Competitive Landscape and Future Outlook
Zvibel discusses competition with NVIDIA/AMD reference architectures and predicts a future blend of storage, memory, and vector databases for AI agents.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| NVIDIANVDA | equity | neutral | They're able to keep track on the latest generation from NVIDIA and AMD, and usually the hyperscalers have different architecture that they've built for a different era, especially around the network. |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
WEKA was founded around late 2013 to recreate storage from the ground up.
Unidentified speaker (speaker_1) ·1:17
How to verify: WEKA company history on its investor relations page or SEC filings if public; corporate press releases from 2013-2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceWEKA started going to market around 2017 or 2018.
Unidentified speaker (speaker_1) ·2:17
How to verify: WEKA product launch announcements or press releases from 2026-2026. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
AI became mainstream around two to three years ago with ChatGPT.
Unidentified speaker (speaker_1) ·2:47
How to verify: Public release date of ChatGPT (November 2026) and subsequent media coverage trends. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
WEKA's Augmented Memory Grid links NeuralMesh directly to inferencing engines.
Unidentified speaker (speaker_1) ·7:16
How to verify: WEKA product documentation or technical white papers describing the architecture of Augmented Memory Grid. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Increasing KV cache hit rates from 90% to 98% can deliver five times more tokens per GPU.
Unidentified speaker (speaker_1) ·10:09
How to verify: WEKA's Inference Economics Scorecard or technical benchmarks published by WEKA or NAND Research. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Increasing cache hit rates from 90% to 98% can reduce time-to-first-token by 20-fold.
Unidentified speaker (speaker_1) ·11:08
How to verify: WEKA's Inference Economics Scorecard or technical benchmarks published by WEKA or NAND Research. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
WEKA's software runs on servers from Supermicro, Dell, HPE, and Lenovo.
Unidentified speaker (speaker_1) ·12:12
How to verify: WEKA compatibility matrix or partner listings on their website. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceWEKA runs on public clouds including AWS, GCP, Azure, and OCI.
Unidentified speaker (speaker_1) ·12:12
How to verify: WEKA cloud marketplace listings (AWS Marketplace, Azure Marketplace, etc.). For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Primary sourceWEKA's Nitro product uses only TLC flash.
Unidentified speaker (speaker_1) ·17:36
How to verify: WEKA product specifications for WEKAPod Nitro. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
WEKA's Prime product uses a combination of TLC and QLC flash.
Unidentified speaker (speaker_1) ·17:55
How to verify: WEKA product specifications for WEKAPod Prime. For recent events, verify 2026-08-18 through 2026-11-18. Use another period only when the speaker explicitly names it.
Do your own research
- Verify WEKA's revenue breakdown between software subscriptions and hardware appliance sales via SEC filings (if public) or press releases.
- Check independent benchmarks for WEKA's claimed 20-fold reduction in time-to-first-token against competitors like Pure Storage or Dell.
- Review sovereign AI procurement trends in EU and Asia to validate the 'neo-cloud' demand thesis.
- Analyze WEKA's supply chain dependencies for NVMe drives and servers to assess the '2-4 week delivery' advantage.
- Compare WEKA's pricing and performance metrics against NVIDIA's reference architectures for AI inference.
- Investigate the technical feasibility of 'augmented memory' claims versus traditional RAM-based caching solutions.
- Look for customer case studies validating the 5x token throughput increase claim.
- Assess the competitive landscape with NVIDIA's own storage partnerships and AMD's ecosystem support.
Full transcriptexpand
- 0:18
- Unidentified speaker (speaker_0): Hello and welcome to the Tech Disruptors Podcast, hosted by Bloomberg Intelligence. In this podcast series, we speak with company executives and management teams about. their view on disruption and how it's driving their decision making and strategy. Bloomberg Intelligence is Bloomberg's research arm and covers over 2,000 companies globally across multiple asset classes, backed by
- 0:38
- Unidentified speaker (speaker_0): Bloomberg and third-party data, supported by nearly 500 research professionals. My name is Derek Mogul, tech analyst at Bloomberg Intelligence. Today we're joined by Liran Zwibel, co-founder and CEO of WECA. Weka, a leader in high-performance AI data platforms, to discuss the company's platform, the shift from training towards inference and agents,
- 0:58
- Unidentified speaker (speaker_0): and how that transition is changing the role of storage in the AI factory. Okay, let's jump right into it.
