OpenProgrammatic Infrastructurethrough 2029
36–44%
Most likely outcome
Layered Coexistence on an MCP Substrate (negotiation + transport stack, governance split)
How AI agents are being wired into the ad buy — MCP, AdCP, the IAB's ARTF, and what agentic buying does to CTV
Read the full research →01The Outcomes — 5 scenarios, priced by the research
Layered Coexistence on an MCP Substrate (negotiation + transport stack, governance split)36–44%
No knockout. MCP is the uncontested substrate (LF Agentic AI Foundation, Dec 2025; ~97M monthly downloads; OpenAI/Apple reference it). AdCP negotiation grammar and ARTF/AAMP transport coexist as stacked layers on OpenRTB network effects, AdCP most likely partly ABSORBED into IAB's AAMP SDKs (watch v3.0, July 2026) rather than displacing OpenRTB or surviving fully independent. The governance war stays adversarial while the stack interoperates. Holdco buyer agents (Omnicom) run ON AdCP, so demand-side adoption FEEDS layering; the four gardens route around both consortia via own MCP servers. Margin effect real but modest: DSP take squeezed a few points, SSP consolidation continues, the spread REROUTED to large publishers, the biggest SSPs running agents, and a buyer-agent class (holdcos + Scope3), not eliminated. CTV: AdCP automates curation/PMP on open supply; ARTF the biddable slice.
How this number was derived
Reference-class anchored, decomposition cross-check. Base rate for ad-tech standards contests resolving as layering/absorption: ~6/8 (OpenRTB, VAST, Prebid, ads.txt, UID2, Privacy Sandbox, AdX, OpenDirect) = ~75%. Driver decomposition: P(MCP stays substrate)~0.88 x P(no winner-take-all|MCP)~0.66 x P(open keeps enough volume)~0.75 = ~0.44. Cross-exam cut the high end twice: the three factors share a hidden volume-conditioning variable (product inflated above ~0.40), and the N=8 class emerged FROM at-scale markets vs AdCP's 'zero dollars' start. It also blocked going below ~30: substrate floor (MCP markets ~95%, LF donation, ~97M downloads) + straddler lock (O'Kelley on AdCP+IAB board; PubMatic/Yahoo/WPP build to both). Member band ~42-44 (base/driver/actor) vs ~30-32 (contrarian/discontinuity); geo-mean-of-odds ~0.37. Banded 36-44.
Drivers
- MCP-as-substrate lock-in: M+N collapse, neutral LF governance, all four gardens adopting MCP directly
- Layer-not-rival: AdCP = async negotiation; ARTF = sync ~100ms transport; both ride MCP
- Straddler lock: O'Kelley (AdCP + IAB board), PubMatic, Yahoo, WPP build to BOTH — knockout serves no pivotal firm
- Absorption path: IAB folds AdCP grammar into AAMP SDKs (v3.0, July 2026)
- Demand-side adoption feeds layering: Omnicom executes real client buys ON AdCP ('diminish not eliminate,' human-gated)
Signals to watch
- IAB AAMP v3.0 (July 2026) incorporating AdCP-style negotiation grammar = absorption confirmed
- AdCP membership past 102 (Jun 2026); any DSP/holdco becoming a VOTING founder vs. pilot-taker
- AdCP Curation Protocol (Q2 2026) shipping with named-streamer adoption vs. stalling
- Continued dual-membership / no major firm forced to pick a side
- Independent (non-IAB) benchmark of ARTF's ~80% latency claim
Horizon · through 2029; absorption/coexistence signals visible by mid-2027
Garden + Model-Layer Enclosure (the open stack governs a shrinking remainder)26–34%
The open-standards fight is real but happens BENEATH two power centers that capture rents first. (1) The foundation-model/discovery layer (ChatGPT ads, Google AI Mode, Amazon Rufus) monetizes selection upstream of any ad-buying protocol. (2) Walled gardens enclose owned inventory via per-platform MCP servers and block rival agents (Amazon v. Perplexity injunction as template; robots.txt; BSA agent-ID rule); Fox's $22B Roku deal (close H1 2027) folds a neutral exchange under a content owner. Spend gravitates inward, the open ~$400B pie stops growing, and AdCP/ARTF resolve over an ever-smaller remainder — over less of the money. CTV's cleanest case is Amazon's full in-garden loop (Fire TV + Prime Video + Rufus + ACR + checkout).
