Key Moments
Arvind Krishna & Nikesh Arora: Quantum Threat, Open vs Frontier AI, AI Cyber Warfare & SaaS Collapse
All-In PodcastKey Moments
Quantum computers capable of breaking current encryption are expected by 2030, yet organizations are too "lazy" to adopt post-quantum cryptography, making them vulnerable now. Open-source AI models will likely dominate enterprise use due to cost and control, while frontier models remain niche.
Key Insights
Quantum computers are predicted to break today's cryptography around 2030, a reality that necessitates the adoption of post-quantum cryptography, which is already understood and has code written.
While toy quantum computers exist, the next evolutionary steps involve demonstrating error correction (by 2027) and scaling up qubit computations (adding a zero each year, potentially reaching 10 million by 2029).
AI models, even without quantum computers, could potentially break cryptography by discovering new number theory insights, though the time it takes to break a single key remains a critical question.
For most practical enterprise use cases, companies will likely train open-source or open-weight models due to significant cost savings (e.g., $1-2 per million tokens vs. $20-50), better auditability, and IP control.
AI models are already showing their capability in identifying code vulnerabilities, with estimates suggesting 60-80% of identified flaws are real and 20% are hallucinations, highlighting their utility for attackers who only need one vulnerability.
The future of enterprise software likely involves significant consolidation, moving away from hundreds of niche vendors towards 5-10 major platforms, a trend also seen in cybersecurity and driven by AI's ability to replace interaction-rich software functionalities.
The looming threat of quantum computing on encryption
The conversation begins with the significant threat posed by quantum computing to current encryption standards. Arvind Krishna explains that DARPA's Quantum Benchmarking Initiative has narrowed down participants, indicating progress in building functional quantum computers. The consensus is that by 2030, these machines will be capable of breaking today's cryptography. This isn't a distant theoretical problem; data encrypted today could be stolen and decrypted years later. While post-quantum cryptography is understood and the code is available, a lack of urgency means many organizations are unprepared. Krishna urges immediate replacement of IT stacks with quantum-proof infrastructure, suggesting that current solutions can already offer some protection as a stopgap. He estimates that half of current IT spending might be on non-quantum-safe infrastructure, indicating a massive, urgent undertaking.
Milestones in quantum computer development
Arvind Krishna outlines the evolutionary path for quantum computers. The first critical milestone is the demonstration of error correction or mitigation on a running quantum computer, which he predicts will be claimed by one or two entities by 2027, though not at full scale. The second milestone is achieving scale, measured by 'gate computations.' Current systems are around 10,000 gates, and the progress is measured by adding a zero each year. If this trend continues, systems could reach 10 million gate computations by 2029, enabling capabilities that would surprise many. The concept of 'logical qubits,' which are error-corrected operating qubits, is introduced as a way to understand the effective capacity of these machines, acknowledging that error correction reduces physical qubit capacity but increases reliability.
AI as a potential crypto-breaker
A counterpoint is raised by a senior AI researcher's view that AI models could break cryptography before quantum computers do, leveraging advances in number theory to discover proofs and algorithms not apparent to humans. While Arvind Krishna acknowledges that AI could discover new quantum algorithms, he stresses that simulating them on classical hardware would not yield efficiency gains, potentially being orders of magnitude slower. The efficiency advantage of quantum algorithms can only be realized on quantum hardware. Nikesh Arora agrees that classical AI could potentially break cryptography by predicting prime number distributions, but the key question is the time it would take. If breaking a single key takes years, it's not a practical threat for most, but this remains an open debate until proven.
The spectrum of AI models: Open vs. Frontier
The discussion shifts to the practical use of AI models in enterprises, distinguishing between open-source/open-weight and frontier (closed) models. Nikesh Arora advises against cost optimization during the experimentation phase, encouraging teams to use the best, most accessible models. He highlights that for many enterprise tasks, a smaller, task-specific, compressed model is more cost-effective than using large, frontier models. For instance, endpoint security relies on machine learning on the endpoint, which is far cheaper than sending all traffic to a large language model for analysis. He posits that frontier LLMs will be reserved for unique, high-value use cases where the business impact of failure is immense, such as drug discovery or breach defense. For day-to-day operations, training open-weight models with proprietary data will likely be the norm for auditability and explainability.
Economic and strategic drivers for open-weight AI
Chamath Palihapitiya argues that economics will heavily favor open-weight models. He notes that with a week of engineering, the cost per million tokens can be reduced from $20-$50 for frontier models to $1-$2 for optimized open-weight models, a 10-40x improvement. This cost difference is compelling for enterprises. Brand and IP protection also drive the need for on-premise or controlled model execution, which is only possible with open-weight models. Geopolitically, while current leading open-weight models are Chinese, there's an expectation that Western equivalents will emerge within 6-12 months. These factors—economics, IP, and geopolitics—lead to the conclusion that 75-90% of token usage could shift to open-weight models. However, David Sacks notes that developer laziness might default to more expensive, easier-to-access frontier models unless controls are in place.