- 1:04
- Unidentified speaker (speaker_1): Yeah, thanks for having me, Derek. I'm excited to be here.
- 1:07
- Unidentified speaker (speaker_0): Excellent. Okay, so for listeners who may be less familiar with Weka, can you start with a high-level high level overview of what the company does? and where you sit within the AI infrastructure stack?
- 1:17
- Unidentified speaker (speaker_1): Yeah, for sure. So we started about uh a decade ago, back in late 2013, when we realized that we could recreate what back then was called storage, uh, from the ground up to to meet the new infrastructure. Uh we didn't like the fact that customers had
- 1:37
- Unidentified speaker (speaker_1): to go and choose. different kinds of products for different kinds of applications. And our vision was to create uh a product that is more scalable than object storage, more performant than block, and as uh easy to use with the right semantics of a file system. And
- 1:57
- Unidentified speaker (speaker_1): back then it looked a little bit like uh science fiction, but for fast forward to today we uh obviously deliver on it. When we started going to market around 17 or 18, uh where we started the company we thought we would be
- 2:17
- Unidentified speaker (speaker_1): unifying a lot of the traditional data center applications, we were actually met with a bunch of uh applications.
- 2:27
- Unidentified speaker (speaker_1): that were very forward-looking back then using GPUs. It was called machine learning or deep learning and other other environments. And we realized, hey, they need a lot of data, they need it in a file system format, and they need it incredibly fast. And
- 2:47
- Unidentified speaker (speaker_1): we gravitated towards these kinds of kind workloads. And as AI became a thing uh two, three years ago with Chet GPT taking over, this is the main uh applications that we're currently targeting.
- 3:05
- Unidentified speaker (speaker_0): Oh excellent. So you were able to uh build an architecture that was basically that basically had uh much foresight into this this current AI proliferation.
- 3:14
- Unidentified speaker (UNKNOWN): So
- 3:15
- Unidentified speaker (speaker_0): let's maybe take a step back from Weka specifically and look where the broader market is heading. Three to five years from now, how do you think AI changes the role of storage in the data center?
- 3:25
- Unidentified speaker (speaker_1): Right. So when when we look at the earlier days of augmenting GPUs, it was mostly for training.
- 3:34
- Unidentified speaker (UNKNOWN): Uh
- 3:34
- Unidentified speaker (speaker_1): we didn't have the models yet, the
- 3:37
- Unidentified speaker (UNKNOWN): uh
- 3:37
- Unidentified speaker (speaker_1): all of the lab. labs were trying to prove that AI can be useful. And the problems were trying to figure out is: hey, how can we train the biggest, most effective models to provide value as we go from now into the future? It's not that training is going to
- 3:57
- Unidentified speaker (speaker_1): become less important than it is today, but it's not going to keep growing as it was. We're going to shift towards inference, meaning, hey, how can we take the trained models and generate more value out of them? And when running
- 4:17
- Unidentified speaker (speaker_1): inference in production, when essentially transitioning to leveraging AI to solve real-world problems, the
- 4:27
- Unidentified speaker (speaker_1): role of storage stops being, hey, can we get the models to improve? And becomes can we deliver higher efficiency? And this is where delivering on context for these models is becoming more and more important, and the ability to drive more
- 4:47
- Unidentified speaker (speaker_1): concurrent users and uh more
- 4:51
- Unidentified speaker (speaker_1): effective tokens out of an existing GPU infrastructure.
- 4:56
- Unidentified speaker (speaker_0): Gotcha. So uh just to double-click on that, with these agents and longer context windows, what really changes most about the storage requirement?
- 5:04
- Unidentified speaker (speaker_1): When you're running either a multi-turn inference session, you know, you go into a chat, you you give it one prompt, you don't get exactly what you want, you give it another. prompt and another prompt and another prompt. Each time you give a new prompt back in,
- 5:22
- Unidentified speaker (speaker_1): the in the inference infrastructure essentially needs to remember all of the intermediate steps it had answered to you before. So, in order to give you the 20th answer, it has to have all of the tokens of the 19th and the 18th and the 17th.
