A single figure sits near the middle of the 85-page complaint Teads Holding Co. filed against Google LLC and Alphabet Inc. on August 3, 2026: 6.88 trillion. That is the number of additional impressions the complaint says rival ad exchanges would have won between 2017 and 2023 had Google Ads bid into non-Google exchanges the way DV360 does. The shortfall works out to more than 25 percent of rival exchange scale against a market where the tie did not exist.

The case was docketed as No. 1:26-cv-06591 in the United States District Court for the Southern District of New York. A copy went to the Securities and Exchange Commission the same day as Exhibit 99.2 to a Current Report on Form 8-K. Kellogg, Hansen, Todd, Figel & Frederick, P.L.L.C. acts as counsel, with John Thorne signing the pleading, and a jury has been demanded. Seven counts are pleaded: two for monopolization under Section 2 of the Sherman Act covering publisher ad servers and ad exchanges, a third for attempted monopolization of the exchange market in the alternative, two unlawful tying counts under Section 1 covering DFP to AdX and Google Ads to AdX, one count under New York General Business Law sections 349 and 350, and one for unjust enrichment.

What makes the filing unusual is what it does not have to argue. Liability is settled. Judge Leonie Brinkema found on April 17, 2025 that Google monopolized the publisher ad server and ad exchange markets for open-web display and unlawfully tied the two together, after a three-week trial with live testimony from thirty-nine witnesses and depositions from twenty more. Then, on October 27, 2025, the same New York court that now holds the Teads case gave preclusive effect to those Virginia findings. Private plaintiffs skip liability and go straight to injury and quantum.

David Kostman, chief executive of Teads, framed the action commercially rather than legally in the announcement accompanying the filing, describing the objective as recovering financial damages and restoring a competitive marketplace for publishers and advertisers.

The mechanics, dated and named

Much of the complaint reads as a chronology of auction rules, each with a year attached, and the effect is cumulative rather than dramatic.

Dynamic Allocation arrived after the 2008 DoubleClick purchase and let AdX bid in real time against static bids other exchanges had to set in advance. Last Look ran from 2014 to 2019 and allowed AdX to see the winning header bid before submitting its own, an arrangement an internal Google note quoted in the filing describes as inherently unfair. Project Bernanke, launched by Google's gTrade team in 2013, deflated the second-highest Google Ads bid and inflated the highest before submitting both into the second-price AdX auction. The complaint walks through the arithmetic on a single impression: bids of 7.06 dollars and 4.71 dollars, submitted as 18 dollars and 1 dollar, clear the auction at 1.00 dollar, pay the publisher 0.80 dollars after the take rate, and still charge the advertiser 4.71 dollars. Profit moves from 1.51 dollars to 3.90 dollars. Bernanke became Global Bernanke in May 2015, spawned a variant called Bell v.2 in 2016, and was rewritten as Alchemist in the autumn of 2019 to survive the shift to first-price auctions. Alchemist, the complaint states, is still running and still being modified.

Project Poirot, operating from 2018, deflated DV360 bids into non-Google exchanges while leaving AdX alone. Google's own estimate, cited in the filing, put the revenue drop at 20 to 30 percent for header-bidding exchanges against a cost to DV360 of 1.9 percent, shifting 7 percent more DV360 spend onto AdX. A second version raised deflation to as much as 90 percent. Poirot and a companion mechanism named Elmo together cost rival exchanges 21 percent of revenue.

Then there is pricing, which the Virginia court already treated as evidence of monopoly power. AdX has held a 20 percent take rate for over a decade. Google Ads charges advertisers up to 32 percent. DFP ad serving fees run at least eight times higher than rivals, and eleven times higher for video specifically, with no difference in functionality alleged. A Google employee is quoted conceding that "20% for just sell-side platform/exchange isn't likely justified by value."

What Teads says it lost

The plaintiff describes itself as a competitor across all three layers. Teads SSP transacts roughly 10 billion impressions a month. Teads Ad Manager averaged about 4,500 active advertisers, 2,000 active agencies and 5,200 campaigns per month during 2025, winning close to 7 billion impressions. The company designs more than 70,000 bespoke advertisements a year for publishers including Sports Illustrated, The Atlantic, The Washington Post and ESPN.

Two exclusions are pleaded specifically. Legacy Teads asked to join Exchange Bidding, Google's server-side alternative to header bidding, and was refused on the stated ground that its proprietary renderer prevented integration. The complaint calls that pretextual. Exchange Bidding also carries a 10 percent fee on video and 5 percent on other formats, charged on top of a rival exchange's own take rate and not applied to AdX. On the ad server side, the story runs from Outbrain's Engage Network in 2009 through to EngageOS, launched in June 2026 to merge editorial and advertising decisioning into a single feed auction, which prospective publisher clients have reportedly called unviable as a full solution without access to AdX or Google Ad Manager.

