Guest: John Bradshaw  (LinkedIn)
Company: Akamai
Show Name: An Eye on AI
Topics: Cloud Computing

Do you really need your AI model to calculate the curvature of spacetime near a black hole’s event horizon just to recommend which sneakers are in season? John Bradshaw, Field CTO Cloud, EMEA at Akamai, poses this provocative question to expose a fundamental inefficiency in how enterprises deploy AI today. The hyperscale model—brilliant in concept, brutal in cost—is forcing businesses to pay for computational overkill while struggling with latency and data migration expenses that compound with every region-to-region transfer.

The Hyperscale Tax: When All Human Knowledge Becomes a Liability

Modern AI models are remarkable precisely because they contain the sum of human knowledge. They can explain quantum mechanics, write poetry, and diagnose diseases with equal facility. But this universality comes at a price that Bradshaw argues most businesses shouldn’t pay.

“These models are so incredible—they essentially contain all of human knowledge to this point,” Bradshaw explains. “But if you’re trying to build something that can tell you what sneakers are in season this month and will work with what pair of trousers, do you also need it to be able to calculate the curvature of spacetime near the event horizon of a black hole?”

The answer, of course, is no. Yet businesses routinely deploy massive foundation models for narrow use cases, processing and paying for tokens that represent capabilities they’ll never access. Every query runs through the entire model, churning through parameters trained on everything from Renaissance art to rocket science, when the business need might be as simple as inventory forecasting or customer sentiment analysis.

The Hidden Cost of Data Gravity

Beyond model inefficiency lies an even more insidious challenge: moving data. As GPU scarcity pushes workloads between hyperscale providers and regions, the economics become punishing.

“The cost of moving these vast datasets to where they need to be to get the most value out of them is a challenge,” Bradshaw notes. “If you’ve moved it to region B, you don’t really want to do a move to region C and then D and then E.”

Each migration incurs direct egress charges, but the real damage is cumulative. Data has gravity. Once settled in a particular cloud or region, extracting it creates friction—technical, financial, and temporal. Retraining models with new data sources requires integration and baking processes that compound delays. The result is an infrastructure that’s simultaneously expensive and slow to adapt.

Latency: The Physics Problem Hyperscale Can’t Solve

Then there’s the constraint no amount of bandwidth can eliminate: the speed of light.

“It doesn’t matter how big the pipes get, it still takes a period of time to actually get it where it needs to be,” Bradshaw observes. A user in Edinburgh interacting with an AI model hosted in California will never have the same experience as someone on the West Coast, regardless of network optimization.

For real-time applications—customer service chatbots, fraud detection, autonomous systems—milliseconds matter. Edge computing collapses this distance by moving compute closer to users, delivering sub-100ms latencies that centralized hyperscale infrastructure simply cannot match.

The Edge Alternative: Purpose-Built Models, Distributed Deployment

Bradshaw’s solution is architectural. Instead of deploying monolithic models that contain all human knowledge, businesses should build purpose-built models trained on domain-specific datasets and deploy them at the edge, close to where users and data sources actually exist.

This approach delivers multiple advantages. Cost drops dramatically because smaller models require fewer computational resources and generate fewer tokens per query. Performance improves because round-trip latency to distant data centers is eliminated. Sustainability increases because data doesn’t ping-pong between regions, each move burning carbon and budget.

Edge computing also enables faster model iteration. When compute is distributed, updating a model with new data doesn’t require orchestrating migrations across global infrastructure. Local deployments can integrate local data sources in real-time, making AI systems more responsive to changing conditions.

What This Means for Enterprise AI Strategy

The hyperscale-first approach to AI made sense when foundation models were novel and edge infrastructure was immature. That era is ending. As AI moves from experimentation to production, enterprises must optimize for economics, performance, and operational simplicity.

Bradshaw’s message is clear: stop paying for capabilities you don’t need, stop moving data you could process locally, and stop accepting latency penalties when edge alternatives exist. The businesses that win the AI race won’t be those with the biggest models—they’ll be those that deploy the right models in the right places.