Jeff Bezos and Vik Bajaj have co-founded Prometheus, a new company building AI tools designed to help engineers design and manufacture physical products faster, according to CNBC. The two serve as co-CEOs. The announcement came on 11 June 2026.

The company's stated focus is the physical economy — the domain of atoms rather than bits. Prometheus's AI tooling targets the engineering workflows that produce hardware: mechanical design, manufacturing processes, and, explicitly, chip development. The chip callout is notable. Semiconductor design is among the most complex, capital-intensive, and time-constrained engineering disciplines that exist, with tape-out cycles measured in months and mask costs running into the tens of millions of dollars. Any meaningful compression of that cycle has asymmetric leverage on the broader technology stack.

Bajaj, who previously co-founded Verily and served as a managing partner at The Column Group, brings a track record bridging deep science and capital deployment. Bezos needs no introduction in terms of operational scale — Amazon's AWS, its logistics network, and Blue Origin's launch cadence all represent long-horizon bets on physical infrastructure at extraordinary complexity. Prometheus is their shared pivot toward AI-augmented engineering tooling as an independent venture.

The framing Bezos and Bajaj used in their CNBC interview with David Faber positions Prometheus as infrastructure for the physical economy rather than a single-product company — accelerating engineering broadly, with chips as an illustrative anchor case rather than the sole addressable market.

The broader context here is worth sitting with. AI's productivity gains to date have been concentrated in software — code generation, documentation, testing, retrieval-augmented workflows. The physical sciences and hardware engineering have been comparatively slower to absorb these gains, partly because ground-truth feedback loops are slower (you cannot unit-test a silicon wafer in milliseconds), and partly because the training data for physical design is proprietary, sparse, and hard to standardize. A company explicitly targeting that gap, with Bezos's capital access and Bajaj's scientific pedigree, is entering a space that has seen earnest attempts — Cadence and Synopsys have both pushed AI into EDA tooling, and startups like Anthropic-backed ventures have probed materials science — but no dominant platform has emerged.

Whether Prometheus is building AI co-pilots layered onto existing CAD and EDA ecosystems, or pursuing a more foundational approach that retrains the underlying design representation, is not yet clear from what has been disclosed. That distinction matters enormously for both the competitive landscape and the adoption path. EDA vendors are entrenched, with decades of IP and deep customer lock-in; displacing them requires either a dramatically superior workflow or a complementary integration strategy that doesn't trigger a defensive response.

In this author's view, the co-CEO structure itself is worth watching. Shared executive authority tends to work when the two leaders have genuinely complementary, non-overlapping domains — a model that held, for example, at early Salesforce. If Bezos is providing strategic weight and external credibility while Bajaj runs the science and product organization, the structure is coherent. If both are operating across the same decisions, the friction typically surfaces within 18 months.

The company is early. No product has been publicly demonstrated, no customer announced, no funding round disclosed beyond what the founders themselves are presumably providing. What exists publicly is a mission statement and two founders with the resources and credibility to attract engineering talent from the EDA, semiconductor, and AI research communities simultaneously. That recruiting surface is real, and in a field where progress is gated on the depth of the people building the models, it is not a trivial advantage.

Prometheus's pitch — that AI can do for hardware engineering what GitHub Copilot started doing for software — is intuitive, but the hard work is in the details of how physical constraints, manufacturing tolerances, and supply chain realities get encoded into models that can generalize. The problem is tractable. It is also genuinely hard.