Invariant Machines is a VC-backed engineering lab. We pick hard problems
and solve them. Every product we ship should be easy, even beautiful, to
use, and a step-function improvement in what it can do or how it feels
to use.
Tough problems rarely fail for lack of ideas. They stall between the
prototype and production: in integration, certification, manufacturing,
and the first real customers. That stretch is the work we’re here
to do.
Once a problem has a workable product, we spin it out into its own
company, with its own name, team, and customers.
Companies
The last meter of the grid. AI racks are about to draw a megawatt
each, at 800 volts DC. At that voltage there is no zero crossing to quench
an arc, so faults have to be cleared in microseconds, by silicon. The parts
exist. The integrated, certified feeder that stands between a
datacenter’s DC bus and the rack does not.
Hephaestus I. SF Electric’s first product is that feeder:
isolation, precharge, and solid-state fault interruption in under 25
microseconds, in one repeatable module. Pilot units arrive July 2027.
The ultimate software foundry. GitHub with the infrastructure
built in: repositories and pull requests, plus CI, deploys, hosting, logs,
traces, and incidents, all run by one platform on one connected graph.
Git is the interface. Every push is a proposal. The verdict comes
back in the push output: tests, a replay of recorded production traffic
against the change, a score, and, when policy allows, the merge and the
deploy.
Large open models, run at the redline. The fastest way to run
large open-weight models, built for long-context agent workloads.
The more you use it, the faster it gets. Redline learns your
workload, generates CUDA kernels for it, and keeps tuning underneath, so
you get the best speed and price with no changes on your side.
What comes next
More problems are in the lab. We name each one when it spins out.
How we work
Start with the hard part. We go straight at whatever makes the
problem tough. That is usually where everyone else stopped, and where the
value is.
Find out early if we’re wrong. The first things we build
and test are the ones most likely to show an idea doesn’t work, and
every number we rely on gets checked against a real source.
Production is the finish line. A demo, a paper, or a prototype on
a bench doesn’t count. A problem leaves the lab once its product
works in production, and not before.