AEC software war games

Higharc sells production homebuilders an end-to-end design-to-sales tool: configurator, options pricing, contract generation, and permit-ready construction documents written into a single per-home record. Pricing is per-home-sold, not per-seat. ~40K homes/year on platform; ~$19B sales volume passed through the system (per 2025 company disclosures and ENR reporting). Funding through 2026 totals over $130M led by Spark Capital, Foundation, USAA, Mucker.

Autodesk Revit + AutoCAD on the design-tool side; Hyphen Solutions, MarkSystems, and Newstar on the homebuilder-ERP side. In practice, the deeper defender is the legacy commercial structure itself — spec-home design firms, change-order revenue, the seat-licensed CAD shop — not any single product.

Foundational: Counter-Positioning. Higharc's per-home pricing makes the cannibalization bind structural for Autodesk. Every dollar of generative-design investment Autodesk makes lands on top of seat licensing and accelerates the bind it cannot afford. Foundational: Data Flywheel — configurator decisions, options-priced sale, locked-in contract, change-order events all written into a per-home corpus the foundation-model layer cannot absorb. Supporting: Switching Costs on builder operations once the configurator becomes the system of record. The key load-bearing claim is that the flywheel is decision-bearing, not event-bearing — what the buyer chose, what the builder priced, what the sales team locked in.

Higharc continues to deepen its top-25-builder ICP. Autodesk has no credible answer; Forma is concept-design-side and explicitly hands off to Revit for documentation. Builder ERPs partner rather than fight (the 2025–2026 integration announcements are the tell). H1 is a clean continuation of the existing trajectory.

The branching question is whether the per-home commercial model travels to multifamily wood-frame (3–5 stories), suburban warehousing, dark-store retail, modular — the typology adjacencies where generative BIM is on track to clear the production-ready bar by H2 (typologies with simple constraint manifolds, light cross-discipline coupling, liability pre-allocated to the developer/builder). Bull case: Higharc reaches IPO-grade revenue ($300M+) by extending the model into one or two of those typologies and remains independent at $5–8B+. Bear case: Autodesk concedes residential-design entirely, acquires Higharc at $1–2B as a GTM accelerant for its own multifamily play, and rebuilds the homebuilder relationship through the acquired channel. Base case: Higharc grows into a $3–8B independent — too valuable to acquire at price, too narrow-typology to threaten the dollar-weighted middle of Autodesk's book.

The per-home commercial model fails to travel to multifamily wood-frame. Tract single-family is structurally narrow enough to ceiling Higharc at $1.5–2.5B revenue; the bull case requires at least one credible typology adjacency by 2028 to reach independent IPO-grade. Inflection signal: any multifamily-wood-frame production-builder logo on the platform by 2027-Q4.

Speckle is the open-source project-graph — the database layer that holds building-project data (geometry, schedules, RFIs, properties) and exposes it to agents and tools through a portable API. It does not generate BIM; it makes existing BIM AI-queryable. Suffolk Technologies led an investment in April 2026; Pomerleau, gbi, and others are deploying it internally as an internal system of record. Open-source codebase, hosted commercial offering, paid enterprise tier.

Autodesk Platform Services (APS) — the tiered API layer Autodesk introduced in December 2025 to gate access to Revit data. Bentley iTwin and Trimble Connect are weaker analogs in adjacent segments.

Foundational: Counter-Positioning via open-core. Autodesk cannot ship an equivalent without breaking its API-monetization model. But this is the SOFTER form of counter-positioning, not the sharper "literally cannot acquire" form — open-core acquisitions are routine. Microsoft acquired GitHub ($7.5B, 2018) and the open-source ecosystem survived; IBM acquired Red Hat ($34B, 2019); Salesforce acquired MuleSoft ($6.5B, 2018). Supporting: Data Flywheel at the integration layer — the more projects that commit to Speckle as system of record, the more agentic tools build natively against it, the harder it is to replatform off. Supporting: emergent Network Economies across the federated tool ecosystem (the chapter 09 federated-network reframe).

