Closer to the Metal

OpenProgrammatic Infrastructurethrough 2030

40–46%

Most likely outcome

Absorbed / Bounded Plumbing (MODAL) — containerized bidding becomes real, multi-fabric open-web infrastructure but a commoditized SSP feature, not a new margin-capturing seat; at-scale workload is…

Containerized bidding — when the buyer's bidder moves inside the exchange's data center, and who captures the margin it relocates

Read the full research →

01The Outcomes — 5 scenarios, priced by the research

How this number was derived
Reference-class M/N x conditional decomposition; the council's strongest convergence (4 of 4 members land 38-48%). P(survives as a durable multi-homing open-web layer beyond Index) ≈ 0.80-0.85 — the 4/5 ad-tech architecture-diffusion rate, backed by 3 live fabrics + an IAB standard + sound egress/QPS economics. x P(stays bounded/enrichment-led, NO defended new SSP rent | survives) ≈ 0.60-0.65 — the contrarian's absorption correction that survived unrefuted: most prior durable layers (Prebid, SPO, HB wrappers) absorbed as plumbing while the margin map barely moved, and ARTF 'deploy-once' portability commoditizes the hosting seat. 0.82 x 0.62 ≈ 0.51, trimmed to 40-46% (a thin vendor slice + CTV-boundary rent leak up; absorption not perfectly exclusive from a bounded-but-real layer). Geomean-of-odds of members' modal points (~.42/.36/.44/.42) ≈ 0.41, confirming center. Mid ~43%.

Drivers

Reference-class gravity: prior open-web shifts (Prebid Server, SPO, HB wrappers) commoditized into plumbing rather than minting a new margin seat — survival without seat-capture is the modal history
ARTF 'deploy-once' portability lowers switching cost AND, by design, destroys any single SSP's hosting rent — the same property that aids adoption caps the rent
Sandbox-not-open-compute (no TEE in ARTF v1.0) gates buy-side data, so the realized at-scale workload is host-signal enrichment (fraud, brand-safety, contextual) + small model-driven DSPs (Bedrock-class), NOT incumbent performance bidders
Exogenous DSP fee war (Amazon ~1%, Google ~4%, TTD negotiable) plus non-container Direct Booking Engines is the real margin compressor — the architecture rides a pre-existing trend rather than originating it
No cross-fabric frequency/dedup standard, so multi-homing reintroduces duplication waste — the layer never coheres into one efficient default venue

Signals to watch

GA announcements vs continued closed beta from Index/PubMatic/OpenX through 2027
Whether any SSP publishes AND sustains a containerized-hosting take-rate >2-3% of routed spend across quarters vs bundling it free to win supply
First at-scale containerized revenue accruing to fraud/contextual/verification agents (Chalice, DoubleVerify, Gracenote-type) rather than bidders
Trade-press framing shifting from 'new category' to 'standard feature every SSP offers'
Reports of buyers bidding against themselves across fabrics / no shared portability or frequency standard by 2028

Horizon · through 2030; distinguishing read (multi-fabric GA but no incumbent tenant, no durable price card) visible by the 2027-2028 GA cycle

Margin Relocation with a Real Seat — in-fabric execution forms a DEFENDED independent-SSP hosting/curation rent and/or an incumbent-scale DSP becomes a tenant; the standalone DSP-infrastructure tier…20–28%

How this number was derived
Fermi, base-rate-anchored, post-cross-examination. P(durable material layer) ≈ 0.80 x P(forms a DEFENDED independent-SSP seat OR incumbent-scale tenancy | durable) ≈ 0.30 — the residual of the absorption conditional (~0.38), trimmed to ~0.30 because portability + SSP-vs-SSP competition + independence-as-moat cut against a defended rent. Driver-dominant's original 27-37% was DEFEATED for double-counting the exogenous fee war: its 0.55 'routes through containers' factor falls to ~0.35-0.40 once Direct Booking Engines (compress fees with no container) are stripped, collapsing it to ~0.22-0.24. Contrarian + base-rate-first converge at 28-36% and 20-28%. Consensus 20-28%; top end needs the confidential-compute unlock or a disclosed sustained term sheet. Counter-weight: doc 05's 'SSP dependency replaces cloud dependency' argues a rent CAN form.

