Amazon on October 28, 2026 published a technical account of Prime 360, the AI highlight system it is using in select NBA on Prime broadcasts this season, and named its own Trainium chip as the hardware that makes the system fast enough for live television. According to Amazon, running Prime 360's diffusion models on Trainium "reduces the latency by more than 40%" - a figure the company published without saying what it was measured against.

In Short

Amazon puts lots of high-resolution cameras around an NBA court and uses an AI model to build new, movie-like replays of real plays for Prime Video. Every frame of those replays takes billions of calculations, so Amazon runs the job on a chip it designed itself, called Trainium, and says that cut the waiting time by more than 40%. If the claim holds, fans watching NBA on Prime get these replays during the game instead of after it, and Amazon gets a chip it mostly sells to AI companies onto a screen viewers can actually see.

What Amazon disclosed

The account appeared on Amazon's corporate news site under the headline "Amazon's most advanced chip helps power the AI behind Prime 360 - here's how." It is credited to Amazon Staff and names no engineer, producer or executive. It is short, and much of its technical content sits in a handful of paragraphs.

Prime 360 itself was not new on October 28. Two weeks earlier, on October 14, 2026, the Sports on Prime account on X introduced it as "Prime Video's proprietary, AI-powered technology that utilizes an array of high-res cameras around the court and AWS Trainium chips to deliver cinematic highlights." What the October 28 piece adds is an explanation of why the chip matters.

According to Amazon, the system was built through a collaboration between Prime Video producers, engineers and AI scientists. Cameras positioned around the court capture the action, and that footage is then processed through AI models that "construct the immersive highlights frame by frame." Amazon locates the hard part of the problem after capture, not during it. "The real technical challenge isn't capturing the footage, but rather delivering broadcast-quality highlights fast enough to be employed within live game coverage," the company wrote.

That framing matters. A replay system that needs an hour of processing is a post-game product, useful for social clips and recaps. One that can produce output within the flow of a live broadcast becomes part of the game coverage itself, sitting alongside the play-by-play and the replays a director cuts to between possessions.

Why diffusion models are expensive

Why would a basketball replay need custom silicon at all? The visual quality of Prime 360 comes from a diffusion model, which Amazon describes as "the same foundational technology behind popular image-generation tools." Such models start from noise and refine it over repeated passes until a finished image emerges. According to Amazon, "the model iteratively refines visual output, transforming noise into a polished frame."

The cost of that approach scales badly with time pressure. Every frame of footage must pass through the model before it can air, and every pass through the model is itself heavy work. "Each iteration through the model involves billions of mathematical operations - matrix multiplications, attention computations, and iterative denoising steps - all of which must complete fast enough to keep pace with a live broadcast," Amazon wrote.

Multiply billions of operations by several denoising steps per frame, then by the number of frames in a highlight, and the scale of the problem becomes clear. Amazon does not supply the variables needed to do that arithmetic. It gives no frame rate, no highlight length, no output resolution and no count of denoising steps, so the total compute behind a single Prime 360 clip cannot be derived from the disclosure.

What Amazon does say is that speed cannot be bought with software alone. "Solving that challenge requires both software optimization and innovation at the chip level," the company wrote. Diffusion-model inference - running an already trained model to produce output - is the workload in question here, not the training of the model.

Tile in, tile out

The most specific technical detail in Amazon's account concerns how the frames are carved up. According to Amazon, Prime 360's diffusion model "processes each frame by dividing it into dozens of overlapping tiles, each of which passes through the model independently." The inputs and outputs at every step are large tensors, which Amazon describes as "massive grids of numbers that AI models use to represent images, motion, and depth."

That structure maps onto how Trainium was built. Amazon says the chip's architecture follows what its engineers call a "tile in, tile out" design, "optimized for workflows whose inputs and outputs are large tensors." Trainium, in Amazon's words, was designed "for workloads that process large blocks of data at high speed."

