Memory vs Storage in AI Infrastructure Design

Great comments Christopher Ratcliffe! PEAK:AIO - When was the last time you heard a vendor say “Buy less” Vendors are not talking about efficiency. Getting the performance you need out of existing infrastructure or optimizing budget/resources you have. Let us show you how to get better performance for less Complexity, Infrastructure, and Cost. PEAK:AIO

What caught my attention in yesterday's keynote was how confidently both Jensen Huang and Michael Dell framed where we are today as an infrastructure problem, not a model problem, not a data problem, not a talent problem. The models are ready. The compute is available, even if it is supply-constrained. The bottleneck is the storage layer that was designed before modern AI workloads existed (i.e. before about 2020). That's the gap PEAK:AIO is talking about in today's post. It's worth a quick read if AI infrastructure is on your radar.

The assumptions that matter most for AI infrastructure DCFs have the least historical evidence behind them. This chart plots ten of those assumptions on two axes: how much each one moves equity returns, and how much historical data exists to anchor it. Bubble size reflects approximate capital exposure. The upper-right quadrant — where high-importance assumptions would be supported by deep data — is empty. The three assumptions with the largest impact on equity returns sit in the upper-left: AI demand durability — projected over a 25-year asset life, from roughly three years of observable market history. GPU residual values — projected across four or five refresh cycles, from a secondary market that barely exists. Chip replacement cycles — projected at an assumed cadence, against a chip roadmap that is now moving on roughly annual timelines. These are the variables that determine whether a 1 GW asset clears financing, supports its capital structure, and generates acceptable long-term equity returns. They are also the variables conventional underwriting has the least machinery to handle. That is the financeability problem named: the model is most sensitive exactly where the evidence is thinnest. Risk transfer is not enough. The first task is risk legibility — making these assumptions explicit, modelable, stressable, and priced. Link to newsletter in comments below. #AIInfrastructure #DataCenters #InfrastructureFinance #RiskTransfer #CapitalMarkets #ReinventingInsurance

Planning AI growth? Scale only when you need to. As AI projects grow, workloads increase — more data, more models, more inference demand. Rather than committing to large-scale infrastructure upfront, consider cluster deployment with multiple ThinkStation PGX systems. Why cluster deployment makes sense: ✅ Scale performance progressively ✅ Expand capacity in response to actual demand ✅ Avoid over-provisioning based on architecture constraints Plan your AI scaling strategy with Edge Computers and expand when demand requires it. 🌐 https://lnkd.in/egudUnZJ #EdgeComputers #ScalableAI #EnterpriseAI #AIInfrastructure

Palantir and Vertiv outpaced Nvidia in market-cap growth during the AI build-out — infrastructure bottlenecks, not compute, drove the biggest gains New data from BestBrokers across 20 publicly listed AI supply chain companies shows that the largest percentage winners between 2021 and 2025 were an enterprise AI software platform and a data centre cooling provider, not a chipmaker. The market priced the cost of removing physical constraints. Read the full piece: https://lnkd.in/eSMiASp6 #CTC #PressRelease #AI #CloudInfrastructure #DataCentre

Most enterprises think they can stitch together their own AI infrastructure—until the integration tax starts compounding every quarter. The real problem isn't buying the GPUs. It's operationalizing them: tenant isolation, multi-tenancy, GPU partitioning, day-two operations, and staying current as the tech evolves every few months. Swapnil Bhartiya sat down with Richard Borenstein, SVP of Growth & Business Development at Mirantis, to unpack why composable, pre-integrated AI infrastructure is becoming table stakes for the Fortune 1000. “The real risk isn’t the licensing cost or even the time to build. It’s that, by the time you finish building, the technology has moved on, and you’re maintaining a deprecated architecture.”  In this full conversation at TFiR, we explore: the hidden DIY tax of AI infrastructure, how Mirantis k0rdent AI delivers a composable GPU cloud in a box, the OpenStack lessons that shaped Mirantis's approach to enterprise AI, digital sovereignty, inference routing, and what mindset shift enterprises need to make in 2026. Check out the discussion on our YouTube page: https://lnkd.in/gzyJnamU #AIInfrastructure #Kubernetes #GPUCloud #OpenSource #CloudNative #Mirantis #EnterpriseAI #DigitalSovereignty #NVIDIA #MLOps

Richard Borenstein, SVP of Growth & Business Development at Mirantis, breaks down why the DIY approach to AI infrastructure is failing enterprises and how Mirantis k0rdent AI, a composable, open-source-rooted platform built on decades of OpenStack and Kubernetes expertise, gives organizations a sovereign, vendor-neutral path to production-grade AI at scale. Check out the discussion: https://buff.ly/wwKFXxH

💡 AI infrastructure is no longer centralized. Inference workloads are now distributed across clouds, edge locations, sovereign environments, and regional GPU clusters. That changes everything about how requests must be routed, managed, and optimized.   Every inference decision now depends on real-time capacity, latency targets, model availability, KV-cache affinity, residency requirements, and inter-site network conditions. Traditional infrastructure was never designed for this level of orchestration. ➡️Read Part 1 of the AINF blog series by Nalin Pai, exploring the challenges shaping the inference fabric era: https://lnkd.in/eWjAVG93   The next era of AI will not be defined by compute alone. It will be defined by how intelligently infrastructure coordinates distributed inference at scale. #Arrcus #NetworkDifferent #AIInfrastructure #AIInference #DistributedSystems #AINF

AI inference is moving across clouds, edge, sovereign environments, and regional GPU clusters. In my latest blog post, I break down why distributed inference needs a new kind of fabric

Managing high-volume AI workloads through real-time inference means rate limits, retries, & expensive synchronous compute that don’t scale economically. There's a simpler way. 👀 Unified Batch Inference on DigitalOcean lets you run large-scale inference asynchronously through a single API, with integrated Spaces storage, a live job queue, & unified provider access. Process millions of requests overnight at up to half the cost of real-time inference. Your production traffic never competes with batch workloads. 🚫🚙🚗