On September 23, 1949, Harry Truman walked to a podium and delivered a 37-word statement notifying the American people that the Soviet Union had detonated an atomic bomb on the Kazakh steppe. Omitted was the fact that the Soviet Union was also building a fleet of long-range bombers capable of dropping nuclear weapons on the United States across the Arctic.

The Pentagon lacked the ability to track these bombers in real time, a gap that created the possibility of a “nuclear Pearl Harbor,” a sudden Soviet atomic strike on US military bases and cities with little or no warning. The shock forced a scientific mobilization that permanently changed our view of what computers could and should be used for.

Before the Cold War, primitive computers were largely viewed as high-speed calculators. During World War II, this brute-force calculation had proved valuable, as machines like the British Colossus chewed through mountains of intercepted communications to break German military encryption. Yet these wartime breakthroughs operated strictly after the fact, relying on recorded data processed in isolated batches. The problem of untrackable Soviet bombers demanded a radically new approach: a system that could ingest continuous radar data, process it instantly, and visually represent this data on a screen for human decision makers. In solving this problem, scientists would convert computers from tools that solved equations into devices that managed reality as it happened.

In 1951, the Massachusetts Institute of Technology was awarded the primary contract to develop these systems, leading to the formation of Lincoln Laboratory, a research and development center focused on air defense. The focal point of the MIT effort was Project Whirlwind, whose original goal in 1944 was to develop a computerized naval flight simulator. Lincoln Laboratory made Whirlwind its computational core and the prototype for the Semi-Automatic Ground Environment, or SAGE, which became the Air Force’s massive computerized air defense system.

The mainframe systems built by the SAGE team became the ultimate hardware sandbox of the era. By 1958, they had constructed the TX-2, a powerful computer designed specifically to handle advanced graphics and continuous data streams. It was the TX-2 that would host Sketchpad, the world’s first interactive computer graphics program and the direct ancestor of all computer-aided design (CAD) programs. Without CAD, designing high-end industrial products such as microchips or fighter jets would be impossible. No team of draftsmen could draw a billion transistors on paper.

Yet despite its foundational role in modern manufacturing, the core mechanics of this software have not fundamentally changed since 1988. That was the year the aptly named Parametric Technology Corporation released the first commercial CAD program featuring parametric modeling. Parametric modeling lets an engineer define the relationships between the parts of a design, so that changing one element automatically updates everything that depended on it. CAD systems remain stuck at the level of capturing geometric intent rather than functional design intent. They can generate a geometric model, but cannot reason about whether an engineering design is correct, optimal, or even physically coherent.

This stagnation matters because CAD sets the clock speed of all physical innovation downstream of it. Design iteration for advanced manufacturing is only as fast as an engineer can encode shapes, run them through multiple separate analysis tools, interpret the results, and manually re-encode.

Some believe AI represents a potential path out of this stagnation. A “Claude for engineers” could enable the pace of iteration in the physical world to match the speed of software. Jeff Bezos has recently come out of retirement to develop precisely this technology. In November 2025, Bezos announced that he would become co-CEO of Prometheus, a San Francisco startup building what he describes as “a very, very modern version of CAD.” But Bezos is not alone in this ambition. The Chinese Communist Party has identified industrial design software as a critical “chokepoint” technology it must circumvent, and AI could enable it to leapfrog its Western rivals and transform its ability to iterate and optimize the designs for complex machines.

The future may belong to whoever can build the machines that can design machines.

The Dawn of CAD

The development of CAD can be viewed as part of a long contest between the great powers to control the skies. This contest accelerated as soon as World War II ended. The United States Strategic Bombing Survey published in 1945 determined “Allied air power was decisive in the war in Western Europe” before darkly concluding that “no nation can long survive the free exploitation of air weapons over its homeland.”

That same year, another report commissioned by the Pentagon, the multi-volume Toward New Horizons, was published. The lead author of the report was Theodore von Kármán, a Caltech aerodynamicist who was one of “The Martians,” a group of Hungarian-Jewish scientists born in Budapest around the turn of the century that shaped mid-century American science. Von Kármán was tasked by General Hap Arnold of the US Army Air Forces to peer decades ahead into the future and outline how science could “secure us the conquest of the air over the entire globe.” The report anticipated supersonic flight, intercontinental ballistic missiles (ICBMs), precision missiles, and wars “fought by airplanes with no men in them at all.”