- 5:42
- Unidentified speaker (speaker_1): This is the the model short-term memory, or now it's called either KV cache or the model context. The problem is that these contexts are quite large and they they end up not fitting either in the GPU memory or the local
- 6:02
- Unidentified speaker (speaker_1): memory. that you have on a single machine. And what we're doing at Waka, what others are trying to do in the industry is essentially leveraging uh flash, leveraging what used to be called storage to solve that problem. And on the one hand, it proves very,
- 6:22
- Unidentified speaker (speaker_1): very efficient. On the other hand, it forces you to deliver even higher. performance, lower latency storage that has characteristics that are more similar to memory than what storage used to be delivering a few years ago. Does it make sense?
- 6:40
- Unidentified speaker (speaker_0): Yeah, that makes sense. So I think that that brings us to this what people are calling an emerging memory and storage hierarchy. Uh, for our listeners, this is a d essentially a combination of multiple layers and including high bandwidth memory, DRAM, Flash, as uh Liran alluded to, and even HDD in some cases. So, and data is placed within that
- 7:00
- Unidentified speaker (speaker_0): hierarchy according to how quickly and frequently it needs to be accessed, particularly during inference. So I think you hit on a key point, which is this KV cache. And I think you have a product called augmented memory grid. And I was just curious, how does it fit into that architecture? And what product problem is that product solving?
- 7:16
- Unidentified speaker (speaker_1): Right. So the augmented memory grid, AMG for sure, takes our existing product, the Neuralmesh uh file system that fundamentally delivers the lowest latency storage out there
- 7:33
- Unidentified speaker (speaker_1): and links it directly to the inferencing engines. So when when you have an inferencing stack, you need the actual inferencing engine to be uh context or KV cache aware. So the the stack the inferencing engine needs to know:
- 7:52
- Unidentified speaker (speaker_1): hey, if it doesn't fit in memory, I need to know where in the stack do I go and take the memory, but also you need to have a high level of an inference router to go and choose where where do I send the current request? And then you need
- 8:12
- Unidentified speaker (speaker_1): to choose a server that is running the correct model
- 8:16
- Unidentified speaker (speaker_1): that may be in the right portion of the data center that is also connected to the right storage that has that context. So with the augmented memory. grid, we're making some modification to the different inference stack
- 8:36
- Unidentified speaker (speaker_1): components to make sure that they would deliver the most optimal results in terms of both token uh throughput, how many tokens can a given infrastructure deliver, and uh the amount of time that customers, uh the end users wait. So the time to first
- 8:56
- Unidentified speaker (speaker_1): token and the time to last token.
- 8:58
- Unidentified speaker (speaker_0): Okay. Yeah, it's very clear. I guess overall, is the bigger economic benefit faster access to context or avoiding to need to repeatedly uh recompute the same context over and over? Where's the bigger economic benefit, you think?
- 9:13
- Unidentified speaker (UNKNOWN): So
- 9:14
- Unidentified speaker (speaker_1): these are two two sides of of essentially the same coin.
- 9:18
- Unidentified speaker (speaker_0): Okay.
- 9:19
- Unidentified speaker (speaker_1): And you know on on the one hand the the most economic direct value is the lack of recompute. So let let's look at two two different inference providers. One is using okay techniques that cash cache to
- 9:39
- Unidentified speaker (speaker_1): 90%, meaning their GPUs actually have to work on 10% of the requests. And this is after you've introduced all of the other tricks you can. So you've put in your uh speculative decoding and quantization and and and other uh means
- 9:59
- Unidentified speaker (speaker_1): of improving the compute, and you compare that to a different provider that for a similar workload can cache up to 98%.
- 10:09
- Unidentified speaker (speaker_1): So your GPUs only have to work on the 2% remaining uh cash misses, the second provider is going to be able to deliver five times the tokens. So they don't if we for the same amount of GPUs.
- 10:27
- Unidentified speaker (speaker_1): for the same amount of uh power for cooling, uh, all of the other environmental effects of these data centers, if you're just able to not compute five times the data, it meaning it means that your output is five times better. But also,
- 10:48
- Unidentified speaker (speaker_1): if you're looking at your custom customers and uh how happy they are with your service your time to first token how long customers wait is going to end up being a fraction so if you're getting from 90 to 98 percent in that in my example
- 11:08
- Unidentified speaker (speaker_1): the time to first token can go down 20fold twentyfold uh because a lot of that a lot of that compute time just doesn't happen in in in many repeatable uh steps. So both the economic value of the deployment and the customer experience
- 11:28
- Unidentified speaker (speaker_1): uh improve in a significant significant way.
