The Stack Collapse
- Rich Washburn
- 3 hours ago
- 6 min read
Jeff Bezos is spending the bulk of his time on a company most people had never heard of six months ago.
Prometheus, founded in November 2025, just raised $12 billion at a $41 billion valuation. Backed by JPMorgan, Goldman Sachs, and BlackRock. Co-CEOs: Bezos and
Vik Bajaj, a Stanford medical school professor who previously co-founded Alphabet's Verily. About 150 employees. Recruiting talent out of OpenAI, Google DeepMind, and Nvidia.
The valuation alone is not the interesting part. The mission is the interesting part.
Bezos is not building another chatbot. He is not building humanoid robots — he said so explicitly, pushing back on early reporting that described Prometheus as a robotics company. "Nothing to do with robotics," he said. He described it instead as a modern evolution of computer-aided design, while acknowledging that description is an oversimplification.
Here is what Prometheus is actually building: what Bezos calls an "Artificial General Engineer."
Think about the progression. OpenAI built AI that understands language. GitHub Copilot built AI that helps write software. Prometheus is trying to build AI that helps invent and manufacture physical things. Imagine telling an AI: design a 1 MW natural gas generator that produces 15% less NOx, costs 20% less to manufacture, uses components available in North America, and can be serviced by one technician.
Instead of giving you a paragraph, it produces CAD models, stress simulations, manufacturing plans, bills of materials, supplier recommendations, production workflows, and testing procedures.
That is the direction. And Bezos framed the productivity claim in concrete terms: something that today requires 100 engineers working for 10 years, reduced to 10 engineers working for one year. If they can deliver even partially on that, the economic implications are staggering.
There is a second project that matters here, and most people are not connecting the two.
Elon Musk announced Macrohard earlier this year, inside xAI. The mission: simulate how a massive software company like Microsoft operates, but run by AI instead of people. A fully AI-run software company. Not a tool that helps developers write code. A system that operates the organizational functions of a software company — product management, coordination, code generation, testing, deployment, commercial strategy — with AI doing the work and a small number of humans directing the machine.
Bezos is targeting the engineering organization. The factory, the laboratory, the design shop. Musk is targeting the business organization. The software company, the administrative layer, the digital workforce.
These look like separate projects. They are not. They are two ends of the same convergence. A Macrohard-like organization could run a Prometheus-like engineering engine. Prometheus could design the products, systems, and machines. Macrohard could manage the business, software, procurement, deployment, and commercialization around them. At that point, the distinction between technology company, manufacturer, engineering firm, cloud provider, and industrial conglomerate begins to disappear.
To understand what is actually happening, you have to look at the technology stack itself. Not the companies. The structure.
The traditional technology economy looked like this: Energy and real estate at the bottom. Data centers and chips above that. Cloud infrastructure above that. Software platforms above that. Applications above that. Companies and workers above that. Products and services at the top.
Each layer supported the layer above it. Different companies occupied different positions, and enormous markets existed in the handoffs between them. The cloud provider sold compute to the software platform. The software platform sold tools to the application company. The application company sold products to consumers. Value moved up the stack, and each handoff was a business.
What the hyperscalers are building now looks different: Energy plus compute plus models plus agents plus engineering plus production. All operating as one vertically integrated intelligence system.
The future hyperscaler may not merely rent servers. It may generate or contract its own power, design its own chips, build its own data centers, train its own models, deploy its own digital workforce, invent products, simulate manufacturing, manage supply chains, operate physical factories, and sell the resulting goods and services directly. That is not a cloud company. It is closer to a private industrial civilization stack.
This is where the infrastructure story I have been writing about all week connects. For two decades, software companies captured the highest margins because software was scarce and hardware was commoditized. AI reverses part of that relationship. Software generation is becoming cheaper. The thing that remains scarce is physical: megawatts, transformers, turbines, switchgear, GPUs, substations, cooling capacity, fiber, entitled land, permitting, gas supply, construction labor, manufacturing throughput.
The intelligence may be digital, but it still has a body. Its body is the power plant, the data center, the semiconductor fab, the fiber network, and eventually the factory.
That is why power and infrastructure work is not adjacent to AI. It is becoming part of the core AI stack. The grid event in Virginia this week — 1,500 megawatts vanishing in milliseconds when data centers disconnected — is not a side effect of AI. It is a structural feature of a world where intelligence has a physical footprint that cannot be abstracted away.
The disruption hits the middle first.
The largest danger is not to electricians, machinists, or field technicians. Physical work remains constrained by reality — by the speed of construction, the weight of materials, the permitting process, the labor market. That stuff is sticky. The most immediate pressure hits the professional middle. Junior software developers, analysts, project coordinators, product managers, CAD operators, procurement teams, proposal writers, researchers, consultants, back-office administrators, middle management, outsourced professional services.
These roles exist to move information between systems, translate requirements, create documents, coordinate people, and maintain organizational continuity. That is precisely what multi-agent systems are being designed to absorb.
The first companies to become dramatically more productive will not necessarily eliminate all employees. They may simply operate with 20 people instead of 200. Or 200 instead of 2,000. Or 2,000 instead of 20,000. That alone remakes labor markets, corporate valuations, and competitive economics. Not through mass unemployment, but through the quiet compression of the organizational layer that sits between capital and output.
There is a deeper transition underneath all of this, and it is the one that matters most.
The critical shift is from AI as a product used by companies, to AI as the operating structure of the company, to AI as the company's productive core. Companies today are collections of employees using software. The software is a tool. The employees are the intelligence. The company is the organizational wrapper around both.
What happens when the software becomes the intelligence and the employees become the directors? You get a different kind of institution. One that remembers, coordinates, designs, tests, acts, learns, allocates resources, produces valuable output, and runs continuously at computational speed. An organization that does not sleep, does not forget, does not lose institutional knowledge when someone retires, and does not require quarterly offsites to maintain strategic alignment.
That is not a software company. It is a machine institution. Bezos is trying to build the invention engine. Musk is trying to build the organizational engine. The hyperscalers already have the compute, capital, and distribution engine. The missing ingredient is joining those pieces into a coherent institution capable of moving from scientific question to engineered design to manufactured deployment at machine speed.
This is where the Bell Labs connection becomes real. Bell Labs worked because it combined immense capital, long time horizons, world-class researchers, engineering capability, manufacturing access, infrastructure, and a clear institutional mission under one roof. The result was an institution that produced transistor-level breakthroughs — not because it was lucky, but because it was structured to convert fundamental research into engineered reality.
The new version replaces the fixed population of human researchers with a hybrid system of humans and scalable machine intelligence. Prometheus looks like an attempt to build the invention engine. Macrohard looks like an attempt to build the organizational engine. The hyperscalers already provide the compute and capital.
Whoever completes that loop — who builds the institution that can move from question to design to manufacturing at machine speed, with a small number of humans directing the whole system — will not merely own a better AI model. They may possess the most powerful productive institution ever created. That is the signal nobody is reading correctly.
This is not two rich guys launching AI companies. It is the first visible sign of a reorganization of the technology stack itself. The AI race is beginning to move beyond intelligence and into ownership of the entire chain that converts intelligence into reality.
The companies that recognize this earliest — the ones that start building the full stack rather than competing at a single layer — will be the ones that matter in ten years. The ones that do not will be the handoffs. And the handoffs are where the margin used to live.
Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital .
Comments