The Last Interface

In the first week of February 2026, approximately $300 billion in market value evaporated from the software sector in 8 trading sessions.^[1] Bloomberg called it the SaaSpocalypse.^[2] The S&P 500 Software & Services Index fell over 4% in a single day. Hedge funds made $24 billion shorting SaaS stocks.^[3] Autodesk, the $52 billion company that dominates the tools used to design and build the physical world, dropped 35% from its 52-week high of $329 to a low of $215 in a matter of weeks.^[4]

The trigger was not a recession or an earnings miss. It was a thesis. Satya Nadella had stated it on the BG2 podcast in December: business applications are fundamentally CRUD databases with business logic, and all that business logic is moving to AI agents.^[5] Ali Ghodsi, CEO of Databricks, sharpened the point: "Millions of people around the world got trained on those user interfaces. And so that was the biggest moat that those businesses have." When the interface becomes natural language, he warned, the moat becomes a puddle.^[6]

But the conversation about the SaaS apocalypse has been asking the wrong question. The debate has focused on whether AI agents will replace SaaS tools — whether a smarter Salesforce or a better Revit will emerge. This is the equivalent of debating whether the next horse will be faster. It misses the car entirely.

The real question is not who builds better construction software. It is who stops selling construction software altogether and starts selling buildings.

What Autodesk Actually Is

Autodesk reported $7.2 billion in revenue for fiscal year 2026, up 18% year-over-year.^[7] Nearly 8 million subscriptions generate 97% recurring revenue at 38% non-GAAP operating margins.^[8] The company holds roughly 65% of the global CAD market and 39% of architectural BIM.^[9] Its file formats — DWG and RVT — are de facto industry standards. Its educational licensing programme has trained a generation of architects and engineers who know no other tools.

Strip away the branding, and Autodesk sells exactly 3 things.

First, user interfaces for geometric authoring. AutoCAD lets a human draw lines on a plane. Revit lets a human place walls and ducts in a 3-dimensional model. Inventor lets a human sculpt parametric solids. In every case: a human sits in front of a screen, manipulates geometry through a graphical interface, and produces a digital representation of something that will eventually be built. AutoCAD costs $1,865 per year. Revit costs $3,005. The AEC Collection — the bundle most firms purchase — costs $3,675 per seat per year.^[10]

Second, collaboration infrastructure. Autodesk Construction Cloud exists because buildings are designed by multiple disciplines working in parallel — architects, structural engineers, mechanical and electrical engineers, plumbers — who inevitably produce conflicting geometry. A duct runs through a beam. A pipe collides with a wall. Construction Cloud provides clash detection, RFI management, document control, and field coordination. It is a communication layer for humans who cannot see each other's work.^[11]

Third, visualisation and documentation. The endpoint of every Autodesk workflow is a drawing — a 2-dimensional representation annotated with dimensions and specifications, formatted according to professional standards, and issued to a contractor who will use it to build something. Drawings are the legal contract between designer and builder, and the most labour-intensive output of the design process.

User interface. Collaboration. Visualisation. 3 pillars. $52 billion in market capitalisation.

Now consider: in a world where agents design, agents coordinate, and agents produce fabrication instructions — what are these 3 pillars worth?

The Faster Horse

The immediate instinct, shared by Autodesk and most AI-for-construction startups, is to make the existing workflow faster. Autodesk has announced Neural CAD — foundation models that generate editable parametric geometry from text prompts, claiming they could automate 80-90% of routine design tasks.^[12] They have invested $200 million in World Labs, the spatial AI company founded by Stanford's Fei-Fei Li.^[13] They have built MCP servers so external AI agents can interact with Revit and Fusion through standardised protocols.^[14] The Autodesk Assistant, their agentic AI, is being embedded across AutoCAD, Revit, Fusion, Civil 3D, and Inventor.^[15]