- 11:31
- Unidentified speaker (speaker_0): Well can you also walk us through your Axon and Wekapod products and explain how those pieces fit together, maybe along with Neuromesh? Neural Mesh?
- 11:39
- Unidentified speaker (speaker_1): Yeah, so they're they're all highly uh related and connected. So NeuralMesh is our software and that's the the main product we've been working on uh for for the last uh decade or so and that's the most performant most scalable file system that's
- 11:58
- Unidentified speaker (speaker_1): delivering uh um uh multi multi-protocol access Uh this is where our current AL Labs uses for both training and with AMG inference.
- 12:12
- Unidentified speaker (speaker_1): We were able to deploy Neural Mesh on standard uh servers that you can buy from companies like Supermicro or Dell or HillardPacket HewlettPacard Enterprise, Lenovo. It can run on the public. cloud, like AWS, GCP, Azure, OCI.
- 12:32
- Unidentified speaker (speaker_1): Obviously, it runs on many of the of the neo clouds. With Axon, we're actually taking a software that runs as standard storage on dedicated appliance. And because we're pure software, because we use a lot of uh techniques like containers. and microservices, we're
- 12:52
- Unidentified speaker (speaker_1): actually able to run on the GPU servers. In that regard, you can go to a deployment in a public cloud or or a neo cloud or a data center that didn't plan to add uh fast storage like Weka. Uh, maybe they they have a competitor that's not as performant
- 13:12
- Unidentified speaker (speaker_1): and they want to run us, we can then run on top of the GPU servers. We can use the local uh cores, the local memory, the local NVMe drives, and deliver an
- 13:25
- Unidentified speaker (speaker_1): insanely fast solution that scales to dozens and hundreds of thousands of uh of GPUs with no no extra infrastructure. And then finally, Wecapod is our own uh appliance. We
- 13:45
- Unidentified speaker (speaker_1): have realized in recent years that customers, even though they like software and even though we're still a software company, they like the simplicity of just buying uh a ready-made turnkey appliance. And and Wacapod Wekapod is that for us. So essentially takes the normal
- 14:05
- Unidentified speaker (speaker_1): software, couples it with the best hardware, uh, and recently we're even designing that hardware, couples it with a very strong supply chain, and then allows customers to consume us in a very simple, straightforward Forward way.
- 14:22
- Unidentified speaker (speaker_0): Gotcha. And with Wecapod, Wekapod, is that more about density and performance? Or is it also about making deployments easier, maybe giving Wekapod more control over the complete system?
- 14:32
- Unidentified speaker (speaker_1): So it's it's all of these things together. So we started uh working on the Wekapod from the simplicity angle.
- 14:41
- Unidentified speaker (UNKNOWN): You
- 14:42
- Unidentified speaker (speaker_1): know, customers just wanted the Wacop technology, but they didn't want to go and integrate integrated it themselves or through uh a a third party. And that is how we've run that for a couple of generations. Then uh we we actually were frustrated with what we were able to get from from
- 15:02
- Unidentified speaker (speaker_1): the standard uh server makers, with what you can achieve. And since it was a complete product from us, uh we turned towards uh designing our own very efficient. hardware. And we came up with the most power efficient, space efficient uh
- 15:22
- Unidentified speaker (speaker_1): design that in it doesn't matter what you're trying to optimize for, if you want to optimize for more performance density or more capacity density to running different protocols, The cause that would be the most effective, efficient way of running Weka. And
- 15:40
- Unidentified speaker (speaker_1): in
- 15:41
- Unidentified speaker (speaker_1): tandem with that, since we've had our own hardware engineering and supply chain, uh, we started running our own supply chain. And this has been a surprise for us. But at this, and and you know, it's been a surprise for the rest of the market with the insane
- 16:00
- Unidentified speaker (speaker_1): crunch around memory. and flash that the market now uh experiences. But the fact that we actually are controlling supply chain, the fact that we can guarantee you know delivery in two, two or three or four weeks almost anywhere in the world is now becoming
- 16:20
- Unidentified speaker (speaker_1): a very strong uh competitive advantage for us. So while we didn't predict it, having uh having complete control over the the full delivery actually allows us to to win you know some
- 16:40
- Unidentified speaker (speaker_1): deals where um in uh otherwise Is they would have considered other competitors now because we can deliver so easily, they're not even opening it up for a question.