Speckle wins mindshare in the AI-integration narrative. AEC-AI tool builders default to it because the alternatives (APS tier-2 pricing, ODA SDK extraction) are either gated or fragile. Autodesk doubles down on APS pricing rather than open the project-graph — the December 2025 move signals this is the chosen defense. The acquirer-vs-independent fight is decided in the H1–early-H2 window: Autodesk either acquires Speckle in 2026–2027 to neutralize the API-perimeter threat, or it lets Speckle accumulate enough mindshare and project lock-in to make later acquisition prohibitively expensive.

If Autodesk acquires by 2027, the deal clears at $300M–$1.5B and Speckle becomes the shipping open-graph layer Autodesk should have built itself. If Autodesk waits, two outcomes diverge: (a) it ships a competing free APS tier and starves Speckle of the commercial logo path, or (b) Speckle reaches mindshare escape velocity and runs independently with a $1–2B trajectory. The bull case requires Speckle to win the AI-integration mindshare war in a narrow 2–3-year window before Autodesk concludes acquisition is the cleanest defense.

Autodesk ships a credibly free APS tier targeted at the AI-integration use case before Speckle reaches mindshare escape velocity. That move starves Speckle of the commercial-logo path without requiring acquisition — the cheapest defense if Autodesk plays it. Inflection signal: APS tier-2 pricing changes or a free-tier announcement at AU 2026.

This is paper-wide asterisk #3 — the cleanest test of counter-positioning at the API perimeter under AI pressure.

Snaptrude is a browser-native architectural modeler with generative concept-design and bidirectional Revit sync. Spring 2026 release pushes the model toward LOD 300–350 (schematic-plus). IFC export, native cloud collaboration, fast iteration loops on early-phase architectural coordination. Used by AE firms primarily in concept-through-schematic; explicitly hands off to Revit for construction documents. Funding through 2026 ~$30M (Series A–B, Accel and others).

Autodesk Revit (production-documentation incumbent) plus Autodesk Forma (concept-design first-party). The fight is structurally different at each end: Forma is the concept-side competitor; Revit is the documentation incumbent Snaptrude explicitly does not try to displace.

Snaptrude has a real but bounded Counter-Positioning on browser-native and generative-first UX that Autodesk's desktop Revit cannot ship without cannibalizing seat licensing. Weak Data Flywheel: concept-design decisions are valuable but bounded — the corpus does not extend into construction documentation, the dollar-weighted decision surface. Snaptrude is on track to clear the production-ready bar in simple-building markets by H2 (similar to Higharc's territory), but stays concept-only in commercial / institutional / healthcare / lab / infrastructure / complex residential. The split outcome is the call: Snaptrude wins where the constraint manifold is small and liability is pre-allocated to the developer/builder; Autodesk holds the rest.

Snaptrude wins on UX and iteration speed in concept-design at progressive AE firms. Forma matures but stays concept-side by Autodesk's own architecture. Snaptrude does not threaten Revit's documentation seat in any segment outside narrow typologies.

The split hardens. In simple-building markets (multifamily wood-frame, suburban warehouse, dark-store retail), Snaptrude can plausibly extend toward production-ready BIM and compete at the document level. In the dollar-weighted middle, it remains concept-only because cross-discipline coupling, evaluation closure, accountability transfer, and tool-graph maturity do not close in 3 years. The likely outcome: Snaptrude is acquired into a generative-design platform — either Autodesk neutralizing the concept-side threat, or a homebuilder-adjacent stack consolidating the simple-building segment.

Snaptrude tries to extend into construction documentation in commercial/institutional and gets caught: the typology economics do not support it, the accountability surface does not transfer to AI-authored docs, and the company burns runway on the wrong wedge. Inflection signal: any Snaptrude-stamped construction-document submission to a major AHJ for a Type-I building by 2028. Negative signal — if it happens, the simple-vs-complex split is breaking earlier than predicted.

Trunk Tools deploys document-AI agents on submittals, RFIs, contract clauses, and spec sections. Gilbane deployment reviewed 2,000+ submittals in 3 months and flagged 72% as non-compliant; ~50% submittal-cycle-time reduction (per company press / ENR coverage 2025). Founder Sarah Buchner: PhD Civil Engineering + Data Science, Stanford GSB, carpenter at 12. $40M Series B led by Insight Partners, July 2025.

