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

IBM CEO Arvind Krishna and Palo Alto Networks CEO Nikesh Arora debate quantum's 2030 encryption threat, open-weight models displacing frontier LLMs, and AI cyber warfare, concluding AI will force major consolidation in cyber and SaaS.

Arvind Krishna, IBM's CEO, argues that quantum computers capable of breaking today's cryptography are likely by roughly 2030, and that enterprises are already years late in migrating to post-quantum encryption. The quantum threat opens the episode, but most of the conversation is AI economics and strategy: why open-weight models will take the bulk of token usage, how frontier labs and vertical integration reshape software, and what AI-driven cyberattacks mean for enterprise defense and SaaS valuations.

The guests are Krishna and Nikesh Arora, CEO of Palo Alto Networks, speaking with hosts Chamath Palihapitiya and Jason Calacanis. Both run large enterprise technology companies and speak from that operator's seat: Krishna on infrastructure, mainframes, and model economics; Arora on security operations, M&A, and the model-vendor relationships his firm sits inside. The reader should come away with a concrete timeline for the quantum migration, a cost-based argument for open-weight models, and a view on where software consolidation is heading.

Quantum breaks today's crypto around 2030, and migration is already late

Krishna's central claim is that a working, error-corrected quantum computer will break current public-key cryptography, and the debate is only the date. He cites DARPA's Quantum Benchmarking Initiative (QBI), which he says started with a couple hundred companies a couple of years ago and is down to 13. Estimates for a break range from 2028 to 2033; he takes the midpoint, 2030, and calls the break "not a supposition, that's just a reality."

He lays out the milestones he would watch, and predicts claims will arrive on a schedule:

He is careful the term "logical qubit," which he says he does not love but accepts as shorthand for effective capacity after error correction: 10,000 physical qubits might yield roughly 200 logical ones. On whether post-quantum algorithms are truly safe, he says they are "quantum proof" at the same level of rigor that today's crypto resists classical computers — grounded in complexity theory and assumptions, not a theorem-level proof.

The practical warning is stored data. An encrypted hard disk stolen today can be held for five years and broken once the hardware arrives. Krishna says post-quantum cryptography is understood and the code is written, but "people have been too lazy to get to it." Arora's operational version: every vendor and endpoint runs crypto keys, so anything not quantum-safe is breakable, and you cannot replace a $5 trillion IT estate in one year. His prescription is to buy quantum-safe infrastructure now and wrap post-quantum-ready keys around traffic as a stopgap. "If he says it's 2029, 2030, we're already late."

Calacanis raises a counter-thesis he attributes to a senior AI researcher: that AI models, not quantum hardware, could break cryptography first, possibly via new results in number theory, with the week's OpenAI math release as an early signal. Krishna concedes an AI could discover a new quantum algorithm but insists it cannot run efficiently on classical hardware — "you cannot simulate something on a different architecture and expect a more efficient result." A classical shortcut might exist, but if breaking one key takes ten years, "most people are going to rest in peace."

Open-weight models will take most tokens on cost, IP, and geopolitics

Krishna predicts 75–90% of token users will move to open-weight models, for three reasons. First, economics: frontier pricing runs $20 per million tokens for a generation-old model and up to $50 for the most extreme, while running an open-weight model such as Kimi K2 (roughly 800 billion parameters) on your own GPUs with a week of engineering work drops the cost to $1–2 per million — a 10x to 40x saving for "effectively what can do 95% of the work." Second, IP: if chain-of-thought is logged and audited by a third party, you are giving away your reasoning. Frontier labs will not hand over weights, so open-weight is the only way to run in-house. Third, geopolitics: the best open-weight models are Chinese today, which he says is low-risk if sandboxed and run locally but different if consumed from a cloud; he expects Western open-weight models within a year or two, since the capability gap is six to 12 months at most.

Arora frames the same conclusion from the buyer's side. In the experimentation phase, let teams use the best, easiest model so 20,000 employees can learn what AI does. In production, match model intelligence to task value. His example: Palo Alto inspects endpoint traffic at $20–25 per endpoint per year using small, task-specific models compressed to 99.9% accuracy — routing all of that through a frontier model on tokens "sounds like the sound of revenue" for OpenAI and is not practical. The exception is an active breach, where the business impact runs into the billions and customers will pay for agent swarms and the most current frontier model to reason through the defense.

