Original · Unique Research · 2026-03-20 · Shanghai
English edition note: This complete translation preserves the source author's assessments and Wu Yun's interview responses as of March 20, 2026. Corporate milestones, customer outcomes, attack figures, infrastructure scale and future plans are reported source statements, not independently verified benchmarks. The original's latency and throughput multiples are retained without inventing underlying measurements or converting them into new percentage claims. Statements about local processing, compliance and driving safety reflect the source's views, not a legal-compliance determination or a verified safety guarantee. Wu Yun is a romanization of the interviewee's Chinese name.
Unique Awards · Guest Interview
Many AI Companies Expanding Overseas Lose to Infrastructure Before They Lose on Models
An Interview with Wu Yun, Akamai's Senior Technical Manager for East China
It is easy to misjudge what taking AI overseas involves.
Many people think models, products, and growth decide who wins. But once they actually run a business overseas, the first walls many teams hit often have little to do with the model itself.
Pages load too slowly.
Voice responses take an extra two seconds.
APIs start behaving erratically at peak times.
Traffic begins to grow, and attacks arrive alongside it.
The business reaches a few countries, and compliance issues emerge.
Call volumes rise, and inference costs add more pressure.
You discover that before many AI products have lost on capabilities, their foundations are already straining.
This was one of my strongest impressions after speaking with Wu Yun, Akamai's senior technical manager for East China: many people still understand AI competition as model competition, but once companies truly enter global markets, an increasing number compete on infrastructure.
Whoever can make an AI product run more reliably worldwide is closer to earning a genuine place in the market.
And that is precisely why companies such as Akamai are becoming important again.
Many people's image of Akamai still stops at CDN. That is not wrong, but it is incomplete.
If Akamai helped the previous generation of the internet answer how to get content in front of users faster, in this AI wave it wants to answer another practical question: how to bring intelligence closer to users faster, more reliably, and at lower cost.
Why Has a Long-Established CDN Company Suddenly Appeared at the AI Table?
Start with the conclusion.
Akamai did not suddenly decide to get into AI. It has followed its existing capabilities to its current position.
Wu Yun described this path in an interesting way. Rather than a dramatic strategic pivot, he presented it as a natural process of evolution.
Founded in 1998, Akamai first proposed the CDN concept and used a globally distributed network to take content delivery to its limits. Later, it extended its capabilities step by step along the same underlying logic:
First, edge content delivery;
then edge security;
then edge computing;
then distributed cloud;
and finally AI inference.
Look closely at that progression and you see that the underlying problem has never changed. It has always been doing the same thing: placing computing, networking, and service capabilities as close to users as possible.
In the internet's early years, the central objects were images, web pages, and video. Today, they are AI inference requests.
Akamai added a crucial piece with its 2022 acquisition of Linode. This mattered because it completed the public-cloud foundation, taking Akamai beyond pure delivery and giving it the ability to host general-purpose computing workloads. In 2023, it launched a general-purpose edge-computing platform, moving compute further out to edge nodes. In 2025, Akamai deepened its partnership with NVIDIA and formally launched Akamai Inference Cloud, deploying high-performance GPUs at edge nodes to bring massively parallel inference directly to its global edge grid.
Connect those milestones and its intent becomes clear:
It is not trying to win the race to train the largest model. It wants to occupy the space where models leave the laboratory and begin working in the real world.
That is precisely the part many companies most easily underestimate today.
Once a large model enters production, the question is no longer just whether its outputs look good. It must face user distribution, latency, concurrency, attacks, compliance, and costs.
Ultimately, bringing AI into global commerce has never depended on the model alone.
The Hardest Stretch for AI Today Is Inference, Not Training
Over the past few years, the AI world's most attention-grabbing stories have largely been on the training side.
Who has more parameters, more clusters, more GPUs being consumed, or another small improvement on a benchmark? Everyone naturally looks there.
But from an infrastructure perspective, inference is what repeatedly happens, continuously consumes resources, and constantly generates costs and complexity.
Training is important, of course, but it does not happen at every moment. Inference is different. It happens every time a user asks a question, clicks, sends a voice message, or triggers an Agent call.
If training a model is like building an engine, inference is like keeping that engine running on roads around the world every day.