Outside Autodesk, the ecosystem is accelerating. Zoo, formerly KittyCAD, has built a text-to-CAD API that outputs parametric solid models with editable source code, reducing error rates from 50% to 16% in a single year.^[16] Leo AI transforms text and sketches into manufacturing-optimised assemblies, claiming 70% reduction in engineering time.^[17] Genusys AI automates MEP design within BIM models, reducing design time from days to minutes.^[18] Blueprints AI claims to be the first autonomous drafting system, trained on millions of architectural data points.^[19]

This is all impressive. It is also all the same category of innovation: making the existing process faster. A human (or an agent acting like a human) still designs a building in software, still produces drawings, still hands them to a contractor, who still builds the physical object through the same fragmented supply chain of subcontractors, suppliers, and tradespeople that has characterised construction for a century.

This is the faster horse. It accelerates the journey without changing the destination.

And it leads to a predictable outcome: a race to the bottom. If agentic software makes design 10x faster, the price of design falls. If it makes coordination automatic, the price of coordination falls. If it generates drawings without human drafters, the price of documentation falls. The value of the software compresses precisely because the software works. Autodesk's own estimate — that AI will automate 80-90% of routine design tasks — is simultaneously a boast about capability and a confession about its own obsolescence. If a mid-sized architecture firm with 50 Revit seats can produce the same output with 10 humans reviewing agent-generated work, Autodesk's revenue from that customer falls 80%.

Bessemer Venture Partners frames this with uncomfortable precision: traditional SaaS captures 1-5% of an employee's value through efficiency gains, while AI captures 25-50% by automating substantial portions of the role itself.^[20] OpenAI is pricing its planned enterprise agents at $2,000-$20,000 per month — benchmarked against the salary of the worker they replace, not the software they substitute.^[21] Foundation Capital calls this the $4.6 trillion "services as software" opportunity — the shift from selling tools to selling labour.^[22]

But even selling labour is still selling the horse. A faster horse. A cheaper horse. A horse that works 24 hours a day. Still a horse.

The Car

The construction industry is a $17 trillion annual market.^[23] Construction software — the entire market that Autodesk, Procore, Trimble, and every startup compete over — is approximately $12 billion.^[24] That is less than 1/10th of 1% of the industry it serves. The ratio between the tool market and the output market is 1,300 to 1.

McKinsey ranks construction as the 2nd-to-last industry on their digitisation index in the US, and dead last in Europe.^[25] Construction labour productivity has grown just 1% per year over 2 decades, compared to 3.6% for manufacturing.^[26] Over 500,000 positions sit vacant while 40% of experienced workers approach retirement.^[27] The industry does not have enough people to build what needs building, and the deficit is widening.

These are the conditions that precede not incremental improvement but categorical transformation. Not a faster horse. A car.

The car, in this analogy, is a company that does not sell construction software. It sells buildings. Or more precisely, it sells the outcome — the designed, coordinated, fabricated, and assembled physical structure — as a product, capturing value across the entire chain from site constraints to occupancy rather than extracting rent from a single link.

This is not a theoretical construct. It has precedent.

Tesla does not sell automotive software. It sells cars. But it does so by vertically integrating the entire value chain — battery production, motor design, software, charging infrastructure, direct sales, insurance, and now AI chip fabrication. Tesla's market capitalisation exceeds $900 billion — greater than the combined market caps of the next 20 largest automakers, despite selling a fraction of their vehicles.^[28] The premium reflects the market's belief that Tesla is not an automaker but a technology company that owns every step from lithium to autonomous mile.

Amazon did not become a $638 billion-a-year company by selling better bookstore software.^[29] It built warehouses, then logistics fleets, then Kiva robotics, then Prime delivery, then Amazon Basics manufacturing — systematically capturing every step of the retail value chain. Today, retail is only 40% of Amazon's revenue.^[30] The rest is services built from capabilities that were originally internal tools.

Anduril does not sell defence analytics. Unlike traditional contractors who bill cost-plus for development, Anduril invests its own capital in R&D and sells finished weapons as products.^[31] Their Arsenal-1, a $1 billion software-integrated weapons factory in Ohio, can turn digital twin designs into physical platforms within weeks.^[32] This is the defence equivalent of selling buildings rather than selling construction software.