By 1953, defense spending had reached 14 percent of GDP, and the Air Force was taking nearly half of it to ensure no foreign power could threaten the skies over the US and to deliver the future that Toward New Horizons had promised. Airframe contractors, electronics firms, and university laboratories set to work on jet interceptors, guided missiles, and a 3,000-mile radar network stretching from Alaska to Baffin Island. American airframe manufacturers would produce 49 new military aircraft designs in the 1950s.

This buildout required inventing new tools because the plane blueprints necessitated creating objects of unprecedented geometric and mechanical complexity, such as turbine blades, swept wings, and complex die molds. These objects were too difficult to be drawn by hand. The aerospace and automotive industries became the earliest adopters of these new industrial design tools because only they had both the geometry problems and the budgets.

The promise of industrial design software was first proven by Sketchpad, which was written in 1963 by Ivan Sutherland, an MIT graduate student. His thesis committee included Claude Shannon, Marvin Minsky, and Steve Coons, academics remembered today as the father of information theory, the father of artificial intelligence, and a foundational figure in computational geometry, respectively. Sutherland would later receive the Turing Award, the Nobel Prize of computer science, for this work.

That same year, General Motors brought into use DAC (Design Automated by Computer), the first CAD system to use interactive graphics for complex industrial molds. By using a wired light pen to sketch directly onto the screen of a room-sized mainframe computer, designers could instantly edit and duplicate complex geometry. An engineer behind DAC, Patrick Hanratty, left the company to launch Integrated Computer Systems (ICS) to commercialize his research.

But Hanratty was a pure engineer who was more comfortable writing software than managing a business; his first company failed. He went on to start Manufacturing and Consulting Services to develop ADAM (Automated Drafting and Machining), licensing the core code base to early industry leaders like Computervision and United Computing. By most industry accounts, the majority of commercial CAD systems descended directly from his code, but it was his licensees who captured the lion’s share of the economic value. While Hanratty sold non-exclusive source-code licenses for flat fees of around $100,000, his licensees packaged that code into proprietary software packages. They sold these to enterprise clients for tens of thousands of dollars per seat, building multi-billion-dollar businesses on the back of his engineering.

Between the 1970s and 1990s, major Western aerospace and auto firms poured millions into custom software, attempting to construct technical moats around their engineering. Yet, not everyone was convinced that this rapid adoption of engineering design software was a good thing. The Lockheed engineer Ben Rich, who would go on to lead the company’s legendary Skunk Works R&D group, wrote a memo in 1972 that called IBM — meaning computers in general — a scourge of the industry because it made engineers lazy and blunted expertise. He believed aerospace design was an art.

But to push the frontier of aerospace increasingly required more powerful design software. Over 18 days in the 1973 Yom Kippur War, Israel lost 109 aircraft — most of them American-designed — to Soviet-supplied radar-guided missiles and anti-aircraft guns. To avoid future losses, the US Air Force made developing jets that were invisible to radar its top priority. A 36-year-old electrical engineer at Skunk Works, Denys Overholser, uncovered the key to stealth technology in an obscure monograph by the Soviet physicist Pyotr Ufimtsev titled “Method of Edge Waves in the Physical Theory of Diffraction,” translated by the Air Force Foreign Technology Division in 1971. The book showed how to calculate the radar return, the echo an object bounces back to a receiver, from simple geometric shapes. Soviet censors had approved the work for open publication because they saw no military application in it. Overholser turned this mathematics into software, developing a program called Echo 1 that could predict the radar cross-section of an object before any physical model was built.

Using Echo 1 to analyze designs, Skunk Works developed a prototype plane called Have Blue, which resembled a flying black diamond. This plane would ultimately become the F-117 Nighthawk, the first operational stealth aircraft. The jet was shaped to scatter radar energy away from whoever sent it, so air defenses could not find it. At the same time, Lockheed’s French rival Dassault Aviation was designing the Mirage 2000 fighter jet. The aircraft’s tailless delta configuration was a single sweep of compound curvature, from its blended wing root to the shock cones in its intakes, geometry that could not be captured in two-dimensional drawings. Dassault had been a licensee of Lockheed’s CADAM, and to solve the Mirage’s design problems it built CATIA, its own three-dimensional surface modeler, on top of the American drafting system.

Dassault spun out a subsidiary, Dassault Systèmes, and signed a distribution agreement with IBM to sell the software globally. CATIA modeled curved surfaces better than anything else on the market, and IBM sold it to the aerospace and automotive giants that already bought everything else from IBM. Boeing chose CATIA as its primary 3D CAD tool in the mid-1980s and became its largest customer. By the late 1980s, CATIA had penetrated the automotive industry as well, with Mercedes-Benz and BMW among its major accounts. By the 1990s, per Dassault’s own accounting, its software was used to develop seven out of every 10 new airplanes and four out of every 10 new cars produced worldwide. The Boeing 777, which made its maiden flight in 1994 as the first commercial aircraft designed entirely in 3D CAD without a full-scale physical mockup, was a CATIA project.