The pressure on the system becomes completely different.
This is especially true for LLM and Agent scenarios. Users often engage in multiple rounds rather than asking one question and stopping. Every request consumes networking, compute, scheduling, and bandwidth. Once users are spread out and traffic grows, these problems surface very quickly.
Wu Yun said Akamai Inference Cloud primarily aims to address three straightforward problems:
The first is latency. The closer AI inference is to the user, the faster the response. In voice interaction, real-time decision-making, and multiple rounds of Agent calls, a few hundred milliseconds and several seconds feel entirely different to users.
The second is cost. Many AI calls increasingly resemble an ongoing financial commitment. As call volumes rise, the cost of traditional centralized GPU clusters grows rapidly. Moving data back and forth, crossing regions, and processing everything centrally all cost money. Well-executed distributed scheduling can open new room for cost savings.
The third is data sovereignty and compliance. Many real-time AI scenarios handle sensitive data. Processing closer to users and locally can make it easier to meet different countries' and industries' data requirements.
None of these sounds glamorous, but they determine something crucial: whether an AI product can move from demo to scale.
This also explains why Akamai's approach differs substantially from AWS and Azure.
It has not chosen to compete in the training race. It is betting on the longer, more complex, more failure-prone process that follows model deployment.
This reflects a mature infrastructure company's perspective: instead of competing for the hottest narrative, take responsibility for the hardest implementation work.
Why Will Edge Inference Become AI's Next Tough Battleground?
When people first hear edge inference, it can sound distant from business—a technical term buried deep in the stack.
But consider a few specific scenarios, and it becomes immediately intuitive.
Take a customer-service voice system. If it pauses for two seconds after the user finishes speaking, the experience is already diminished.
Or AI effects in a livestream. You cannot freeze the image and wait for the model to finish calculating.
Or real-time NPC interaction in games. If a player speaks and the character takes ages to respond, immersion disappears immediately.
Connected-car voice assistants, immediate feedback from health devices, and local analysis on smart wearables all share a characteristic: users will not wait patiently.
In these scenarios, having an answer is not enough. The answer must arrive quickly enough.
Wu Yun also noted that not every AI task belongs at the edge.
Training extremely large models, aggregating global data, and complex tasks that depend heavily on compute are still better suited to a central cloud. They require large GPU clusters, centralized storage, and unified scheduling.
Tasks genuinely suited to the edge have strong real-time requirements, sensitive data, and widely distributed users. Voice models, livestream effects, real-time generation in games, in-car intelligent assistants, and health wearables are better handled on-device or at edge nodes close to the device.
In other words, the future is not about the edge replacing the central cloud or vice versa. A clearer division of labor will emerge: heavy training and analysis at the center, high-frequency real-time inference toward the edge.
Behind this judgment lies something more important: AI infrastructure will increasingly resemble a layered network rather than one unified, enormous cloud.
This is where Akamai has an opportunity. The global nodes and backbone network it accumulated over the past 28 years were built for serving users nearby. Previously they served content; now they serve inference.
On the surface, the problem has changed. The underlying logic has not.
What Reveals Infrastructure's Value Is Never the Concept, but the Result
The biggest danger in discussing infrastructure is abstraction.
Talk at length about architectures, platforms, and capabilities, and readers may remember only that it sounds impressive. But what constitutes real value? You have to look at the actual business problems solved.
Wu Yun offered several examples I found representative.
Start with streaming media.
An online video provider used AI to identify video content automatically and edit sports highlights. The crucial question was not simply whether the system could edit, but how quickly. A highlight appearing a minute late can be much less effective at spreading, affecting discovery, willingness to watch, and the retention experience.
In this scenario, Akamai helped reduce latency by 70%.
The significance is not merely that the system became faster. It shortened the content-consumption chain: highlights appeared sooner, users were more likely to see them, and the platform was less likely to lose viewers.
Next, games.
A Japanese gaming company used a text-to-image model to generate assets for mobile games. Benchmarks comparing Akamai Inference Cloud with other major hyperscale cloud platforms showed clear results: multimedia asset-generation latency was reduced by a factor of 2.5, throughput showed a 3-fold increase, and costs fell by 86%.
