The recently concluded FIFA World Cup in the US, Canada, and Mexico offers much more to discuss beyond just the scores on the pitch.
As a post-80s veteran, I, Xiao Zao Jun, also followed this year's World Cup on and off (though mostly through replays, of course, as I'm getting too old to stay up late).
The most intuitive takeaway from watching is that the "way we watch" the World Cup has changed. It used to be the domain of TV stations, with everyone gathered in front of the television listening to commentary and waiting for replays. This year, Xiaohongshu (Little Red Book) secured the broadcasting rights for the first time, bringing the live streaming room into its own community, turning a football match into a social interaction.
AI-edited goal highlights went viral in minutes, fans played meme relay games in the comment sections, AI commentators could answer "what is offside" in real time, and vertical short videos allowed people to catch up on the action during their commute... For the first time, the operational strategies of internet platforms and the powerful interactive capabilities of AI were integrated into the World Cup.
The brands in the spotlight are visible to everyone. What remains unseen is how these smooth, high-definition visuals are transmitted from the stadiums in New Jersey, USA, to mobile phone screens in China at 3:00 AM.
In fact, standing behind this is an "invisible" cloud service provider: Tencent Cloud.
According to publicly disclosed data, for this World Cup, Tencent Cloud supported the official live broadcasts of the tournament across 17 countries and regions globally. Behind the broadcasting pipelines of mainstream domestic platforms like Xiaohongshu and CCTV Video are its services.
█ Tencent Cloud and the World Cup: Actually Old Friends
Before discussing this year's events, I would like to take a moment to share the story of Tencent Cloud and the World Cup over the years. Many friends might not know that this is not Tencent Cloud's first time handling World Cup live broadcasts, but rather its third.
The first time was the 2018 Russia World Cup.
That year, the broadcasting rights for the World Cup in mainland China had not yet been distributed to the internet, and people mainly relied on television to watch the matches. At the time, few believed that an internet company could successfully execute a live broadcast of such a massive event.
According to Li Zhicheng, General Manager of Tencent Cloud Audio and Video, recalling the event: About a week before the World Cup opened, the Tencent Cloud team started building the solution step by step, from moving servers to the client's machine room and laying network cables, to going to Zhongguancun to buy routers. Finally, they completed all preparations just one day before the opening match.
As a result, after the opening match live broadcast ran on Tencent Cloud's solution, people online were discussing, "Why is the picture quality of tonight's live broadcast so good?" The reviews were overwhelmingly positive.
This battle established Tencent Cloud's reputation in the sports live broadcasting circle for the first time.
The second time was the 2022 Qatar World Cup.
The situation that time was different again. CCTV took the lead in distributing the rights to the internet, and short video platforms entered the arena for the first time. Tencent Cloud became a partner for CCTV Video and internet distribution platforms in live broadcasting, video on demand, and media processing, providing end-to-end solutions.
During that World Cup, domestic audiences generally felt that the picture quality and latency were better than in previous editions, making it the best overall experience in history. At that time, Tencent Cloud took the lead in using its self-developed "ultra-low latency fast live streaming" solution in the World Cup, reducing live streaming latency from the traditional dozen or so seconds to under 800 milliseconds. For football fans, this difference in experience, especially on mobile devices, was very obvious and direct.
The third time is the recently concluded US-Canada-Mexico World Cup.
This year, behind the broadcasting pipelines of multiple platforms including Xiaohongshu and CCTV Video are Tencent Cloud's end-to-end solutions. The ultra-high-definition and ultra-low latency live streaming experience has been continued in this edition.
Building on this foundation, there are two other "firsts":
The first large-scale application of AI: In this World Cup, AI truly transformed from a "nice-to-have" addition into an integral part of the broadcasting pipeline. Capabilities such as AI directing consoles, AI intelligent editing, AI horizontal-to-vertical conversion, and AI multi-language translation were deployed on a large scale at a top-tier event like the World Cup for the first time. Besides the content side, AI was also applied to operations and maintenance. Tasks that previously required manual monitoring and on-duty handling, such as adaptive bitrate, adaptive picture quality, and cross-region scheduling, can now be handed over to AI Agents for automated O&M, eliminating the need for so many colleagues to stay up late staring at screens.
