AI Infrastructure & Workflows
Russ Artzt has built his career around moments when computing changes shape, when yesterday’s standard stack starts to look like a constraint. A co founder of CA Technologies and later executive chairman and head of R&D at RingLead, acquired by ZoomInfo, Artzt has watched infrastructure evolve from mainframes, to client server, to cloud. In our conversation, he argued we are now in the middle of another re-platforming, one that is reorganizing AI infrastructure into distinct layers, and creating a fast growing category of compute providers in the process, neoclouds.
The timing matters. The AI infrastructure story has shifted in the news from who has the best model, to who can actually supply, power, and deliver capacity. Reuters has chronicled the capital intensity of this new compute layer, CoreWeave alone has described plans to spend tens of billions on infrastructure and data center capacity to meet demand. Reuters has also documented the rise of European neocloud players like Nebius signing multi billion dollar contracts to supply AI infrastructure to hyperscalers, alongside the surge in capacity being sold to neocloud providers.
Artzt’s thesis is simple, almost architectural, the generative AI boom did not just create new applications. It created a new industrial supply chain, chips, networking, data centers, and workflow integration, and that supply chain is reshaping who wins.
Artzt traces the current wave to the moment large language models went mainstream.
“A couple of years ago when OpenAI started, ChatGPT was the first real large language model that everybody saw great value in,” he told me. “It was incredible. It is a real breakthrough, and they just keep getting better.”
Once that value became obvious, the market did what markets do, it replicated the pattern at scale. “Everybody else came in and started writing their own large language models,” he said, sometimes building models outright, sometimes building applications on top of them, tuned for specific workflows.
“The AI landscape is just growing like crazy,” he said.
But in Artzt’s framing, the most consequential part of the story is what sits underneath. “It is like building a house,” he said. “You gotta build a strong infrastructure, otherwise your house is gonna fall apart.”
That house is now a visible constraint. Analysts and investors are increasingly treating the power and infrastructure side of AI as the bottleneck worth investing in, Reuters recently reported on a BlackRock survey where clients favored power providers and infrastructure as their top AI investment angle.
Artzt’s account of NVIDIA is not just that it sells GPUs. It is that NVIDIA assembled a coherent platform, hardware, networking, and a developer ecosystem, at exactly the moment AI workloads demanded it.
“It started with the video game environment,” he said. NVIDIA spent years dominating graphics, then turned that capability into the modern GPU. And the GPU, in his view, is not simply a faster CPU, it is a different kind of computing machine.
“GPUs are very, very different,” he said. “It can multitask much better than the typical Intel CPUs, it ends up doing parallel processing and ends up being very fast for the kind of compute you need when you are doing a large language model processing.”
Then comes the move he thinks truly separated NVIDIA from the field, they did not stop at the chip.
“NVIDIA said, we are going to build the GPUs and we are going to build the whole network,” he said. AI at scale is a clustering problem, connecting accelerators into systems that can move data quickly and shift workloads dynamically. “You need network switches so you can network from one GPU to another,” he explained.
On top of that sits the software layer developers rely on to extract performance from expensive hardware, “a software layer called CUDA,” he said. “Everybody uses it because they want to get the optimal use and performance out of NVIDIA chips, because they are expensive.”
This platform framing matters because it explains why demand concentrates. If the base layer is standardized around NVIDIA’s stack, then everyone above it, hyperscalers, neoclouds, and enterprises, has a gravitational pull toward the same supplier.
In Artzt’s telling, neoclouds are the clearest sign that AI infrastructure is reorganizing into layers. They are not full service hyperscalers. They are GPU specialists, providers built primarily to deliver AI compute as a service.
Real estate services firm JLL recently estimated the neocloud ecosystem at approximately 190 distinct operators, with major providers including CoreWeave, Nebius, and Crusoe. Other analyst lists name similar leaders, CoreWeave, Lambda, Nebius, Vultr, and more, reflecting how quickly the category has expanded.
So why did they emerge now, because the AI boom produced a specific mismatch, GPU demand spiked faster than most companies could procure hardware or justify hyperscaler economics. Artzt described it bluntly.
“It is hard to get GPUs because of this demand,” he said. “What is happening is what is called neocloud. Providers are popping up, companies like CoreWeave and Lambda, competing with Amazon Cloud, Google Cloud, Microsoft Cloud.”
CoreWeave is a useful case study because its arc mirrors the market. Reuters reported as far back as 2023 that CoreWeave raised 2.3 billion dollars in debt financing collateralized by NVIDIA chips, financing meant to acquire more GPUs, invest in data centers, and hire, which shows how early GPU capacity became bankable in its own right.
By 2025, Reuters reported CoreWeave signed a five year 11.9 billion dollar contract with OpenAI, a landmark deal that formalized the notion of a specialist compute supplier becoming core infrastructure for one of the most important AI companies in the world. By late 2025, Reuters reported another large capacity deal, 14.2 billion dollars with Meta through 2031, as CoreWeave diversified beyond its early concentration.
At the same time, Reuters reported European neocloud providers rising rapidly, including Nebius signing a $3billion dollar deal with Meta and pursuing aggressive expansion to meet demand.
This is where Artzt’s rent vs buy framing becomes important. Neoclouds are not merely cheaper, they change the capital decision.
“If I want to rent GPUs, I do not have to buy them,” he said. “I could rent them, use them for a period of time, and when my project is done, get rid of it.”