How this number was derived
Decomposition. P(gardens keep enclosing + blocking through 2029)~0.80 (four gardens shipped own MCP servers, zero joined consortia; Amazon won the Comet injunction Mar 2026; US refusal-to-deal narrow post-Trinko). x P(enclosure + model-layer capture is DOMINANT | open contest subordinated)~0.40 (base + driver independently land ~0.40 vs a persisting ~$400B open market). = ~0.32. Cross-exam DEFEATED the contrarian's 33-42% modal: its F2 (~0.6, holdcos bypass via own agents) refuted by doc 12 — Omnicom buys ON AdCP, 'diminish not eliminate' — pulling F2 to ~0.45, landing ~0.30; and corrected its 'gardens >32% of CTV' driver (only 3 sellers clear 10%: YouTube 11.9%, Amazon >10%, Disney 10.8% — Disney a streamer), so enclosure runs through model-layer + holdco routing more than CTV o&o. Brokerage-open-standard proxy ~37% supports high-20s/low-30s. Banded 26-34, co-modal with Layering.
Drivers
- Concentration predicts gatekeeping: Amazon blocks+litigates, Google fuses ads into AI answers, Meta/TikTok ship write-capable inward servers
- Model-layer discovery chokepoint upstream of any ad-buying protocol
- Four gardens on own MCP servers, zero consortium participation
- Fox+Roku ($22B, close H1 2027) creating a content-owner CTV garden; spend gravity pulls budget inward
- Margin-from-opacity: enclosure protects the spread an open transparency standard would vaporize
Signals to watch
- OpenAI's 'answer independence' firewall holding vs. breaking under monetization pressure
- Whether the Comet injunction survives to final merits and is replicated/cited
- Share of discovery starting inside Rufus/Gemini/ChatGPT vs. open CTV apps
- Fox keeping Roku 'open, partner-friendly' post-close vs. tilting to Fox/Tubi demand (H1 2027)
- eMarketer CTV concentration: top gardens' share rising vs. more sellers clearing 10%
Horizon · through 2029; enclosure trajectory readable by 2027-2028
Agentic Plateau — pilots don't scale; the standards war stays pre-volume14–20%
The standards war turns out to have been fought over a market that stays small. The pilot-to-production gap is vast: AdCP went from 'zero dollars' to a handful of campaigns; the IAB agent registry had ~10 entries by Mar 2026; TTD's agents are 'not buying ads — at least not yet'; AdCP keeps approval architectural; Omnicom/WPP keep human sign-off ('nothing's overnight'). Paparo's structural blockers hold: publishers won't put rate cards in a shared protocol, and premium CTV is relationship-priced in the upfronts. Agents stay useful for pre-buy workflow and narrow curated-PMP automation; autonomous agent-to-agent buying at scale never becomes dominant this decade. Both AdCP and ARTF persist as niche tooling; incumbents (DSP ~20%, the largest SSPs) keep the spread, lightly AI-assisted. Vendor efficiency figures (40-50% cut, 87% faster) prove real but bounded to curation.