The first large-scale expansion overseas: Tencent Cloud simultaneously supported the official live broadcasts of the tournament in 17 countries and regions globally, including a series of markets such as South Korea, Indonesia, and Bangladesh.
There are also some quite interesting little stories behind this.
Take South Korea, for example.
Before the South Korean team was eliminated, the service consumption in the local market was massive. But once the team was knocked out, local audiences suddenly stopped watching, and the traffic dropped significantly in an instant.
For cloud providers, this kind of "roller-coaster" bandwidth fluctuation is actually the hardest to handle: if they cannot handle the peak, users experience stuttering; if they handle the peak but cannot scale down resources, costs are wasted.
Tencent Cloud's approach relies on its globally deployed infrastructure for flexible resource scheduling: bandwidth and computing power are tilted towards whichever country is still playing; whenever a team is eliminated, resources are freed up accordingly to supply other markets. The entire process dynamically scales according to the schedule and match results, keeping the balance between cost and experience stable.
Then there is Indonesia.
Indonesia is an archipelago of thousands of islands. Its network infrastructure is actually not very good, but it has a population of over 200 million, a high proportion of young people, and a very strong willingness to watch football.
Just before the World Cup kicked off, Tencent Cloud co-built a CDN with the largest local operator, just in time for the launch of the event, which significantly improved cross-network quality and user experience.
Finally, let's look at Bangladesh.
An app in Bangladesh, Rockstreamer, originally used a competitor's solution, but after launch, user feedback indicated a very poor experience.
Later, the operator approached Tencent Cloud. In just 5 days, Tencent Cloud migrated the entire system from the original solution. After the switch, it ran very smoothly. The reason this migration could be completed so quickly is that the OTT media protocols supported by Tencent Cloud were fully aligned with those of the original vendor, meaning the client did not need to change its system architecture.
Behind these stories is actually a "hard job" that Tencent Cloud has accumulated over seven or eight years in the overseas OTT industry: supporting the entire set of overseas media protocols. We will talk more about this topic later.
From building a solution in a week to save the day in 2018, to supporting live broadcasts in 17 countries and regions in 2026. Over eight years, Tencent Cloud's role has transformed from a "novice sharpening their spear right before battle" to a "behind-the-scenes giant in global sports live broadcasting."
A few days ago, Tencent Cloud held a media communication meeting focused on audio and video AI. Besides discussing the event itself, the meeting introduced a topic that surprised many peers: they launched an overall audio and video AI solution called "WAND," borrowing a popular term in the AI circle this year—"Harness"—to describe what they are doing.
As a noun, "Harness" means "tack or harness for a horse"; as a verb, it means "to control or utilize." Around April, with the explosive popularity of AI coding tools like Claude Code and Cursor, the industry began to realize that what truly does the work behind these tools is not a single large language model, but an entire scheduling framework. Therefore, the term "Harness" began to gain popularity in the large language model and programming circles.
Tencent Cloud's point is: the multimodal field also needs something similar. And WAND is the answer they provide.
Once the news broke, many peers started wondering: What exactly is the WAND solution? Why launch it now? And what does it have to do with the World Cup?
Next, I will sort this out for you.
█ What Exactly Does WAND Want to Solve?
To understand WAND, we must first return to the World Cup live broadcasting we just discussed:
Behind a single live broadcast, there are numerous technical links involved—collection, encoding, transmission, transcoding, distribution, picture quality processing, multi-language translation, vertical screen adaptation, ad insertion, CDN acceleration, terminal playback... Behind every link, different models and algorithms are at work.
Here comes the question: with so many models, who is responsible for stringing them together? After stringing them together, who ensures that no link drops the ball? If a bad case occurs (such as translation hallucinations, picture tearing, or incoherent sound), who provides the fallback?
This is exactly what WAND aims to solve.
Let's put it more straightforwardly. Can a single large model solve the problem? Yes, but only for a small segment.
For example, in a short drama overseas project, using a video generation model, a single prompt can generate 10 seconds of footage. Want to generate a full episode? No, you have to piece it together. After piecing it together, the transitions are unnatural, the sound is incoherent, the subtitles have garbled text, the picture quality fluctuates, and the translation is prone to hallucinations.
Another example: for a live broadcast, relying solely on an AI directing model is useless. You still need encoding, transmission, and picture quality enhancement upfront, and editing, distribution, and ad insertion downstream.