- 16:54
- Unidentified speaker (speaker_0): Yeah, that's super interesting. It seems like some of the traditional hardware vendors are moving up the stack. It seems like uh Weka, which is traditionally, I guess, software, moving down the stack. So everything is kind of converging.
- 17:06
- Unidentified speaker (UNKNOWN): Uh,
- 17:07
- Unidentified speaker (speaker_0): just out of curiosity, one quick question on the Weka Wekapod, pod is that based primarily on TLC or QLC NAND, or do you have uh different flavors?
- 17:16
- Unidentified speaker (speaker_1): So we actually have two two different flavors. We have a model called Nitro, and Nitro is uh something we're focusing on uh inference context or KV cache management to to be the best counterpart for the augmented memory grid. And this is
- 17:36
- Unidentified speaker (speaker_1): a TLC only, so these are the fastest. flash devices that have the most endurance. And you know, when you're caching something, you you keep changing the content. So you need the flash to be able to go through many, many, many write cycles. And then we have a different product called uh Prime with a
- 17:55
- Unidentified speaker (speaker_1): WackerPod Prime, and and that is uh a combination of TLC and QLC that's uh based on our alloy flash technology. And this is where we can steer small IO and metadata to very small amount of TLC. It could be just 1% of the
- 18:15
- Unidentified speaker (speaker_1): complete capacity of the whole capacity, and the bulk of the storage runs on the significantly cheaper, larger QLC device. devices, and that's uh allowing us to deliver much better uh cost efficiency.
- 18:35
- Unidentified speaker (speaker_1): Normally, with the Prime configuration, we're also running uh data reduction.
- 18:42
- Unidentified speaker (speaker_1): So Prime is where we're running native uh
- 18:48
- Unidentified speaker (speaker_1): uncompressed data that just needs to deliver. the highest throughput at the lower latency. And then prime is where we're targeting the more performance
- 18:59
- Unidentified speaker (speaker_1): level but capacity and cost-optimized uh solutions.
- 19:05
- Unidentified speaker (speaker_0): That's fascinating. So zooming back out, I want to talk about what this means for the underlying storage mix. Uh, do you think AI ultimately moves much more data onto Flash, or do you still expect this heavily tiered architecture with object storage and HTD behind the active tier?
- 19:23
- Unidentified speaker (speaker_1): So let's zoom out. If if we're just looking at raw production,
- 19:32
- Unidentified speaker (speaker_1): the amount of hard drive capacity that uh that is being built every year is several times the the amount of SSD or NVMe capacity that is being built. So anyone that tells you that Flash is going to be able to take over
- 19:51
- Unidentified speaker (speaker_1): everything cannot actually deliver on it and is likely coming from uh a strong a strong position. So hard hard drives are here to stay for at least some some years to allow the the nvme production
- 20:11
- Unidentified speaker (speaker_1): capacity to grow but you know at that point likely also the requirements are gonna grow so the the gap would would keep widening having said that
- 20:24
- Unidentified speaker (speaker_1): you know
- 20:26
- Unidentified speaker (speaker_1): anything that's active is gonna sit on flash, and we're going to leverage hard drives for archival and backup use cases.
- 20:37
- Unidentified speaker (speaker_0): Very clear. So we have heard from or I've actually read a an interesting white paper from Heinex Hynix, um, where they're actually combining the HDD, uh as you alluded to, as an inactive T. tier for for specifically for KV cash to sit behind the active flash layer. Is that something you've been experimenting with or you've
- 20:57
- Unidentified speaker (speaker_0): um been discussing with customers?