The Form 8-K carries a new risk factor that is blunter than most, naming retaliation by Google as a specific risk to the business and noting that a meaningful portion of Teads revenue is generated through transactions involving Google's advertising technology. That disclosure lands two days before second-quarter results, scheduled for August 6. First-quarter revenue came in at 266.0 million dollars, down 7 percent year over year, with connected TV growing more than 50 percent. The company employs around 1,700 people across more than 30 countries following the 900 million dollar combination of Outbrain and Legacy Teads completed in February 2025.

A queue, not a case

Teads is not first and will not be last. OpenX filed on August 4, 2025, PubMatic on September 8, 2025 and Magnite on September 16, 2025. Raptive, representing more than 6,000 websites, filed in October 2025, and The Atlantic Monthly Group followed in January 2026. Advertiser-side claims have run in parallel, including a mass arbitration announced in May 2026 on behalf of thousands of United States businesses.

Remedies in Virginia remain unresolved. The Department of Justice sought divestiture of AdX within twelve months of final judgment and open-sourcing of DFP auction logic, while Google proposed behavioural undertakings and a monitoring trustee. Judge Brinkema signalled scepticism about a forced sale at closing arguments on November 21, 2025. Brussels moved on its own track, fining Google 2.95 billion euros in September 2025, receiving a rejection of the divestiture prescription in November 2025 and publishing a redacted decision on January 14, 2026. Damages suits do not wait for any of that.

The other docket: who owns the words

The Southern District of New York had a busy week. Three days before the Teads filing landed, Judge Paul Engelmayer ruled that Reddit can proceed with claims that Perplexity and the data scraper SerpApi wrongly obtained copyrighted Reddit posts from Google search results. The complaint, filed in October 2025 and amended in February 2026, alleges circumvention of Google's SearchGuard, a CAPTCHA-based anti-scraping system, in violation of the Digital Millennium Copyright Act's anti-circumvention provisions. Engelmayer dismissed some claims but let the core theory stand, finding that Reddit's allegations, if proven, could support a DMCA conclusion, and that privacy-related allegations about deleted posts surviving in scraped datasets could support reputational injury.

That ruling gives context to remarks Reddit's chief executive made a month earlier. In an interview recorded at Cannes Lions on June 22, 2026 and published on July 3, Steve Huffman put a number on how much of modern language models rests on his platform. Asked how much of the output from Gemini, Claude or ChatGPT originates on Reddit, he pointed to the only public anchor available: OpenAI's last public research paper, which he said described Reddit as roughly a third of the training set. He placed that paper at the GPT-2 or GPT-3 stage and acknowledged nothing comparable has been disclosed since.

His formulation of the current position was compact. "Commercial use requires commercial terms," he said, drawing a line between academic research, which Reddit still supports, and commercial deployment, which it now litigates over. The adversarial track opened on June 4, 2025, when Reddit filed a 28-page complaint against Anthropic in San Francisco Superior Court alleging breach of contract, unjust enrichment, trespass to chattels, tortious interference and unfair competition, and claiming the company crawled the platform more than 100,000 times after saying publicly that it had stopped.

The economics behind the position are lopsided. Reddit signed a data licensing agreement with Google in February 2024 reported at roughly 60 million dollars a year, followed by an OpenAI partnership announced on May 17, 2024. Data licensing revenue in a recent quarter came to 36 million dollars, a fraction of what the same conversations earn as advertising inventory. Advertising revenue in the first quarter of 2026 reached 625 million dollars, up 74 percent, against 127 million daily active uniques. Roughly 40 percent of conversations on the platform are commercial in nature.

Europe supplied a harder version of the same question on August 3. A Munich court ruled against Suno in a case brought by the German collecting society GEMA, setting a penalty of 250,000 euros per breach. The judges applied United States fair use doctrine themselves and rejected it, finding that model weights reproduce the works they were trained on. Generative model providers operating in Europe now face a licence bill rather than a defence.

Underneath all of it sits a verification problem that search itself is now testing. Search Engine Roundtable reported on August 3 that Google Search is asking some users to sign in to prove they are human rather than serving a captcha, with the prompt appearing in Chrome Incognito after several pages of results. Barry Schwartz noted he could not replicate the behaviour across multiple browsers and queries, and the report rests on a single observation posted by Kamlesh Shukla. Huffman, for his part, said Reddit wants to verify humanness without verifying identity, preferring phones, passkeys and device biometrics to government documents, which he noted can simply be borrowed for a fee. His platform also declines to ban users from writing posts with AI assistance, on the argument that machine-assisted prose is how people are going to write and that community downvoting supplies the enforcement a rule could not.