Krishna adds a behavioral point: developers default to the most expensive option unless controlled, so cheap models belong in production and expensive ones in experimentation. On data, Arora notes that many Claude Code users have no zero-data-retention arrangement and are effectively training the models on their code — which also teaches a model what bad code looks like, useful to attackers who need only one vulnerability while defenders must find and patch all of them.

Defense wins by making attack expensive, not by writing perfect code

Arora describes the current state of AI code review bluntly: models find real vulnerabilities, but roughly 20% of findings are hallucinations (60–80% real by his estimate of the room), and nobody patches without re-testing. The asymmetry favors attackers — one unpatched hole is enough — so defenders must find everything.

Krishna maps this onto three lessons from warfare. Alliances can outweigh a single attacker, which is why sharing vulnerabilities and distributing patches centrally beats every company patching open-source flaws alone. Asymmetry decides wars of attrition: if the attacker is cheaper, the defender runs out of money and concedes, so the goal is to make defense equally cheap. And technological superiority tends to win outright, which is the real objective — find vulnerabilities, patch them, and raise the cost of getting through.

On agentically written code, Krishna says it is already superior to human-written code at IBM, where agents write roughly half the code, because agents also run evals and fix what they find before release. But full verification is "almost computationally infeasible" today, and the loop only finds flaws we know how to look for. Arora's accountability line is simple: a commercial enterprise taking economic value for a product must own what it ships — "you can't unleash and say oops I don't know how to control it."

Nobody is slowing down, and regulation is driven by politics

Calacanis asks whether the inevitability of an AI arms race changes the "stop AI" conversation, and relays a Washington conversation with a senior member of Congress who said one major CEO is "completely freaked out" cyber and floated a possible 50-state ban on data centers by Q4. Arora doubts 50 states ever agree on anything and says the industry has over-rotated onto AI's dark side at the expense of its benefits. Krishna is more direct: the data-center and AI regulation debate is "driven by populism, not by any reality," and a slowdown in one country only lets Japan, China, or India catch up. He thinks the current balance is roughly right but concedes an election can move it, and puts the odds of a 50-state data-center ban at essentially zero while allowing that 30 states could act, with the other 20 "rubbing their hands in glee."

On the haves-and-have-nots question, Arora says the binding constraint is compute supply, not price: the further a company sits from the US or China epicenter, the less compute it gets, regardless of intent. Krishna frames it as the recurring labor-versus-capital trade-off — capital-intensive economies have won each prior cycle — and argues that countries protecting every existing job at all costs will lose to those that accept displacement in exchange for better jobs.

Vertical integration lets the owner move profit between layers

Asked the conglomerate pendulum, Krishna uses IBM's mainframe, which has lasted 65 years, as the case study. In the 1960s IBM went "all the way to dirt" — sand to wafers to fabs to chips to software — which prevented IP leakage and kept the full stack's profit inside the company. Today IBM designs chips and most software but does not fab, because fab scale is a commodity game. The rule: integrate where advantage is decisive, buy where it is commodity.

Arora calls Nvidia the modern version, with Jensen Huang operating from the dirt to the application layer and funding the ecosystem (including Hugging Face) to keep participants on his stack. The strategic payoff of owning the stack is the ability to choose which layer to commoditize: if a rival attacks one layer, destroy the profit pool there and move profit elsewhere.

Applied to the application layer, Arora argues the open-versus-closed debate is really open stack versus closed stack. As models grow in memory and context, separating reasoning from context and memory stops making sense, so an "open router" that just passes reasoning tokens around becomes impractical. A personal agent trained on your context, APIs, and connectivity builds a moat because moving it to another stack means losing all of that. He expects every frontier lab to run a full-court press on consumer agents in the three months, following Meta's demonstration, and predicts many iPhone apps become headless.

SaaS faces consolidation as AI erodes terminal value

Arora's cyber M&A logic: customers should not have to stitch together 40 security products. He compares it to CRM, where 23 systems at Fidelity decades ago consolidated into off-the-shelf platforms like Salesforce, Microsoft Dynamics, and Oracle — once products converge, differentiation disappears and platforms win. Cyber valuations are elevated now partly on the AI security scare, which he expects to normalize even as customers spend heavily on defense.

