The Everywhere Problem: Why “Data Anywhere” Is Becoming the Defining Infrastructure Challenge of the AI Era

The most consequential question in AI today is not which model is smartest. It is where the data lives, and whether the intelligence can reach it.

For the better part of a decade, the AI industry operated under a reassuring premise: centralize the data, centralize the compute, and genius would follow. The hyperscaler model — centralizing vast training datasets in large cloud clusters and applying massive GPU compute to compress them into model parameters — produced extraordinary results, but also an architecture that is now straining under the weight of its own success.

Call it the “data anywhere” problem. As AI bleeds out of the research lab and into the operating fabric of hospitals, factories, financial institutions, and sovereign governments, the data that must inform these systems is inherently distributed, jurisdictionally constrained, and operationally immovable. Regulators in Europe insist that its citizens’ financial records never leave the continent. A pharmaceutical company’s clinical trial data in Basel cannot legally cohabit a cloud bucket with a genomics dataset from Seoul.

Whatever the case is, the intelligence must go to the data. The data, emphatically, will not come to the intelligence.

The Economics of the Shift

This structural tension is made acute by a simultaneous revolution in AI economics. The industry is undergoing a tectonic rebalancing from training-centric to inference-centric spending, and the implications for data architecture are profound.

Deloitte estimated that inference workloads accounted for half of all AI compute in 2025, a figure that will jump to two-thirds in 2026. The ratio is inverting with astonishing speed. Analysts estimate that by 2026, inference demand will outpace training demand by 118 times. By 2030, inference could account for 75% of total AI compute, driving $7 trillion in infrastructure investment.

The cost math is equally vertiginous. For every $1 billion spent training an AI model, organizations face $15–20 billion in inference costs over the model’s production lifetime: a ratio illustrated starkly by GPT-4, whose training cost roughly $150 million, yet cumulative inference costs reached $2.3 billion by end of 2024. Training, once the headline obsession of AI investors and procurement officers, is being reframed as the one-time tuition fee. Inference is the perpetual operating cost of intelligence, and it is now the dominant line item.

Yet here lies the paradox: inference costs for a GPT-3.5-level system fell more than 280-fold between November 2022 and October 2024, with hardware costs declining roughly 30% annually and energy efficiency improving 40% per year. Prices fall; consumption accelerates faster. Per-unit inference costs dropped 100 times, while Microsoft and Google reported AI workloads growing 31 times in half that period.

The Jevons Paradox, where efficiency gains drive greater resource use, has found a modern expression in GPU clusters.

Where Data Lives, Intelligence Must Follow

The inference economy fundamentally reshapes infrastructure requirements, and nowhere more so than around data gravity. Inference, unlike training, is not a batch job run once in a data center. It is a continuous, latency-sensitive, geographically distributed service, and only as good as the data it can reach at the moment of query.

This is the core of the data anywhere challenge.

For example, a language model reasoning over a patient’s live ICU telemetry cannot afford a 200-millisecond round trip to a hyperscaler’s eastern seaboard cluster. A financial services fraud model running inference at the point of transaction cannot exfiltrate account data to a jurisdiction where it would violate GDPR. A sovereign AI deployment cannot be predicated on infrastructure owned and operated by a foreign commercial entity.

The frontier labs are acutely aware of this. Anthropic’s agreement with Google Cloud for up to one million TPUs, delivering more than a gigawatt of AI compute capacity by 2026, signals how leading labs are investing at unprecedented scale to shape the global infrastructure footprint of inference.

A Taxonomy of Data Intensity

Not all AI systems confront this challenge identically and it is instructive to consider a rough taxonomy as there are various types of AI models and complexity. Let’s break that down with three core examples: LLMs, image, and physical models.

Large language models — the Claude, GPT, and Gemini families — deal primarily in language tokens: relatively lightweight, compressible, and amenable to privacy-preserving techniques like differential privacy or federated learning. Their data anywhere problem is very complex.

Generative visual models present an even harder case. Systems like Black Forest Labs’ FLUX.2 can produce high-resolution, photorealistic images in under a second on powerful hardware, but generating a single image requires far more data and compute than generating text. As visual AI moves beyond creative tools into industrial inspection, medical imaging, and satellite analysis, the underlying data is often large, sensitive, and difficult to move, increasing the need to run AI where the data already resides.

The more complex category is physical AI. NVIDIA’s Jensen Huang has declared that “physical AI has arrived, and every industrial company will become a robotics company.” New models such as NVIDIA’s Cosmos 3 aim to give machines a generalized understanding of the physical world by combining simulation, vision, and reasoning, while companies like Physical Intelligence are training robots on real-world sensor data – including force, motion, and visual inputs – to enable more adaptable, autonomous behavior.

The same scaling dynamics that improved large reasoning models are now being applied to real-world data such as vibration, sound, and sensor inputs. But this information is inherently local. A robot on a factory floor cannot send real-time visual and touch data to a distant cloud for processing without introducing delays that could create safety risks, meaning AI must increasingly run at the edge, close to where the data is generated.

Trust, Explainability, and Outcomes

This is where the data anywhere challenge moves beyond infrastructure and becomes a governance issue. As AI is applied to high-stakes decisions – from healthcare diagnoses to financial risk models to physical control systems – questions about where data resides are increasingly tied to who is accountable for outcomes.

In today’s regulatory environment, explainability is not optional. The EU AI Act, for example, requires that high-risk systems demonstrate the basis for their outputs, which is difficult if the data informing those decisions is distributed across multiple systems, jurisdictions, and regulatory frameworks.

Trust, therefore, becomes the prerequisite for adoption at scale. Control over the data environment is becoming just as important as control over the models themselves.

The Next Generation of AI Infrastructure

The resolution of the data anywhere challenge will define the competitive map of AI for the next decade. Federated inference, secure data-processing environments, edge-optimized models, and orchestration systems that account for where data is allowed to reside are not niche technical features, but preconditions for AI’s expansion beyond the use cases where data can be freely centralized.

The companies and governments that build infrastructure capable of delivering trusted, explainable, sovereign inference — intelligence that reaches the data rather than demanding the data travel to it — will command the AI era’s most durable moat. Training a smarter model is increasingly a solved and commoditized problem. Deploying it responsibly, at the edge, across jurisdictional boundaries, against data that cannot move, is the problem that remains.

Data anywhere is not a slogan. It is the hardest unsolved problem in enterprise AI. And it will determine whether the extraordinary capability unlocked by the past decade of training investment ever translates, at scale, into outcomes the world can trust.