Stripe did not remain a payment processor. It expanded from payments to company incorporation (Stripe Atlas now creates 25% of all Delaware corporations^[33]), lending (Capital), banking (Treasury), and tax compliance — systematically capturing the financial lifecycle of a business. Its valuation reached $107 billion^[34] by owning the chain, not a link.

In construction itself, FINFROCK has operated for decades under the tagline "Delivering Buildings as a Product" — integrating design, manufacturing of structural components, sub-trade coordination, and construction under one entity.^[35] Persimmon, the UK's largest housebuilder, vertically integrated bricks, tiles, and timber frame manufacturing, achieving industry-leading margins of 27% and cutting site building time by nearly 2/3rds.^[36] Mighty Buildings, a Y Combinator company, 3D-prints prefabricated building panels with 80% automation, producing panels for 2 homes per day with 1% construction waste versus the industry standard of 30%.^[37]

And then there was Katerra. It raised over $2 billion from SoftBank to revolutionise construction through vertical integration and technology. It made 20 acquisitions, hired thousands, built factories. In June 2021, it filed for bankruptcy.^[38] The thesis was sound — the execution was not. Katerra tried to do everything at once, acquired companies it could not integrate, and never proved unit economics at any single step. As industry analysts noted: "the California factory set up to produce its core product was never actually activated."^[39] The lesson is not that vertical integration in construction fails. FINFROCK and Persimmon prove it works. The lesson is that it must be built incrementally, proving each link before adding the next.

The $50 Billion to $500 Billion Thesis

Here is the argument I find most compelling about the future of construction technology.

The conventional framing of the SaaS apocalypse is zero-sum: AI agents will cannibalise seat-based revenue. Autodesk's $52 billion in market capitalisation will compress as agentic software replaces human users. The winners will be whoever builds the best AI design tools, and they will capture roughly the same market — a slightly larger construction software TAM growing from $12 billion to perhaps $25 billion by the mid-2030s.^[40]

This framing is wrong because it assumes the market stays the same size. It assumes the car is just a faster horse.

The alternative framing: construction technology goes from a $12 billion market to a multi-hundred-billion-dollar market by changing what it sells. The company that can take a set of requirements — a site, a program, a budget, a timeline — and deliver a finished building, using agentic design, automated coordination, robotic fabrication, and integrated supply chain management, is not competing with Autodesk. It is competing with Skanska, Bechtel, and Bouygues. It is not selling a $3,000-per-year software seat. It is selling a $300 million hospital.

The math is striking. If construction software captures 0.08% of a $17 trillion industry, a company that captures 1% of the value chain addresses a market of $170 billion — 14x the entire construction software market.^[41] At 5%, the addressable market is $850 billion. The construction software industry is fighting over $12 billion. The construction industry itself is $17 trillion.

Companies like Buildworld are positioning for exactly this thesis — not building better Autodesk, but building the agentic layer that automates industry workflows end-to-end, creating their own tools rather than licensing incumbents'. When a company can orchestrate the entire design-to-fabrication pipeline using AI agents, open APIs, and open formats, the question is not whether it can replace Autodesk's software. The question is whether it can replace the entire fragmented process that Autodesk's software serves. And if it can, the revenue is not $3,675 per seat. It is a percentage of the building.

Bret Taylor, co-founder of Sierra and former CEO of Salesforce, articulates the business model implication: "If you're selling software that completes a job, what is the secular business model for that? It felt like, let's pay for a job well done."^[42] His company, Sierra, reached $100 million in annual recurring revenue in under 2 years by charging for resolved customer service interactions — outcomes, not seats.^[43] But even outcomes-based pricing for software is the intermediate step. The endpoint is selling the output of the industry itself.