Outside cyber, he is bearish on software terminal value: it is unclear which SaaS products survive the five to eight years as AI resets product-market fit, and "there's going to be huge disruption in the SaaS space over the few years." Krishna extends the point: enterprises run 500 to 5,000 software vendors, a massive attack surface, and AI capabilities make it possible to replace hundreds of them with five to ten main platforms — consolidation follows because independent distribution cannot be afforded on falling software value. He draws one line: interaction-rich software is easy to replace with AI, but systems that must be reliable to 10^-15, like a retail-banking database lookup or closing the books, will be hard to replace, and rebuilding them offers no cost advantage. On M&A timing, Krishna says multiples looked attractive early in the year and less so a month ago, and he is waiting to see.

The through-line across the episode is that capability is arriving faster than the institutions meant to absorb it. Quantum gives a hard deadline that most enterprises have not started planning for; open-weight models give buyers a 10x–40x cost escape hatch from frontier pricing; and AI-written code cuts both ways, helping defenders find flaws while handing attackers a cheaper path in. The unresolved tension is timing: Krishna's 2030 quantum estimate, his 75–90% open-weight prediction, and Arora's SaaS-disruption window are all forecasts from executives with positions in the outcome, and none will be settled until the milestones arrive. The concrete markers to track are third-party verification of error-correction claims by 2027, whether gate counts add a zero per year through 2029, and whether Western open-weight models close the six-to-12-month gap to Chinese ones.

Nvidia is described as the modern vertically integrated stack, funding ecosystem players to keep them on its stack and using control of the stack to shift profit between layers and destroy profit pools where competitors attack.

Nikesh argues cybersecurity should consolidate into a single platform so customers stop stitching dozens of point solutions, and warns that outside cyber the terminal value of SaaS is unclear, with product-market fit being disrupted by AI over the five to eight years.

Because every layer of enterprise IT runs on crypto keys, Nikesh argues companies should begin replacing infrastructure with quantum-safe systems now and wrap traffic in post-quantum-ready keys as a stopgap, rather than waiting for a quantum break date.

Arvind argues quantum algorithms can only deliver their advantage on quantum hardware, so enterprises should treat post-quantum cryptography as a known, already-written codebase that simply needs to be adopted.

If you control the vertical stack, you can decide which part of the stack to commoditize. So you can decide to move profit between layers.

If you control the vertical stack, you can decide which part of the stack to commoditize. So you can decide to move profit between layers.

We start saying open models, open weight. It's literally an open stack versus a closed stack.

We start saying open models, open weight. It's literally an open stack versus a closed stack.

The average enterprise uses anywhere from 500 to 5,000 distinct software vendors... could I replace 490 of them with minor AI capabilities and do I need to go back and leverage five to ten main platforms?

The average enterprise uses anywhere from 500 to 5,000 distinct software vendors... could I replace 490 of them with minor AI capabilities and do I need to go back and leverage five to ten main platforms?

You can't wait for 2029 or 2030 because it's not like one day all of us are going to go out and say, let's go burn $5 trillion of IT infrastructure and let's go build it again in one year because now we need quantum safe infrastructure.

You can't wait for 2029 or 2030 because it's not like one day all of us are going to go out and say, let's go burn $5 trillion of IT infrastructure and let's go build it again in one year because now we need quantum safe infrastructure.

You cannot simulate something on a different architecture and expect a more efficient result.

You cannot simulate something on a different architecture and expect a more efficient result.

I think it is going to be possible, but if it takes 10 years to break one key, most people are going to rest in peace. That's not a practical answer.

I think it is going to be possible, but if it takes 10 years to break one key, most people are going to rest in peace. That's not a practical answer.

Satya Nadella on the AI Doomer Slowdown, Microsoft’s Master Plan & Who Wins AI

The Real Reason AI Founders Are Calling For a Slowdown | Conversations

153. A Conversation with Zeng Ming on Historical Industry Perspectives: Harsh Truths, Companies That Will Die, Good ≠ Great, and Why OpenAI and Anthropic Likely Won't Be the Big Winners of the Native Era

Databricks CEO on What We’re Getting Wrong Existential Risk

Aligning Multi-Agent Systems and Financializing Compute

Turing CEO on Super Intelligence, Open Source, Agents

NVIDIA Crushes Quarter & Buys Hugging Face | Instinct Hits $2.5B Valuation, Cognition Raises at $46B

Pacing Gets Political + Andon Labs & an AI Religion?

Aravind Srinivas on the 1-Person Companies Making Millions With AI

Ep. 031 - EMERGENCY EPISODE: Are We Doomed? | Jordan Nanos, Doug O'Laughlin, Max Kan, Joey Brookhart

How the AI Boom Could Reindustrialize America

Andrew Feldman on Building Cerebras and the Future of Chips | Ep. 57

IBM CEO Arvind Krishna and Palo Alto Networks CEO Nikesh Arora debate quantum's 2030 encryption threat, open-weight models displacing frontier LLMs, and AI cyber warfare, concluding AI will force major consolidation in cyber and SaaS.