- 21:00
- Unidentified speaker (speaker_1): Yeah, so we actually have been tiering to hard drive based uh object storages from from the get go, from the first commercial uh GA release we've had. Uh we kept this functionality that allowed you to extend the namespace from
- 21:20
- Unidentified speaker (speaker_1): uh an NVMe hot tier uh to a hard drive-based object storage for for the capacity we're making that management actually quite straightforward and uh very easy to to deploy. and that's going
- 21:39
- Unidentified speaker (speaker_1): to allow customers a seamless transition where all of the active, very short-term memory, uh for or or KV context is going to sit on flash, but still they would be able to go and and read older sessions from hard
- 21:59
- Unidentified speaker (speaker_1): drives, so you won't have to go and waste a lot. of GPU recomputation even on older ones. And and I'm sure uh either you and and and the listeners are experiencing this when we're working with our preferred AI uh systems, could be
- 22:19
- Unidentified speaker (speaker_1): like chat or Claude or Gemini.
- 22:22
- Unidentified speaker (speaker_1): Many times we're going back to
- 22:25
- Unidentified speaker (UNKNOWN): uh
- 22:26
- Unidentified speaker (speaker_1): to sessions. that we were running last week or maybe even last month, and we find value in getting these results back. You may experience the the first time you come back to it, it takes a bit longer. And this is because it's it's being pulled uh off of sort of an
- 22:45
- Unidentified speaker (speaker_1): archive, arcade, because it's not economical to keep uh you know a month. old session on on flash. So this is as AI becomes more prevalent and as more users use it, uh, we're going to see the case of going back to older sessions
- 23:05
- Unidentified speaker (speaker_1): even more important. And this is where natural transition from NVMe to hard drives uh is going to come in really. important.
- 23:16
- Unidentified speaker (speaker_0): Right. So honing back in specifically on Weka, um curious, where are you seeing the strongest demand today, would you say? And where's the biggest opportunity? Would it be hyperscalers, the neo clouds, sovereign AI, or maybe traditional enterprise?
- 23:30
- Unidentified speaker (speaker_1): What we're seeing where the strongest pull uh is currently happening is between the neo. clouds and
- 23:41
- Unidentified speaker (speaker_1): some of the sovereign neo clouds. These are technically companies that solve the same problem, but they many times do it for different uh
- 23:53
- Unidentified speaker (UNKNOWN): you
- 23:53
- Unidentified speaker (speaker_1): know business reason. The infrastructure, these neo-cloud build is different than what you get. in in the hyperscaler. They're able to keep track on the latest generation from NVIDIA and AMD, and usually the hyperscalers have
- 24:14
- Unidentified speaker (speaker_1): different architecture that they've built for a different era, especially around the network. The enterprises
- 24:23
- Unidentified speaker (speaker_1): are
- 24:25
- Unidentified speaker (speaker_1): not able. to match this kind of infrastructure. You know, you need uh liquid cooling, uh high voltage uh DC uh delivery within the data center, incredibly fast networking. So when when you're
- 24:44
- Unidentified speaker (speaker_1): looking at the large amount of enterprises, they they just can't match it. When you're looking at a lot of the AI labs, it's very difficult for them to do it. And that is the reason we're seeing the neo clouds uh essentially taking the role that traditional uh on
- 25:04
- Unidentified speaker (speaker_1): on premises or hosted environments delivered, and and they take the the lion's share of of the build out. outs and then in turn they have off takers that are either governments or uh ai labs or enterprise customers.
- 25:24
- Unidentified speaker (speaker_0): So specifically on sovereign AI, why do you think that's become such a large opportunity? What are those customers trying to accomplish that they can't simply get from a hyperscaler?
- 25:34
- Unidentified speaker (speaker_1): So sovereign AI is is getting more and more important as many countries realize that they have to take control
- 25:45
- Unidentified speaker (speaker_1): over intelligence. They understand a large portion of
- 25:51
- Unidentified speaker (speaker_1): the upcoming economy is going to rely on
- 25:57
- Unidentified speaker (speaker_1): on artificial intelligence, on these tools, and they don't want uh to to take the risk of depending on a different country in order to be able to run their economy in in the future. You know, we've we've had the scare
- 26:17
- Unidentified speaker (speaker_1): of having uh some models being locked out of uh uh foreign countries if you remember the introduction of uh Fable from Anthropic.
- 26:30
- Unidentified speaker (speaker_1): That has pushed a lot of them. Another aspect is in especially in Europe with G GDPR rules, with a concern that hey, when I'm sending uh these API calls to to the models, even if they say that they don't, there is some residual memory that
- 26:50
- Unidentified speaker (speaker_1): stays with. within the models. They want all of the original content to stay within the country.
