AI Product Economics Maturation

Sources#

• 2026 State of Scaling: The Great Sorting
• Accenture and Google Cloud Deepen Partnership with Formation of New Accenture Gemini Enterprise Business Group
• Forward-deployed engineers are the AI industry's latest talent obsession
• Ramp's latest data on China vs. the American AI Labs
• Startup ARR is less secure than ever, new research shows
• State of AI 2026: The Builder's Economy
• The ICONIQ Pacesetter Index

Summary#

ICONIQ Growth's State of AI 2026: The Builder's Economy (third bi-annual builder survey, ~305 executives at software companies building AI products, Q2 2026, empirical) makes one thesis its spine: the market has moved from proving AI works to proving AI pays. AI products are approaching half of revenue at the surveyed companies, gross margins are expanding, and the operators pulling ahead treat "pricing, cost, and org design as product decisions rather than afterthoughts." This page collects the unit-economics half of that maturation — revenue mix, margin trajectory, pricing-model shift, provider mix, internal-AI cost, and forward-deployed-engineer (FDE) monetization. The org-restructuring half lives at AI-Native Organization; the revenue-per-employee thread at AI Investment Story, Not Efficiency Story.

Evidence note — split the tiers inside one empirical source. The raw doc is tagged empirical (a real survey of ~305 executives), but ICONIQ's headline numbers mix two grades. The 2025 figures are self-reported actuals; the 2026P / 2027P figures are self-reported projections — closer to prediction-grade than measured empirical, and doubly so because a projection and a survey (respondents forecasting their own businesses, with a VC-publisher's optimistic framing around them). Projected numbers are marked "projected" inline throughout. Where ICONIQ's forward trajectory corroborates a measured finding elsewhere in the vault (e.g. rising RPE), it is the weakest evidence for it, not the strongest.

AI as a revenue and margin story#

The core "proving AI pays" exhibits (all averages across the surveyed cohort, excluding pure AI-native companies where noted):

• Share of revenue from AI products: 32% (2025 actual) → 42% (projected 2026), +10pp (N=265, excludes companies with 95%+ AI-driven revenue as likely AI-native). AI products approaching half of revenue is the deck's "no longer an experiment" headline.
• Gross margin on AI products: 45% (2025 actual) → 53% (projected 2026, +8pp) → 59% (projected 2027, +6pp) — +14pp of projected total expansion over two years (N=287). High-growth companies project 64% vs peers' 58% margins in 2027. Margin expansion is attributed to revenue scale plus optimization levers (reducing inference costs #1, routing strategies #2, growing revenue for cost leverage #3, OSS-model switching, price raises, provider-price negotiation).
• The margin story is projected, not banked. Notably this runs more optimistic than the measured margin picture at Emergence Capital, whose cap-table data shows the fastest-growers running 6–16pp below slower peers on gross margin (absorbing AI-infra cost for growth). ICONIQ's respondents forecast margin expansion; Emergence measures margin pressure at the growth frontier — a survey-projection-vs-receipts tension worth holding (see Telemetry vs. Survey Measurement).

The same publisher's measured margin, two months later — and why it does not grade the projection#

ICONIQ's Pacesetter Index (2026-09-17, empirical) is the rare case of a publisher shipping a measured number into the space its own projection occupies. It is not a survey: quarterly financial and operating data, 2024 – Q2 2026, from public software companies plus ICONIQ's venture and growth portfolio. Its company-level gross-margin row, by ARR band (median / top quartile):

ARR band	Gross margin, median	Top quartile
<$10M	55%	80%
$10M–$25M	60%	70%
$25M–$100M	80%	85%
$100M+	75%	85%

Three reasons this is not a verdict on the 59%-by-2027 projection above, and they should be checked before anyone treats it as one. (1) Different metric — the survey forecasts gross margin on AI products; the index measures gross margin of the company. For a company whose revenue is only 32–42% AI, those are not the same line, and for a genuinely AI-native one they converge only approximately. (2) Different cohort — the survey covers ~305 software companies building AI products, most of them incumbents adding an AI line; a Pacesetter is selected on top-quartile three-year revenue growth plus ICONIQ's own AI-Native/AI-Driven label, with no control group published. (3) Different axis — the projection runs along calendar years on a fixed panel; the index runs along ARR bands across different companies in one pooled window. A cross-section cannot falsify a forecast.