How this number was derived
Reference-class, adopted as the convergent estimate of four of five members (driver 15-22, contrarian 14-20, actor 10-16, discontinuity 8-14; base-rate folds it into human-gated Layering). Class of hyped 'next-paradigm' ad-tech pushes (programmatic-direct, automated-guaranteed, blockchain, identity currencies, prior universal-ID): ~2/8 plateaued as assistive/niche within 3-4 yrs = ~25% raw stall. Discounted for the well-footed MCP substrate AND real Omnicom buys: 0.25 x ~0.65 = ~0.16. Cross-exam trimmed the top (real human-gated holdco buys lower FULL-stall odds) and noted overlap with human-gated Layering and complementarity with Disintermediation's volume gate, so mass moves between them. Market cross-check: Zvi '>5% agentic checkout AND $5B AI ad spend by end-2026' ~27% (~73% missed) supports near-term-disappointment mass compressing to mid-teens by 2029. Banded 14-20%.
Drivers
- Volume reality: a handful of campaigns vs. hundreds of billions of daily transactions
- Rate-card / price-taker resistance (Paparo): premium CTV relationship-priced via upfronts
- Human-in-the-loop by design: TTD 'not yet'; AdCP approval architectural; Omnicom 'nothing's overnight'
- Trust/liability/measurement gaps unresolved; shoppable-CTV checkout still a QR/send-to-phone hand-off
- Agentic-checkout stumbles (OpenAI Instant Checkout retreat; Walmart ~1/3 conversion)
Signals to watch
- Whether AdCP/AAMP production spend crosses from 'campaigns' to a reported material % of CTV/programmatic by 2027
- Independent (non-vendor) audit of the 40-50% / 87%-faster figures
- TTD/Omnicom/WPP moving the human-approval boundary toward true autonomy
- AdCP Curation Protocol (Q2 2026) shipping AND getting streamer adoption vs. stalling
Horizon · through 2029; plateau-vs-scale readable 2027-2028
Material Disintermediation — agents compress the open-market CTV stack by 202910–16%
Contrarian-on-magnitude: agentic buying does not merely reroute the spread but materially and visibly compresses the open-market DSP+SSP+curation stack by 2029. Seller-agent direct buys, agentic SPO, and a live take-rate price war (Amazon DSP 0%/1% CTV fees, Universal Ads 0 fee, Azerion Hawk ~80% cut) drag blended open-market take toward the low teens for the contestable slice; TTD's ~20% take, flat since 2017, finally breaks; SSP consolidation accelerates past the sub-10 survivors. Holdco buyer agents (Omnicom recapturing the DSP toll for clients) are the accelerant. Contained to the open market — gardens' in-wall spread does not compress, so this is a compression of the remainder, not of all spend.
How this number was derived
Reference-class, double-gated; converged across contrarian, base, driver, discontinuity. Take-rate compression after a who-touches-the-dollar rewiring: ~4/5 shocks since 2012 (header bidding, SPO, ID-resolution, transparency, ads.txt) produced double-digit-pt erosion within 3-4 yrs, CONDITIONAL (~0.80) on scaling. Gate 1: P(agentic open-market buying reaches material scale by 2029)~0.25 ('piloted not production'; TTD 'not yet'; Paparo). 0.80 x 0.25 = ~0.20. Cross-exam added Gate 2: doc 12's 'diminish not eliminate / rerouted not removed' (Claim 12.5) + Paparo 're-concentrates to large publishers' — P(compress not reroute|scale)~0.55. 0.80 x 0.25 x 0.55 = ~0.11. Same cross-exam DEFEATED the contrarian's 15-22%: enclosure CONTAINS compression to the shrinking open market, capping upside. Member band centers low-teens; Zvi proxy ~27%. Banded 10-16%.