Brand safety, meanwhile, has moved from adjacency to behaviour. A MediaPost piece published August 3 argued that risk in interactive environments rarely appears as one isolated post and instead shows up as patterns across users, sessions, comments and repeat behaviour, citing a 2026 NYU Stern Center for Business and Human Rights paper on what gaming's community-health model offers other platforms. Huffman's own account tracks that shift from the other side: a decade ago brand safety was the first and last question marketers asked Reddit, and now, by his description, it rarely comes up.

Who sets the ceiling on a budget

A quieter change arrived at the newest large ad platform, and its significance is arithmetic rather than editorial. OpenAI is converting the fixed daily spending limits in ChatGPT Ads Manager into seven-day averages, applied to running campaigns without advertiser action. The change was flagged in a weekly product update email to beta advertisers, reported on July 24, 2026, as arriving the following week.

The note describes the shift as moving "from fixed daily limits to average daily budgets," defined as an average spend per day across seven days. It adds that no action is required, that total budgets are unchanged, and that spend may vary by day while staying within daily and seven-day limits.

Three properties of that sentence carry weight. There is no opt-out described. The seven-day window is shorter than the monthly reconciliation used by incumbents, which limits cumulative drift but constrains nothing about any single day. And the reference to daily and seven-day limits leaves half of itself undefined: the seven-day limit is calculable at seven times the average, while the daily limit is asserted and never quantified.

Comparison makes the gap legible. Google Ads has operated average daily budgets for years with published multipliers: up to twice the average on a single high-traffic day, with monthly billing capped at 30.4 times the average, a commitment restated in advertiser communications ahead of a March 1, 2026 pacing change. An advertiser holding a 500 dollar average daily budget on ChatGPT can derive a 3,500 dollar weekly ceiling. The maximum exposure on any individual Tuesday is not stated anywhere in the material circulated.

The direction of travel also runs against the incumbent. Google spent the first half of 2026 building hard ceilings rather than removing them, expanding campaign total budgets to Search, Performance Max and Shopping campaigns in January 2026 after sustained requests for absolute spend control, and reporting a 66 percent average reduction in manual budget adjustments among advertisers using them.

Timing sharpens the point. Daily budgets shipped on May 22, 2026 as hard per-calendar-day caps. Sixty-three days later they were redefined. The same product note carried seven additions that broaden the platform rather than change how it spends: conversion-optimized cost-per-click campaigns, geographic exclusions, AppsFlyer and Adjust integrations across seven markets, automatic advanced matching on hashed customer data, a bulk API, refreshed product cards carrying price and star ratings, and automatic intraday pacing.

Averaged budgets meet an auction that has been anything but stable. Adthena counted 7,378 distinct advertisers inside ChatGPT during the week of July 13 to 20, 2026, with the United States accounting for 60.1 percent of them and placements reported to have grown 97 times between April and June. Pricing has already moved a long way: CPMs fell from 60 dollars at launch to as low as 25 dollars by mid-April as inventory outpaced demand. When the United Kingdom auction opened, brands recorded cost-per-click swings ranging from 39 percent to 278 percent with no visibility into which competitors had entered.

Automation is arriving with the same shape elsewhere in the stack, and the usage data is narrower than the marketing. StackAdapt logged 15,000 weekly workflows through its Ivy Studio agent layer, with nearly 70 percent of that activity concentrated in only 20 workflows; OpenAI models now feed the Toronto platform's core stack. Twenty repeated tasks is a real result and a narrow base at the same time, which is roughly where the agency conversation has landed as well. AdExchanger reported on August 3 that holding companies are exploring becoming a futures market for tokens, cutting direct deals with model and cloud providers for wholesale rates and reselling to clients at a markup pitched as a discount. Digiday's Media Buying Briefing the same day found the underlying reason unglamorous: holdcos still have not worked out how to sell AI tools and services, while principal media has become a major source of profit margin.

A quarter of a television platform routes strangers' traffic

Samsung said on August 3 that it has restricted new app registrations containing residential proxy functionality and is working to remove existing apps carrying the same components, after the Norwegian security consultancy Mnemonic traced proxy code from Bright Data inside a Pac-Man game the manufacturer had promoted in its Editor's Choice section.

A residential proxy network rents out household internet connections. Traffic from a paying customer enters the network and leaves through a consumer device, arriving with a residential IP address attached rather than a data centre one. Harrison Sand, an offensive security consultant at Mnemonic, rooted the software on a Samsung television and captured every packet moving through it. What he found was a tunnel that keeps running after the app is closed and stops only when the app is deleted. The Bright Data code loaded when the game opened but stayed dormant until a consent screen was accepted.