What it does supply is a sanity band. A 59% AI-product gross margin in 2027 would sit below what $25M+ Pacesetters already post at the company level today (75–80%) and between the two smaller bands' medians (55%, 60%) — so the projection is not numerically outlandish, and the low margins ICONIQ's own respondents are forecasting their way out of are the ones its own measured cohort shows concentrated at small scale. Two details that cut against reading the ladder as a clean margin-improves-with-scale story: it is non-monotone (80% → 75% crossing into $100M+), and the sub-$10M band's 25pp median-to-top-quartile spread says early AI-forward companies do not share one margin structure. Figures are rounded to the nearest 5.

The pricing-model shift#

Rising margins are partly a pricing story. Subscription/platform components remain the most common model, but usage- and value-aligned models are climbing as builders align price to the cost and value of consumption:

• Consumption-based pricing at 42%, outcome-based at 23% of respondents, both up over six months.
• Companies now blend ~1.7 pricing models on average (up from 1.5 in Q4'25) — pricing architecture is itself becoming a composed product decision.
• Rationale from the ICONIQ network: match price to the cost (inference scales with usage) and the value (outcomes) of AI, rather than to seats. For consumption-priced products, token/inference cost is often a shared expense between provider and customer.

The seller-side share, re-measured on operating data three months later. ICONIQ's own State of Scaling reports that 70% of Pacesetters now tie part of their revenue to consumption (hybrid or consumption-based pricing, Q2 2026 firmographics) — on a growth-selected cohort rather than the ~305-company survey panel, and measuring "tie part of revenue to" rather than "use this model," so it is not a restatement of the 42% above. The operators' own advice is the useful part: "Launch pricing gradually" — phase it into existing customers, apply it to all new ones; "Don't have humans forecast consumption" — ML models "consistently outperform human judgment on usage-based revenue"; and keep new-logo acquisition first-order longer than feels natural, because de-emphasizing it "around $500M ARR" can "undermine the revenue model before the business is ready for that transition."

The buyer side of the same shift, and the gap between them (a16z via TechCrunch, September 2026)#

ICONIQ measures what builders charge. Julie Bort's TechCrunch piece (2026-09-03, practitioner-opinion — a news article reporting an unread VC survey) supplies the first reading of what buyers want, and the two do not line up: a16z surveyed 50 technical AI buyers and more than half want AI fees tied to the work produced or other outcomes rather than to usage like tokens consumed. Against ICONIQ's 23% outcome-based on the seller side, buyer-side demand runs roughly 2× ahead of supply — on two different cohorts, two different instruments and two different questions (what would you prefer, vs what do you bill), so the ratio is directional at best and both numbers are self-report.

a16z partners Tugce Erten and Sarah Wang give the argument rather than the measurement: pricing "around the recognizable work" — reports processed, tickets closed, leads generated — is what proves worth to the customer and makes the product "economically valuable to both sides", while token-metered pricing is a SaaS-era seats-and-storage reflex applied to a product whose value is not proportional to its consumption. Their post was not fetched; the quotes reach the wiki through Bort.

Two qualifications the enthusiasm has to survive, both already on this page. (1) It is the opposite of what the cost side wants: this page's own overrun ranking puts token spend first, with one workflow projected at $0.10/run reaching $1.50+ once agents retried and self-corrected. Outcome pricing hands the seller that variance in full — a fee denominated in closed tickets is a fee that does not move when a run costs fifteen times its estimate, which is precisely why the 23%/1.7-models picture is a blend rather than a switch. (2) The demand is being expressed by a buyer with a six-month renewal review over the seller's head (Seven Powers Applied to AI), so "tie it to outcomes" is a request for a pricing model and an accountability instrument; the two arrive together and the second is the one that binds.

A third buyer reading, on a larger panel and with the guardrail attached (2026-09-22). ICONIQ's State of Scaling carries its own Enterprise Buyers of AI-powered Software survey (June 2026, n = 132): 40% of enterprise AI buyers prefer either hybrid consumption + subscription (25%) or fully consumption-based (15%) — and "~60% of respondents require a guarantee that costs won't exceed a threshold and real-time visibility into usage" before they will accept it. That is the missing half of the a16z reading and it changes the picture: buyers are not asking to bear consumption risk, they are asking to shift it, with a cap. A capped consumption contract with real-time metering puts the overrun variance back on the seller exactly as outcome pricing does — so on the cost side the two buyer preferences are the same preference, and both point at the token-spend line ranked first in the overrun list below.

