Akamai’s Cloud CTO Is Betting on Customer Gravity. AI Agents Pull Both Ways.

Bangkok and Singapore sit 1,434 kilometers apart. Light in fiber covers 1 kilometer in about 5 microseconds. So a round trip takes at least 14.3 milliseconds, on a perfectly straight cable that does not exist.

A person loading a web page won’t notice. An AI agent will. Jon Alexander, Akamai’s senior vice president of product, says an agent may loop 100 to 200 times to finish one task. At 100 loops across that border, 1.4 seconds go to network travel alone. Real internet routes run about three times slower than the speed of light allows.

That arithmetic sits behind the bet Jay Jenkins, the Singapore-based chief technology officer for Akamai Cloud and AI, calls customer gravity. In a November 2025 essay, he argued that compute should move closer to where customers interact, rather than sit next to large central datasets. It is a direct challenge to data gravity, the idea Dave McCrory proposed in 2010 that applications drift toward large stores of data.

He also works for a company with a stake in that answer. Akamai sells more than 4,400 edge locations, small sites that sit close to users, and has signed more than $14 billion in multiyear cloud contracts this year. Jenkins has a theory. Akamai owns the footprint the theory needs.

I put the bet through four tests: writes, machine customers, model memory and money. Each shows agents pulling compute toward the user, and back toward the data.

A placement rule, not a law

“Data gravity is a constraint. Customer gravity is an aspiration,” Jenkins says. “Data gravity isn’t there because it’s good for everybody. It’s not. It centralizes computing. It actually restricts it.”

Where McCrory later drafted a formula around network latency and bandwidth, Jenkins’s version works like a scheduler. For each job, it weighs five things: latency, where the data sits, which processor fits the job, price and carbon. A batch job can run wherever compute is cheapest, he says. A fraud check cannot wait.

He does not claim the edge always wins. “If there is data that can’t be moved, sure, make those decisions centrally, but at least do that as a decision, not just because it’s a default, because it’s the status quo,” he says.

That line matters most to CDOs. Many have spent years moving data into a lakehouse, a single central platform for analytics and AI. Customer gravity does not ask them to tear it down. It asks them to list which AI workloads sit in the center by decision, and which sit there by default.

Akamai’s AI Grid orchestrator is designed to make that choice for every request. That works well when a request only reads data. The first test is what happens when it writes.

Reads travel. Writes wait for a vote.

Jenkins’s essay used a flash sale, with inventory checks, fraud screening and pricing running near the shopper. But checking stock is a read. Selling the last unit is a write, and agents that book, buy or pay will write often.

“It’s essentially a form of eventual consistency,” he says. “Where it is transactional, we want to make sure that you maintain the integrity of that transaction. … You can’t just rely on the cache and just update the cache. You do need to establish a quorum.”

A quorum is a group of copies that must agree before a write counts. Engineers know the pattern from Amazon’s 2007 Dynamo paper: keep N copies, write to W of them, read from R, and set R + W > N so every read overlaps the latest write. The catch: a write is only as fast as the slowest copy it waits for. Spread the copies apart for safety, and every sale pays the speed of light in glass.

Customer gravity can move the read to the shopper. The write still has to collect its votes.

The customer is a machine

The second test is harder. If the customer is an AI agent running in a large cloud region, doesn’t customer gravity pull compute back to the center?

“Another word for an agent is a proxy,” Jenkins says. “That proxy you want as close to the edge as possible, because it’s going to be talking to multiple agents.”

“Data gravity is a constraint. Customer gravity is an aspiration.” — Jay Jenkins @ Akamai

The problem is what the agent talks to. Tools are the databases, APIs and ledgers an agent queries, and they do much of the work. Researchers at Georgia Tech and Intel found that in tool-heavy agent workloads, tool processing on CPUs can take up to 88% of end-to-end latency. If those tools sit in one central region, an agent at the edge still pays the long trip on every call.

Akamai’s own Jon Alexander, senior vice president of products, concedes the point. “Data has a lot of gravity. Moving data is expensive and slow, so you want to have execution sit near where the data is,” he told Fierce Network. Both forces pull on the same agent at once.

Nobody knows how to move a model’s memory

The third test is the model. When it reads a prompt, it builds a key-value (KV) cache, its working memory of the conversation. That memory is large. Meta’s Llama 3.1 70B has 80 layers and 8 key-value heads of 128 dimensions each, according to Meta’s technical paper. At 16-bit precision, that works out to 327,680 bytes of cache per token, so a 32,768-token conversation carries about 10.7 gigabytes of state. Sending that over a 10-gigabit link takes more than eight seconds. The alternative is rereading the whole prompt. Either way, the model’s memory pulls the conversation back to where it started.

