Arvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse

NVIDIA

Across the shows: 27 speakers in other episodes, 23 bullish and 10 bearish calls, past 30 days

• Nikesh Arora · Palo Alto NetworksBullish

  Nvidia's Jensen Huang is building a vertically integrated AI stack and funding the ecosystem to keep it there.

  Nikesh describes Nvidia as the modern version of the vertically integrated stack, reaching down to the silicon and up to applications. He says Huang understands that a balance will emerge between frontier models and other models. Nvidia is funding others across the ecosystem so that they do not stray from the stack he is building. Nikesh argues that with control of the full stack, a company can choose which layer to commoditise and move profit between layers. He adds that attacks on one layer can be absorbed by shifting profit to another, which gives the company powerful market leverage.

  “If you look at what Jensen is doing, he's a modern day version of the vertically integrated stack... He's trying to fund everybody in the ecosystem to ensure that they don't stray from the vertical integrated stack that he's producing.”at 32:03

Cybersecurity

• Nikesh Arora · Palo Alto Networks

  Enterprises must start replacing IT stacks with quantum-safe infrastructure now, not in 2029 or 2030.

  Nikesh argues that nearly every enterprise runs on cryptographic keys, and every vendor has keys at each end, so anything that is not quantum-proof will be breakable first. Companies cannot wait for a single cutover date, because replacing trillions of dollars of IT infrastructure cannot be done in one step. He says that the timeline is already late, since the threat is about three years away. He estimates that about half of enterprise IT spend is on infrastructure that is not quantum-safe, and that a few hundred million dollars per firm is already earmarked this year. As a stopgap, companies can wrap traffic in post-quantum-ready keys now, and the replacement should begin immediately.

  “you can't wait for 2029 or 2030... So you should be replacing your IT stack today with quantum safe or quantum proof infrastructure right now.”at 5:47

• David Friedberg · The Production Board

  AI math advances could enable classical cryptanalysis before quantum computers can break encryption.

  Friedberg relays a senior AI researcher's view that AI models could break encryption before quantum computers do. He links it to the recent OpenAI math result, which he sees as an early sign. He then argues the mechanism: on the open frontiers of number theory, better models could yield better predictors of how primes are distributed. Classical models built on such results could then start to break cryptography through an arithmetic route rather than a quantum one. Arvind rebuts this, saying the question is how long it would take per key.

  “But just using classical computing, you can actually have a different path, an arithmetic path. For example, if you had a better predictor of the distribution of prime numbers... We can start to have classical models that can start to break cryptography.”at 8:27

• Nikesh Arora · Palo Alto Networks

  AI finds many real flaws but hallucinates often; attackers need one bug, so defenders should pool vulnerability fixes.

  Nikesh says frontier models trained on code can tell good code from bad, and they find vulnerabilities. The results are noisy, though: he cites hallucination rates near 30%, and a show of hands among peers suggested roughly 60 to 80% of findings were real. Every flagged issue therefore has to be tested before it can be trusted. The asymmetry favours attackers, who need only one vulnerability, while defenders must find and patch every one. He proposes that trusted parties pool vulnerabilities and distribute patches across companies, rather than each firm testing and patching alone. Even AI-written code is not error-free, so it still needs human verification.

  “the reason it's perfect for attackers is they just need one vulnerability. They don't need 10 vulnerabilities. If you're a defender, you need find every vulnerability and patch it.”at 17:07

• Arvind Krishna · IBM

  Defenders must make defence as cheap as attacks and build technological superiority to win the AI cyber contest.

  Arvind applies three lessons from warfare. Alliances can outmatch an attacker, but the bigger point is asymmetry: attackers are often cheaper than defenders, so defenders run out of money and concede. Defence must become as cheap as offence to reach symmetric conflict. Technological superiority has decided most wars. The aim is to find vulnerabilities, patch them, and set boundaries that make penetration expensive for attackers. He adds that coding agents already write about half of IBM's code and do better than before because they run evals before shipping. Yet they only catch the flaws they know how to find, and full verification is computationally infeasible for arbitrary inputs.

  “often an attacker is cheaper than the defender, so the defender runs out of money and at some point then concedes. Or can you make defense equally cheap...”at 19:21

• Nikesh Arora · Palo Alto Networks

  Cyber valuations are inflated by the AI shock and will rationalize, since buyers will spend heavily on cyber.

  Arora notes cyber stocks are enjoying much better valuations lately because of the AI shock. He calls this fine but expects it to rationalize over time. His reason is that all these companies will spend unlimited amounts of money on cyber, which supports the sector's demand but does not justify the current valuation premium.