The provider mix reshuffled — Anthropic to #1#

Multi-model is now the default: builders run ~3.3 model providers on average (up from 3.1 in Q4'25), and the top of the leaderboard reordered in six months (Q4'25 → Q2'26, % of respondents, select-all):

• Anthropic 51% → 81% — jumped from #3 to the top provider among these AI-building software companies.
• OpenAI 77% → 71% (now #2); Google 56% → 50% (#3); Azure 30% → 26%; AWS 27% → 22%; Meta 21% → 21%; then Databricks 9%, Mistral 8%, DeepSeek 7%, Alibaba 6%, Moonshot 4%, xAI 3%.
• Selection criteria held their order: model reliability & accuracy #1, cost #2 — but security/privacy climbed (#4→#3) and SOC2/enterprise SLAs climbed (#10→#8), signaling enterprise-readiness pressure as products mature. The ICONIQ network frames security as "becoming a switching cost" and observability/regulatory-explainability as the emerging blockers (a CISO at a global insurer "can't yet describe to regulators what deployed agents are doing").

This is a market-share reading (breadth of adoption among ~305 builders), not a spend or token-volume reading — but it is the vault's first datapoint on Anthropic's builder-side provider position, and it corroborates the Anthropic entity's "$11B ARR, rapid growth" narrative from the demand side.

Independently corroborated by payment records (added 2026-08-04). Ramp's July 2026 AI Index (empirical, corporate-card and bill-pay records across Ramp's customer base) finds the same reordering over the same six months on a completely different population and instrument: Anthropic passed OpenAI in May 2026 and reached 42.4% of US businesses in June against OpenAI's 39.5%, with OpenAI down ~1.9pp from its November-2025 peak of 41.4%. The two readings are not comparable in level — ICONIQ asks ~305 AI-building software companies which providers they use (select-all), Ramp counts which vendors get paid by every business on its rail — but a self-reported survey and a receipts-based panel agreeing on direction and timing is stronger than either alone, and this page's own evidence note is a warning about the survey side. See Telemetry vs. Survey Measurement, where the convergence is the notable part.

Ramp's data also puts a bound on the deck's Chinese-model tail (DeepSeek 7%, Alibaba 6%, Moonshot 4% of ICONIQ respondents): across all AI-spending US businesses, only 5.8% pay any model-serving or inference platform — Ramp's proxy for open-source and Chinese model access — and 96.4% of those still pay OpenAI or Anthropic directly. ICONIQ's higher percentages are consistent with that, because its cohort is the AI-intensive tail Ramp isolates; on both instruments the Chinese providers sit an order of magnitude below the American labs and are being added to portfolios rather than swapped in. Full treatment at The Open-Weight Frontier Gap.

Internal AI: rising spend, unpredictable true cost#

The deck's other economic frontier is internal AI — what it costs a builder to make its own workforce AI-productive:

• Internal AI-systems spend is projected to rise from 11% to 16% of revenue in 2026, with further growth expected in 2027 — up from the 1–3% of revenue ICONIQ measured in prior analyses. The figure is deliberately broad: it folds in indirect spend (change management, upskilling, data governance) to capture the "true cost of AI," which respondents say is hard to predict.
• Where the overruns come from (ranked): (1) token spend — moving from single-turn calls to multi-step agentic pipelines scales cost non-linearly (one workflow projected at $0.10/run reached $1.50+ once agents retried/self-corrected); (2) data infrastructure — production-grade RAG, permissioning, structuring; (3) organizational enablement — governance, usage standards, sustained training that "rarely appears in initial business cases." Token count itself "fails fast as a metric — once consumption becomes the target, the practice undercuts the goal"; charging AI usage back to cost centers works better.
• Realized productivity is real but modest and use-case-jagged. Self-reported productivity gains lead in coding assistance (48% at high-growth vs 32% at peers, +16pp gap), then documentation/knowledge-retrieval and product & design (~40%), down to sales/HR (~25%). AI agents specifically deliver <30% gains across every revenue band — lower than non-agentic AI — and "often require human intervention" for multi-step or context-heavy tasks. High-growth firms also ramp new AI tools faster (2.5 vs 3.5 months to value) and write more AI-assisted code (59% vs 47%). The internal-productivity-as-moat exemplar is Ramp (99% internal AI adoption; 350+ reusable workflows shared company-wide, Git-backed, versioned, and reviewed like code) — the internal-productivity-as-moat and skillify-it disciplines observed in one company.

What the cost line actually looks like, measured elsewhere (ICONIQ State of Scaling, September 2026)#

Every figure above is a builder's self-report of its own internal spend. 2026 State of Scaling (ICONIQ Venture & Growth, September 2026, empirical) — the 52-page annual report whose cohort work the Pacesetter Index excerpts — approaches the same cost line from three outside angles, and none of them measures AI spend as a share of revenue, which is worth stating plainly because the report is the most likely place a reader would look for it.

