Practitioner Validation
Voices from the field — how eighteen operators independently reinforce the five-thesis architecture.
Sources 18 expert interviews
Clusters 9 convergent themes
Voices 56 verbatim quotations
A note from Pragmatic Capital
We believe in learning from those doing the work.
At Pragmatic, our convictions are sharpened by the people closest to the build. For this volume, we interviewed dozens of practitioners who have spent their careers in technology — hyperscaler product leaders, silicon executives, frontier-lab veterans, power engineers, data-center operators, and the financiers who underwrite what gets built.
What follows is a distillation of the most informative perspectives. The voices are verbatim, organized into nine convergent clusters, and chosen because they describe — from nine different vantage points — the same architecture we have been arguing from first principles.
Across ten interviews spanning hyperscalers, silicon vendors, and frontier labs, the operators describe the same trajectory: the training-inference mix is inverting, inference demand is accelerating in late 2025 and into 2026, and the per-query economics that matter most to cloud P&Ls sit on the inference side of the ledger.
“Over the next five years, the workloads from training will go from 80–90% today, probably down to inversion — 10% inference or 15–20% down to probably 70/30 or 80/20 inference to training.”
Equinix Capacity & AI Demand
“This is what I can tell you. Everything is moving towards inference. I believe previously, it was maybe 70% training, 30% inference. Moving forward, I believe it's going to be more 50/50 in the next year or so, but moving after this, it's going to tilt much more towards inference.”
Lead Data Executive, AI/IoT Cloud at Microsoft
“Right now, the utilization is split 70/30 — 30% for inference, 70% for training. McKinsey and others had contemplated that by the end of 2026, that inflection will go where there will be 4x more inference as training.”
Former Senior Product Executive, AWS
“If we focus on inference, inference is starting to pick up Q2, Q3, Q4 '25 and entering '26 is when inference would accelerate. The inference demand is driving further addition to your existing GPU clusters.”
Former Senior Executive, AWS SageMaker
“They need to lower the cost of inference, because ChatGPT has around 850 million weekly active users, and that's not cheap. When we think about future models, it's going to be less about 'our model is more than our last' and more of 'our model is able to handle specific popular workloads cheaper and more efficiently than our prior model.'”
Former Senior Executive, ChatGPT at OpenAI
“I expect that just inferencing alone will be roughly $300bn in 2026. Again, just looking at almost an exponential 100% growth.”
GPU Sector Architecture & Capex
“It's still towards the training right now more. If I were to put a number, it's somewhere between 67% and 33% inference. If I have to say five years from now, it's going to flip — it's going to be on the reverse side.”
Enterprise Architect, Google
The operators who actually site, power, and build data centers describe an unusually hard constraint: megawatt pricing has risen ~20% post-COVID, lead times for grid interconnection stretch to four years, and the market has moved structurally from a buyer's to a seller's posture. Multiple independent voices converge on the same scale — from 100MW minimums up to single sites at 1GW.
“The old standard for building a data center for the last 10–12 years has been about $10m a megawatt. That was until COVID. In today's market it's about 20% overall, so it's about anywhere from $12m to $14m per megawatt to build a data center. In some cases, it goes higher.”
Former Senior Executive, Data-Center Site Selection at Nvidia
“In today's market where very few utilities and grid areas have available capacity of the size that we want for our new data centers — which means 100MW minimum up to 600–800MW — there's so few sites that have that kind of power. It's very much a seller's market.”
Former Senior Executive, Data-Center Site Selection at Nvidia
“If you can't get power, you've got to wait four years for new power. That's unsolvable because it still doesn't pencil out to build your own power generation. It's still very expensive and it takes so long that you're willing to pay for a site with power.”
Former Senior Executive, Data-Center Site Selection at Nvidia
“Vacancy is very low right now. In certain markets, it's almost non-existent. If you have access to power and land, then that's the number one constraint right now around the world.”
Equinix Capacity & AI Demand
“The lead times and interconnection lead times in various utility networks around the world could be as short as 1–2 years to as much as three or four years.”
Equinix Capacity & AI Demand
“When you hear Andy Jassy or Sundar or Satya sharing they added 1GW this quarter or 1.2GW a specific quarter, it doesn't mean they have their data centers ready. It's just a number which they are throwing there because they signed a land lease deal, but they have to yet figure out how the power would get there.”