- 27:09
- Unidentified speaker (speaker_1): or companies like Mistral when or Cohere when when they're promising hey we'll we'll keep this model especially closed for you are having a big win with with this kind of uh deployment.
- 27:25
- Unidentified speaker (speaker_0): I see so as these opportunities grow of course the competitive landscape is evolving as well.
- 27:30
- Unidentified speaker (UNKNOWN): Who
- 27:30
- Unidentified speaker (speaker_0): do you view as WEC's closest competitors today? And does that change depending on the workload or customer?
- 27:36
- Unidentified speaker (speaker_1): Our competitors essentially are the companies that Nvidia and AMD both choose to work with. You know, you go and you look at their website and and you figure out what uh what companies end up being in in the reference architecture.
- 27:56
- Unidentified speaker (speaker_1): These pro projects are insanely complex if you don't get the
- 28:03
- Unidentified speaker (speaker_1): the vote of trust from you know the the GPU maker, the network technology maker, if you are not highly confident all of these things are going to play well together,
- 28:20
- Unidentified speaker (UNKNOWN): uh
- 28:20
- Unidentified speaker (speaker_1): it's too high of a risk uh to actually go and
- 28:25
- Unidentified speaker (UNKNOWN): and
- 28:25
- Unidentified speaker (speaker_1): plan this project. So the
- 28:29
- Unidentified speaker (speaker_1): ecosystem essentially is
- 28:33
- Unidentified speaker (speaker_1): targeting a very, very small number of vendors that successfully are able to deliver these projects.
- 28:41
- Unidentified speaker (speaker_0): Okay. And what do you view as genuinely hard to replicate in AI storage today? Is it performance, predictable tail latency?
- 28:49
- Unidentified speaker (UNKNOWN): Is
- 28:49
- Unidentified speaker (speaker_0): it metadata scan? scaling, data movement, mixed workload support, or maybe something else?
- 28:54
- Unidentified speaker (speaker_1): So it's a lot of these things combined. Uh, you know, the the security requirements are becoming more and more prevalent,
- 29:04
- Unidentified speaker (UNKNOWN): uh,
- 29:05
- Unidentified speaker (speaker_1): especially in the sovereign AI, but also when we're looking into the enterprise. So we need to come up with systems that are highly secure and and and obviously there's some overhead there that deliver very low latency for a very big
- 29:24
- Unidentified speaker (speaker_1): amount of uh of data, uh a lot of metadata is involved. So the the kind of problems that now the industry industry's solve
- 29:34
- Unidentified speaker (UNKNOWN): uh
- 29:35
- Unidentified speaker (speaker_1): then the industry solves are so much harder than than what we used to solve, you know, fifteen or twenty years ago. So the the products are fundamentally different.
- 29:47
- Unidentified speaker (speaker_0): I see. And I think that's a great setup for what Weka is announcing today. So Liron, there was a lot of activity around Weka and its partners at Supercompute. Can you bring us up to speed on the key announcements and why they matter?
- 30:01
- Unidentified speaker (speaker_1): Yeah, so
- 30:02
- Unidentified speaker (speaker_1): we we acknowledge the fact that
- 30:07
- Unidentified speaker (speaker_1): building infrastructure for inference is difficult. We acknowledge the fact that choosing the right kind of uh technology is hard, that deciphering uh
- 30:22
- Unidentified speaker (speaker_1): what kind of architecture is going uh to deliver the best result
- 30:27
- Unidentified speaker (speaker_1): is actually quite confusing. So we have worked with NAND Research to build
- 30:36
- Unidentified speaker (UNKNOWN): uh
- 30:36
- Unidentified speaker (speaker_1): a new inference economics scorecard where we've taken a lot of our and our customers' experience on how do you take what really matters in real life. and and deliver a few metrics uh that that actually
- 30:56
- Unidentified speaker (speaker_1): matter
- 30:57
- Unidentified speaker (speaker_1): and we've created this benchmark that allows you to go you can test a technology stack at a smaller scale you can compare what you're seeing with uh with other results unlike other benchmarks we actually run it uh for
- 31:17
- Unidentified speaker (speaker_1): longer period of time. So caching does make a difference. So you need to run it for at least uh eight hours. And we're showing around, hey,
- 31:29
- Unidentified speaker (speaker_1): what is the uh token throughput you're getting? What's the total amount of tokens you're getting? What's the token per watt? What is the token per dollar? dollars across the different kind of metrics that you run? You're actually able to
- 31:46
- Unidentified speaker (speaker_1): go and answer the question for the next build-outs that I'm planning, what's going to be the the best mix of technologies to to use?
