Arvind Krishna: IBM's Reinvention, AI Bets and Quantum | Podcast

ELI5/TLDR

IBM was a company everyone trusted but assumed was a relic. Its CEO, Arvind Krishna, spent six years cutting the slow-growing parts away and betting on software, cloud, and artificial intelligence — and the company is now growing again. In this interview he explains why he thinks the AI building boom is partly overbuilt, why the old mainframe computer refuses to die, and why he is putting hundreds of billions of dollars of hope on quantum computing arriving by 2029. The throughline of all of it is one idea: get a frightened, declining organization to start taking risks again.

The Full Story

What IBM actually is now

Most people picture IBM as a hardware company — big machines in cold rooms. Krishna corrects that early. Hardware is only a fifth of the business. Roughly half is now hybrid cloud and AI software, and about a third is consulting that helps clients move into the digital and AI era.

When he became CEO in 2020, his diagnosis was blunt. IBM was trusted but seen as the past, not the future. His fix was to do things that pay off over years, not quarters.

“They’re not always the biggest revenue in a month or in 3 months or in 1 year, but they become the big revenue over the next many years.”

The portfolio surgery

Two decisions defined the turnaround. The first was buying Red Hat in 2019. The logic was coldly economic. IBM was far behind on public cloud, and catching up would mean spending five to ten billion dollars a year only to remain fifth place. Think of it like deciding not to open a fourth coffee chain when three giants already own every corner — better to sell beans to all of them. Red Hat let IBM become a useful partner to every big cloud provider rather than a losing competitor.

The second was spinning off the IT services business — roughly a third of the workforce. It was declining at 5% a year, low-margin, and a place where customers wanted stability rather than innovation. Krishna’s reasoning was almost arithmetic: if one piece is shrinking, every other piece has to grow twice as hard to hit the company’s overall target. Cut the anchor and the rest can sail.

He frames acquisitions through three lenses. Keep the engineering team free to build their own product. Fully integrate the sales and support side, because IBM’s real edge is reach — it operates in far more countries than anyone it buys. And absorb the dull machinery (HR, taxes, contracts, cash) immediately, because there is no value in running those twice.

The biggest lever: making fear go away

Asked for the best thing he has done, Krishna doesn’t name a deal. He says he made the culture willing to take risk again.

His theory of decline is almost biological. When an organization starts shrinking, people turn inward and quietly optimize for personal survival — and the safest move is to never stick your head up.

“If the culture is in decline, then people begin to say, ‘I survive by not raising my head, by not looking like an outlier.’”

The fix was to actively ask for risk. He tells people not to bring him a 90%-confidence plan but a 50%-confidence one, then builds in a buffer because half of those will slip. His view is that risk-aversion in survivors is usually learned, and learned behavior can be unlearned. He reckons he is about halfway through this culture shift.

Where the AI bubble is

Krishna is careful with the word bubble. His phrasing is that some will disappoint, many will thrive, but not all will thrive. The part he thinks is overbuilt is the physical infrastructure.

His back-of-envelope math: one gigawatt of data-center power needs roughly 60 to 80 billion dollars of chips to fill it. The world has committed to over a hundred gigawatts, which implies six to eight trillion dollars of buildout. To pay that back in five to seven years you would need an extra one to two trillion dollars of new revenue every year — and he doesn’t believe that much demand exists yet. He’d be comfortable if the buildout were half its current size.

He also thinks the largest AI models will become commodities — valuable, but with low switching costs, which means no protective moat around the margin. His prediction: of the dozen or so companies trying to build frontier models, maybe two or three survive.

On who wins, he’s clearer about consumers than enterprises. Companies with a huge existing consumer base have a built-in distribution advantage if they point it at AI. On the enterprise side, he thinks it’s wide open.

Why IBM isn’t building giant models

IBM rents chips, mostly avoids building frontier models, and instead builds small open-weight models (the Granite family — none over 100 billion parameters, while the giants are now in the trillions). The bet is that most companies will end up using multiple models, and that small ones are cheaper to run and can live on-premise where data privacy matters.

This separates IBM from Microsoft, Google, and Amazon, who are chasing one consumer product billions of people might use. IBM’s question is narrower: can Nestlé, Pepsi, or Bank of America use this to fix procurement, accounts payable, or internal data decisions?

The mainframe that wouldn’t die

Here is the surprise. The mainframe — the very thing AI coding tools were supposed to kill — is IBM’s fastest-growing hardware, up every year for six years after a decade of decline. The reason is that mainframes run the workloads you cannot afford to get wrong: credit-card authorizations, retail banking, airline reservations. They need near-perfect reliability (“six to nine nines” of uptime), and moving them to the cloud would cost roughly three times as much.

The new Z17 machine does three notable things. It puts a small GPU right on the main processor, so a fraud-check model can run inside a transaction instead of sampling some transactions afterward. An add-on card lets a full machine run 450 billion AI inferences a day with no extra data movement and effectively no delay. And it ships with post-quantum cryptography — encryption designed to survive future quantum attacks — baked in.

The quantum bet

Krishna’s big long-range wager is quantum computing. His one-sentence version: a quantum computer harnesses quantum mechanics to do a new kind of math. Ordinary computers do arithmetic very fast. GPUs do matrix math, which is what unlocked AI. Quantum machines do something different again.