Drivers

QPS-decoupling lets a sub-scale buyer see a full exchange's bidstream — the most credible structural disintermediation lever, from cost structure not a case study
DSP take was ALREADY compressing pre-containerization (Omnicom/Publicis TTD fee audits; Hyundai in-housing) — containerization can accelerate an in-motion trend
SSP-as-supply-gateway has structural pull to become the compute host IF it can defend the rent against portability; doc 05 frames the pricing-power counterparty moving to the SSP
Independent decisioning vendors multi-homing (Chalice across Index/OpenX/PubMatic; SWYM across five) is the strongest evidence a real market, not isolated betas, exists
Confidential-compute/TEE in a future ARTF version, if it ships, would redeem richer-signal AND remove the neutrality conflict, unlocking buy-side budget

Signals to watch

Any TTD/Yahoo/DV360/Amazon move from ARTF-contributor to actual fabric tenant with independent corroboration
A disclosed fabric commercial model / term sheet that undercuts incumbent DSP take AND holds across quarters
ANA/Adalytics benchmark showing the DSP slice falling specifically where containerized/direct routes are used (vs market-wide compression)
MiQ-type cross-DSP intelligence confirmed running production volume inside a fabric
MRC or equivalent moving toward lift/incrementality accreditation for in-fabric decisioning

Horizon · through 2030; the incumbent-tenant or measurable-take-compression signal is the 2027-2029 fork

AdCP Topology Bypass — direct buyer-agent/seller-agent negotiation routes around the impression auction, shrinking the problem containerized bidding solves; relocated margin flows to publishers and…11–17%

How this number was derived
Reference-class + adjustment, reconciled across members who split ~2x. Class base for a challenger topology route-around at material scale ≈ 0.5, adjusted hard: P(direct agent-negotiation reaches material open-web/CTV scale by 2030) ≈ 0.30-0.35 (Scope3-led, 20+ firms incl. PubMatic/Yahoo + AAMP SDKs, but an ~8-month spec missing TTD/DV360/Amazon vs a 15-yr auction) x P(sidelines containers rather than composing above them | scales) ≈ 0.40 (layers CAN compose; PubMatic, an AdCP founder, calls them complementary). ≈ 0.12-0.14. The PMP=81.6% anchor was DEFEATED as bypass evidence: PMP still clears through the existing auction/deal-ID rails. Consensus 11-17% — above driver-dominant's 9-14% on the discontinuity principle, below discontinuity's 13-20% because composability evidence is real.

Drivers

AdCP is a genuine topology change (direct agent negotiation) not a speed optimization — the rivalrous branch is existential for the architecture
CTV crown-jewel data favors direct/private/clean-room-style deals over open auctions; general AI-agent/MCP momentum is a tailwind specific to AdCP's substrate
Holdco buyer-agents (WPP Open Buyer Agent for Video) build the negotiation layer ABOVE the fabric, not inside it — deliberate restraint paired with parallel construction
AdCP shifts margin toward publishers/orchestrating SSPs; Yahoo (most forward incumbent) is a founding member, an explicit hedge against in-exchange containers

Signals to watch

AdCP membership additions, especially any demand-side incumbent
Trade reporting of live AdCP-negotiated deals at volume (vs spec announcements)
Whether an end-to-end transaction composes AdCP-negotiation + ARTF-execution — its ABSENCE supports this scenario; a documented composed transaction undercuts it
Share-of-programmatic moving from open auction toward agent-negotiated direct in ANA/eMarketer

Horizon · through 2030; early divergence ~2027, decisive 2029-2030, full effect possibly beyond the horizon

Stall / Abandonment — the category stays pilot-bound and never escapes beta, or collapses into one dominant proprietary fabric; not even absorbed as a standard feature8–14%