This is the biggest difference between the multimedia field and the pure text field—in the text field, a large model plus a Chatbot interface closes the loop on user experience. In the multimedia field, the chain is much longer; if any single link drops the ball, the end-user experience collapses.
Therefore, the multimodal field needs a "commander-in-chief" to string all these links together: when users make requests, it calls the models; when models have issues, it fixes them; when models are insufficient, it supplements with algorithms; when costs need balancing, it schedules resources.
This is exactly what Tencent Cloud's newly launched WAND does.
According to Li Yutao, Vice President of Tencent Cloud, at the communication meeting: "We propose that the Harness philosophy should also be applied in the multimedia field to string all models together. This includes not only large language models but also multimodal models, and even some small-parameter picture quality enhancement models and small-parameter application models, all of which will be integrated."
In plain language, it means: if various large models are the engines, then WAND is the transmission + chassis + steering wheel. With just an engine, the car won't run.
█ Why Is a Solution Like WAND Needed Now?
Some might ask, now that generative video models are so powerful, will they become even stronger in the future, to the point where this middle layer is no longer needed?
I believe the exact opposite is true. The stronger the generative models, the more important this middle "orchestration" layer becomes.
Large models are the "productive forces," and the orchestration layer is the "application outlet." An increase in productive forces will stimulate synchronous growth in application demand, and the breadth of demand growth will always outpace the speed of productive force improvement. This supply-demand gap can never be filled by a single large model.
In addition, there are several very realistic driving factors:
First, model selection paralysis.
Currently, there are no fewer than dozens of multimodal large models globally—domestic, overseas, open-source, closed-source, general, and vertical—changing almost every day. For enterprise clients, every new model requires integration, testing, and evaluation, which is time-consuming, labor-intensive, and does not necessarily yield good results.
WAND does this work for the clients: selecting, evaluating, tuning models, and patching shortcomings, all encapsulated into APIs for immediate use.
Second, cost-effectiveness is the hard truth.
Many clients do not pursue the "most extreme effects" but rather the "most cost-effective effects." For short drama overseas projects, medium picture quality is sufficient, and good cost control is key. Using a medium model plus post-production repair actually yields a higher ROI than using the most expensive top-tier model.
The value of WAND is enabling clients to flexibly combine models based on scenarios and calculate costs clearly.
Third, bad cases from models need a fallback.
None of the multimodal models currently on the market are 100% stable. Garbled text in videos, abnormal transitions, incoherent speech, and hallucinations in multi-language translation are recognized pain points in the industry. Without a fallback solution layer, clients can only assign personnel for visual inspection, regeneration, and trial-and-error, multiplying costs several times over.
WAND's approach is to introduce a "check-repair" closed loop: one model generates the result, another model checks it, and if issues are found, it repairs them instead of starting over.
Taking the translation scenario as an example, according to Tencent Cloud's actual test data—this end-to-end processing flow can improve the stability of translation effects by approximately 20%. For overseas businesses like short dramas and comic dramas, a 20% improvement in stability means deploying half the manpower for QA, synchronously improving production costs and success rates.
█ What Qualifies Tencent Cloud to Build WAND?
Tencent Cloud is not the only vendor capable of launching such solutions, but its choice to bet on this at this specific time is well-founded.
The first reason is a decade of foundational work in audio and video PaaS.
Tencent Cloud's audio and video product line has been incubated for nearly ten years—cloud live streaming, cloud video on demand, media processing, real-time communication... Each module has served a massive number of enterprise clients. This means it has a very deep understanding of the audio and video production chain, user interaction, and industry pain points.
And what do multimodal models learn? Audio and video data. The experience Tencent Cloud has accumulated in this vertical industry is even "older" than general large models.
The second reason is a sufficiently large customer base.
For short and comic drama overseas expansion, nearly over 90% of top-tier clients use Tencent Cloud. For video processing, super-resolution, enhancement, and compression capabilities of top domestic short video and live streaming platforms, Tencent Cloud has almost achieved the first place in industry scale.
With many clients, massive traffic, and rich live network data, this is the best fuel for training and iterating WAND.
The third reason is end-to-end solution capabilities.
From my observation, the threshold for such solutions does not lie in how many models are strung together, but in whether the integrated result can be "delivered as a whole."