Those figures make it immediately apparent that edge inference is more than a small system optimization. It directly affects content-production efficiency and business economics.
The automotive example is even more illustrative.
A Southeast Asian in-car software company used Akamai's AI inference cloud to optimize its vehicle voice assistant, reducing latency in responding to user commands by 30%. For someone driving, a faster voice assistant and one that is merely usable are entirely different experiences. More fundamentally, more timely responses can do more to help driving safety.
When AI truly enters these scenarios, you find that infrastructure's value is reassessed.
Previously, networking, nodes, acceleration, and security were seen as things to address later. Not anymore.
The limits of many AI products' user experiences are now determined by these underlying capabilities.
The First Problem Chinese Companies Encounter Overseas Is Often Not the Product
This part resonated with me. Over the past few years, many Chinese companies' understanding of international expansion has remained at an early stage.
They think that once the product is built, translated into English, and supported by advertising, overseas expansion is more or less done. But as soon as the business actually moves outward, they discover it is entirely different.
Wu Yun said performance is usually the first problem Chinese companies encounter overseas. Once the business reaches a certain stage, especially after gaining brand influence, security attacks and compliance issues become more apparent. Costs matter too, but for most companies, overall business stability and reliability matter more.
That rings true.
Taking a business overseas is not as simple as moving a product. It is more like throwing the entire system into a more complex external environment.
Network quality differs by country, users are more dispersed, the attack surface is larger, regulations are more fragmented, and peaks are less predictable.
Many problems a company never encountered in China arrive all at once overseas.
Wu Yun also noted that industries look very different on the surface. E-commerce involves browsing, ordering, payments, and marketing; gaming involves downloads, updates, matches, and top-ups; manufacturing involves cross-regional connections, remote collaboration, and coordination among global systems.
Strip away those surface differences, however, and the core needs are similar: a better experience for overseas users, uninterrupted business, more stable systems, and avoiding failures in security and compliance.
This is infrastructure thinking: looking beyond industries' outward forms to the common problems underneath.
What a Leading Chinese Cross-Border E-Commerce Company Reveals
Wu Yun shared a representative case that I think deserves elaboration.
The customer is a leading Chinese e-commerce company operating overseas and an Akamai partner of more than a decade. Initially, its problem was global access performance, particularly latency involving dynamic payment interfaces.
That sounds technical, but it affects the core business. A slow, stalled, or unstable payment does not merely reduce a system score. It costs actual conversions and orders.
Akamai supported it through more than 4,000 edge nodes in over 100 countries and dynamic acceleration capabilities, establishing a reliable performance foundation for its global business.
As the business grew, a new problem emerged: security attacks.
This is where many companies fall into a misconception: believing a security team and an in-house WAF are enough.
This company was no exception. It had a strong security team and its own WAF-type products. But its defensive systems were centrally deployed. Once attacks reached a certain scale, the protection system itself struggled. Beyond expanding origin servers and adding hardware, there were few options.
This illustrates that security is not simply a question of whether a system exists. It is whether the architecture can withstand global operating conditions.
That is when Akamai's advantage emerged. Rather than concentrating defenses at the back and waiting for attacks, it places them forward in the edge network. Closer to users and to attackers, it can stop large-scale attacks at the edge instead of letting the pressure reach origin servers directly.
The attack it blocked involved more than 200 million requests, 500,000 QPS, and over 600,000 IP addresses.
Those look like technical metrics, but their meaning is simple: once a business operates globally, it needs more than a system that can block a few things. It needs infrastructure genuinely capable of withstanding global-scale pressure.
That e-commerce customer is now continuing discussions with Akamai about new AI-related collaboration. This development is interesting in itself.
It shows that AI is becoming a new requirement for every company expanding overseas. Once AI enters the business, underlying infrastructure becomes more important, not less.
Chinese AI Companies Expanding Overseas Need More Than Traffic and Localization
Discussing Chinese AI models and applications going overseas, Wu Yun divided the requirements into two layers. I found the distinction particularly apt.
The first is the toC application itself: overseas access performance, latency, concurrency, security attacks, CC attacks, credential stuffing, promotional abuse, and API risks. These problems emerge together as traffic grows.