The document/spec/RFI/submittal layer is fragmented — no clear single incumbent. Procore (workflow), Autodesk Construction Cloud (ACC, federated models), Bluebeam (markup), PlanGrid-via-ACC (legacy), and internal GC processes all compete for slices. Document Crunch is the closest peer competitor (contract-review-heavier).

Foundational: Evaluator Power at Priya's layer-3 — the institutional standing to refuse a non-compliant submittal with weight backed by accountability transfer. The 72%-correct-refusal rate is the moat, not the 28%-processed throughput. Customer pays for the calibrated “no.” Supporting: Data Flywheel on a custom construction-document LLM, decision-bearing because the corpus records what the GC's submittal reviewer accepted, rejected, or escalated. Switching exposure: Procore revoked API access in September 2025; resolution unknown as of mid-2026 — the platform-risk vector is real.

Trunk Tools deepens enterprise-GC logo penetration (Gilbane, Suffolk, peer firms). Procore native AI Copilot + Agent Studio launches but at lower precision in narrow document categories. Trunk holds via depth and custom-LLM advantage. Document Crunch coexists in adjacent specialty.

The accountability-stamping layer is essentially untouched at H2 per Priya. Trunk's positioning — calibrated refusal under E&O-adjacent contractual transfer — survives delegation to frontier models. The buyer is not paying for “the AI ran the submittal”; they are paying for “Trunk caught the 72% the project would have eaten.” Independent path likely if Procore-API platform risk resolves; ACQUIRED if it does not (Procore is the buyer, not the killer, in that scenario).

Procore API revocation extends and Trunk loses the workflow-embedded data path. Without that integration, the flywheel substrate erodes and the company becomes an Insight-Partners-priced acquisition target rather than an independent. Inflection signal: status of Procore API access by 2026-Q4.

Construction-progress computer vision: 360-degree jobsite cameras, drones, helmet-cams capturing weekly walks; visual-to-BIM comparison at scale. OpenSpace cumulative project corpus: 75,000+ projects (per company disclosures). Buildots leans on Skanska / Suffolk / Turner deployment depth. Combined funding through 2026: ~$200M+. The bundle is fair because both run the identical flywheel mechanism in slightly different field-capture form factors.

Autodesk Construction Cloud (PlanGrid lineage + native field-capture), Procore (native field-capture features bundled with workflow). Both incumbents will increasingly bundle the capture function as table-stakes.

Claimed: Data Flywheel on visual-to-BIM corpus. The Q3 thesis-grade discipline applies here — no third-party benchmark exists on per-customer marginal flywheel value past year 1. Cumulative corpus depth is impressive in absolute terms; the load-bearing question is whether per-customer learning curve compounds, or whether year-2+ marginal data has decreasing value. Supporting: Switching Costs on jobsite hardware deployment and field-team workflow integration. Supporting: Cornered Resource on the deployed customer base — but this is a weak version of cornered, not a strong one.

Distribution incumbents (ACC, Procore) ship native field-capture as bundled features. Buildots and OpenSpace hold via depth and customer-integration cost. Frontier vision models (GPT-4V class on construction imagery) improve but do not yet match the corpus-depth advantage on construction-specific signal.

Two scenarios diverge sharply. (a) If frontier vision models cross a construction-specific accuracy threshold on public corpora, the corpus-depth advantage compresses and the flywheel thesis breaks — both companies become acquisition candidates at modest multiples. (b) If a “living-graph” competitor emerges that extends data-flow into operations and FM, today's frozen-at-handover field-capture becomes the wrong layer of the stack. Most likely path: ACQUIRED into Autodesk or Procore at $400M–$1.2B each, as field-capture becomes a feature of the GC-platform suite.

Frontier vision models close the construction-specific accuracy gap on public corpora before the per-customer flywheel proves out. If the moat is corpus-depth and the corpus loses its privilege under foundation-model absorption, the flywheel is the substrate, not the moat. Inflection signal: any 2027 third-party benchmark showing a frontier vision model matching OpenSpace accuracy on a held-out construction-imagery test set. [Pass 3 deep-dive will sharpen] on per-customer learning-curve data if any vendor publishes it.