He puts the arrival of a genuinely useful machine at 2029, and rates his confidence at “100” — then immediately clarifies he means 100% sure they’ll have one, not 100% sure how useful it will be on day one. The basis for that confidence is concrete: IBM already runs machines at the scale of hundreds to low thousands of qubits. To hit 2029 they need to scale up tenfold and improve error correction tenfold. Hard, but not a leap into the unknown.

The first three use cases he expects:

“The first one I think nobody debates is going to be in the world of materials.”

Better corrosion coatings, better drugs, better fertilizer, possibly better magnets (which matter for EVs and electrification) — all by simulating molecules instead of running physical lab experiments. Second, financial risk: pricing complex derivatives or bonds in milliseconds instead of using day-old data. Third, optimization: smarter routing — he notes 30% of truck miles run empty because good routing is too hard to compute today.

He values the opportunity for IBM at “hundreds of billions,” using GPUs as the analogy: in 2015 they were a niche product, and seven years later demand was insatiable.

National security and the dark side

Quantum carries a sharp edge. There’s an algorithm (Shor’s) that can break most of today’s encryption, meaning a sufficiently powerful quantum computer could read encrypted communications in the clear — an offensive military capability. Combined with economic advantage and defense applications, that makes it a top-down national priority for governments, China included.

On regulating AI

Krishna is skeptical that the technology can be regulated, though he thinks use cases can be. His reasoning is physical: tangible goods have weight and borders, so you can control them. A digital good crosses any boundary freely. He points to authoritarian regimes that try and fail to fully control internet access as the tell.

On leadership

Krishna has a PhD and 15 patents, but he waves away the “is being clever useful” question. His framing is that a leader’s first job is to empower the team and to see “around the corner” on technology, while deliberately hiring people stronger than him in law, politics, and finance.

His standout idea is the “getting fired mentality,” handed to him by a mentor 15 years ago.

“Arvind, you should live in the pleasure of being fired.”

Not picking fights — but being genuinely unafraid of losing the job, which frees you to do the right thing. He says he’s come close to being fired a couple of times, most recently in 2014 over an unpopular decision.

He also credits COVID with making IBM’s painful restructuring easier: when the whole market is already disrupted, you can take all the pain at once. He reckons it let him do in one year what might have taken three or four.

Key Takeaways

IBM today is ~50% hybrid cloud and AI software, ~33% consulting, ~20% hardware — not a hardware company.
The Red Hat acquisition was a decision to partner with all cloud giants rather than spend billions losing to them.
Spinning off IT services removed a 5%-declining, low-margin drag so the rest of the company could hit growth targets.
M&A playbook: leave engineering free, fully integrate go-to-market (IBM’s edge is global reach), absorb back-office functions immediately.
His single biggest move was cultural — making a frightened, declining org take risks again; he asks for 50%-confidence bets, not 90%.
Decline makes people optimize for personal survival by not standing out; that risk-aversion is learned and can be unlearned.
AI infrastructure math: ~100GW committed implies $6–8T of buildout, requiring $1–2T of new annual revenue to pay back — which he doubts exists.
Frontier models will likely become low-moat commodities; maybe only 2–3 builders survive.
IBM builds small open-weight models (Granite, under 100B params) on a bet that the future is multi-model, cheap, and on-premise.
Mainframes are growing because they run mission-critical workloads where cloud costs ~3x more; the Z17 runs AI inference inline and ships post-quantum encryption.
Quantum = “a new kind of math.” He’s confident IBM has a useful machine by 2029; needs 10x scale and 10x better error correction from today’s hundreds-to-thousands of qubits.
First quantum use cases: materials science, financial risk pricing, and logistics optimization.
Shor’s algorithm makes quantum a national-security matter — it can break most current encryption.
He believes AI tech can’t be regulated (digital goods cross borders freely), but use cases can.
“Getting fired mentality”: being unafraid of losing your job frees you to do the right thing.

Claude’s Take

This is a high-signal interview, and the interviewer (a sovereign-wealth-fund CEO who runs the same turnaround playbook himself) asks sharper questions than most. Krishna is genuinely worth listening to on AI economics because he has skin in the game but no incentive to pump the bubble — IBM deliberately sits out the frontier-model arms race. His gigawatt-to-dollars-to-required-revenue calculation is the most useful thing here: it’s a clean, falsifiable way to think about whether the buildout pencils out, and it doesn’t require believing AI is fake.

The BS filter does flag a few things. This is a CEO telling a flattering version of his own turnaround on a friendly podcast, so the culture-change narrative is tidier than reality surely was. The quantum “100% confident by 2029” is doing a lot of work — he immediately hedges it down to “we’ll have a machine,” and useful-in-production he pegs at the 2030s. Vendors have been promising useful quantum for a while; treat the timeline as advocacy, not forecast. And “hundreds of billions” for IBM’s quantum opportunity is the kind of number that’s safe precisely because no one can check it for a decade.

Score is 8. The infrastructure math, the mainframe-survival logic, and the model-commoditization thesis are durable mental models you can reuse. It loses a point or two for the inevitable self-serving gloss and a quantum section that’s more sales than science. Worth the watch.

Arvind Krishna: IBM's Reinvention, AI Bets and Quantum | Podcast | In Good Company