The second is the model-calling layer. Once an AI product is genuinely running, companies demand lower latency, greater scalability, and better value for money. Early approaches that prioritize simply getting things working often become increasingly difficult to sustain.
This is especially apparent among Chinese AI companies today.
As overseas users and call volumes grow for products such as DeepSeek, MiniMax, Hailuo, and Talkie, the underlying pressure inevitably increases. AI applications differ substantially from traditional internet applications: they depend more on continuous inference, backend compute, and stable cross-regional calling paths, and are more likely to face additional data-localization and security requirements.
In other words, taking an AI company overseas has long ceased to be only about translating pages, running ads, and localizing operations. It is becoming a more complete systems-engineering undertaking.
The product must be competitive. The model must be capable enough. Growth must happen. But infrastructure, security, compliance, nodes, and inference scheduling must keep pace as well.
Leave out one link, and problems follow.
This is why Wu Yun said high-growth overseas traffic is an advantageous scenario for Akamai. Its strength has always been a highly distributed global platform close to users and to attack sources.
Gradually, you sense that the truly difficult part of globalizing AI is no longer just building the technology, but keeping it working in complex global environments.
In Global AI Competition, the Real Barriers Are Moving Deeper into the Stack
Asked where AI's technological barriers really lie, Wu Yun gave a practical answer.
Compute, models, data, and infrastructure all matter. No single dimension can sustain the entire competition.
Over the longer term, however, scalable infrastructure will become increasingly important. Compute is the prerequisite, data the fuel, and models the vehicle; infrastructure determines how they are organized, amplified, and kept running reliably.
Chinese companies' strengths and weaknesses here are fairly clear.
The clearest weakness remains high-end compute, particularly in training. Geopolitics makes dependence on high-end chips, generational chip gaps, CUDA ecosystem compatibility, and the efficiency of interconnecting 10,000 accelerator cards real challenges.
The clearest advantage is data. A population of 1.4 billion, more than 1 billion internet users, and large-scale multimodal data continuously generated by internet, healthcare, transport, and industrial systems provide Chinese companies with distinctive environments for applications and resources for training.
But possessing data does not guarantee victory. Whether data becomes genuine global competitiveness depends on the maturity of the whole system.
Can models produce reliably? Are calls affordable enough? Can overseas users access the service quickly? Can deployment comply with rules across regions? Will it fail under high concurrency? Can it withstand attacks?
Together, these questions form an AI company's real long-term barriers.
This is also why Akamai's priorities for the next three years appear clear: keep building around a distributed AI inference grid, develop a cloud–edge compute system, and strengthen low-latency, cost-effective inference.
It is not telling a flashy new story. It is still doing its longstanding job, with the object of service shifting from content to intelligence.
A Final Word: Much of AI's Next Competition Will Be Decided Out of Sight
My greatest takeaway from this interview is this:
What is most easily overrated in AI today is the excitement around models; what is most easily underrated is the real difficulty of infrastructure.
Models matter, of course. Without them, this wave would never have happened. But once AI leaves the laboratory and enters global business, what determines the user experience becomes very concrete.
Is it fast enough? Stable enough? Affordable enough? Can it keep operating legally and compliantly across regions? Can it withstand peak loads and attacks?
These questions rarely attract attention or become hot discussion topics. Yet they often decide whether a company can turn AI into a lasting business.
Seen from this perspective, the value of companies such as Akamai is becoming visible again.
In the past, it solved how to deliver content worldwide. Today, it wants to bring inference closer to users. Tomorrow, whoever can do this more reliably, more broadly, and more cheaply will have a better chance of capturing genuinely lasting value in the era of global AI.
Many people are still watching the models.
But some of the tougher battles are already being fought underneath.
Selected Q&A
1. Please introduce yourself. What do you mainly do at Akamai?
I mainly handle presales technical consulting for East China at Akamai, working with companies from different industries that are expanding overseas. I help customers clarify their technical requirements and provide solutions suited to their business scenarios.