Former Senior Executive, AWS SageMaker
“Today, there are severe constraints on the chips, on the networking side and also on the power side. Even that capacity is low — it takes one year versus 3–4 years for a nuclear facility to be up and running from conception.”
Former Power Engineer, AWS
“Different companies are trying to find pockets of power that they have available in their facilities and trying to maximize it as they can. Those data centers are sold out, but they're trying to find pockets of power in facilities to fulfill those demands today. It's not as easy as it was before.”
PE Data-Center Infrastructure
The "are they spending too much?" question is where the skeptics camp. The operators who sat inside the Azure, AWS, and Anthropic capex conversations describe the opposite: documented, recurring revenue already justifies the bulk of the spend, and the hyperscalers that can lock in capacity today are doing so because demand is outrunning supply.
“About 70%, about $57bn for backward-looking FY '26 is justified with revenue. It's justified by documented recurring revenue from Azure OpenAI Service, from OpenAI production hosting, from Copilot subscriptions, and from Azure migrations that consume AI — so 71% is justified.”
Microsoft AI/Cloud Strategy
“Microsoft is in actually really good shape right now compared to their competitors in terms of their CAPEX spend being justified by likely sources of revenue moving forward. I don't think it's an elephant in the room anymore.”
Microsoft AI/Cloud Strategy
“I think that the frontier-level intelligence have just a massive level of demand and willingness to pay, and that's where I think there will be a significant margin for frontier-level labs — and also why they're willing to make the $100bn bets on compute in order to stay on the frontier.”
Former Senior Executive, Anthropic
“The hyperscalers, every one of them is adding capacity, expanding their sites with a philosophy that you build and the customers will come, and that has been the case so far. This demand will sustain.”
Former Senior Executive, AWS SageMaker
“I have confidence that this is going to be a good return on investment starting '27, '28, '29 when most of the enterprises would be logged into an ecosystem for all of their AI spend.”
Former Senior Executive, AWS SageMaker
“Despite skepticism, it seems we're seeing real demand right now from the revenue reporting from OpenAI and others.”
OpenAI Enterprise Partnerships
“If any of the demand forecast materialized, there's no way that we're going to have the power in place to support the growth plans across the board based on the utility plans in various countries around the world.”
Equinix Capacity & AI Demand
“It would definitely be bigger companies such as hyperscalers or enterprise companies or AI technology companies that can secure a building before it's built. The ones that need high-density environments are pre-buying and they're locking in ahead of time.”
PE Data-Center Infrastructure
Frontier labs, hyperscaler product leaders, and IBM all land on the same observation: enterprises — and the AI labs themselves — systematically refuse to commit to a single cloud. The pattern is not sentimental; it is about optionality, pricing leverage, regional scale, and resilience. This is the demand condition that makes a credibly-neutral infrastructure layer structurally valuable.
“Diversification in cloud providers and in hardware access is definitely the smart way to go. That's why you're starting to see Anthropic have partnerships with not only just Amazon, but then also Google.”
Former Senior Executive, ChatGPT at OpenAI
“As for the deals OpenAI has made with Oracle and other hyperscalers, it's part of a broader strategy to diversify infrastructure needs and create a reliable multi-cloud setup. These types of deals allow OpenAI to scale while also maintaining flexibility, but also help mitigate risk if one provider becomes constrained.”
Former Senior AI Engineer, OpenAI
“The larger enterprises are definitely going to prefer the interoperability — having access to multiple models and the flexibility given agentic workflows to pick the economically right or efficient or cheap model for the right kind of complexity and nuance of workflow or task.”
OpenAI Enterprise Partnerships
“It's going in the direction more of how enterprise customers today are preferring to go with more of a hybrid cloud model rather than locked into one cloud service. You can think of that as a similar expectation from Anthropic's point of view.”
Former Power Engineer, AWS
“It is a platform that we view as one of portability and to avoid lock-in. Clients that want to move to an all-AWS-all-the-time or an all-Azure-all-the-time strategy is just building themselves into another technical debt scenario.”
