The race to crash Nvidia's party

Today: why AI infrastructure could evolve in a lot of different directions, Microsoft removes the extra fee for some security services, and the latest enterprise moves.

A data center worker examines a server cabinet designed by AI chip startup Cerebras.
Cerebras will operate a massive AI supercomputer for G42 using its own chip designs. Credit: Cerebras

Welcome to Runtime! Today: why AI infrastructure could evolve in a lot of different directions, Microsoft removes the extra fee for some security services, and the latest enterprise moves.

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

So many instances

With the rise of Nvidia-centric AI cloud startups like CoreWeave and Lambda Labs alongside ambitious in-house chip design projects from Tesla and Cerebras and the hybrid approach taken by cloud providers, the generative AI boom is leading to a wide-open competition to win AI workloads.

This week Tesla said it had begun putting its Dojo supercomputer into production to train the models behind its autonomous vehicle software (which could use some training). On Thursday, Cerebras introduced its Condor Galaxy 1 AI supercomputer, which it will operate on behalf of G42, a cloud provider based in Abu Dhabi.

  • Both Tesla and Cerebras designed their own chips for their AI supercomputers, rather than relying on Nvidia's AI chips.
  • “It’s a challenge to Nvidia. It’s a challenge to hyperscalers," Cerebras CEO Andrew Feldman told EETimes. "It’s saying, ‘Look, you don’t need to be one of these [big] guys to stand up extraordinary compute.’”
  • Neither effort is cheap: Tesla said it expects to spend $1 billion getting Dojo up and ready, and Cerebras' deal with G42 could be worth nearly as much assuming plans to build future supercomputers for the company come to pass.

It's very early, but there are signs AI cloud services and on-premises efforts like Tesla's could evolve around several different types of AI chips, unlike the rise of general-purpose cloud computing.

  • Tesla's Full Self Driving Chip was in development for over a year before its coming-out party in 2019.
  • But this week Elon Musk suggested that Tesla would pursue a hybrid approach of its own, mixing its own chips with Nvidia's.
  • Cerebras has a unique approach to chip design: It prints its core designs directly onto silicon wafers "the size of a pizza," rather than cutting individual chips from those wafers like most chip designers and manufacturers.
  • And, of course, the cloud companies have their own AI chip designs in the works while enjoying the purchasing power to rank among Nvidia's best customers.

Nvidia's chips control an eye-popping 90% of the market for AI workloads, and that's a lot of eggs in one company's basket.

  • That's similar to the control Intel enjoyed over the data-center chip market during the cloud computing boom, before years of design and manufacturing missteps led cloud providers to start looking for alternatives toward the end of the last decade.
  • Nvidia can credit years of investment in the software tools needed to get up and running on its powerful chips for its lead in this market, which may feel like it sprung up overnight but has been building for a long time.
  • Companies relying on alternative designs will have to do much of that heavy lifting themselves, and train their software developers on their custom architectures.
  • But given the supply constraints around Nvidia's chips, they might not have any other choice; it could be easier to learn how to write software for a new AI chip than to find enough of Nvidia's chips to complete the task at hand.

Given how expensive it can be to run modern AI workloads on someone else's computer, it's not hard to see why companies are evaluating their alternatives.

  • After 15 years of centralized cloud computing, the intense demand for AI infrastructure could allow decentralized approaches and custom system development to re-emerge.

Security logs, as a treat

Following last week's disclosure of a serious flaw in Microsoft's authentication software, the company bowed to pressure from CISA and others and announced that it would allow standard customers to access certain logging services that were previously only available as part of a premium package.

"As our expanded logging defaults roll out, Microsoft Purview Audit (Standard) customers will receive deeper visibility into security data, including detailed logs of email access and more than 30 other types of log data previously only available at the Microsoft Purview Audit (Premium) subscription level," the company said in a blog post Wednesday. The attack on several U.S. government agencies carried out by Chinese hackers was only detected because they were using a premium version of Microsoft 365 that provided the additional log data, CISA said last week.

The incident raised long-standing questions about the tension between protecting cloud customers and selling security software features, and now that Microsoft has changed its policies other companies might face similar scrutiny. What were once considered advanced security features for the truly paranoid might now be everyday concerns as attacks continue to grow in sophistication, and customer expectations might evolve in turn.

Enterprise moves

Abhishek Sinha joined AI startup Cohere as vice president of product after 11 years at AWS in various product roles involving analytics.

Merritt Baer, who also worked at AWS in the Office of the CISO, is now Field CISO at Lacework.

Ilan Avital was hired by AI chip company NeuReality as chief R&D officer after almost 30 years at Intel.

The Runtime roundup

Kevin Mitnick, who served several years in prison on hacking charges before becoming a well-respected cybersecurity consultant, died of cancer at 59.

Dell acquired Moogsoft, an AIOps startup that had raised $90 million in funding, for an undisclosed amount.

Postman acquired observability startup Akita Software, which had raised a small amount of funding from investors like a16z and journeyman basketball player Kevin Durant.

IBM missed revenue expectations for the second quarter with almost no growth year over year, but said it still expects to eke out a 3% to 5% gain in revenue this year.

SAP hit revenue expectations for the second quarter but lowered its guidance for the year, which sent its stock down more than 6% in after-hours trading.

TSMC's plans to expand its presence in the U.S. will be delayed after the company said it was finding it hard to hire qualified workers, a problem that Max Cherney and Anna Kramer highlighted for Protocol last year.

We've come full circle: Google will try to address security breaches by asking some workers to use computers disconnected from the internet, as an experiment.

Thanks for reading — Runtime is off for the weekend camping in the Deschutes National Forest — see you Tuesday!

Great! You’ve successfully signed up.

Welcome back! You've successfully signed in.

You've successfully subscribed to Runtime.

Success! Check your email for magic link to sign-in.

Success! Your billing info has been updated.

Your billing was not updated.