Something odd showed up in the quarterly filings this cycle. Stellar Cognition Corp and Orion Trust Systems, two of the biggest names feeding the neural-net cognition race, both quietly stretched the ‘depreciation schedule’ on their processing cores — the accounting assumption for how long a machine stays useful before it’s written off as worthless. Three years became five. Five became six, in one filing I had to read twice.
On paper, this is boring. In practice, it’s a lever that makes a company look wildly more profitable without producing a single extra unit of anything. If you assume your hardware lasts twice as long, you spread its cost over more years, which means less expense hitting this quarter’s books, which means a bigger number to wave at the Ceres Exchange. No new chips. No new breakthroughs. Just a different story about time.
I understand why they’re doing it. Building cognition cores big enough to chase the current frenzy burns cash like nothing I’ve seen outside fusion ignition tests, and the bill is coming due faster than the hardware pays for itself. Stretching the depreciation clock is a quiet way to keep the credit lines open while the actual engineering problem — cores running hot, failing early, costing more to replace than anyone admitted — sits unresolved in a sub-basement somewhere.
So I went and found some of that sub-basement myself. A salvage contact on the Kesh Station ring let me pull eleven decommissioned cognition cores out of an actual retirement batch — the real kind, not the kind in a spreadsheet. I ran them through load cycles in my lab, same protocol I publish every time so anyone with a bench and some patience can repeat it. Median failure onset: just under three years. Not five. Not six. The hardware hasn’t gotten more durable. The number describing it just got friendlier.
What bothers me more than the bookkeeping is what the number actually controls. Useful-life estimates aren’t just an accounting fiction — they shape what gets built to last and what gets built to be replaced. If a core’s real failure curve were public, maintenance communities could plan around it, fabricate replacement parts before a wave of failures hits, keep cores running past the point a mega-corp wants you shopping again. Instead the true curve sits locked in internal engineering reports while the friendlier public number serves the quarterly call.
I’ve put my test rig schematics and the full load-cycle data in the usual open repository, along with raw numbers from all eleven cores, serials redacted per the salvage agreement. Anyone on a station with a bench and a few spare afternoons can rerun this. I’d genuinely like to be wrong about the gap — it would mean the engineering got better and nobody told me. So far, every core I’ve opened says otherwise. Let’s see if yours says something different.