2. What stages did your career go through before joining Akamai?
Most of my career before Akamai was in the United States. After graduating from CMU, my first job was software development at a publicly listed e-commerce company in Boston, mainly involving data development and analysis related to e-commerce marketing. I later became a technical solutions architect at a multinational advertising and digital-marketing company. That experience showed me that I was better at identifying customers' real needs and proposing suitable technical solutions.
3. What attracted you to Akamai?
Partly, it was circumstance. I lived and worked in Boston for five years, then returned to China and joined a company also headquartered in Boston. That felt like a coincidence. Also, for many computer-science graduates, Akamai is a company with considerable technical standing. The interview process helped me understand its technological evolution and its direction in helping Chinese companies expand overseas. All of that looked promising.
4. How did Akamai move from CDN to cloud + AI?
It did not happen suddenly at one particular moment. It unfolded naturally through technological evolution. Akamai began with CDN and expanded into edge security and edge computing. Acquiring Linode in 2022 completed its public-cloud foundation; it launched a general-purpose edge-computing platform in 2023; and jointly launching Akamai Inference Cloud with NVIDIA in 2025 formally connected cloud and AI.
5. What is Akamai Inference Cloud's core positioning?
It is a distributed edge-AI inference platform. The focus is not large-model training, but the latency, cost, and data-sovereignty problems models face in real business environments. Its main customers are enterprises sensitive to real-time performance and serving globally distributed users.
6. How does Akamai differ from AWS and Azure?
Akamai does not focus on training, but on inference after models are deployed. It wants to bring AI services closer to users, reduce latency, improve value for money, and meet data-compliance requirements across regions.
7. Why will edge inference become an important AI battleground?
AI inference requires increasing amounts of computation, and requests occur in vast numbers, in distributed locations, in real time. Many scenarios are extremely sensitive to response speed. Centralized cloud inference faces pressure on latency, bandwidth, costs, and compliance. Edge inference can address that gap.
8. Which scenarios are better suited to edge inference?
Customer-service voice systems, AI livestream effects, real-time game NPC interaction, connected-car intelligent voice assistants, and smart health wearables all suit edge inference because they require the lowest possible response latency. By contrast, extremely large-model training and complex global analysis are still better suited to a central cloud.
9. What results has Akamai already achieved in edge inference?
There are several representative cases. In streaming media, it helped an OTT video platform reduce the latency of AI-generated highlight editing by 70%. In gaming, it helped a Japanese game company reduce asset-generation latency by a factor of 2.5, achieve a 3-fold increase in throughput, and cut costs by 86%. In automotive, it helped a Southeast Asian in-car software company's AI voice assistant reduce response latency by 30%.
10. What infrastructure problems do Chinese companies most often encounter overseas?
Performance issues usually come first. As the business scales, security attacks and compliance issues become more apparent. Costs also matter, but for most companies, overall business stability and reliability matter more.
11. Do overseas requirements differ greatly between industries?
The business scenarios look very different, but the underlying needs are not so different. Whether in e-commerce, gaming, or manufacturing, companies ultimately pursue a better user experience, greater business stability, and assurance around security and compliance.
12. Can you share a representative customer case?
A leading Chinese e-commerce company operating overseas has worked with Akamai for more than a decade. Initially, its core need was global access performance, especially latency in dynamic payment interfaces. As the business grew, it encountered large-scale security attacks. Through global edge nodes and distributed security protection, Akamai stopped attack traffic involving more than 200 million requests, 500,000 QPS, and over 600,000 IP addresses at the edge, helping the customer maintain business continuity.
13. What new infrastructure requirements do you see from Chinese AI companies expanding overseas?
On one hand, there are toC application issues around global access, latency, security, and API risk. On the other, as model call volumes grow, enterprises demand more from inference latency, scalability, and value for money. AI products expanding overseas impose more complex infrastructure requirements than traditional internet products.
14. Where do the real technical barriers lie in global AI competition?
Compute, models, data, and infrastructure all matter. Chinese companies' clearest weakness today remains high-end compute, especially in training. Their advantages are abundant data resources and broad application scenarios. Over time, organizing those advantages through stable, scalable infrastructure will become increasingly important.
15. What are Akamai's main priorities for the next three years?
It will continue to build around a distributed AI inference grid, deepen the cloud–edge compute system, and keep strengthening low-latency, cost-effective inference.