Taking horizontal-to-vertical conversion as an example, almost every cloud vendor in the industry claims they can do it. But when it comes to actual implementation, you'll find that if the picture contains scoreboards, subtitles, and logos, how should they be reasonably arranged after converting to vertical? The vast majority of vendors have not solved this problem and can only perform horizontal-to-vertical conversion on "clean pictures."
However, the WAND solution includes logo extraction, subtitle extraction, seamless removal, translation, overlaying, multimodal picture understanding, AIGC generation... With a complete set of capabilities integrated, horizontal-to-vertical conversion can truly be implemented on live footage with scores and subtitles.
This is the value of end-to-end—it's not about piecing together a list of capabilities, but about the ability to truly run through the business.
█ What Does WAND Look Like?
According to information disclosed by Tencent Cloud, WAND's capability matrix is roughly divided into three layers:
The bottom layer is the model layer: It includes Tencent's self-developed media-specific models and third-party multimodal models from external partnerships, totaling dozens of services that form a "model pool," calling different models based on scenarios.
The middle layer is the orchestration layer: The core of the entire solution, covering over 60 audio and video capabilities, responsible for demand recognition, model selection, parameter tuning, result evaluation, defect repair, and cost scheduling. This layer is essentially the modernized encapsulation of Tencent Cloud's decade of accumulated audio and video PaaS experience.
The top layer consists of scenario-based solutions: For different scenarios, WAND assembles different end-to-end solutions—live broadcasting, short and comic dramas, e-commerce, advertising... These solutions are all built around the three words "end-to-end." Clients do not need to piece together multiple vendors; one vendor handles it all.
WAND sounds a bit like an "audio and video middle platform," but the positioning of the two is actually very different: the middle platform emphasizes "the precipitation of common capabilities," while WAND emphasizes "dynamic orchestration around AI large models and end-to-end solutions."
█ Going Global: Another Battlefield for WAND
After discussing the technology, let's briefly touch upon another chip Tencent Cloud holds—going global.
Running WAND solely domestically has limited value. The real explosive point is overseas.
The reason is simple: the iteration of the audio and video technology stack overseas is slower than domestically. The entire set of media protocols used by the traditional OTT industry is very troublesome to make compatible. For overseas vendors wanting to switch to new solutions, just protocol adaptation takes several years.
Tencent Cloud has been holding back a major move in this area for seven or eight years—basically supporting all media protocols in the overseas OTT industry. This means that overseas clients switching from other vendors to Tencent Cloud do not need to touch their own system architecture. The aforementioned 5-day migration for Bangladesh's Rockstreamer is exactly how it was achieved.
Coupled with WAND's AI capabilities, Tencent Cloud's strategy overseas becomes—"I am fully compatible with your original stuff, and you don't yet have what I've newly built." This is a posture of completely asymmetrical competition.
According to disclosures, Tencent Cloud's infrastructure investment overseas has at least doubled every year. Almost all top-tier OTTs in South Korea use Tencent Cloud services, and over 60% of live streaming clients in the Asia-Pacific region use Tencent Cloud's solutions. The services for 17 countries and regions during this World Cup also verified that this strategy can withstand event-level traffic.
█ Final Words
Audio and video cloud services in the multimodal era are shifting from "selling single-point capabilities" to "selling end-to-end solutions."
The more models there are and the longer the chain, the harder it is to guarantee the end-user experience. At this time, someone needs to do the "stitching work." How well it is stitched directly determines whether AI applications can truly reach the hands of users.
Li Yutao, Vice President of Tencent Cloud, analyzed that the upcoming competition among cloud vendors in the multimodal field will roughly focus on three directions—training computing power, data storage, and inference applications. The first two compete on capital and infrastructure, while the third competes on whether model capabilities can truly be transformed into services that clients can afford and calculate costs for.
WAND happens to be the answer Tencent Cloud has submitted for the "third direction."
As for whether it can truly become the new "key" in the multimodal era, let us wait and see.
References:
1. Tencent Cloud Audio and Video Media Communication Meeting (July 2026);
2. "Tencent Cloud Escorts World Cup Live Broadcasts in 17 Countries and Regions Globally, with Online Peak Exceeding 50 Million on Final Night";
3. Tencent Cloud Official Website, WAND Solution Introduction Page;
4. Public interview materials and industry reports.