• The input price fell by half, and the mechanism is routing, not discounting. The Silicon Data SDLLMTK token-expenditure index — realized blended cost-to-serve of LLM inference per million tokens, weighted across 20+ models, daily close 2025-12-01 to 2026-08-17 — runs $1.02 → $1.79 → $1.26 → $2.07 → $1.05, peaking around June 2026 and ending roughly where it started. ICONIQ's own reading of the ~50% fall from peak: "model routing is becoming a margin lever: companies progressively optimize workloads across providers based on cost, performance, and task complexity." The chart's phase labels attribute the swings to model releases — a "premium mix shift" (GPT-5.2, Claude Opus 4.6), an "open-weight migration" (Kimi K2.5, Qwen3.5), then "frontier quality premium re-expands" (GPT-5.4/5.5, Claude Opus 4.8). This is a price index, not a spend index: a company can face a halving unit price and still double its bill.
• Public companies started talking about the bill. In earnings-call transcripts across calendar Q2 2025 – Q2 2026, the share of substantive AI commentary devoted to revenue/monetization tripled 5% → 15%, and compute/infrastructure cost doubled 3% → 6%, "concentrated in inference-heavy companies" (p.19; six categories normalized to 100%, with ~56% of all AI mentions excluded as generic).
• The only company-level number is an anecdote, and it is denominated in payroll. A CFO at a $500M+ ARR fintech: "Our token spend went from near zero to ~5-10% of payroll, so we pulled that out of the future headcount plan. Mid-tier model defaults and caching then took us well below that even as usage climbed." One unnamed company, a share of payroll rather than of revenue, and a figure the speaker says he then cut. But the framing is the sharpest statement of the substitution this page's spend section implies without measuring — token spend funded out of the headcount line, which is the spend-to-headcount trade observed from inside a single P&L. His other claim is the durable one: "The hard part is finding the value, not controlling the cost."

Parse note. The payroll figure reaches the vault as "~510% of payroll" in the docling-parsed raw — a hyphenated range broken across a line (~5- / 10% of payroll) and welded into a plausible-looking nonsense number. Recovered from pdftotext -layout; no table check can see this class of damage because it happens in prose.

Forward-deployed engineers as a monetized GTM motion#

The deck's talent spotlight is the forward-deployed engineer (FDE) — an embedded, customer-facing engineering role that AI-builders are institutionalizing as a revenue function, not a cost of delivery.

The role now has its own page. Forward-Deployed Engineering as a Delivery Layer carries the delivery-layer treatment — ICONIQ's operator definition and org placement, the 5–10x fully-loaded-cost economics gate and the per-market staffing template, the supply estimates, buyer reception, and the three competing answers to who owns the layer (model vendor / customer-internal / systems integrator). This section keeps the monetization reading, which is this page's subject, and the two open questions below stay here for the same reason.

The deck's own figures:

• ~50% of companies are scaling up their FDE model as a permanent part of the GTM motion (not a temporary implementation crutch); enterprise customers are expected to see the highest FDE integration by 2027.
• Primary role FDEs play (single-select, N=200): Revenue Driver 38% (materially contribute to expansion/retention/strategic-account outcomes) > Product Intelligence Loop 24% (customization deliberately informs the core roadmap) > Delivery Necessity 22% (needed for implementation, not a growth driver) > Product Gap Coverage 17% (compensating for capabilities not yet built). A majority frame FDEs as growth or product-signal, not mere services.
• Compensation ~70/30 base/variable, variable usually tied to customer retention/renewal — an explicitly outcome-aligned comp structure. Monetization remains fragmented: bundled into subscription, billed as separate professional-services fees, or hybrid.

The FDE-as-revenue-driver framing extends the "keep GTM close to the product" instinct (Founder-Led Sales Discipline) into a scaled, post-PMF form, and the Product-Intelligence-Loop role (24%) is a GTM-side customer-signal-into-roadmap loop.

The supply side, and who owns it (TechCrunch, July 2026)#

ICONIQ's read is demand-side and builder-side: ~305 companies say they are scaling FDEs. TechCrunch's 2026-07-30 report (Rebecca Bellan, vendor-claim) covers the other side of the market — and its numbers must be handled at arm's length, because every one of them comes from an unpublished proprietary study by Christian & Timbers, an executive search firm that recruits for the role its study declares scarce, shared exclusively with TechCrunch. Method as reported: 250+ C-suite hiring executives across 180 companies, 80 Fortune 500 executives surveyed, 300+ FDEs interviewed, January–June 2026. Survey-shaped, but the seller is also the source.