Index-everything-permission-aware retrieval over project graphs, RFIs, specs, sub-contracts, change orders, meeting minutes, daily reports, financial transactions. Browser sidebar + Slack/Teams integration + native chat surface. Wedge: top-25 GCs and large AE firms with sprawling document landscapes (Gilbane, Suffolk, Mortenson; HOK, Gensler, Stantec). Per-seat pricing tiered with retrieval volume. Founders: ex-Glean / ex-Procore / ex-Autodesk hybrid team. Pre-seed in 2026 — the company does not yet exist at this scope.

Autodesk Construction Cloud and Procore both ship retrieval but neither is permission-aware across the full enterprise document landscape; both are bound to their own workflow surface. The deeper defender is the layer-fragmentation itself — nobody is yet clearly incumbent at enterprise-knowledge for AEC.

Foundational: Data Flywheel on workflow-embedded retrieval patterns — what query a project engineer asked, what document the agent surfaced, what the engineer accepted as the answer. Decision-bearing in the Helmer-plus sense (chapter 09's definition). Foundational: Agentic Workflow Lock-in via tool-graph — once GleanCo-AEC has learned which downstream tools (Procore, ACC, Bluebeam, Newforma, internal SharePoint, Egnyte) to call in what sequence for which project type, the calibration is opaque to a competing agent runtime. Supporting: Switching Costs on the integration plumbing.

The category does not yet exist. First-mover wins enterprise reference logos in the top-25-GC segment. Procore and Autodesk ship narrower retrieval-within-platform but cannot solve cross-platform without breaking their own walled-garden value props. Microsoft Copilot on enterprise tenant data is the structural alternative — but Microsoft does not understand the AEC document corpus or the permission model.

If GleanCo-AEC establishes daily-default at 5–10 enterprise GC and AE firms by 2028, the calibration depth + the cross-platform tool-graph position the company as an independent at $1–3B revenue trajectory. If it fails to ship before Procore + Autodesk realize the cross-platform retrieval is the actual battleground, it gets boxed in by walled-garden retrieval and becomes a feature, not a company. The Glean-vs-Microsoft asterisk is the horizontal precedent: it is winnable, not pre-decided.

Procore + Autodesk close their walled gardens (pricing API access out, restricting retrieval) before GleanCo-AEC has anchored daily-default at enterprise customers. Inflection signal: any 2027 announcement of cross-platform retrieval restrictions from either incumbent.

A vertical workflow co-pilot for licensed AEC professionals. Three sub-wedges to consider: (a) structural review (calibrated checking of structural engineers' output before stamping); (b) jurisdiction-specific code compliance review (egress, energy, accessibility, structural code); (c) contract administration during construction (interpreting RFIs, change-order scope, schedule-impact analysis against contract). Per-evaluation pricing or share-of-savings on E&O premium reduction. The strongest sub-wedge is (b) code compliance because the evaluation surface is well-defined, the stakes are pre-allocated to the licensed professional, and insurer underwriting is a credible 3-year path.

The fragmented professional-services layer itself: licensed structural engineers, code consultants (Jensen Hughes, Code Consultants Inc.), specialty plan reviewers (ICC Plan Review, third-party AHJ contractors). Plus Solibri, UpCodes, Semble at the rule-engine layer. None of them have the calibrated-refusal-at-enterprise-scale moat HarveyCo-AEC would build.

Foundational: Evaluator Power stacked across Priya's layer-2 (engineering judgment) and layer-3 (accountability/stamping). The product is calibrated refusal backed by E&O-grade contractual liability transfer. Foundational: Data Flywheel on jurisdiction-specific code interpretations — what the model surfaced, what the licensed professional approved, what the AHJ accepted. The corpus is privileged because it is cross-firm and cross-jurisdiction in a way no single AE firm or city can build.

HarveyCo-AEC anchors 2–3 mid-market AE firm logos and 1–2 progressive AHJ pilot programs. Mechanical-evaluation tools (UpCodes, Semble) are at lower precision and not stamping-grade. Insurer conversations begin but do not yet underwrite.