The fit is straightforward to see. Tiles that pass through the model independently do not have to wait for one another, so the work on a single frame can be spread out and run in parallel rather than processed as one large block. Overlapping tiles are a common way of avoiding visible seams where separately processed regions meet, though Amazon does not explain its own reasons for the overlap.

The reference to "depth" is the only hint in the account of what the model handles beyond flat camera images. A product called Prime 360, fed by an array of cameras around the court, suggests viewpoints that no single camera recorded. Amazon does not describe how camera angles are combined or how a viewpoint is chosen.

According to Amazon, "the result is diffusion-model inference that reduces the latency by more than 40%, making it fast enough to generate these highlights for in-game use and enabling producers to enhance storytelling and the viewing experience for fans."

What the account leaves out

The 40% figure is the only performance number Amazon published, and it arrives without a reference point. A reduction relative to graphics processors, to an earlier version of Prime 360, or to unoptimized code running on Trainium itself would each mean something quite different. Amazon does not say which applies.

Several other gaps limit how far the claim can be assessed:

  • Chip generation. The headline calls Trainium "Amazon's most advanced chip," but the text never names a generation. Amazon's Project Rainier cluster, for comparison, was built on Trainium2.
  • Scale. No count of cameras, chips or servers is given.
  • Location. Amazon does not say whether the processing runs at the arena or in an AWS data center.
  • Timing. "Fast enough to keep pace with a live broadcast" is not quantified. How many seconds or minutes separate a play from its Prime 360 version is not disclosed.
  • Schedule. The highlights appear within "select" NBA on Prime broadcasts. Which games, and how many, is not stated.

There is also a tension in how Amazon describes the output. The opening of the account says Prime 360 delivers "cinematic highlights of actual game footage." Its closing says that "those cinematic highlights that traditional cameras cannot replicate are being generated by Amazon-designed hardware." Both descriptions can be true at once - the source material is real play, the rendered frames are model output - but the account does not say whether viewers are told on screen that what they are watching was produced by a generative model.

Amazon also describes the project's origins in modest terms for such a confident piece. According to the company, Trainium was designed "from the ground up for exactly these kinds of workloads - even though this type of application was just a moonshot idea a few years ago."

Trainium's day job

Prime 360 is an unusual public showcase for a chip whose main customers are AI developers rather than sports fans. Amazon acknowledges as much. "Amazon has been building custom chips for years, but most of that work has powered cloud infrastructure behind the scenes - making computing faster and more efficient for the businesses and developers who build on AWS," the company wrote.

Amazon is one of the large hyperscaler operators that design their own AI silicon, running both Trainium and Inferentia chips. The Trainium name points to model training, yet the Prime 360 workload Amazon describes is inference. That is consistent with how the largest Trainium customer uses the chip. When AWS activated Project Rainier with nearly 500,000 Trainium2 chips on October 29, 2025, the cluster was to be used to build and deploy Claude, covering both training and inference, and Anthropic planned to scale to more than 1 million Trainium2 chips by the end of 2025. Amazon put Trainium2's price-performance advantage over competing options at 30% to 40%.

The commercial weight behind the chip has grown quickly since then. In its third-quarter 2025 results, Amazon described Trainium2 as a multibillion-dollar business growing 150% quarter over quarter. By the first quarter of 2026, Amazon said Anthropic would secure up to five gigawatts of Trainium capacity through AWS, and OpenAI had committed to roughly two gigawatts starting in 2027, as AWS revenue rose 28% to $37.6 billion. Anthropic's IPO filing, made public this month, showed a commitment of at least $110.0 billion to Amazon running from May 2026 to April 2036, according to ZeroHedge and Cape Fear Advisors.

All of this has a cost. Amazon raised its full-year 2026 capital expenditure forecast to approximately $220 billion from about $200 billion, after cash capex of $53.1 billion in the second quarter. Against that backdrop, a consumer-facing use for the company's chips serves a purpose beyond basketball. "With Prime 360, that investment is showing up in something millions of fans will see on screen," Amazon wrote, without quantifying the millions.