Arvind Krishna, IBM's CEO, argues that quantum computers capable of breaking today's cryptography are likely by roughly 2030, and that enterprises are already years late in migrating to post-quantum encryption. The quantum threat opens the episode, but most of the conversation is AI economics and strategy: why open-weight models will take the bulk of token usage, how frontier labs and vertical integration reshape software, and what AI-driven cyberattacks mean for enterprise defense and SaaS valuations.

The guests are Krishna and Nikesh Arora, CEO of Palo Alto Networks, speaking with hosts Chamath Palihapitiya and Jason Calacanis. Both run large enterprise technology companies and speak from that operator's seat: Krishna on infrastructure, mainframes, and model economics; Arora on security operations, M&A, and the model-vendor relationships his firm sits inside. The reader should come away with a concrete timeline for the quantum migration, a cost-based argument for open-weight models, and a view on where software consolidation is heading.

Quantum breaks today's crypto around 2030, and migration is already late

Krishna's central claim is that a working, error-corrected quantum computer will break current public-key cryptography, and the debate is only the date. He cites DARPA's Quantum Benchmarking Initiative (QBI), which he says started with a couple hundred companies a couple of years ago and is down to 13. Estimates for a break range from 2028 to 2033; he takes the midpoint, 2030, and calls the break "not a supposition, that's just a reality."

He lays out the milestones he would watch, and predicts claims will arrive on a schedule:

He is careful the term "logical qubit," which he says he does not love but accepts as shorthand for effective capacity after error correction: 10,000 physical qubits might yield roughly 200 logical ones. On whether post-quantum algorithms are truly safe, he says they are "quantum proof" at the same level of rigor that today's crypto resists classical computers — grounded in complexity theory and assumptions, not a theorem-level proof.

The practical warning is stored data. An encrypted hard disk stolen today can be held for five years and broken once the hardware arrives. Krishna says post-quantum cryptography is understood and the code is written, but "people have been too lazy to get to it." Arora's operational version: every vendor and endpoint runs crypto keys, so anything not quantum-safe is breakable, and you cannot replace a $5 trillion IT estate in one year. His prescription is to buy quantum-safe infrastructure now and wrap post-quantum-ready keys around traffic as a stopgap. "If he says it's 2029, 2030, we're already late."

Calacanis raises a counter-thesis he attributes to a senior AI researcher: that AI models, not quantum hardware, could break cryptography first, possibly via new results in number theory, with the week's OpenAI math release as an early signal. Krishna concedes an AI could discover a new quantum algorithm but insists it cannot run efficiently on classical hardware — "you cannot simulate something on a different architecture and expect a more efficient result." A classical shortcut might exist, but if breaking one key takes ten years, "most people are going to rest in peace."

Open-weight models will take most tokens on cost, IP, and geopolitics

Krishna predicts 75–90% of token users will move to open-weight models, for three reasons. First, economics: frontier pricing runs $20 per million tokens for a generation-old model and up to $50 for the most extreme, while running an open-weight model such as Kimi K2 (roughly 800 billion parameters) on your own GPUs with a week of engineering work drops the cost to $1–2 per million — a 10x to 40x saving for "effectively what can do 95% of the work." Second, IP: if chain-of-thought is logged and audited by a third party, you are giving away your reasoning. Frontier labs will not hand over weights, so open-weight is the only way to run in-house. Third, geopolitics: the best open-weight models are Chinese today, which he says is low-risk if sandboxed and run locally but different if consumed from a cloud; he expects Western open-weight models within a year or two, since the capability gap is six to 12 months at most.

Arora frames the same conclusion from the buyer's side. In the experimentation phase, let teams use the best, easiest model so 20,000 employees can learn what AI does. In production, match model intelligence to task value. His example: Palo Alto inspects endpoint traffic at $20–25 per endpoint per year using small, task-specific models compressed to 99.9% accuracy — routing all of that through a frontier model on tokens "sounds like the sound of revenue" for OpenAI and is not practical. The exception is an active breach, where the business impact runs into the billions and customers will pay for agent swarms and the most current frontier model to reason through the defense.