So when is it worth moving a conversation to another site?

“The true answer today is nobody knows, because we’re all moving through this together,” Jenkins says.

The research supports him. Inside a single cluster, NVIDIA Dynamo can already send each request to the server that holds its cache. Doing the same across thousands of sites is the open problem.

Semantic caching brings a second risk. A semantic cache returns a stored answer when a new prompt is close enough in meaning to an old one.

“The hit rate on a semantic cache is set by a customer,” he says. “If somebody is doing something as important as fraud, I would expect that they would have some data that they can run through to test and adjust that to the level of risk which they are willing to wear across those decisions.”

That setting matters. In one banking test, a similarity threshold of 0.7 produced false-positive rates as high as 99% for some models. For a CISO, the threshold is a risk control and needs an owner.

When the grid fails, agents stop

“Outages are inevitable,” Jenkins says. If AI Grid’s orchestrator goes down, do agents fall back to a central region, or stop? “Right now, it fails closed,” he says, meaning they stop. He adds: “A lot of the orchestration is based around latency. It is not based around cost or around resource availability.” The AI Grid launch announcement promises lower cost per token. Ask which one you are buying.

Data residency, Jenkins says, is a hard rule in the router: workloads run only where the customer is allowed to run them. Proving it is harder. “The location is still a promise in the contract,” he says. NVIDIA’s GPU attestation reports the hardware model and measurements of its driver and firmware, not where the chip sits. Jenkins says the real answer is better logging.

In Asia, that evidence has to satisfy rules that differ by country and by sector. India’s central bank requires payment data to be stored only in India. Vietnam’s Decree 53 requires domestic companies to store certain user data in Vietnam. Indonesia lets most private companies keep data offshore, but sector rules, such as those for financial services, can require local storage. Singapore does not demand local storage, but the Monetary Authority of Singapore’s outsourcing guidelines expect banks to secure audit and inspection rights over their cloud providers. A router that respects borders is only half the job. The other half is proving it to a regulator.

The biggest deal runs on CPUs

The last test is the order book. In September 2026, Akamai signed an USD11.6 billion, seven-year agreement with Anthropic, a large expansion of an existing relationship. Is that customer gravity, or a rush for scarce GPUs?

“You would think that that Anthropic deal would be GPU-based,” Jenkins says. “But that Anthropic deal is not GPU-based. It is mostly based on CPU workloads.”

He expects agents to split their work across several models, each on the cheapest processor that can run it. “If you’re doing multi-model inference, then that allows you to say, ‘Okay, well these can run on GPUs and these others can run on CPUs.’ And then overall, we can save costs,” he says.

“The true answer today is nobody knows, because we’re all moving through this together.” — Jay Jenkins @ Akamai on moving a model’s KV cache between sites

That matches how agents behave. Alexander says up to 90% of an agent’s run time happens off the GPU, in tool calls, file writes and code execution, close to the Georgia Tech and Intel figure. Jenkins also admits the edge’s limit: “Right now, edges are basically CPU-bound. There aren’t a lot of GPUs at these edges.”

Put the tests together and customer gravity becomes more precise than its slogan. The agent’s loop, meaning its planning, tool calls and code, moves toward the user on CPUs. The model stays in a few large GPU sites.

Akamai’s own survey of 200 practitioners found that 60% call proximity to users critical, yet 46% still serve inference from a single cloud region. Akamai sells the fix for the gap it measured.

What to ask before you buy

• For the CDO: List the defaults. Mark which AI workloads sit in the center by decision and which sit there by default.
• For the CAIO: Profile the loop, not the model. Trace one agent task: its model calls, its tool calls and where each tool runs.
• For the CTO: Price the write. Ask for quorum settings and the 99th-percentile write time from your users’ cities.
• For the CISO: Own the similarity threshold. Tune it on your own fraud or claims data, and log every cache hit.
• For the CDO and CISO: Demand placement logs. Location is a contractual promise today. Ask for per-request logs and the right to audit them.

A test of speed, not location

What would show, three years from now, that a company got customer gravity wrong?

“It will be when you cannot meet your customer’s expectations because they can get a better service somewhere else, and you cannot actually change,” Jenkins says.

That is a test of speed, not location. Agents will keep pulling both ways. Their loop moves toward the user. Their model, memory and systems of record stay near the data. Customer gravity does not ask CDOs to move everything to the edge. It asks them to stop treating the center as the default.

Light will still need 14.3 milliseconds to reach Singapore and come back. Jenkins is betting the agent won’t have to make the trip.

Image credit: iStockphoto/SvetaZi