• Independent directional corroboration of ICONIQ's demand finding, on a different population and instrument: companies planning to hire FDEs went from 5–10% at the start of 2026 to 70% by the end of Q2, with the largest consulting and services firms reporting a need for 10× their FDE headcount (teams of 20–100). ICONIQ's ~50% of AI-builders scaling FDEs as a permanent GTM motion is a different question asked of a different cohort, and the two agree on direction and timing. Weight accordingly: empirical survey vs vendor-claim.
• The scarcity claim, which is the part with the seller's thumb on it: ~17,000 US FDEs on the market, of whom only ~2,000 — "Not 2,000 available… 2,000 total" — have the sector knowledge plus applied-AI experience to deliver ROI, now defined as "multiple tens of millions of dollars of ROI impact." A recruiting firm's estimate of how few people can do the job it recruits for is the least independent number in this page; the qualitative version from a second, also-interested party (Chris Taylor, CEO of Ode with Anthropic) is more usable: "Many FDEs are well equipped to help you roll Claude Code out to your workforce. Very few are capable of building your flagship AI product feature."
• The frontier labs are now the FDE supply. Anthropic's Ode and OpenAI's Deployment Company are staffed with FDEs "whose sole purpose is to go forth and spread their tech around the enterprise" — vertical integration into the delivery layer, motivated by the labs' own path to profitability against cheaper open-weight competition.
• Which produces the strategically interesting finding, and it is a moat argument, not a talent one. Enterprises are building internal FDE teams rather than renting them from Ode or Deployment Co., explicitly to avoid handing their proprietary business processes to their model vendor: "Everybody's concerned that if they give up their proprietary business processes, [the AI firms] can compete with them, which is true in many different areas." This is workflow lock-in read from the customer's side — the same encoded-process asset that is a startup's moat is, for the enterprise, an asset it declines to expose to a supplier who could climb the stack with it. Taylor confirms the phrase "internal forward-deployed engineers" is circulating while noting his own clients aren't yet asking Ode to build such teams for them — i.e. the motive is documented, the behaviour mostly isn't yet.
• It does not touch the monetization question below. The article never discusses how FDE work is priced, bundled, or booked, and carries no margin analysis. The open question stands untouched, which is itself worth recording: the loudest coverage of the role in the corpus is silent on whether it pays.

The self-undercutting bit, kept because it is the honest part. Christian's own forecast is that the role may not last: "Maybe in two years, everything's automated, and agents are automating agents as opposed to humans automating agents", with the FDE possibly gone in five to ten years — a prediction from the person whose business depends on the opposite, and the reason the 2,100%-by-year-end demand projection should be read as a marketing number rather than a forecast to hold this page to.

Connections#

• The Price of Fixed Capability — a measured bound on how long a capability premium lasts: the cost of a newly-SOTA score falls ~66% in its first quarter and ~32% per quarter two years on. Epoch concludes that "supra-normal profits in LLM service provision may be fleeting for any given model," which is the input-price side of this page's margin projections
• Forward-Deployed Engineering as a Delivery Layer — the role this page monetizes, treated as an industry-wide delivery layer: the operator definition, the 5–10x fully-loaded-cost gate that bounds the Revenue-Driver framing, the per-market staffing template, and the three competing owners of the layer (model vendor, customer, systems integrator)
• AI Investment Story, Not Efficiency Story — the RPE half of the same ICONIQ deck: ARR/revenue-per-FTE is the vault's fourth revenue-per-head instrument (survey + forward projections), sorted into the instrument frame there; this page carries the margin/pricing/cost economics the RPE page references
• Organizational Complements to AI — ICONIQ's internal-AI-spend jump (1–3% → 11%→16% of revenue) and its three overrun sources (tokens, data infra, enablement) are a priced-out enumeration of the complements that gate AI's value — the cost of the workflow/skill/governance redesign that page argues productivity gains depend on
• AI-Native Organization — the org-restructuring half of this deck (flatter orgs, role-mix shift, function-level headcount, FDE as a new hiring category); the "internal productivity is a moat / skillify it" discipline (Ramp's 350 versioned workflows) is Tan's thesis measured across ~305 builders
• AI-Native Startup Lifecycle — this is what the Scale stage's P&L looks like once AI products are the revenue: margin optimization, pricing composition, provider portfolio, and FDE-driven expansion
• Telemetry vs. Survey Measurement — the instrument caveat: ICONIQ is a survey with forward projections, so its optimistic margin/RPE trajectory is self-report about the future, weighted below measured receipts where the two disagree
• Compounding Data Moat — Ramp's internal-productivity moat and the "model quality is rented, you own how you wire the work" reading of the provider reshuffle: switching providers is cheap (3.3 in the portfolio), so durable advantage sits in the internal workflow layer, not the model
• Seven Powers Applied to AI — why the pricing architecture matters more than it used to: with the moat of inertia gone from enterprise AI contracts, outcome-denominated fees are the mitigation that page's switching-cost row prescribes, expressed as a billing unit
• Cost-per-Task Over Cost-per-Token — the same billing-unit argument one layer down, at the model API rather than the application: what the seller pays per token against what the buyer wants to pay per unit of finished work
• ICONIQ — the publisher of both decks behind this page, and the reason its two ICONIQ readings must be weighed separately rather than pooled: a self-reported exec survey with 2026P/2027P projections, and a measured operating-data benchmark on a growth-selected cohort with no control group
• Anthropic — the builder-side market datapoint: Anthropic moved to the #1 model provider (51%→81%) among the surveyed AI companies over six months