Krishna adds a behavioral point: developers default to the most expensive option unless controlled, so cheap models belong in production and expensive ones in experimentation. On data, Arora notes that many Claude Code users have no zero-data-retention arrangement and are effectively training the models on their code — which also teaches a model what bad code looks like, useful to attackers who need only one vulnerability while defenders must find and patch all of them.

Defense wins by making attack expensive, not by writing perfect code

Arora describes the current state of AI code review bluntly: models find real vulnerabilities, but roughly 20% of findings are hallucinations (60–80% real by his estimate of the room), and nobody patches without re-testing. The asymmetry favors attackers — one unpatched hole is enough — so defenders must find everything.

Krishna maps this onto three lessons from warfare. Alliances can outweigh a single attacker, which is why sharing vulnerabilities and distributing patches centrally beats every company patching open-source flaws alone. Asymmetry decides wars of attrition: if the attacker is cheaper, the defender runs out of money and concedes, so the goal is to make defense equally cheap. And technological superiority tends to win outright, which is the real objective — find vulnerabilities, patch them, and raise the cost of getting through.

On agentically written code, Krishna says it is already superior to human-written code at IBM, where agents write roughly half the code, because agents also run evals and fix what they find before release. But full verification is "almost computationally infeasible" today, and the loop only finds flaws we know how to look for. Arora's accountability line is simple: a commercial enterprise taking economic value for a product must own what it ships — "you can't unleash and say oops I don't know how to control it."

Nobody is slowing down, and regulation is driven by politics

Calacanis asks whether the inevitability of an AI arms race changes the "stop AI" conversation, and relays a Washington conversation with a senior member of Congress who said one major CEO is "completely freaked out" cyber and floated a possible 50-state ban on data centers by Q4. Arora doubts 50 states ever agree on anything and says the industry has over-rotated onto AI's dark side at the expense of its benefits. Krishna is more direct: the data-center and AI regulation debate is "driven by populism, not by any reality," and a slowdown in one country only lets Japan, China, or India catch up. He thinks the current balance is roughly right but concedes an election can move it, and puts the odds of a 50-state data-center ban at essentially zero while allowing that 30 states could act, with the other 20 "rubbing their hands in glee."

On the haves-and-have-nots question, Arora says the binding constraint is compute supply, not price: the further a company sits from the US or China epicenter, the less compute it gets, regardless of intent. Krishna frames it as the recurring labor-versus-capital trade-off — capital-intensive economies have won each prior cycle — and argues that countries protecting every existing job at all costs will lose to those that accept displacement in exchange for better jobs.

Vertical integration lets the owner move profit between layers

Asked the conglomerate pendulum, Krishna uses IBM's mainframe, which has lasted 65 years, as the case study. In the 1960s IBM went "all the way to dirt" — sand to wafers to fabs to chips to software — which prevented IP leakage and kept the full stack's profit inside the company. Today IBM designs chips and most software but does not fab, because fab scale is a commodity game. The rule: integrate where advantage is decisive, buy where it is commodity.

Arora calls Nvidia the modern version, with Jensen Huang operating from the dirt to the application layer and funding the ecosystem (including Hugging Face) to keep participants on his stack. The strategic payoff of owning the stack is the ability to choose which layer to commoditize: if a rival attacks one layer, destroy the profit pool there and move profit elsewhere.

Applied to the application layer, Arora argues the open-versus-closed debate is really open stack versus closed stack. As models grow in memory and context, separating reasoning from context and memory stops making sense, so an "open router" that just passes reasoning tokens around becomes impractical. A personal agent trained on your context, APIs, and connectivity builds a moat because moving it to another stack means losing all of that. He expects every frontier lab to run a full-court press on consumer agents in the three months, following Meta's demonstration, and predicts many iPhone apps become headless.

SaaS faces consolidation as AI erodes terminal value

Arora's cyber M&A logic: customers should not have to stitch together 40 security products. He compares it to CRM, where 23 systems at Fidelity decades ago consolidated into off-the-shelf platforms like Salesforce, Microsoft Dynamics, and Oracle — once products converge, differentiation disappears and platforms win. Cyber valuations are elevated now partly on the AI security scare, which he expects to normalize even as customers spend heavily on defense.

