Amazon gives away the tool that fixes programmatic's QPS waste problem
Amazon donates Dynamic Traffic Engine to IAB Tech Lab, enabling DSPs to signal bidding priorities to SSPs and cut QPS waste across the programmatic supply chain
Data
IAB Tech Lab today announced that Amazon Ads has donated its Dynamic Traffic Engine to the organization, making the framework available as an open-source project and placing it under industry-wide governance. The announcement, published April 15, 2026, describes a file-based signaling system designed to let demand-side platforms communicate bidding priorities directly to supply-side platforms, reducing the volume of unnecessary bid requests that flow through the programmatic supply chain.
The problem the system targets is straightforward in theory but expensive in practice. Every second, SSPs forward millions of bid requests to DSPs that have no intention of bidding on the vast majority of them. The resulting queries per second load consumes infrastructure capacity on both sides of the auction, drives up operational costs, and does little to improve match rates. According to Anthony Katsur, CEO of IAB Tech Lab, the framework addresses "the industry's QPS waste and escalating infrastructure costs" directly.
What the Dynamic Traffic Engine actually does
The system is structured around three components. DTE Cloud, hosted by the buyer, holds signal files and configuration files that describe the types of bid requests the DSP values. DTE Evaluator Library, integrated on the seller side, polls those files periodically and generates filtering recommendations for each incoming bid request. DTE Filtering, the enforcement layer, then allows the SSP to either forward or discard a request based on those recommendations.
Signal files are defined as tuples - combinations of features extracted from OpenRTB requests. A low-value model tuple might combine delivery channel, format, country code, publisher ID, slot size, slot position, and device type. For example, the GitHub repository published alongside the announcement shows a rule expressed as app|banner|537075271|MEX|320x50|a|4, where each pipe-delimited value maps to one of those dimensions. The evaluator library constructs the same tuple from each incoming OpenRTB request and checks it against the signal output file, returning a filterDecision value of either 0.0 (filter the request) or 1.0 (forward the request).
Two model types are supported. LowValue models evaluate requests as low value when no tuple matches; HighValuemodels evaluate requests as high value when no tuple matches. This distinction matters because buyer priorities differ by use case. A buyer running a high-priority deals campaign might configure a HighValue model identifying specific deal IDs, device types, and countries it wants to prioritize. A buyer focused on broad QPS reduction might configure a LowValue model to shed traffic that historically generates no response.
The configuration layer itself is JSON-based and divided into two files per seller. The Model Configuration file defines signal schema and feature extraction rules, specifying how features are pulled from OpenRTB fields using JSON path expressions and a set of named transformations: Exists, ConcatenateByPair, GetFirstNonEmpty, IncludeDefaultValue, and ApplyMappings. The Experiment Configuration file controls traffic splitting - separating requests into treatment and control groups so buyers can maintain a learning population that is never filtered, preserving model training data even while filtering operates on the majority of traffic. The control group allocation, identified by learning=1 in the extension fields, is always forwarded regardless of model output.
The evaluator library, available in both Java and Golang, syncs with the cloud storage location every five minutes to pull updated signal and configuration files. For the Java library, Amazon used Corretto 17 in testing; the Golang library was tested with Go 1.21. Memory consumption for signal output files is estimated at between 1 and 10 megabytes depending on supply variability. Evaluator latency for the current rules-based model is documented as under 1 millisecond. The library uses Guava Map for its in-memory structures.
Sellers not on AWS can establish access through an IAM role and SigV4-signed requests to the S3 bucket. The reference cloud implementation uses AWS S3, but the specification notes that other cloud storage solutions can be adapted. Onboarding time estimates from the documentation suggest 1 to 2 days for permission setup and an additional 1 to 2 weeks for full integration using the provided Java or Golang library, or approximately one month for organizations building a custom evaluator.
Monetization insights as an additional signal layer
Beyond filtering recommendations, the system includes a separate monetization insights layer. Buyers may optionally provide additional model files containing RPMA - Revenue Per Million Ad Requests - data broken down by delivery channel, format, country, publisher ID, slot size, slot position, device type, OS, device make, and browser. This data allows sellers to identify their highest and lowest performing traffic patterns and optimize within QPS allocation caps, independent of the core filtering mechanism. The documentation is explicit that this data is for seller optimization insights only and is not used by the DTE evaluator library for filter decisions.
