As data centers grow, engineers at Meta, Google, and other companies worry about the reliability of ever-smaller chips and hardware failures with unclear causes
As the largest computer networks continue to grow, some engineers fear that their smallest components could prove to be an Achilles' heel. Tweets: @doctorkarl , @heatwavekgns , @nytimestech , @markoff , @trippmickle , @nytimestech , and @hpcdan Tweets: Dr Karl / @doctorkarl : “Early in the semiconductor era, engineers worried about the possibility of cosmic rays occasionally flipping a single transistor and changing the outcome of a computation. Now ... switches themselves are increasingly becoming less reliable.” https://www.nytimes.com/... Richard Heatwave Berler / @heatwavekgns : The tiny volume of high speed computer chips packed with billions of transistors are posing problems in detection and addressing: https://www.nytimes.com/... @nytimestech : As the switches in computer chips have shrunk to the width of a few atoms, the reliability of chips has become another worry for the people who run the biggest networks in the world. https://www.nytimes.com/... John Markoff / @markoff : The limits of hardware reliability are exposed at scale.... https://www.nytimes.com/... Tripp Mickle / @trippmickle : Remember the recent server outages at Facebook and Amazon? Tiny transistors are behind some of these big headaches in computing, says @markoff https://www.nytimes.com/... @nytimestech : Managers at giant data centers have a new worry to ponder: tiny flaws in computer chips that only become noticeable when used in very complex environments. https://www.nytimes.com/... Dan Reed / @hpcdan : An atom here, an atom there and pretty soon you're talking real errors https://www.nytimes.com/...
Context & Ripple Effects
This story lands at the end of a long arc the coverage has been tracking: engineers flagged the slowdown of Moore's Law back in 2015, and capital responded — semiconductor startups drew more than $12B from equity investors in 2020, up 8x since 2016, hunting for alternatives to straightforward scaling.
The new wrinkle is that the hyperscalers themselves — Meta, Google, and peers running the largest networks — are now reporting that their smallest components are becoming less reliable, with failures whose causes they cannot clearly attribute. It reframes the fundamental limits on making ever-smaller chips from an ASML-era manufacturing question into an operations question for the companies buying the chips.
First-order effects
- Engineering teams at Meta and Google must divert resources toward diagnosing unexplained hardware failures and building redundancy, raising the effective cost per unit of compute even before chip prices move.
Second-order effects
- Hyperscaler procurement gains a reliability criterion alongside cost and density, strengthening demand for proven designs and mature fabrication capacity — a market where the scarcity of secondhand older-node machines has already constrained supply during the chip shortage.
- Chipmakers face pressure to justify each shrink not just on performance but on error rates, shifting vendor competition toward validation and fault-attribution tooling.
Third-order effects
- If reliability degrades faster than density improves, the industry's center of gravity moves from transistor scaling to system-level fault tolerance — mirroring how the slowdown of Moore's Law pushed innovation toward architecture rather than raw miniaturization.
- A downturn like the inventory correction the chip market later entered would compound the problem: operators stuck with less-reliable dense parts have fewer economic reasons to refresh, extending the life of fragile hardware.
The trend: As transistor scaling approaches physical limits, data center reliability is shifting from a component property guaranteed by fabs to a systems problem owned by the hyperscalers who operate the fleets.