Person
5 episodes · 100 quoted claims
IBM announced its Red Hat acquisition in 2018, took a year to close it after regulatory review, and it cost about 30% of IBM's market capitalization, according to Arvind Krishna.
“remember when we made that decision, that was 2018 that we announced. It took a year to get through regulators to close it. It was 30% of our market cap.”
Arvind Krishna defines the deep learning era as roughly 2010 through 2022, a period of major AI progress.
“So in the deep learning era, let's call that 2010 through 2022, that twelve years. Deep learning made incredible progress, no question about it.”
Arvind Krishna describes Moore's law as a doubling of what chips can do every two years.
“That was the beginning of Moore's law, really, right? Every two years you get a two x advantage in what you can do.”
Arvind Krishna says IBM's early results suggest Groq's inference chips are about ten times cheaper than alternatives.
“And Groq is looks like, at least at our first uh steps, looks like it's ten x cheaper.”
Arvind Krishna argues that 10x gains in chips, chip design and software could combine for AI compute that is 1,000 times cheaper.
“So there's a ten x from the software side. You put those three tens together, that's a thousand times cheaper.”
Arvind Krishna says many investors in fiber-optic networks built around 2000 did not earn a return.
“if I look at fiber optics in the ground back in the year 2,000, not all of those people got a return.”
Arvind Krishna says about 60% of total technology spend is outside the United States, and regulated US industries tend toward hybrid cloud.
“About 60% of the total spend is outside the US. Even inside the US, anyone in a regulated industry is going to be hybrid in some sense.”
Arvind Krishna says IBM has never misused client data in its 114 years.
“we tend to have the trust with those people that we have never, in 114 years, misused that data, not even once.”
Arvind Krishna says HSBC tested IBM quantum computing for bond trading pricing.
“That's why HSBC proved to itself we could do bond trading pricing on it.”
Arvind Krishna says IBM's open-source quantum software has about 650,000 users worldwide.
“The fact that there's 650,000 people globally who use it, hey, it tells me that there is excitement and there is a movement”
Arvind Krishna says about 300 clients work with IBM on quantum computing in a research rather than commercial mode.
“We have 300 clients, not commercial. There are 300 people working with us in a, let's call it, research mode.”
Arvind Krishna gives roughly 20-30% odds that IBM's quantum computing reaches something remarkable in three to three-and-a-half years.
“Will we get to something remarkable in three, maybe three and a half years? I'm gonna give it low odds. It's possible, but it's maybe twenty, 30%.”
Arvind Krishna estimates that filling a one-gigawatt AI data center costs about $80 billion at today's prices.
“It takes about $80 billion to full to fill up a one gigawatt data center. Okay, that's today's number.”
Arvind Krishna estimates global AGI-chasing commitments at about 100 gigawatts, implying roughly $8 trillion of capital spending.
“the total commits in the world on this space of the chasing AGI seem to you like a hundred gigawatts. At these announcements, that's 8 trillion of capex.”
Arvind Krishna puts the odds that current known technologies reach artificial general intelligence at zero to 1%.
“I am not convinced, or rather, I give it really low odds, like we're talking zero to 1%, that the current set of known technologies gets us to AGI.”
Arvind Krishna dates the foundational transformer-based approach behind LLMs to a 2017 paper.
“this stuff is eight years old at this point. That's the first paper I think was 2017 when intention and this approach came to be.”
Arvind Krishna predicts up to 10% of total US employment could be displaced by AI over the next couple of years.
“Could there be up to 10% job displacement? I'll use that phrase. I actually believe that's going to be likely over the next couple of years.”
Arvind Krishna says an IBM team of 6,000 using an in-house coding tool became 45% more productive within four months.
“Within four months, the team who embraced it, which is 6,000 people, so not a tiny number, was 45% more productive just to compare”
Arvind Krishna says he formed his hybrid cloud thesis for IBM in 2017.
“So I made the hybrid observation deeply in 2017. As I was making that, I said, Okay, how do I test this?”
Arvind Krishna says hybrid cloud and AI software make up almost half of IBM's total revenue.