Reporting requirements are also specified. Sellers integrated with DTE must provide automated weekly reports delivered to an S3 bucket, covering total bid requests forwarded to Amazon, total spend, total impressions, bid request volumes in both treatment and control groups, requests filtered by DTE, DTE filter rate, spend per million ad requests, fill rate, eCPM, and separate treatment and control breakdowns for fill rate, spend per million ad requests, and bid rate. The reporting template requires one CSV file per data center - US-IAD, US-PDX, EU, and FE-SIN - broken out by overall, banner, video, site, app, and CTV.
The history behind the open-source move
The Dynamic Traffic Engine did not originate with today's announcement. According to the GitHub repository, version 1.1 of the specification was released on January 14, 2025, as a closed beta for Amazon Ads and directly integrated testing with sellers. The version history runs from that closed beta through a series of incremental updates: version 1.2 on January 22, 2025, renamed the model identifier and updated S3 paths; version 1.3 on January 29 added Java and Golang version documentation; version 1.4 on February 28 added the bucket parameter and clarified failure handling; version 1.5 on March 17 added a new OpenRTB request map option; version 1.6 on April 29 added detail on reporting metrics; version 2.0 on July 1 added a Deals model type and experiment configuration updates; version 2.1 on September 24 added the RPMA Threshold Sharing monetization insights model; version 2.2 on October 5 added new FAQs; and version 2.3 on February 5, 2026, generalized language across all sections.
TripleLift was among the first SSPs to publicly disclose integration with the system, announcing its early participation in the DTE beta on February 4, 2025. According to coverage at the time, the integration eliminated the need for complex predictive models that TripleLift had previously relied on to determine which traffic to forward to Amazon Ads, replacing historical bidding behavior modeling with direct demand signals.
According to Pieter de Zwart, Director of Engineering at Amazon Ads, the shift represents a structural change in how supply chains operate: "For years, sellers have had to make educated guesses about what buyers want, without access to the signals that actually drive decisions. We're shifting from guessing to knowing, using a broader range of real high-value signals from across the programmatic buy brought directly into the process."
Agentic integration is part of the plan
The donation does not position DTE as a standalone tool. IAB Tech Lab's announcement explicitly links the framework to the organization's broader Agentic Advertising Management Protocols initiative - AAMP - which IAB Tech Lab formally named on February 26, 2026, consolidating its agentic work under three pillars: Agentic Foundations, Agentic Protocols, and Trust and Transparency. According to the press release, initial proposals include operationalizing DTE within AAMP, specifically using the Agentic Real-Time Framework to update DTE components agentically in real time.
That connection matters because AAMP itself has been building infrastructure around how autonomous agents interact with the programmatic supply chain. The IAB Tech Lab Agentic RTB Framework version 1.0 entered public comment on November 13, 2025, defining how containerized agents participate in real-time bidding infrastructure. The AAMP Agent Registry reached 10 entries by March 11, 2026, including Amazon among the registered participants.
If DTE components can be updated via ARTF in real time, the implication is that an agentic system managing a buyer's campaign could modify the signal files defining which supply the DSP values - adjusting filtering rules in response to campaign performance without human intervention. The DTE's five-minute refresh cycle for signal and configuration files already supports dynamic updates; the agentic layer would provide the mechanism for generating those updates autonomously.
Why this matters for the marketing ecosystem
The programmatic supply chain has faced pressure around transparency and efficiency from multiple directions. The Incorporated Society of British Advertisers found that publishers receive only 51 cents per advertising dollar, with 15 percent disappearing through supply chain inefficiencies. The Association of National Advertisers documented that 42 percent of programmatic spending directs toward nonworking media. Against that backdrop, an open-source tool that lets buyers communicate directly with sellers about what traffic has value carries structural implications beyond cost reduction.
The margin dynamics in programmatic have already been shifting, with fee differentials and supply path optimization driving significant budget reallocations. DTE addresses a different dimension of the efficiency problem - not which path to route a bid through, but whether to generate the bid request at all. For publishers, the monetization insights layer offers a path toward understanding which segments of their inventory generate revenue and which consume infrastructure without return.