The structural-exposure question pivots on insurance. If Travelers, Berkley, or Beazley underwrites E&O on AI-assisted code-compliance review by 2028, HarveyCo-AEC's independence is locked in — the moat is the calibrated refusal under insurance-priced confidence. If the insurance layer does not move, the company stays valuable but bounded as a per-evaluation tool sold to AE firms; the independent path narrows toward acquisition by a specialty-services consolidator (Jensen Hughes, IMEG, peer roll-up).

The insurance layer does not move. Without insurer underwriting on AI-assisted output, HarveyCo-AEC cannot price share-of-savings on E&O premium reduction; without that pricing, it is a tool, not an evaluator-power business. Inflection signal — this is paper-wide implication: does any insurer underwrite E&O on AI-authored or AI-assisted design output by 2026, even narrowly? That signal is load-bearing for the entire Evaluator-Power play in AEC.

Per chapter 09's federated-network framing, this is not a verticalized stack. It is a set of specialist agent nodes — one for submittals, one for RFI generation/response, one for change-order analysis, one for daily reports — each deep at one job, each composable into the broader project-lifecycle agent network through MCP-style protocols. Per-task or per-token consumption pricing. Wedge: mid-market GCs (50–500 employees) priced out of full Procore implementations but document-heavy enough to want automation.

Procore (workflow incumbent), Autodesk Construction Cloud (federated-model + workflow), Bluebeam (markup + workflow), and the manual-process status quo at mid-market GCs. Procore is the primary defender because the consumption-priced wedge directly counter-positions Procore's per-customer-volume pricing.

Foundational: Agentic Workflow Lock-in. Each node accumulates a tool-graph and calibration that competing agent runtimes can read in name but not in calibration. Supporting: Counter-Positioning — consumption pricing makes mid-market GCs viable customers in a way Procore's annual-volume-based pricing does not. The Procore cannibalization bind is real: shifting to consumption pricing would torch the upmarket land-and-expand model. Supporting: Data Flywheel on each node's per-task corpus.

The mid-market GC wedge is real and unserved. Procore's pricing keeps it out; Autodesk does not target the segment. DecagonCo-AEC ships 3–5 nodes, lands 50–100 mid-market GC logos, accumulates tool-graph against off-the-shelf email/Slack/SharePoint/Excel rather than against an integrated platform.

The federated-network moat compounds on calibration depth. By 2028, the question is whether DecagonCo-AEC moves upmarket (replacing or layering above Procore at enterprise) or stays in the mid-market wedge. Independent base case: it stays disciplined in mid-market and reaches $200–500M revenue at $1.5–3B valuation. Acquired bear case: Procore acquires for distribution and the mid-market wedge, around $500M–$1B.

Procore reprices for mid-market with a consumption-tier offering before DecagonCo-AEC is anchored. The Procore cannibalization bind is real but not absolute — if Procore decides the mid-market is worth the partial price-model fracture, the wedge closes. Inflection signal: any Procore consumption-pricing or self-serve mid-market tier launch by 2027.

Real-time multiplayer coordination canvas across all design-and-build disciplines — architects, structural engineers, MEP, civil, owner reps, GC preconstruction, key subcontractors. Browser-native. Comments and resolutions on geometry, schedules, RFIs, decisions. Every coordination decision captured. Per-seat pricing initially, evolving toward per-project-volume.

Bluebeam Revu (markup and review — the dominant tool for design-review meetings). Procore (coordination workflow). Autodesk Construction Cloud (federated-model coordination). All three are the partial defender; none of them is multiplayer in the Figma sense.

Foundational: Network Economies — the value of the platform compounds with the number of disciplines on it. AEC is harder than software design because the disciplines are siloed by firm (architect at HOK, structural at Thornton Tomasetti, MEP at Buro Happold, owner at the developer, GC at Suffolk). The classic two-sided network problem multiplied by 5+ sides. Foundational: Agentic Workflow Lock-in — every coordination decision captured becomes calibration for the agent layer that summarizes, escalates, and routes. Supporting: Counter-Positioning against Bluebeam and the static markup workflow.