IBM Enterprise Modernization
The cloud architects are explicit about the rule that governs where AI workloads actually run: compute follows data, because moving data is expensive, slow, and regulated. The IBM mainframe voice and the Databricks voice both describe the downstream consequence — enterprise data is sticky, inference wants to be near it, and the data-infrastructure spend compounds as AI usage deepens.
“The rule of thumb that most of these companies will follow is: ideally, you don't want to go to the models — you want the models to come to the data.”
Former Senior Product Executive, AWS
“Right now, I think the latest thing that I saw out there was that 70% of all the business in the world touches a mainframe somewhere along that path, if you will, to completion.”
IBM Enterprise Modernization
“To convert COBOL code to a distributed language, something like Java, it could take anywhere from $10 to $15 per line. If you've got an application that's got 20 million lines of code, you're going to be spending $200m to convert that — and that is without any additional functionality.”
IBM Enterprise Modernization
“You keep all the data where it resides — even if that's in a third party — and we will provide you the tools, a data fabric, to have it virtually look like all that data is local to the servers that need to do the inferencing.”
IBM Enterprise Modernization
“At later stages, the data infrastructure spend rises because the AI system just touches data far more often and in more complex ways. How frequently, how deeply, how many times the data is accessed and transforms — that's really the driver here.”
Senior Executive, Databricks Go-to-Market
“From a customer point of view, they have to have a data platform. If they don't have a data platform, there is no AI. Forget about AI. You have to educate the customer that first things first, you've got to have your data platform.”
Enterprise Architect, Google
“They had to build distributed inference models proximate to where their traders and their other people are located. That would be one example where it demanded that inference be proximate to the user base that they were talking to.”
Equinix Capacity & AI Demand
The operators are unusually consistent on this point: 2026 is the year agentic workflows move from demo to production, and the move is a step-function, not a trend line. Agents loop, they call other agents, they burn context, and they do this at a concurrency that enterprise finance teams have not previously modeled. The compute footprint of a single "task completed" is an order of magnitude above what a SaaS query once consumed.
“Late 2025 and 2026 and through 2027 will be agentic. Agentic is one inflection between versions.”
Former Machine Learning Engineer, OpenAI
“The reason for the massive growth on inferencing is the AI agents. Many of the companies — call it Claude, call it OpenAI and many others like Salesforces of the world — all these companies are heavily investing in AI agents, and the majority of the enterprises rather than investing in RAG systems.”
GPU Sector Architecture & Capex
“Agents also fall into inference. Agents are essentially just the scaffolding that you build on top of the inference in order to get an outcome that you want.”
Former Senior Product Executive, AWS
“You're still going to use a large model to do the orchestration and distribution of the request, and smaller models that would represent agents — agentic models that will be able to complete one task.”
Lead Data Executive, AI/IoT Cloud at Microsoft
“Early pilots, they might look cheap, but once you're in production, CFOs get spooked by the API usage — what if it's unbounded, are there spiky inference costs, are there all these agent loops that are multiplying compute, these cloud bills, are they hard to forecast?”
Senior Executive, Databricks Go-to-Market
“Now you're at a point of scale where you have very high concurrency and these agent workflows that are quite sophisticated, multi-step algorithms. You have all these feedback loops running. You have all of this historical context that's accruing and compounding.”
Senior Executive, Databricks Go-to-Market
“Maybe it happens just once a day to start, because this requires so much computing to process all the information and then provide the output. Then imagine that personalized warning memo happening twice a day — and then 10x a day.”
Former Senior Executive, Anthropic
One of the quietest but most consequential findings in the interview corpus is a numerical claim made explicit by the Databricks executive: at scaled production, every $1 spent on API model calls is accompanied by $4–6 in cloud compute and $3–4 in data infrastructure. Multiple independent voices describe the same structure — the surface layer (tokens) is the smallest, most visible line, while cloud and data compound faster beneath it.
“Scaled production ratio, I would actually say still $1 API, maybe $4–6 in cloud, and then maybe like $3–4 in the data stuff.”
Senior Executive, Databricks Go-to-Market
“The model call still stays roughly linear, but everything around it does not. This is actually where cloud and data platforms start compounding faster than the model spend, because once you get to scaled production AI starts to become more of a full-blown operating system, not just some feature.”