Drivers
- Live take-rate price war: Amazon DSP 0%/1% CTV fees, Universal Ads 0 fee, Hawk ~80% cut
- Seller-agent 'true direct buy' bypassing exchange/DSP; agentic SPO compounding SSP consolidation (20+ to <10)
- Holdco buyer agents (Omnicom #1 post-IPG, Acxiom data) recapturing the ~20% toll for clients
- Margin-from-opacity erosion: an open transparency standard is attractive to BUYERS who keep the spread
- Vendor-reported supply-chain-cost cuts (40-50%, 87% faster) IF they survive independent audit
Signals to watch
- TTD reported take rate breaks below ~18% / Green abandons '15-20%', tied to agentic direct buying
- A major streamer reports CTV spend shifting to seller-agent direct buys away from DSP-mediated PMP
- Independent audit confirming or refuting the 40-50% / 87%-faster figures
- Curation fees visibly compressing rather than relocating
Horizon · earliest visible 2028; resolves 2029
Regulatory / Substrate Shock (wildcard — antitrust rupture OR MCP fragmentation)8–14%
A discontinuity outside the market equilibrium becomes decisive by 2029 via one of two routes. REGULATORY: public law redraws the board — Brinkema orders AdX divestiture and the remedy reaches Google's agentic front door (Ads MCP/Ask Advisor/AP2); the FTC/DOJ Feb-2026 inquiry treats AdCP/AAMP's registry + dues + voting classes as an exclusionary SSO gate; RealPage's decree makes shared algorithmic reliance on non-public data enforceable without explicit agreement, pointing at agent-to-agent pricing; or the EU designates AI a 'virtual assistant' CPS, cracking garden ad-agents. SUBSTRATE FAULT LINE: the council's near-certainty on MCP is the most-likely-wrong premise — a forked/competing agent-substrate gains ad-tech traction, gardens' 'MCP directly' servers prove to be incompatible per-platform dialects, or the model layer subsumes the substrate so 'whose model' beats 'which MCP.'
How this number was derived
Two near-independent shock paths combined. REGULATORY (surviving split 5-10% vs 12-18%): driver aggregation, doc 11 §3 — Google remedy reaching agentic plays ~0.22, FTC/DOJ SSO-guidance ~0.08, RealPage-style agentic-pricing ~0.07, EU AI-as-CPS ~0.07 (EU DECLINED Apr 2026), agent-blocking reversal ~0.06: P(>=1)~0.45 x P(reshaping not marginal)~0.27 = ~0.12. Neither floor (single-proceeding base + EU deferral) nor ceiling (five proceedings + tight SSO-fit lifting P(decisive|fires)) was refuted, so the band spans it. SUBSTRATE FAULT LINE (6-12%): single-vendor (Anthropic) standard adopted by rivals is a governance fault line; swept 'STOP investing 2026' ~48% measures one-year withdrawal, not multi-year fragmentation; base rate for such a standard fragmenting within ~5yr ~0.08-0.10. Two ~0.08-0.12 paths, de-duped vs other triggers, land ~8-14%.
Drivers
- Five antitrust proceedings mid-flight at once (Brinkema remedy, FTC/DOJ Feb-2026 inquiry, RealPage doctrine, EU DMA AI-CPS, agent-blocking split)
- AdCP/AAMP registries + dues + voting classes map onto SSO/algorithmic-coordination targets (Allied Tube/Hydrolevel)
- Structural-separation spillover: a Google AdX remedy reaching Ask Advisor/Ads MCP/AP2
- MCP authored by Anthropic, adopted by rivals — governance asymmetry; per-garden 'MCP directly' interop asserted, not proven
- Model-layer power center forming above the substrate ('whose model' could dominate 'which MCP')
Signals to watch
- Whether Brinkema orders structural AdX divestiture reaching agentic surfaces
- Final FTC/DOJ competitor-collaboration guidance (post-May 21 2026) naming SSO consortia / agent registries / algorithmic pricing
- EU moving from 'assessing' to designating AI a virtual-assistant CPS within the window
- Documented MCP dialect incompatibilities across garden servers, or a major vendor forking/deprecating MCP
- Anthropic monetizing MCP/Claude discovery; MCP download/registry growth stalling or splintering
Horizon · through 2029; antitrust rulings 2026-2027 with appeals to back of horizon; substrate signals 2027-2029
Probabilities are the research's own scenario bands, priced as outcomes. The full argument — history, current state, drivers, and sources — lives in the corpus: read the Future Trajectory chapter.