Open Questions#

• ICONIQ's respondents project gross margins expanding to ~59% by 2027, while Emergence's cap-table data measures the fastest-growers running 6–16pp below peers today. Does the projected margin expansion materialize, or is it survey optimism that regresses toward the measured growth-margin tradeoff as these companies scale? Checked against the same publisher's own measured benchmark and still not answered (2026-09-22): The ICONIQ Pacesetter Index is measured rather than projected, which is the upgrade this bullet wants — but it is measured by ARR band, not by year (55% / 60% / 80% / 75% median across <$10M / $10–25M / $25–100M / $100M+), on a different cohort (top-quartile-growth AI-forward companies, no control group) and at a different level (company gross margin, not gross margin on AI products). A cross-section of different companies at different sizes cannot falsify a forecast about one panel across time, and this one is not even monotone — median margin falls 80% → 75% entering the largest band. It does confirm, on receipts rather than self-report, that the growth-margin tradeoff the bullet's other half asserts is real and is concentrated in the smallest, fastest-growing bands. The trigger event is unchanged: 2027 actuals for the surveyed panel, or any same-companies-across-years margin series. Extended, and the missing comparator arrived without helping (2026-09-22): the full report the index excerpts, 2026 State of Scaling: The Great Sorting, plots gross margin by year and against a non-Pacesetter comparator — the two things the index lacked. Company-level medians: under $100M, 72% / 74% / 70% / 70% (2022-23 → 2026), with Pacesetters at 71% against Others' 70%; above $100M, 78% / 79% / 79% / 78%, with Pacesetters at 73% against Others' 78%. So a time axis now exists and it is flat — no aggregate margin expansion anywhere in four years — and the only real gap is the ~5pp ICONIQ attributes to Pacesetters' 'higher inference, infrastructure, and deployment costs' at scale. This is still not the quantity the bullet asks about: 70-78% is company gross margin, and the 45%→53%→59% projection is margin on AI products specifically, a sub-line that a blended company margin can hide entirely. What it does add is the direction of the drag — at the top of the ladder AI-forward companies give up margin rather than gaining it, which is the opposite of the projection's slope. n = 4-12 company-quarters on the Pacesetter cells. Trigger event unchanged.
• FDEs are monetized fragmentedly (bundled / separate PS fees / hybrid) and comped on retention. Does a dominant FDE monetization model emerge, and does the "Revenue Driver" self-framing (38%) survive a margin analysis — i.e. are FDEs actually accretive, or a services drag reclassified as growth? (Still open, and pointedly so: the corpus's most prominent July-2026 coverage of the role discusses supply, scarcity, and vendor structure at length and says nothing about pricing, bundling, or margin. The answer will come from a filing or an operator's P&L, not from role coverage.) A third staffing model appeared with no price attached, which adds a column and no evidence (2026-09-22). Accenture and Google Cloud Deepen Partnership with Formation of New Accenture Gemini Enterprise Business Group (Accenture newsroom press release, 2026-09-08, vendor-claim) announces a global systems integrator staffing a 1,000-person FDE workforce on a model vendor's platform — neither the vendor-side delivery arm nor the customer-internal team this page's other bullets track. Accenture and Google Cloud state the workforce is built on "nearly 50,000 Google Cloud-skilled professionals" and that Google Cloud will train them; the release states no billing model, no employer for the FDEs, no timeline and no margin. So the count of plausible monetization models goes from three (bundled / separate PS fees / hybrid) to four (an SI's own services contract, where the platform vendor books consumption and the integrator books time), while the evidence on which dominates stays at zero — and the largest single FDE commitment in the corpus is the one that discloses least. The nearest thing to an economic test arrived from a different source and is a hurdle rate, not a result: ICONIQ's operators in 2026 State of Scaling: The Great Sorting state FDE economics "typically only makes sense when deployed against opportunities that can drive a 5–10x return on fully loaded cost," with capacity scaling "linearly with headcount" and viability only "at meaningful ACV levels." That cuts against the Revenue-Driver self-framing in a specific way the bullet should keep: if the motion only clears its cost against large accounts, then 38% "Revenue Driver" is a statement about the accounts FDEs are assigned to, not about the motion's margin, and the two are separable. Demand-side, the same report's n=132 buyer survey has 73% viewing FDEs positively — the layer is wanted, which is not the same as accretive. Trigger unchanged, with one new candidate: an SI segment disclosure that breaks out agentic-deployment services.