“IBM is largely a hybrid cloud and A I software company. We have made the transition to that's almost half our total revenue.”
Krishna says hardware accounts for about 20% of IBM's business.
“And then we have about 20% that is hardware. I realize many people think that we are largely a hardware company, but that is just a fifth of the company”
Krishna says he was convinced in 2019 that AI would be big, about three years before the wider world noticed.
“back in 2019, I was convinced that A I would be big. It took another three years for the world to wake up to that.”
Krishna says IBM would have needed to spend roughly five to ten billion dollars a year to catch up in public cloud.
“When you're that far behind, you would have to spend multiple billions, five to 10 billion a year to try and catch up.”
Krishna says IBM acquired Red Hat so it could partner with all the big cloud providers instead of competing in public cloud.
“there were only two or three companies I thought that could help give us a portfolio. That makes us a great partner for all of them”
Krishna says IBM spun off its IT services business because it was declining about 5% a year and would have made the company's growth targets harder to meet.
“If you say and conclude revenue growth is essential, then something which is itself declining at 5% is something that should not be part of it.”
Krishna describes Confluent, IBM's latest acquisition, as the best in the world at making data real-time using Kafka.
“Confluent with its Kafka backbone is the best in the world at taking data and making it real-time for all the purposes that you might want to use it for.”
Krishna says IBM's mainframe business has grown every year in the six years since he became CEO, after declining for the ten years before.
“in the last six years mainframe has grown every single year.”
Krishna estimates that populating one gigawatt of AI data-center capacity costs 60 to 80 billion dollars of semiconductors.
“about a gigawatt of power, you can debate, but it costs you sixty to $80 billion worth of semiconductors to go populate it.”
Krishna says commitments of over 100 gigawatts of AI data centers imply 6 to 8 trillion dollars of spending.
“people have committed over a hundred uh gigawatts of uh a I data center build-out, that points to six to 8 trillion worth of a build-out.”
IBM built Watson in 2011 to show that AI could handle natural-language questions, including winning the game show Jeopardy.
“In 2011, we wanted to prove that A I can do and solve problems that were unimaginable.”
Krishna says IBM's Watson effort failed partly because it targeted healthcare, the hardest industry, with monolithic applications.
“we unfortunately picked the vertical that is the hardest health. That was the mistake.”
IBM's Granite family of open-weight AI models has no model above 100 billion parameters.
“We have open weight models called the granite family, but in there, I'll note we have not built a single model that's over 100 billion parameters”
Krishna recalls the Morris worm, created by a Carnegie Mellon student, as an earlier example of clever people exploiting software vulnerabilities.
“I remember when I was in graduate school, this is thirty five years ago, uh, we had the Morris worm that came out of a really clever, uh, kid at Carnegie Mellon.”
Krishna says IBM's stock was down about 25% in the software sell-off, versus 40 to 60% for many software companies.
“we are down about twenty five. So so, in some sense, we are taking a hit”
IBM put AI on its mainframe platform in 2021, added more in 2024, and added still more in the latest generation.
“We put it on the platform in 2021, then we put more in 2024, and then we put even more, uh, this year.”
Krishna says the Spyre card lets a fully populated IBM Z17 mainframe run 450 billion AI inferences per day in line with transactions.
“That lets you, in a fully populated mainframe, do 450 billion inferences per day on the platform at zero latency and right in line”
The IBM Z17 mainframe includes post-quantum cryptography to protect data against future quantum attacks.
“we have post quantum cryptography built right into the platform.”
Krishna says IBM's team recently showed quantum computers can predict magnetic properties of materials.
“our team just showed that you can predict magnetic properties of materials using quantum computers.”
Krishna cites a figure that about 30% of truck miles and containers are empty, as an optimization problem quantum computing might address.
“30% of all truck miles and containers are empty as opposed to use.”
Krishna says IBM needs to scale its quantum computers tenfold and improve error correction tenfold by 2029.
“we have them at a scale of hundreds to low thousands today. So that's not a future statement. So we got to get up in scale by a factor of ten”