The last major advance in CAD came from Samuel Geisberg, a Jewish mathematician from the Soviet Union who emigrated to the United States in 1974 and joined Computervision, then a leading CAD developer. In 1985, he founded Parametric Technology Corporation to develop Pro/ENGINEER, the world’s first commercially successful parametric modeling software.

Before parametric modeling, if an engineer increased the diameter of a jet engine fan, the surrounding nacelle, mounting structure, and airflow geometry all had to be updated manually. With Pro/ENGINEER, those interconnected features updated automatically, dramatically reducing design time and errors. In 2010, the company rebranded as PTC.

Today, the high-end tier of industrial design software is a triopoly: Dassault’s CATIA, Siemens NX, and PTC’s Creo, with CATIA and NX alone holding up to 75% of the market by some estimates. NX itself descends from Unigraphics, the product of United Computing, one of Hanratty’s licensees.

The triopoly’s position rests less on technological superiority than on vendor lock-in. Their moat is the fact that their customers’ accumulated assets (models, workflows, and government regulatory certifications) are specific to the vendor’s proprietary format. By holding their clients’ assets hostage, they can extract monopoly rents. The interests of the incumbents are to maintain their high profit margin profiles while making incremental improvements to their software. If there is going to be truly transformative engineering design software innovation, it likely must come from outside the industry.

The Outsiders

Any outsider attempting transformative improvements to CAD must solve three problems.

The first is the knowledge-in-the-model problem. A CAD file reduces an engineering design to instructions for rebuilding a shape. The reasoning behind a specific design decision, such as why a fillet radius is three millimeters instead of two, lives exclusively in the engineer’s head.

The second is the parametric trap. Because a parametric system builds geometry through a hierarchy of parent-child relationships, altering a foundational assumption can cause thousands of dependent features to cascade into failure. This makes large assemblies brittle. Companies routinely inherit legacy models they are afraid to touch because nobody understands the underlying dependency tree.

The final issue is the simulation gap. CAD defines what shape something is, but predicting whether that shape will survive its operating environment requires exporting the geometry to separate simulation tools for stress, fluid dynamics, or thermal analysis. This creates a discontinuous loop between designers and simulation specialists. Because the two systems do not share a unified model, engineers routinely build parts that pass software simulations but fail in the field under actual boundary conditions.

Perhaps the biggest obstacle to building a Claude for engineers is training data. Coding models learned from GitHub and StackOverflow, which together hold billions of lines of public code and decades of annotated answers. Hardware has no equivalent. Design files are the product itself, so companies guard them as trade secrets, store them in proprietary formats, and in aerospace, cannot legally export them. Prometheus’ immediate answer is to obtain models from alternative sources, such as expired patents, and create the training data directly. The company claims to have assembled the largest body of engineering design data in history and is hiring engineers near New Delhi to produce new 3D models by hand.

Bezos has described Prometheus’ goal as creating an “artificial general engineer.” The phrase means more than a better CAD system. CAD records engineering decisions; what Bezos wants to build is software that makes them. He believes wealth generation is driven by invention, and the most transformative inventions are general purpose technologies like electricity, the computer, and the Internet, tools that raise productivity across every sector of the economy rather than improving any single one. An artificial general engineer, therefore, would be a general purpose technology for accelerating technical invention itself. “The cycle from dream, to manufacturing at rate, to having it out in the world can be very long,” Bezos told Axios. “What we’re doing is building a set of tools that will empower engineers to compress that cycle time and make that dream-build loop be ten times faster or even more.”

One of the biggest potential beneficiaries is Blue Origin. Bezos has said his space company could eventually surpass Amazon in size and importance, but it trails SpaceX badly in its launch cadence and revenue. SpaceX’s biggest edge is the speed at which the company moves through cycles of design, test, and revision. A tool that compresses that cycle tenfold would give Blue Origin a way to change the terms of its rivalry with SpaceX.