• Enterprises are building internal FDE teams specifically to avoid exposing proprietary business processes to their model vendor (C&T via TechCrunch, vendor-claim, motive documented via one recruiter and one vendor CEO; the behaviour mostly not yet observed — Ode reports no client asking it to build such a team). Does that in-housing actually happen at scale, and if it does, does it cap the FDE-as-revenue-driver motion at exactly the accounts worth the most — i.e. is the labs' delivery-layer integration self-limiting? Falsifiable from job-postings data (internal FDE-titled roles at non-vendor enterprises) against vendor-services revenue disclosures. A genuinely new axis, and it does not dissolve the concern (2026-09-22): Accenture and Google Cloud Deepen Partnership with Formation of New Accenture Gemini Enterprise Business Group (vendor-claim) puts a fourth party in the frame. A global systems integrator staffing 1,000 FDEs on Google Cloud's Gemini Enterprise is neither of the two options this bullet contrasts: the FDE relationship, and with it the encoded process knowledge, is owned by a firm that sells no competing software and whose interest in the account outlives the platform choice. Taken literally that answers the stated fear — the model vendor does not get the processes. It does not answer the fear's actual content, and the release says so in its own words: the group's second priority is "building repeatable, industry-specific solutions," i.e. process knowledge acquired at one client and resold to its competitors, by a firm serving ~9,000 clients. So in-housing and the SI route are not substitutes, and a buyer holding this concern has no reason to be reassured by the third model. Two further caveats keep this from being progress on the bullet's falsifiable half: it is a vendor-claim announcement with no timeline and no employer named for the 1,000 engineers, so no behaviour has yet been observed; and it says nothing about whether enterprises are building internal teams, which is still the quantity the bullet asks about. What it does change is the job-postings test named above — an internal-FDE count is now ambiguous between in-housing and SI-supplied staff placed on client sites, so the instrument needs to separate employer from work location.
• Internal AI spend jumped from 1–3% to a projected 16% of revenue with respondents calling true cost hard to predict. Is 16% a transient enablement bulge that falls as tooling matures, or a durable new cost line for software companies? (Partially answered on the price leg, untouched on the spend leg (2026-09-22). 2026 State of Scaling: The Great Sorting is the same publisher's operating-data report and it measures the input price, not the spend: the Silicon Data blended token-expenditure index runs $1.02 → $1.79 → $1.26 → $2.07 (Jun'26) → $1.05 (Aug'26), roughly halving from peak, with ICONIQ attributing the fall to model routing across providers rather than to list-price cuts. A halving unit price is a necessary condition for the bulge deflating and nowhere near a sufficient one — spend is price × volume, and the same report's agentic-workflow evidence has volume rising fast. The one company-level datum is an anecdote and is in the wrong unit: a $500M+ ARR fintech CFO says token spend went 'from near zero to ~5-10% of payroll' before mid-tier defaults and caching took it 'well below that even as usage climbed' — one unnamed firm, a payroll denominator, and a figure the speaker cut himself, but it is a direct observation of a cost line first inflating and then being engineered down, which is what 'transient bulge' predicts. Public-company earnings calls corroborate the line becoming salient without sizing it: compute/infrastructure-cost mentions doubled 3% → 6% of substantive AI commentary Q2'25 → Q2'26. The trigger event is unchanged and now sharper: the same panel re-surveyed on AI spend as a share of revenue for 2026 actuals — ICONIQ publishes no such figure in 52 pages, so this cannot be settled from the operating-data side.)