AI and cyber warfare: Vulnerabilities and defense strategies
The conversation touches upon AI's role in cyber warfare and defense. Models are increasingly capable of identifying code vulnerabilities, with estimates suggesting 60-80% of identified flaws are real, while 20% are hallucinations. This makes AI a powerful tool for attackers, who only need to find one vulnerability, whereas defenders must find and patch all of them. The strategy for defense involves collaboration, pooling identified vulnerabilities to develop common defenses, similar to how AI models are updated with new attack techniques to protect all customers. From a warfare perspective, the principles of powerful alliances, asymmetry, and technological superiority are discussed. The goal is to achieve technological superiority in cybersecurity, making it prohibitively expensive for adversaries to find ways through defenses. While agentic code can be superior for tasks like code generation and initial flaw detection, full verification remains computationally infeasible.
The inevitability of an AI arms race and regulation
There's a consensus that an international AI development race is inevitable, with no country likely to voluntarily halt progress. The analogy of airplane development is used: security measures like TSA came after the initial invention. The focus has overly shifted to AI's perils rather than its benefits. The responsibility for AI safety lies with the creators and deployers; the argument of not being able to control the software they built will not hold. Regulation is seen as driven more by populism and political advantage than by immediate, practical realities, though long-term economic advantage is also a consideration. A potential ban on data centers by some US states is discussed, but a complete slowdown of AI development globally is deemed impossible, as nations would use any pause to gain an advantage.
Vertical integration, SaaS collapse, and the future of enterprise software
The discussion concludes with the concept of vertical integration and its implications for the SaaS market. Arvind Krishna draws a parallel to IBM's historically vertically integrated mainframe business, which controlled the entire stack from chip design to software, thereby retaining profit and preventing IP leakage. While IBM no longer controls fabs, it maintains control over chip design, engineering, and much of the software. This model is contrasted with NVIDIA's current strategy, which controls key parts of the AI ecosystem to maintain leverage. The prevailing trend is towards consolidation in enterprise software. Companies currently use hundreds of distinct vendors, creating a massive attack surface and complexity. AI's ability to automate many tasks and offer interaction-rich functionalities is expected to disrupt many current software products. This disruption, coupled with the need for cost efficiency and lower distribution costs, suggests a future with far fewer, larger software platforms, possibly 5-10 key players, rather than the current fragmented market. While highly reliable, low-level functions like database lookups may be harder for AI to fully replace, the majority of interactive software is ripe for disruption and consolidation.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
The general consensus is that quantum computers capable of breaking today's cryptography could emerge around 2030, with debates centering on specific years like 2028, 2029, or 2033.
Topics
Mentioned in this video
DARPA is running a competition called QBI (Quantum Benchmarking Initiative) to identify companies capable of building quantum computers.
The regulatory approach of the EU in AI is mentioned as an example of imposing regulations out of fear of possibilities, potentially hindering a vibrant tech economy.
The Transportation Security Administration is used as an analogy for how security measures (like those for planes) often develop after the initial technology is created.
Mentioned as an example of a consolidated CRM platform.
OpenAI's recent math dump is mentioned as a potential early indication of AI developing proofs not apparent to humans, and their models are discussed in the context of token usage and cost.
Mentioned in the context of Jensen Huang's efforts to fund the ecosystem around NVIDIA's vertically integrated stack.
NVIDIA is presented as the modern-day equivalent of a vertically integrated stack, with Jensen Huang actively funding the ecosystem to maintain control.
Mentioned as an example from Nikesh Arora's early career where numerous disparate systems (23) existed, which have since consolidated into singular platforms like CRM.
Mentioned as a potential company pursuing vertical integration strategies at the application layer.
Mentioned as an example of a consolidated CRM platform, highlighting the cost associated with such solutions.
The IBM mainframe is used as a prime example of successful vertical integration, having lasted 65 years due to controlling the entire stack from chip design to software.
His work is referenced in the context of simulating complex systems, stating that one cannot simulate something on a different architecture and expect a more efficient result.
He is cited as an example of a CEO aggressively implementing zero data retention (ZDR) for customer privacy.
Mentioned in the analogy that security measures (like TSA) were not considered by the Wright brothers when they first invented planes.
He is described as the modern-day version of a vertically integrated stack, funding the ecosystem to ensure it doesn't stray from NVIDIA's integrated stack.
Discussed in the context of geopolitical competition in AI development, with its open-weight models being currently leading, and potential future access to compute.
Mentioned in the context of the AI arms race and competition with other global powers like China.
Referenced as an example where technological superiority was a key factor in winning the war.
Mentioned as a country that would not slow down AI development if others did, highlighting the global race.
Mentioned as a country that would not slow down AI development if others did, highlighting the global race.
Referenced as an example where technological superiority was a key factor in winning the war.
Mentioned as a tool used by individuals who do not have concerns about data training models, contrasting with ZDR practices.
Mythos is mentioned as a potential AI solution for cybersecurity that was initially thought to solve everything, but practical applications are discussed. It's also referenced in the context of AI models discovering vulnerabilities.
Mentioned as an example of a consolidated CRM platform.
Mentioned as a tool that can identify vulnerabilities in code but requires re-testing before use for patching.
More from All-In Podcast
View all 467 summaries
41 minDylan Field: Is AI Replacing Designers?, Design is the New Code & Democratizing Creativity
95 minCoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer
42 minElizabeth Holmes' Husband Billy Evans: The Trial, Prison & Nathan Fielder’s Documentary
26 minMartin Shkreli: Big Pharma's Blind Spot, Betting Against Quantum & What Prison Taught Him
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free