Taylor also identifies why incumbents struggle with this transition: "Closing a technology gap in your product is hard, but not impossible. Changing your business model is really hard."^[44] Autodesk cannot easily pivot from selling seat licences to selling buildings. Its entire corporate structure — sales, marketing, customer success, financial reporting — is built around recurring subscription revenue from named users. The company that captures the construction value chain will not be the one with the best CAD engine. It will be the one with the least legacy to protect.

What Autodesk Cannot Do

Autodesk knows the threat is real. To their credit, they are investing with genuine technical ambition. Neural CAD generates actual parametric geometry with editable feature trees — a meaningful achievement over generic AI that produces approximations.^[12:1] The $200 million World Labs investment signals they take spatial AI seriously.^[13:1] Their MCP servers allow external agents to interact with Autodesk products, which is an unusually open move for a company built on proprietary formats.^[14:1]

But there is a structural constraint that no amount of investment can resolve. Autodesk earns $3,675 per seat per year from nearly 8 million subscribers.^[10:1] Every initiative that reduces the number of humans who need those seats reduces revenue. Every tool that makes an agent self-sufficient outside the Autodesk ecosystem reduces lock-in. The innovator's dilemma is not abstract for Autodesk. It is a line item on every quarterly earnings call.

The open standards movement accelerates the erosion. The buildingSMART consortium's Global openBIM Mandates report shows expanding government requirements for IFC — the ISO-standard open BIM format — across Germany, France, the Netherlands, and the Nordic countries.^[45] Bonsai, the open-source Blender add-on, works natively on IFC files as its primary format.^[46] FreeCAD is integrating native IFC support.^[47] Every government mandate that requires open formats weakens the RVT lock-in by one more jurisdiction.

An agent does not care what format a file is in. An agent can read IFC, output STEP files for manufacturing, work natively in Zoo's KCL text-based geometry language. The proprietary format — Autodesk's most durable moat — is not a barrier to an agent. It is not even a speed bump.

And the educational moat — students trained on Revit's ribbon interface who demand it at employers — dissolves when the next generation learns to prompt, not to click.

Building 3.0

The construction industry is approaching what might be called its 3rd major technological transition. Building 1.0 was hand-drafted — T-squares, parallel rules, ink on vellum. Building 2.0 was computer-aided — AutoCAD replaced the drafting board, Revit replaced the flat drawing with an intelligent model, BIM 360 replaced the filing cabinet with cloud collaboration. Each transition preserved the fundamental workflow: a human designer creates a representation, communicates it to a human builder, who interprets it and constructs the physical object.

Building 3.0 does not preserve this workflow. It replaces it.

The designer is an agent. The representation is not a drawing but a data structure. The communication is not a document but an API call. The builder may be a robot, a CNC machine, or a 3D printer operating from the same data structure the agent created. The human's role shifts from creator to curator — defining requirements, reviewing proposals, approving solutions, and intervening when judgment exceeds computation.

But here is the critical distinction. Building 3.0 is not about making software that does design faster. It is about companies that own the entire chain — from the agent that designs the structure to the factory that fabricates the components to the site that assembles them. Companies like Buildworld, Mighty Buildings, and FINFROCK are not competing with Autodesk for software revenue. They are competing with general contractors for project revenue. They are not selling the tool. They are selling the building.

This is why the construction software market could expand from $50 billion in total enterprise value to $500 billion or more. Not because software gets more expensive, but because software companies become construction companies — or rather, because the distinction between a software company and a construction company ceases to be meaningful when the software is the construction.

The precedent is clear. Tesla is not a car software company. It is worth $900 billion because it sells cars. Amazon is not a retail software company. It generates $638 billion in annual revenue because it delivers packages. The construction technology company that captures even a fraction of the $17 trillion construction value chain will make Autodesk's $52 billion market cap look like a rounding error.

Autodesk's revenue grew 18% last year. Free cash flow increased 54%.^[7:1] They are guiding for $8.1-$8.2 billion in fiscal 2027.^[48] By every financial metric, the company has never been stronger.

The strongest quarter in a company's history is often the one just before the disruption becomes visible in the numbers.