Stall / Abandonment — the category stays pilot-bound and never escapes beta, or collapses into one dominant proprietary fabric; not even absorbed as a standard feature

How this number was derived
Reference-class M/N. Class = hyped ad-tech architecture/standard bets at 4-5yr: Privacy Sandbox/Topics (stalled), most unified-ID (never universal), blockchain-in-adtech (failed), DMP onboarding (faded) vs ads.txt/sellers.json (succeeded) → ~1-in-4 to 1-in-3 stall, base ~0.25-0.30. But the sound egress/QPS utility floor + 3 live fabrics + a real IAB standard pull ABANDONMENT (vs mere absorption) to the LOWER end, and the absorption mass that previously inflated this is now its own modal scenario (the contrarian's split, conceded by base-rate-first). Net 8-14%. A regulatory shock (FTC Mobilewalla bidstream-data; EU AI Act high-risk) is folded in here and into the modal as a stall accelerant, NOT modeled as its own scenario — discontinuity's standalone 5-9% regulatory slice judged a cross-cutting modifier (manufactured-spread guard).

Drivers

Sandbox-not-open-compute leaves the strategic performance-bidding case unredeemed; only host-held-signal use cases work, a smaller market that may not sustain standalone pilots
Identity-substrate decay (13-16% IP-linkage accuracy) undercuts the CTV performance thesis; containerization fixes none of it
Single-fabric lock-in + DSP-neutrality conflict (no TEE guarantee) deters scaling buyers; no published commercial model + zero MRC accreditation = no underwritable reason to shift budget
A macro ad-spend downturn removing tolerance for unproven infra, OR a fabric-specific regulatory inquiry as an accelerant

Signals to watch

Persistent closed-beta status and tester-roster churn (InterMedia, Bay Street, Navigator) through 2027
Absence of any independent (non-vendor) performance validation by 2028
No disclosed price card / commercial model; no second full DSP joining after Bedrock
ARTF v2.0 shipping without confidential compute and adoption flatlining; any regulator naming a container fabric

Horizon · through 2030; if no production scale by end-2028, this scenario is in play

Incumbent-Fabric Capture — containerization scales but a hyperscaler/walled garden (Amazon via AWS RTB Fabric + DTE, or Google), not an independent SSP, keeps the relocated spread; open-web…8–13%

How this number was derived
Decomposition, sized between the contrarian/driver-dominant range after cross-examination trimmed it. P(in-fabric bidding becomes a material architecture) ≈ 0.42-0.45 x P(dominant fabric is a hyperscaler/garden vs independent SSP | material) ≈ 0.25 — LOWERED from an initial 0.30 by the doc-04 revealed-preference evidence that survived as the sharpest fact: AWS RTB Fabric is a managed NETWORK not in-exchange co-location, and Amazon donated DTE (a QPS-signalling file format) to AVOID tenancy and keep the auction in place — so the hyperscaler-fabric path requires Amazon/Google to REVERSE a stated strategy, partly offsetting capital/integration advantages and Index's owned-metal lead. 0.43 x 0.25 ≈ 0.11. Range 8-13%. Preserved as a distinct who-captures fork the extrapolative members under-modeled.

Drivers

Hyperscaler control of physical compute (AWS RTB Fabric) + bidstream-efficiency tooling (DTE) positions Amazon to own the fabric layer IF it reverses its current managed-network stance
Walled-garden share growth (78% → ~83% by 2027) starves independent open-web SSPs of spend to fund a defended fabric
Capital + integration advantages let incumbents out-build independent SSP fabrics
Retail-media growth pulling premium performance spend toward Amazon-adjacent rails

Signals to watch

AWS RTB Fabric / Amazon moving from managed-network toward true in-fabric container hosting (a reversal of current posture)
Independent SSP fabrics migrating onto hyperscaler infrastructure (Index's owned-metal stance breaking)
A walled garden opening (or tightly controlling) a container interface on its own terms
A hyperscaler/garden acquiring an independent fabric-builder

Horizon · through 2030; legible 2028-2030

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.