The review problem is structural rather than technical. Many affected apps are shells of a few lines whose only function is to load content hosted elsewhere. "What was reviewed is not necessarily what is running," Sand wrote, warning that a single code change on a remote web server could activate hundreds of millions of televisions at once.

Samsung's statement contains three separate commitments: a registration gate that has already been applied to new submissions containing proxy functionality, a platform-wide developer policy explicitly banning residential proxy software development kits, and a retrospective sweep of the existing catalogue. No completion date was attached to the third.

The measurement behind both manufacturers' responses came from the security firm Spur and was published by KrebsOnSecurity on July 2, 2026: more than 42 percent of apps available on LG smart TVs contained SDKs turning the set into a proxy node indefinitely, and more than a quarter of Tizen apps carried similar components, with Bright Data accounting for a majority across both platforms. LG Electronics USA committed to suspending non-compliant webOS apps on July 21. Bright Data responded the following day, saying every peer opts in through a dedicated screen and that its practices have undergone a second independent audit by PwC. Spur's Trevor Sutter framed the objection differently, arguing that a single consent prompt buried inside a television app substitutes for neither ongoing control nor platform oversight, and that the risk grows when the person tapping accept is a minor in the household.

The advertising consequence is a measurement one. These are the same devices on which both manufacturers have been building inventory businesses. Samsung Ads opened Smart TV home screens to programmatic buying through The Trade Desk and DV360 on June 10, 2026 with Magnite SpringServe handling ad serving, and remote-enabled interactive advertisements arrived on Samsung TV Plus through Amazon DSP two weeks later, with the free streaming service reported past 100 million monthly active users. Verification systems assess traffic quality partly by checking IP addresses against databases of data centre ranges and known proxy networks. When the exit node and the connected TV impression share a household connection, those signals become harder to separate, and proxy traffic is generally encrypted, placing its contents beyond inspection.

That layer was already strained. DoubleVerify recorded a 140 percent rise in connected television fraud schemes and variants in the first quarter of 2026 against the same quarter a year earlier, while only 21 percent of surveyed advertisers used invalid traffic detection as a key performance indicator for CTV. In July, HUMAN Security disclosed a scheme reaching two billion bid requests a day. Much of what Sand watched leave his own television was large-scale harvesting of LinkedIn profiles and collection of AI training data, which loops the hardware story back into the licensing disputes: European Data Protection Board guidelines adopted in July 2026 concluded that consent is unlikely to work as a legal basis for large-scale scraping used to train generative models, and blocking only functions against traffic that identifies itself.

Consent design on televisions has drawn enforcement attention before. In December 2025, the Texas attorney general sued Hisense over automatic content recognition affecting approximately 1.27 million residents, alleging the technology was presented during setup under a name that disclosed nothing about what it collected. Days ago, the Video Advertising Bureau cautioned buyers that ACR measurement depends on opt-in rates that vary by market and demographic. Accepting ACR terms produces a viewing record; accepting a proxy prompt produces an open network path. Neither Samsung nor LG has published figures on how many households accepted the latter.

The regulatory ledger, three continents at once

Four separate instruments moved within the coverage window, and none of them is about advertising directly.

The European Data Protection Board raised doubts about the legal foundation of transatlantic data flows. Chair Anu Talus wrote to Commissioner Michael McGrath on 31 July 2026, citing Article 45(2)(b) of the GDPR and paragraphs 58 to 60 of the relevant assessment, after the Federal Trade Commission's five commissioners lost the protections that underpinned the adequacy finding. Ad tech firms moving European Economic Area data westward now await a Commission answer.

Enforcement capacity in Brussels expanded on a parallel track. A consolidated text dated 27 July 2026 gives the AI Office power to impose daily penalties of up to 5 percent of average daily worldwide turnover on providers including Google and Meta, alongside authority to seal offices and recover its own enforcement costs.

Kenya went further on scope than most jurisdictions have. New rules force mandatory impact assessments on recommendation engines and seven other categories of AI use, including hiring tools and credit scoring, and give AI processors only 48 hours to report breaches to a controller. Recommendation engines are the load-bearing component of every social and retail media surface in the market.

In the United States, the pressure is fiscal and state-level. MediaPost reported on August 3 that a lawsuit has been filed challenging Utah's new tax on targeted advertising, asking the Third Judicial District Court in Salt Lake County to declare the tax unlawful and unconstitutional and to bar the Utah State Tax Commission from enforcing it permanently. AdExchanger's data privacy roundup the same day covered New Jersey's data broker law, introduced on June 28, signed on June 30, and effective immediately except for a registration deadline trade groups pushed back to 2027. Charlie Simon of RTB House observed that the speed made the original passage of the California Consumer Privacy Act look leisurely by comparison.

Four instruments, four legal traditions, one common feature: each governs an input that programmatic buying treats as ambient rather than contested.

Also noted