Chamath Palihapitiya just did something he has not done since he left Facebook in 2011: took an operating job. In late June, 8090 raised a $135M Series A led by Salesforce Ventures, and he stepped off the board to run it full-time. The company sells something it calls a "Software Factory."
The easy read is "another AI-writes-code startup, this one with a famous CEO." That read misses the interesting part. Look at what the factory actually sells, and there's a decision buried in it for every founder shipping software with AI — not just the regulated enterprises 8090 is chasing.
What a "software factory" actually is#
Strip the branding and 8090's product is a control plane over the software development lifecycle: it refines requirements, captures architecture decisions, turns intent into structured work orders for AI agents, and then validates every feature against its requirements — with an audit trail for each step. It orchestrates multiple models, but the models are the interchangeable part. The platform is the process wrapped around them.
Notice what is not the pitch. The pitch is not "our agents write better code." It is documentation, governance, and validation. 8090 runs an actual delivery business next to the platform and aims it at healthcare, aerospace, financial services, energy, and the US government — buyers who pay for accountability, not for keystrokes. Earlier this year it launched the EY.ai PDLC with Ernst & Young, a "collaborative mesh" of agents under human oversight. Same shape: humans specify and validate, agents fill the middle.
The hard part moved#
Here is the one idea worth taking away, even if you never touch an enterprise platform.
When an agent can emit a plausible diff for almost any request, code generation stops being the constraint. The scarce work moves upstream — can you specify intent precisely? — and downstream — can you prove the output meets it? Generation is the cheap middle.
That is why a $135M round goes to a company selling governance rather than generation. Generation is commoditizing in public: every model release undercuts the last on price-per-token and every coding tool can now produce a reasonable first draft. Value doesn't stay in the commoditizing layer. It migrates to the parts that are still hard, and right now those parts are at the two ends of the pipe — the spec and the check.
You can see the same bet elsewhere in the same week. Bespoke Labs raised $40M to build the environments that make agents reliable — not bigger models, better verification. Two rounds, one thesis: reliability is the product now.
Three shapes of shipping with AI#
For a founder this is a decision, and it is not about company size. There are three shapes of AI software delivery, and you choose by blast radius and regulatory exposure:
| Copilot | Coding agent | Software factory | |
|---|---|---|---|
| Unit of work | a line | a task / PR | a work order from a requirement |
| Your role | author | reviewer | specifier + validator |
| Audit trail | none | git history | requirement-to-code |
| Earns its overhead when | low stakes | scoped features | wrong changes are costly or must be defensible |
A copilot fits low-stakes work you'd review line-by-line anyway. A coding agent fits well-scoped features behind a human gate. The full factory pattern — specs, work orders, validation against requirements, audit trails — is pure overhead until the cost of a wrong change is high: money, an outage you can't roll back, or a regulator asking why the system did what it did. Then the overhead is the cheapest thing you'll buy.
Steal the discipline, skip the invoice#
You do not need Salesforce Ventures money to run this pattern. The factory's discipline is four moves you can do with the tools already open on your screen:
- Write the requirement as a checkable assertion. Not "add rate limiting" but "no client can exceed 100 requests/minute; the 101st returns 429." An assertion you can test is a spec; a vibe is not.
- Turn it into a scoped work order. One requirement, clear boundaries, the context the agent needs. Vague prompts are where factories — and solo builders — generate confident garbage.
- Validate the output against the assertion, not against your gut. Did it actually enforce 100/minute? Prove it with a test, not a glance at the diff. (This is its own skill — here's how to decide when an AI feature is reliable enough to ship.)
- Keep the trail. Which requirement produced which change. Future-you, and anyone you ever have to answer to, will need it.
That is most of what $135M buys, running on a laptop. The lesson of the software factory is not that you should become an enterprise; it's that the era where typing was the bottleneck is over, and the founders who win the next one are the ones who get good at the two things agents still can't do for you: saying exactly what they want, and checking that they got it.