Outside cyber, he is bearish on software terminal value: it is unclear which SaaS products survive the five to eight years as AI resets product-market fit, and "there's going to be huge disruption in the SaaS space over the few years." Krishna extends the point: enterprises run 500 to 5,000 software vendors, a massive attack surface, and AI capabilities make it possible to replace hundreds of them with five to ten main platforms — consolidation follows because independent distribution cannot be afforded on falling software value. He draws one line: interaction-rich software is easy to replace with AI, but systems that must be reliable to 10^-15, like a retail-banking database lookup or closing the books, will be hard to replace, and rebuilding them offers no cost advantage. On M&A timing, Krishna says multiples looked attractive early in the year and less so a month ago, and he is waiting to see.

The through-line across the episode is that capability is arriving faster than the institutions meant to absorb it. Quantum gives a hard deadline that most enterprises have not started planning for; open-weight models give buyers a 10x–40x cost escape hatch from frontier pricing; and AI-written code cuts both ways, helping defenders find flaws while handing attackers a cheaper path in. The unresolved tension is timing: Krishna's 2030 quantum estimate, his 75–90% open-weight prediction, and Arora's SaaS-disruption window are all forecasts from executives with positions in the outcome, and none will be settled until the milestones arrive. The concrete markers to track are third-party verification of error-correction claims by 2027, whether gate counts add a zero per year through 2029, and whether Western open-weight models close the six-to-12-month gap to Chinese ones.

Nvidia is described as the modern vertically integrated stack, funding ecosystem players to keep them on its stack and using control of the stack to shift profit between layers and destroy profit pools where competitors attack.

Nikesh argues cybersecurity should consolidate into a single platform so customers stop stitching dozens of point solutions, and warns that outside cyber the terminal value of SaaS is unclear, with product-market fit being disrupted by AI over the five to eight years.

Because every layer of enterprise IT runs on crypto keys, Nikesh argues companies should begin replacing infrastructure with quantum-safe systems now and wrap traffic in post-quantum-ready keys as a stopgap, rather than waiting for a quantum break date.

Arvind argues quantum algorithms can only deliver their advantage on quantum hardware, so enterprises should treat post-quantum cryptography as a known, already-written codebase that simply needs to be adopted.

If you control the vertical stack, you can decide which part of the stack to commoditize. So you can decide to move profit between layers.

If you control the vertical stack, you can decide which part of the stack to commoditize. So you can decide to move profit between layers.

We start saying open models, open weight. It's literally an open stack versus a closed stack.

We start saying open models, open weight. It's literally an open stack versus a closed stack.

The average enterprise uses anywhere from 500 to 5,000 distinct software vendors... could I replace 490 of them with minor AI capabilities and do I need to go back and leverage five to ten main platforms?

The average enterprise uses anywhere from 500 to 5,000 distinct software vendors... could I replace 490 of them with minor AI capabilities and do I need to go back and leverage five to ten main platforms?

You can't wait for 2029 or 2030 because it's not like one day all of us are going to go out and say, let's go burn $5 trillion of IT infrastructure and let's go build it again in one year because now we need quantum safe infrastructure.

You can't wait for 2029 or 2030 because it's not like one day all of us are going to go out and say, let's go burn $5 trillion of IT infrastructure and let's go build it again in one year because now we need quantum safe infrastructure.

You cannot simulate something on a different architecture and expect a more efficient result.

You cannot simulate something on a different architecture and expect a more efficient result.

I think it is going to be possible, but if it takes 10 years to break one key, most people are going to rest in peace. That's not a practical answer.

I think it is going to be possible, but if it takes 10 years to break one key, most people are going to rest in peace. That's not a practical answer.

Satya Nadella on the AI Doomer Slowdown, Microsoft’s Master Plan & Who Wins AI

The Real Reason AI Founders Are Calling For a Slowdown | Conversations

153. A Conversation with Zeng Ming on Historical Industry Perspectives: Harsh Truths, Companies That Will Die, Good ≠ Great, and Why OpenAI and Anthropic Likely Won't Be the Big Winners of the Native Era

Databricks CEO on What We’re Getting Wrong Existential Risk

Aligning Multi-Agent Systems and Financializing Compute

Turing CEO on Super Intelligence, Open Source, Agents

NVIDIA Crushes Quarter & Buys Hugging Face | Instinct Hits $2.5B Valuation, Cognition Raises at $46B

Pacing Gets Political + Andon Labs & an AI Religion?

Aravind Srinivas on the 1-Person Companies Making Millions With AI

Ep. 031 - EMERGENCY EPISODE: Are We Doomed? | Jordan Nanos, Doug O'Laughlin, Max Kan, Joey Brookhart

How the AI Boom Could Reindustrialize America

Andrew Feldman on Building Cerebras and the Future of Chips | Ep. 57