  “For some reason, you know, cyber is enjoying much better valuations lately because of the AI mid horse event, which is fine, which will rationalize over time because all these guys are going go spend unlimited amounts of money on cyber.”at 36:23

Software & SaaS

• Nikesh Arora · Palo Alto Networks

  Huge SaaS disruption coming as AI makes it unclear which software survives the next five to eight years.

  Arora says it is hard to value the terminal value of software because it is unclear what will survive the next five to eight years given where AI is. He argues product-market fit is being disrupted for many software companies. His test: he knows exactly how half the packaged software he buys should work if AI were applied well, so someone will build it. That leads him to expect huge disruption in SaaS over the next few years.

  “It's not clear what software is going to survive the next five to eight years with where we are from an AI perspective... So I think there's gonna be huge disruption in the SaaS space over the next few years.”at 37:06

• Arvind Krishna · IBM

  Enterprises will consolidate from hundreds of software vendors onto five to ten platforms as AI replaces them.

  Krishna says the average enterprise uses 500 to 5,000 distinct software vendors, which is a massive attack surface. He asks whether companies could replace 490 of them with minor AI capabilities and rely on five to ten main platforms. That implies massive consolidation because the spend will go there. He expects the consolidation to play out, though he gives no firm timeline, saying it could take six months or six years. Independent distribution per product becomes unaffordable as the value of that software falls.

  “The average enterprise uses anywhere from 500 to 5,000 distinct software vendors... That implies massive consolidation because that's where the spend goes.”at 37:57

• Arvind Krishna · IBM

  Interaction-heavy software is easily replaced by AI; reliability-critical deterministic software is hard to replace.

  Krishna distinguishes software by type. Interaction-rich software, where much of the value is in user interaction, can be replaced with a lot of AI, as has been seen clearly. Software that must be reliable to an extreme standard, like a database lookup for retail banking or closing the books at 10 to the minus 15 error rates, will be hard to replace easily with AI-generated software. He sees a real distinction between the two categories.

  “I need you to be reliable at 10 to the minus 15... Which is a database lookup for somebody doing retail banking or closing the books, I think that's gonna be hard to replace that easily... With AI generated software.”at 38:57

Quantum computing

• Arvind Krishna · IBM

  Quantum computers will break today's encryption around 2030; post-quantum cryptography exists, so migration must start now.

  Arvind argues that the timing of a cryptographically relevant quantum computer is debated between roughly 2028 and 2033. He takes 2030 as the midpoint and says that at that point today's cryptography will break, which he calls a reality rather than a supposition. He cites DARPA's Quantum Benchmarking Initiative, which started with a couple hundred companies and is now down to about 13, as a sign of what is real. Post-quantum cryptography is well understood and its code is written, so the gap is enterprise inertia, and he calls the present a wake-up moment. He also flags harvest-now-decrypt-later risk: an encrypted disk stolen today could be broken years later. He qualifies the point: post-quantum security rests on complexity-theory assumptions rather than a proof. He also argues that classical AI breaking encryption is possible in principle but would take too long per key to matter in practice.

  “if you pick the middle point, let's call it 2030, well, it will break today's cryptography. That's not a supposition. That's just a reality.”at 1:10

• Arvind Krishna · IBM

  Quantum gate counts could reach 10 million by 2029 if they grow tenfold yearly from about 10,000 today.

  Arvind proposes two milestones for judging quantum progress. The first is a demonstration of error correction or mitigation on a running machine rather than in simulation. He predicts someone will make such claims by next year, that third parties will evaluate them, and that one or two players will show real progress by 2027. The second is scaling gate computations, which are now around 10,000. If that count adds a zero each year, it reaches about 10 million by 2029. He says this could surprise everyone. He also notes that error-corrected logical qubits are a helpful but imperfect way to measure effective capacity.

  “if you put a zero per year, that means you could be at 10,000,000 by 2029. That's remarkable.”at 3:59

Open-source AI models

• Nikesh Arora · Palo Alto Networks

  Most enterprise AI use cases will end up on trained open-source models, with frontier models reserved for high-value cases.

  Nikesh argues that enterprises should match each task to the level of intelligence and cost it needs, rather than using a Ferrari for everything. Customers will not pay for frontier-level capability on routine tasks. Frontier models will be used only where the value at risk is large, such as a live breach or drug discovery. For most enterprise use cases, companies will train an open-source model on proprietary data and manage its weights. He gives auditability and explainability as the reasons, and says his own defender model would likely follow this path.

  “I think for most practical use cases in enterprises, we're gonna end up training some version of an open source model.”at 12:19

• Arvind Krishna · IBM

  Expects 75-90% of token users to move to open-weight models within a year or two.