The network-cold-start problem is brutal. Single-firm pilots show value but cross-firm adoption requires coordinated buy-in across architect, engineer, GC, owner. Procore and ACC have the integrated-workflow share; FigmaCo-AEC has to win on multiplayer UX in a way the incumbents cannot match without rebuilding their products. Slow but real H1 traction in tech-leaning AE firms (CRTKL, Stok, NBBJ) and tech-leaning GCs (Suffolk, DPR).

The cold-start problem is the dominant force. Cross-firm AEC coordination requires owner mandate; owners do not mandate tools. Single-firm pilots show value but the 5+-sided network does not compound to the de facto coordination layer without external orchestration. Most likely H2: Bluebeam (under Nemetschek) or Autodesk acquires for the multiplayer-UX patterns and folds them into the existing Bluebeam Revu / ACC product stack. The acquirer pays for the UX assets and the engineering team, not the network. Bull case: a single hyperscaler, healthcare system, or repeat-builder REIT mandates the platform across their project portfolio, the network compounds within that owner's ecosystem, and FigmaCo-AEC reaches independent escape velocity. The base case is acquired; the bull case is the exception.

Acquired-bear (closer to $400M) if the company runs out of runway before any owner mandate materializes, and Bluebeam/Autodesk acquire defensively for the patents and the team. Inflection signal that flips to bull-INDEPENDENT: any hyperscaler, healthcare system, or REIT mandating the platform across their 2026–2028 project pipeline.

Standardized API integration across BIM authoring tools (Revit, Archicad, Bentley, Vectorworks), IoT/sensor stacks (KMC, Distech, Schneider EcoStruxure, Honeywell Forge, Siemens Building X), construction-management systems (Procore, ACC, Bluebeam), and FM systems (Tririga, FM:Systems, Planon). Sells to AI-native AEC startups that need fast integration with the closed-platform ecosystem; also sells to enterprise GCs and owners running multi-vendor stacks.

Closed-platform vendors directly: Autodesk APS, Bentley iTwin, Schneider EcoStruxure, Honeywell Forge, Siemens Building X. Each defends its own perimeter; none of them has motive to ship a competitor to PlaidCo-AEC because the result would be the cross-platform integration they each individually want to prevent.

Foundational: Cornered Resource on the contractual relationships with closed-platform vendors. Plaid's horizontal moat is not the technology — it is the bilateral agreements with banks. PlaidCo-AEC's analog: signed agreements with BIM-tool vendors, BMS OEMs, building-system OEMs that allow it to expose data through standardized APIs. The moat is the rolodex and the deal terms, not the schema. Supporting: Switching Costs — once a customer's integration plumbing runs through PlaidCo-AEC, replacing it is a many-month engineering effort. Supporting: emergent Network Economies on the federated-network side — more vendors connected means more demand-side pull from customers, means more leverage to sign the next vendor.

The category is uncrowded. ODA (Open Design Alliance), Speckle, and IFC each cover slivers; none integrates the whole BIM + IoT + CM + FM surface. PlaidCo-AEC anchors agreements with 3–5 BIM vendors and 2–3 BMS OEMs in H1. The hard part is signing the closed-platform incumbents who do not want to be commoditized.

The structural risk is MCP. If Anthropic-led MCP standardizes the AI-agent-to-tool integration layer well enough that vendors expose MCP endpoints natively, the cornered-resource moat compresses because the integration is no longer a contractual rolodex — it is a protocol any agent runtime can speak. PlaidCo-AEC has to ship its commercial moat (data-cleansing, normalization, telemetry) before that protocol matures. Independent base case: it does, reaches $200M+ revenue at $2–5B valuation; bear case: MCP commoditizes the integration layer and PlaidCo-AEC becomes a feature inside someone else's platform.

MCP matures faster than PlaidCo-AEC's commercial moat ships. The horizontal precedent (Plaid) had FedNow as the analogous protocol risk and survived because the bank-relationship rolodex stayed valuable; PlaidCo-AEC needs the BIM/BMS vendor rolodex to stay similarly valuable under protocol commoditization. Inflection signal: native MCP endpoints from Autodesk APS, Schneider EcoStruxure, Honeywell Forge by 2027.