Senior Executive, Databricks Go-to-Market
“Over a three-year horizon, who ultimately captures more of the AI dollar? Is it the model providers? Is it the infrastructure, the data platform? I think probably infrastructure actually outpaces models after maybe year one or year two.”
Senior Executive, Databricks Go-to-Market
“Amazon is making money by selling them GPU, CPU, memory, storage and then some ancillary telemetry management services. The other way Amazon can monetize the enterprise inference use case is by providing model endpoints. That's what Bedrock is.”
Former Senior Product Executive, AWS
The skeptic's read is that enterprise AI is still a set of pilots. The operators' read is that the pilot-to-production ramp has begun, the dollar magnitudes are moving an order of magnitude at a time, and the consulting ecosystem is acting as a systematic conversion funnel. Coding tools, in particular, are treated as the hidden wedge into enterprise compute budgets.
“Back in 2024, these folks are doing $5bn–20bn in revenue, their AI spend was anywhere from probably $200,000–1m total, which is quite small actually — just a lot of proof of concepts. 2025, we're seeing some of these same enterprises with 20–50 POCs and pilots and maybe some early production workloads, but that total AI spend has climbed to probably $3m–5m, maybe even a bit more.”
Senior Executive, Databricks Go-to-Market
“2026, if I were to predict where we're going to accelerate, I think these same enterprises are probably going to do $10m–20m annually, mainly because they're going to have a lot more production AI workloads.”
Senior Executive, Databricks Go-to-Market
“I would say we're around 20% right now. My 40–50% 12 months from now includes the recent development and popularity and scale of agentic workflows.”
OpenAI Enterprise Partnerships
“OpenAI counts in that bucket too. That's a very big part of revenue. That's about 20–25%, about $19bn–24bn — absolutely the fastest growth bucket so far. Contributed 13pp of Azure's 39% growth.”
Microsoft AI/Cloud Strategy
“Developer tools doing really well with GitHub Copilot. GitHub is the smartest acquisition Microsoft has ever done, ever. Getting those libraries, getting those developers and mind share. It's a GTM dream right now.”
Microsoft AI/Cloud Strategy
“It actually allows us then to, in some ways, provide a funnel for our software products as well as our infrastructure products. Not a lot of margin within consulting, but incredibly important within our relationships with our enterprise clients.”
IBM Enterprise Modernization
Finally, the operators who build, sell, and finance these assets converge on the same conclusion: the AI data center is not merely a larger version of the traditional data center. It is a different business — one with higher capex per megawatt, more expensive content inside the shell, stickier customer contracts, and a GPU refresh pattern that does not retire hardware so much as demote it down the workload stack.
“If you're spending $15m a megawatt, you're probably spending $20m a megawatt for your servers, your cabling, your network, your storage devices. You're spending more on the hardware that goes into the data center than you're spending on the data center itself.”
Former Senior Executive, Data-Center Site Selection at Nvidia
“To build an air-cooled traditional IT data center, you're able to get it to about $7m per megawatt. Now when you do high density, you're looking between probably $10m at a minimum per megawatt. There's a 30% increase when you build a facility with high-density AI water-cooled infrastructure.”
PE Data-Center Infrastructure
“The first data center colocation services is very sticky, meaning it's difficult for a customer to just pick up and move. Once the contract is done, you're pretty much locked in for a good amount of time. A lot of these contracts are very long term.”
PE Data-Center Infrastructure
“Normally, at data center, CPU was roughly five years in average. You refresh the hardware completely. But with GPUs, we are not seeing that. There's a new procurement happening, but at the same time, there's a displacement of the existing GPUs with lower needs like inferencing.”
GPU Sector Architecture & Capex
Σ · Meta-Synthesis
The architecture the operators describe
The nine clusters do not stand in isolation — they interlock. Inference dwarfs training because agents are a step-function in per-task compute and enterprise adoption is compounding. That demand runs into a physical supply ceiling: capacity is structurally scarce and capex is rationally backed by recurring revenue, not speculative. The enterprises buying that capacity refuse to be locked in — multicloud is a mandate — and their data will not migrate, so models come to the data. At scaled production the economics resolve into three revenue lines, with the model call as the smallest. What emerges underneath is a structurally superior asset class: sticky, high-density, and refresh-resistant. Eighteen operators, speaking independently, arrive at the same architecture the thesis has been arguing from first principles.