Why this matters for advertisers

For the advertising market, the more relevant question is what Prime 360 does to the value of the NBA package Amazon sells.

Amazon secured 11-year NBA and WNBA rights in July 2024, starting with the 2025-26 season and covering 66 exclusive regular-season games per year in the original announcement. The second season opened on October 23, 2026, with the Knicks at the Celtics and the Rockets at the Spurs, part of a slate Amazon described as 67 games. A second Friday doubleheader, Knicks at 76ers and Lakers at Warriors, is scheduled for October 30.

The inventory is already in demand. When Amazon reported second-quarter 2026 results on July 30, Andy Jassy said advertising on Thursday Night Football, the NBA, the WNBA and NASCAR had sold out, as advertising revenue rose 26% to $19.8 billion. Alan Moss, Amazon's vice president of global advertising sales, said in May that the NBA deal brought 30 new advertisers in its first year and that the Prime Video-exclusive play-in tournament drew more than 3 million viewers. Prime Video reached an average of 315 million ad-supported viewers globally as of the fourth quarter of 2025, up from about 200 million in April 2024.

Sold-out inventory changes the calculation. When there is no unsold supply to fill, the next lever is price, and broadcast features that set one NBA broadcaster's coverage apart from another's become part of that argument. Amazon has not said whether Prime 360 segments will carry sponsorship, brand integration or adjacent ad placements. Branded replay segments are a long-standing format in sports broadcasting.

Prime 360 also extends a line of AI features Amazon has built into its NBA coverage. For the first season, Prime Video introduced Key Moments and Rapid Recap, using proprietary AI models to find significant plays and to assemble two-minute highlight reels for late arrivals, alongside advanced stats that Amazon said were powered by AWS. In December 2025, Amazon added AI scene search on Fire TV for Prime Video films, built on Amazon Bedrock. Those tools select, summarize or locate existing footage. Prime 360 is different in kind, because it renders new frames.

Amazon's page also points readers to Prime Insights, which it describes as real-time analytics powered by custom Amazon AI models for Thursday Night Football, NASCAR and NBA viewers. Taken together, the features show where Amazon is placing its bets on the CTV sports broadcast: more of the viewing experience produced or shaped by models running on Amazon infrastructure.

There is a vertical-integration story here that marketers following Amazon will recognize. Amazon owns the rights, the streaming service, the ad sales operation, the cloud and now, on its own account, the chip that renders the replays. Whether that integration translates into anything advertisers can measure - longer viewing sessions, higher completion rates, better recall around highlight segments - is a question Amazon's account does not attempt to answer. It offers one number, 40%, and leaves the rest to the broadcasts.

Timeline

Summary

Who: Amazon, through its Prime Video sports division and its custom chip program, with Trainium as the hardware named in the account. Viewers of NBA on Prime and advertisers buying Amazon's live sports inventory are the audiences affected.

What: Amazon published an explanation of how Prime 360, an AI system that builds cinematic highlights from an array of high-resolution courtside cameras, runs its diffusion models on Trainium. Each frame is split into dozens of overlapping tiles processed independently, and Amazon says the setup reduces inference latency by more than 40%, without disclosing the baseline, chip generation, number of chips, or which broadcasts carry the feature.

When: Amazon published the account on October 28, 2026. Prime 360 was introduced on X on October 14, 2026, and the 2026-27 NBA on Prime season began on October 23, 2026.

Where: In select NBA on Prime broadcasts on Prime Video during the 2026-27 season. Amazon does not say whether the processing runs at arenas or in AWS data centers.

Why: Diffusion models need billions of operations per pass for every frame, and Amazon says only chip-level work made them fast enough for live coverage. For Amazon, Prime 360 puts a chip mainly sold to AI developers such as Anthropic and OpenAI in front of sports viewers, while adding a distinctive feature to an NBA package whose ad inventory sold out in the second quarter of 2026.