Making the framework open source and placing it under IAB Tech Lab governance broadens the pool of potential integrations beyond the Amazon Ads ecosystem. The work will proceed through IAB Tech Lab's Open Source Initiative, with implementation priorities defined through working groups. The Prebid transaction ID dispute that IAB Tech Lab engaged with in August 2025 illustrated the tensions that arise when standards bodies and implementations diverge. Placing DTE under the same governance structure used for OpenRTB, ads.txt, and the Open Measurement SDK is a deliberate attempt to avoid that pattern.
According to Katsur, the goal extends beyond infrastructure savings: "This isn't just about technical efficiency and cost savings; it's about establishing and managing a refined model of a broad supply chain that optimizes campaign performance by ensuring that every bid request represents a high-value opportunity for both buyers and sellers."
Timeline
- January 14, 2025 - Dynamic Traffic Engine version 1.1 released as closed beta for Amazon Ads and directly integrated seller testing (TripleLift announced participation February 4, 2025)
- January 22, 2025 - DTE version 1.2 released; model identifier renamed and S3 paths updated
- January 29, 2025 - DTE version 1.3 released; Java (Corretto 17) and Golang (Go 1.21) versions documented
- February 4, 2025 - TripleLift publicly discloses early participation in Amazon Ads DTE beta
- February 28, 2025 - DTE version 1.4 released; bucket parameter added, failure handling clarified
- March 17, 2025 - DTE version 1.5 released; openRtbRequestMap added as new evaluation input option
- April 29, 2025 - DTE version 1.6 released; reporting metric definitions expanded
- July 1, 2025 - DTE version 2.0 released; Deals model type and experiment configuration updates added
- September 24, 2025 - DTE version 2.1 released; RPMA Threshold Sharing monetization insights model added
- October 5, 2025 - DTE version 2.2 released; new FAQs added
- November 13, 2025 - IAB Tech Lab releases Agentic RTB Framework version 1.0 for public comment
- February 5, 2026 - DTE version 2.3 released; language generalized across all sections
- February 26, 2026 - IAB Tech Lab formally names its agentic initiative AAMP, clarifying three pillars and announcing Agent Registry
- March 11, 2026 - IAB Tech Lab Agent Registry reaches 10 participants including Amazon
- April 15, 2026 - IAB Tech Lab announces Amazon Ads donation of Dynamic Traffic Engine as open-source project under IAB Tech Lab Open Source Initiative
Summary
Who: IAB Tech Lab, the global digital advertising technical standards body established in 2014, and Amazon Ads. Key named individuals include Anthony Katsur, CEO of IAB Tech Lab, and Pieter de Zwart, Director of Engineering at Amazon Ads.
What: Amazon Ads donated its Dynamic Traffic Engine - a file-based framework enabling DSPs to share demand signals with SSPs to filter low-value bid requests - to IAB Tech Lab's Open Source Initiative. The framework, which has been in closed beta since January 2025, includes a DTE Cloud for hosting signals, a DTE Evaluator Library in Java and Golang, and a DTE Filtering layer for SSP enforcement. Initial proposals link DTE to IAB Tech Lab's AAMP initiative and the Agentic Real-Time Framework for real-time agentic updates.
When: The donation was announced on April 15, 2026. The underlying technology has been in development since at least early 2025, with version 1.1 released as a closed beta on January 14, 2025, and subsequent updates through version 2.3 on February 5, 2026.
Where: The framework operates within OpenRTB-based programmatic advertising infrastructure globally. The reference cloud implementation uses AWS S3, with signal files available across four regional data centers: US-IAD, US-PDX, EU, and FE-SIN. The source code is available in the IABTechLab GitHub repository.
Why: Programmatic supply chains generate substantial QPS waste as SSPs forward bid requests to DSPs that have no intent to bid on them. This consumes infrastructure capacity and drives up costs on both sides of the supply chain. By enabling buyers to share machine-readable signals about which inventory they value, DTE allows SSPs to filter requests before they are sent, reducing unnecessary load. The open-source donation extends these efficiency mechanisms beyond the Amazon Ads ecosystem, making the framework available to any DSP-SSP integration aligned with OpenRTB.