Did Biden just loosen the AI chip shortage?

Welcome to Runtime! Today: The Commerce Department places new restrictions on shipments of AI chips to China, a critical vulnerability in Cisco router software is being actively exploited, and the latest funding rounds in enterprise tech.

(Was this email forwarded to you? Sign up here to get Runtime each week.)

Smaller, slower, cheaper

At some point in the future, they'll call it "Raimondo's Law," referring to the observation that the restrictions placed on AI chip shipments by the U.S. government expand at regular intervals.

New export controls handed down Tuesday will prevent Nvidia, Intel, and AMD from shipping their B grade data center AI chips to China, almost exactly a year after the Biden administration imposed controls on the sales of their best AI chips to entities inside China. The U.S. wants to deny China "access to advanced semiconductors that could fuel breakthroughs in artificial intelligence and sophisticated computers," Commerce Secretary Gina Raimondo told reporters Tuesday, according to the Wall Street Journal.

CNBC reported that "the U.S. will simply restrict the export of data center chips if they exceed a performance threshold set last October, or exceed a new performance density threshold benchmark measured in flops per square millimeter."
Nvidia designed new, limited versions of its H-series and A-series AI chips to comply with last year's rules, but the new "performance density" requirement closes that loophole.
Intel and AMD also came up with new designs to get around last year's requirements that will have to be re-examined in light of the new rules.

Nvidia said that it doesn't expect the new rules to have a material impact on its earnings in the short term, given that it can't make AI chips fast enough to satisfy demand in the middle of the generative AI boom.

The 800-series chips earmarked for China could actually be an interesting option for companies that can't afford or even find the 100-series chips for their generative AI training workloads.
As more businesses warm to the idea that they can get useful results from smaller models that require less training time, they might be willing to pay a little less for chips that are a little slower than the most-performant options.
The 800-series chips can also be linked together for better interconnect performance by using the "chiplet" manufacturing strategy, which is exactly what the new rules were designed to prevent Chinese companies from doing, Reuters reported.

Chip designers are going to have to think very carefully about their AI roadmaps over the next several years as restrictions evolve along with the technology..

"Overly broad, unilateral controls risk harming the U.S. semiconductor ecosystem without advancing national security as they encourage overseas customers to look elsewhere," the Semiconductor Industry Association said in a statement.
Nvidia's recent announcement that it was transitioning from a two-year development cadence to a single-year cadence has to have been informed by debate over the new restrictions.
That will force Nvidia to execute faster, but as new generations of topline AI chips arrive enterprises could find it much easier and cheaper to obtain earlier-generation chips for model training.

Network effects

Cisco is urging customers to apply mitigations to protect themselves against a newly discovered vulnerability in its IOS XE router operating system that is under active attack.

The vulnerability, which does not yet have a patch, was disclosed by Cisco on Monday and given the highest severity score possible under the Common Vulnerability Scoring System (CVSS). It appears that it affects any Cisco product running IOS XE with a web user interface that is exposed to the internet, and customers should disable the web UI until a patch is available, according to a Cisco advisory.

Left unaddressed, the vulnerability could allow a remote attacker to gain full control over those devices and cause incredible mayhem to internal systems. It's not clear how easy that exploit process is, but according to Dark Reading more than 10,000 devices around the world have already been compromised in the last 36 hours.

Enterprise funding

Procurify raised $50 million in Series C funding to help companies buy the things they need to run their businesses more easily and with better oversight.

ScyllaDB landed $43 million in Series C funding as it looks to open up a new competitive front against MongoDB with its NoSQL database.

Cleanlab raised $25 million in Series A funding to help companies clean up and structure their data for AI workloads.

Pantomath scored $14 million in Series A funding for its data observability technology.

The Runtime roundup

Microsoft is about to sign a huge Microsoft 365 deal with Amazon, according to Business Insider, and that's kind of funny.

The U.K.'s Competition and Markets Authority defined the main questions behind its investigation into cloud infrastructure computing competition, which has the potential to force big changes at AWS and Microsoft.

The University of Washington is having a lot of trouble integrating a new Workday system, forcing it to delay payments to suppliers for months, according to the Seattle Times.

Widespread and growing use of GitHub's Copilot is hitting Stack Overflow hard, and the company behind a popular repository of software development information laid off 28% of its staff this week.

Microsoft thinks it might be able to use its Project Silica technology — which can store 7TBs of data for up to 10,000 years on a pane of glass — as an enterprise cloud storage device.

Thanks for reading — see you Thursday!