For now, SpaceX’s Falcon 9 rocket, which carries the majority of all mass delivered to orbit globally each year, is the best example of the benefits of rapid design iteration. Falcon 9 is one name for what was effectively four rockets. Version 1.0, flying in 2010, lifted 10.4 tonnes to low Earth orbit at a list price of roughly $54 million, about $5,200 per kilogram. Version 1.1 stretched the tanks and rearranged the engines in 2013, reaching 13.1 tonnes at $61 million, or $4,700 per kilogram. Full Thrust arrived in 2015 with subcooled propellants and 22.8 tonnes of capacity at $62 million, cutting the figure to $2,700 per kilogram. Block 5 locked the design in 2018 for rapid reuse. In eight years, payload more than doubled while the price of a launch rose barely 15 percent. Cost per kilogram fell nearly in half before rocket reuse was even factored in. SpaceX bought that speed with a culture that tolerated crashed boosters and exploded prototypes, a culture no competitor would accept. Prometheus proposes to buy the same speed in simulation.

The costs of stagnant design tools are easy to understate because the losses are invisible: they are the things that were never built. The gap between what physics permits and what companies are willing to build has widened as the cost of attempting novel designs has grown. This is visible across aerospace. No novel airliner configuration has entered service since the swept-wing jet era of the 1950s because certification of a truly new design would likely take over 20 years and tens of billions of dollars. The Boeing 777X, a derivative of an already-certified airframe, has already cost roughly $10 billion, largely because of certification-driven delays since 2019. The plane is not expected in service until next year — 14 years after launch. State-of-the-art CAD models have only incrementally refined the standard tube-and-wing design that has carried virtually every passenger on Earth over the last 70 years.

Concepts offering up to 30% better fuel burn, such as the blended wing body, in which the fuselage and wings merge into a single lifting surface, or the transonic truss-braced wing, in which long, thin wings are held up by external struts, have remained in wind-tunnel testing or demonstrator phases for decades precisely because manufacturers cannot confidently predict full-envelope behavior without building and flying the aircraft. A design environment that resolves these questions reliably in simulation would remove the primary source of that delay and risk. The FAA would still test, but physical trials would shift from an unpredictable process of discovery to a predictable exercise in verification. Prometheus would also have to prove its AI models aren’t black boxes. Certification requires systems whose physics can be traced and verified, and software that cannot demonstrate this will not satisfy the FAA.

If Prometheus can truly unify design and physical simulation in one system, it would fundamentally alter the capital requirements of advanced manufacturing, rendering the financial risk of building novel designs manageable. But significant technical and institutional hurdles remain. Reliably embedding multi-physics simulation and safety-critical reasoning into generative models, validating those models across the full operating envelope, and securing regulatory acceptance of AI-assisted or AI-generated designs will require extensive real-world testing and new certification frameworks that do not yet exist. The nation that builds such a system first would develop weapons, aircraft, and energy systems on cycles its rivals could not match, and because invention compounds, the lead would grow.

China has already identified industrial design software as one of the technological choke points it must overcome. Domestic vendors still hold only a fraction of the high-end market inside China, and ZWSOFT, the largest of them, trails far behind the West. If China can obtain a more comprehensive set of engineering design training data than its Western rivals, it could use AI to leapfrog the incumbents. Beijing has a data-collection tool no Western company can legally match: civil-military fusion, under which the government can require manufacturers to hand over design data. It has also spent decades collecting Western intellectual property through industrial espionage. Even so, the quality and consistency of that data for training reliable engineering models are far from clear.

Prometheus’ ultimate answer to its data problem doubles as its business model. Bezos and Prometheus’ co-CEO, Vik Bajaj, are seeking up to $100 billion for a buyout fund described in investor documents as a “manufacturing transformation vehicle.” The fund would acquire manufacturers in chipmaking, defense, and aerospace, train Prometheus’ models on their data, and station forward-deployed engineers inside them. Prometheus’ tools would cut design cycles and widen margins inside the portfolio. The portfolio, in turn, would feed the models. A person familiar with the plan compared the result to a Berkshire Hathaway-style holding company.

A holding company, no matter its size, transforms only what it owns. The Chinese state does not have that constraint. It could push artificial general engineering technology down to its 10,000 “Little Giants,” midsized industrial suppliers fostered with state subsidies. These companies have been groomed to displace Germany’s famed Mittelstand firms that dominate the narrow technical niches, such as precision components and materials, that all advanced manufacturing depends on.

CAD was born from a moment when the United States looked at the sky and realized it could not see what was coming. Today the gap is different. We can see the designs we want, but it is too costly and risky to attempt them. The artificial general engineer is the machine that closes this gap, and someone will build it. Whoever does will control the pace of everything downstream of design, which is to say everything physical.