Sources#

• State of AI 2026: The Builder's Economy — ICONIQ Growth, State of AI 2026: The Builder's Economy (2026-07-08, empirical): §"AI Go-to-Market & Economics" (revenue-from-AI %, gross-margin trajectory, pricing-model mix, unit-economics levers), §"AI Models & Infrastructure" (provider mix, selection criteria), §"AI for Internal Productivity" (internal AI spend, cost overruns, productivity-by-use-case, agent gains, Ramp case study), §"Talent & Organization → Spotlight: Forward Deployed Engineering" (FDE role, comp, monetization). Most chart values read from the deck's images (two-pass): revenue % (image_000037), gross margin (image_000046), provider mix (image_000020), productivity-by-use-case (image_000089); FDE-role split and internal-spend figures from native text/tables.
• Forward-deployed engineers are the AI industry's latest talent obsession — Rebecca Bellan, TechCrunch, 2026-07-30 (vendor-claim — every figure is from an unpublished Christian & Timbers study, a search firm recruiting for the role it declares scarce, "shared exclusively with TechCrunch"; method as reported: 250+ C-suite interviews across 180 companies, 80 Fortune 500 surveyed, 300+ FDEs interviewed, Jan–Jun 2026): the 5–10%→70% Q1→Q2 hiring-intent jump, the ~17,000-total / ~2,000-elite supply estimate, Ode with Anthropic and OpenAI's Deployment Company as lab-owned delivery arms, the in-housing-to-protect-proprietary-process motive, and Christian's own prediction that the role may disappear within 5–10 years. Silent on pricing, bundling and margin
• The ICONIQ Pacesetter Index — ICONIQ Venture & Growth, The ICONIQ Pacesetter Index (2026-09-17, empirical, not a survey — quarterly financial and operating data 2024 – Q2 2026 from public software companies plus ICONIQ's own venture and growth portfolio). Cited here only for the company-level gross-margin row by ARR band (55% / 60% / 80% / 75% median; 80% / 70% / 85% / 85% top quartile) and ICONIQ's prose that "Pacesetters often operate at lower margins." Selection caveat on every figure: a Pacesetter is selected on top-quartile three-year revenue growth plus ICONIQ's own AI-Native/AI-Driven label, and the publication carries no control group; COI: the benchmark sample is partly the publisher's own portfolio. Parse note: the whole benchmark table exists only as a page image (); all 56 cells were legible and re-verified cell-by-cell at compile time. Figures rounded to the nearest 5; the report PDF is lead-form gated and was not fetched. Instrument provenance at ICONIQ
• Startup ARR is less secure than ever, new research shows — Julie Bort, TechCrunch, 2026-09-03 (practitioner-opinion): the a16z half only — a survey of 50 technical AI buyers, over half wanting fees tied to outcomes rather than tokens, with the "pricing around the recognizable work" argument attributed to partners Tugce Erten and Sarah Wang. Press reporting of a VC post that was not fetched; n=50 with no sampling frame published, and the article's own framing (enterprise ARR insecurity) is the reporter's synthesis, handled on Seven Powers Applied to AI
• Accenture and Google Cloud Deepen Partnership with Formation of New Accenture Gemini Enterprise Business Group — Accenture newsroom press release, 2026-09-08 (vendor-claim, ~1,120 words; mirrored on googlecloudpresscorner.com, ingested from the Accenture copy only). Cited here only in the two FDE open questions, for the third staffing model (a systems integrator committing a 1,000-person FDE workforce to Google Cloud's Gemini Enterprise, built on ~50,000 Google Cloud-skilled Accenture professionals) and for what it withholds: no billing model, no employer for the FDEs, no timeline, no curriculum, no margin. First-party throughout — quotes from Julie Sweet and Thomas Kurian, with an Everest Group analyst endorsement supplied inside the release. No figure from it is load-bearing on this page's economics. Full handling at Forward-Deployed Engineering as a Delivery Layer
• 2026 State of Scaling: The Great Sorting — ICONIQ Venture & Growth, 2026 State of Scaling: The Great Sorting (September 2026, empirical; 52-page PDF, docling-parsed). The annual report the Pacesetter Index excerpts. Cited here for the Silicon Data token-expenditure index (p.18, hand-transcribed at ingest from an image-only chart), the earnings-call AI-commentary shares (p.19, likewise), the 70%-consumption-pricing firmographic and operator pricing advice (p.26, p.28), the gross-margin-by-year-and-comparator chart (p.36), the enterprise-buyer survey (June 2026, n = 132) and the CFO token-spend-as-share-of-payroll quote (p.47). Parse warning: that quote is welded in the docling body as "~510% of payroll"; the true text is "~5-10% of payroll", a hyphenated range broken across a line — recovered from pdftotext -layout, and a failure mode none of the table checks can see. Selection, COI and the company-quarter n trap at ICONIQ