That logic is especially appealing when ROI is uncertain, procurement lead times are long, and GPU generations move fast enough that ownership risks obsolescence. Neoclouds are the infrastructure expression of a broader shift, compute is becoming more liquid, and capacity is becoming a market.
Neoclouds exist because compute is scarce and expensive. But building neoclouds introduces a new constraint, physical infrastructure and operational execution.
“They have to create data centers to support their customers,” Artzt said. “But in some cases, they have not been able to build the data centers fast enough to accommodate the needs.”
He described customer requirements that sound almost absurd until you remember the scale of modern clusters, a thousand GPUs, a hundred gigawatts of power, air conditioning, plumbing, and a demand to deliver quickly.
That aligns with what Reuters has been reporting about the capital intensity and build urgency of the category, CoreWeave’s stated spending plans and the surge in capacity sold to neocloud providers reflect the same underlying reality, the market wants infrastructure faster than it can be delivered.
But Artzt’s emphasis is not only money. It is the operational talent required to stitch the whole system together.
“I think the biggest problem is gonna be expertise,” he said. “People who know how to do it.”
Modern AI data centers are not just GPU warehouses. They are heterogeneous systems, CPUs, GPUs, networking fabric, and storage, and the integration is non trivial. “These data centers need CPUs and GPUs connected together with networks, with network switches, with storage, it all has gotta come together,” he said. Then he asked the question that sits under many AI infrastructure debates, “Who knows how to do it, nobody.”
In the neocloud model, the compute layer is increasingly flexible, you can rent GPUs from a hyperscaler, a specialist, or a regional provider. The data layer is the opposite, it needs to be stable, affordable, and always available, because AI workflows repeatedly pull data into compute and push results back out.
In modern AI stacks, storage is the home base for unstructured data, PDFs, images, audio, video, logs, and internal knowledge bases. Compute is where that data gets processed, for training, fine tuning, RAG indexing, embedding generation, batch inference, and document parsing. Outputs then flow back to storage, embeddings, indexes, transformed datasets, logs, checkpoints, and results. The same dataset can move many times across stages of the pipeline and across vendors.
This is also why object storage services like Backblaze are increasingly being used as a stable data layer in multi provider stacks. Teams keep large datasets in object storage, then push data into GPU compute wherever it makes the most sense, hyperscaler instances for some workloads, neocloud GPU capacity for others.
“Backblaze should be where a customer should be able to store all their videos and audio and images and PDFs, all go into Backblaze’s repository,” Artzt told me. “Put them all in Backblaze. We will manage it for you.”
The practical reason this pattern matters is economics. In many cloud environments, moving data out of a platform triggers egress fees, and those costs compound when workflows repeatedly shuttle data to GPU clusters and back.
“If you store the data in Amazon and you move it around, you pay what is called an egress charge,” Artzt said.
Those data movement economics are now a competitive battleground. Reuters reported that Google Cloud removed certain network fees for customers migrating to another provider, and AWS announced a similar leaving focused waiver, but both come with conditions and process steps that keep egress economics very real for everyday multi cloud workflows.
Artzt’s point is that storage vendors become strategic when they plug directly into neocloud compute, so data movement becomes a workflow feature, not an integration project.
“You want to egress it, probably a neocloud provider, analyze it, and when he is done processing, he will send it all back,” he said. “So what we need to do is build integration between the two.”
That word, integration, shows up repeatedly in Artzt’s thinking. In this stack, storage wins not by being cheaper, but by becoming workflow infrastructure, the layer that makes the data to GPU loop predictable and frictionless.
“They need connectors, seamless integration,” he said. “I hit a button and it goes. It runs. I want you to do some GPU processing, boom.”
Artzt also highlighted a practical detail that determines adoption, standards compatibility.
“There’s a standard, it’s called S3,” he said. “Amazon created it and all the storage players have some level of compatibility to S3.”
S3 compatibility is how a multi provider stack becomes feasible. Tools think they are talking to AWS, but can route to compatible storage elsewhere. In Artzt’s words, “It thinks it is talking to Amazon, but it is talking to Backblaze.”
But compatibility is rarely perfect. “It doesn’t support it a hundred percent,” he said, and that is where competitive differentiation increasingly lives, in closing the right gaps for AI workflows, not in claiming compatibility in the abstract.
Artzt’s stack view produces a few practical takeaways for infrastructure buyers.
First, expect specialization to increase. Hyperscalers will remain dominant, but neoclouds are rapidly expanding as a compute layer focused on GPUs and AI workloads.
Second, treat rent vs buy as a strategic choice. Renting GPUs via neoclouds can reduce procurement friction, avoid obsolescence risk, and match spend to uncertain AI ROI, but it still demands thoughtful integration and data movement planning.
Third, plan for a systems delivery bottleneck. Chips matter, but the buildout is gated by power, facility readiness, and expertise.
Fourth, treat storage as workflow infrastructure, not an endpoint. The value is not just in storing data, it is in enabling movement and processing across the AI pipeline with minimal friction and predictable cost.
The AI infrastructure stack may be reorganizing into layers, but adoption does not happen in diagrams. It happens when a workflow works, when the economics are clear, and when the integration is good enough that a team can hit a button and trust what happens next.
Vaibhavi is an investor at GTMfund, focused on Pre-Seed and Seed B2B software. She’s known for her founder-first approach, strong GTM instincts, and ability to connect early teams with the right operators at the right moment.