  Arvind builds his case on three drivers. The first is economics: frontier tokens cost about $20 per million, while a Chinese open-weight model such as Kimi K3, about 800 billion parameters, could run on the user's own GPUs for $1 to $2 per million after about a week of optimisation. That is 10 to 40 times cheaper for roughly 95% of the work. The second is IP: if chain-of-thought can be logged by a provider, enterprises will want to self-host, which only open-weight models allow. The third is geopolitics: the best open-weight models are Chinese, but sandboxed use is low risk, and the Western gap is only six to twelve months, so capable Western open-weight models should follow. He admits developers default to expensive models, and says production use cases beyond a small set remain doubtful.

  “I don't know whether it'll be 75 to 90%, but that's the range I would put. Of all token users, we'll go to open [weight] models for the reasons I mentioned...”at 14:25

AI regulation & safety

• Nikesh Arora · Palo Alto Networks

  Nobody is slowing AI; builders are accountable for what they ship, and the debate is over-focused on dangers.

  Nikesh says nobody is slowing AI development, and that everyone is working frenetically to build better and faster models. He argues that the public conversation has over-rotated to the dark side and perils of AI, with too little on its benefits, which he calls a narrative problem. He also holds that companies are accountable for what they release, and that they cannot say the software wrote itself. A company that takes economic value from selling software accepts responsibility for it. In his view, existing regulation is already enough, and perhaps too much, to hold firms to account.

  “if you build something, you're accountable for it. If you build something and unleash it, you can't unleash and say, oops, I don't know how to control it. That argument doesn't work.”at 23:29

• Arvind Krishna · IBM

  AI and data-centre regulation is driven by populism, not reality; slowing AI down is impossible globally.

  Arvind says Congress and the White House have debated AI and data-centre questions since 2022, with hearings in both administrations, and he calls the regulatory push populist. Politicians seeking votes have incentives to look tough, though enough people care about long-term economic advantage to keep the balance in place for now. He thinks a 50-state data-centre ban is implausible, but around 30 states banning is possible. On global slowdown, no country will stop: Japan, China and India would use any pause to get ahead. Digital technology cannot realistically be slowed, and Europe shows what happens when regulation comes first.

  “The regulation question on AI and data centers is driven... by populism, not by any reality.”at 26:22

AI infrastructure & data centers

• Nikesh Arora · Palo Alto Networks

  Compute access will be uneven: companies far from the US and China will lack supply, regardless of their budgets.

  Nikesh fears that fast compute advances will concentrate in two large countries, and that the benefits will not be spread evenly across the rest of the world. Compute is scarce right now, and it is hard to get enough to run anything at scale. The further a company is from the centre of gravity, the less likely it is to get capacity. The constraint is supply rather than price. A European company with good intentions could find it gets less access than US or Chinese companies.

  “I fear that we will see fast advances in compute development in two large countries in the world... compute is very dear right now... It's purely because of supply constraints.”at 27:58

AI & jobs

• Arvind Krishna · IBM

  Nations that protect every existing job will lose to those that accept capital-heavy, labour-light AI adoption.

  Arvind frames AI as the latest round of a labour-versus-capital trade-off that has recurred since the 1700s. Each cycle, the side that spent more on capital and worried less about labour has won. He argues that the two leading nations are willing to be capital-intensive and to accept that labour matters less. Nations that insist on protecting every existing job at all costs will lose. Displacement may be acceptable where it leads to better jobs. Cheap AI means a worker in one country can use the capability, so the social choice between protecting jobs and adopting AI will decide the winners and losers.

  “this is the classic labor versus capital trade off that is happening... People who spend more on capital than worried about labor tended to win every previous one of the seven or eight cycles like this.”at 28:54

Enterprise AI adoption & AI software

• Nikesh Arora · Palo Alto Networks

  Memory and context will lock users into AI agent stacks, so model routing services will struggle.

  Nikesh argues that context and memory are now inseparable from reasoning as models grow. Routing services such as OpenRouter would therefore only move reasoning tokens, and running a model-agnostic router will become impossible. Over time the economics will favour end-to-end stacks priced per action. Personal agents illustrate the point: once an agent is trained on a user's context, APIs and connectivity, switching it to another stack means losing that memory. Friction will be high, so application-layer vendors will build vertical stacks too.

  “memory will require or make it easier and there'll be too high of friction for you to move your task from one stack to the other stack because it has all your context and memory.”at 33:42

Consumer AI products

• Nikesh Arora · Palo Alto Networks

  AI agents will force apps to go headless, making consumer control the largest AI battle.

  Arora argues that frontier labs will push a full-court press to make AI assistants the interface for consumers. He says this has been tried six to nine months ago without success, but Facebook's approach now shows how it can work, so he thinks nearly everyone will pursue it. Apps will expose APIs and have to become headless, which he calls the largest battle of consumer control.

  “Now with Facebook showing us how it can work, I think pretty much everybody has it. There'll be APIs. It'll be the largest battle of consumer control.”at 34:37