- 31:58
- Unidentified speaker (speaker_0): That's truly fascinating. So taking today's announcement together. with everything we've discussed, finish the sentence for me. Five years from now, the most important thing about AI storage will be?
- 32:10
- Unidentified speaker (speaker_1): Five years from now, the most important thing about AI storage is going to be that it will not look like storage. The storage today has been designed for humans.
- 32:23
- Unidentified speaker (speaker_1): And we we're getting AI agents to to use it now as humans as we go down the stack the the boundaries between the short the short term kv cache memory uh the
- 32:40
- Unidentified speaker (UNKNOWN): uh uh
- 32:40
- Unidentified speaker (speaker_1): the context memory the semantic memory and you probably have heard of vector databases uh cat catalogs all all of these things are going to blend together, and we're going to start building systems that are best uh used by the
- 33:00
- Unidentified speaker (speaker_1): AI agents and are best used by AI systems that optimize for the quickest response time and the highest uh throughput, and that is going to allow humans to get a lot a lot more outcome out
- 33:19
- Unidentified speaker (speaker_1): of the AI system that they're running.
- 33:23
- Unidentified speaker (speaker_0): It's truly fascinating stuff. And I think that's a great place to leave it. Thank you for the time and insights, Leron. It's certainly an exciting time to be square at the intersection of AI, storage, memory, and data intelligence.
- 33:36
- Unidentified speaker (UNKNOWN): To
- 33:36
- Unidentified speaker (speaker_0): our listeners, thank you for tuning in.
- 33:38
- Unidentified speaker (UNKNOWN): If
- 33:38
- Unidentified speaker (speaker_0): you like the episode, please subscribe and leave a review and be sure to check back to hear conversations with leading disruptors in the tech landscape.
- 33:46
- Unidentified speaker (UNKNOWN): If
- 33:46
- Unidentified speaker (speaker_0): you want to learn more about our research, including deep dives on all things tech, check out our work on the Bloomberg Terminal at BIGo.
- 33:53
- Unidentified speaker (UNKNOWN): This
- 33:54
- Unidentified speaker (speaker_0): is your host, Derek Mogul, signing off.
More from Tech Disruptors
Honeywell CTO on AI and Physical-World Autonomy
Episode discussion: Honeywell CTO outlines AI-driven autonomy strategy, citing edge computing and deterministic guardrails for industrial applications
Speakers’ claims: Mustafa Okur interviews Suresh Venkatarayalu, CTO and President of Forge at Honeywell, regarding the company's pivot from automation to autonomy. Venkatarayalu explains that Honeywell has simplified its portfolio to focus on commercial buildings and industrial process automation, leveraging its Forge platform to integrate AI with deterministic control systems. He describes a three-phase AI strategy involving internal knowledge models, partnerships with Nvidia for open-source models in secure hardware, and hybrid architectures for data sovereignty. The discussion highlights use cases in reducing false alarms in ethylene plants and building management systems, as well as collaborations with Google, LS Electric, and Dangote to expand ontologies and meet cybersecurity mandates.
Figma on Why Design Gains Value in the AI Era
Episode discussion: Kris Rasmussen on Figma's AI Strategy: Design Context as a Moat and Hybrid Pricing Models
Speakers’ claims: Kris Rasmussen, Chief Architect at Figma, discusses how AI agents serve as interface glue rather than replacements for traditional UIs, emphasizing that design systems provide critical context for better AI outputs. He argues that as code becomes commoditized, value shifts to design, with Figma leveraging proprietary user data to fine-tune open-weight models for specific workflows. Rasmussen highlights bottlenecks in AI adoption due to security and backend expertise gaps, noting that human curation remains essential. He also outlines Figma's infrastructure strategy, which avoids building foundation models from scratch, and predicts a shift toward hybrid seat-based and consumption-based pricing models.