The Frontier

Ken Griffin says AI is almost certainly in a bubble because companies still question whether AI bills produce productivity; he remains bullish long term. Thursday’s selloff hit data-center and AI-spending stocks, including CoreWeave, Meta, AMD, and SpaceX.

Chamath estimates that 1% of companies account for 90% of AI spending, with much of that concentrated spend cycling through unprofitable VC-backed startups. Ramp’s chief economist told Chamath his estimate was broadly correct, reinforcing concerns about circular revenue rather than broad enterprise demand.

Chamath sees AI diffusion beyond startups moving slowly, while top AI spenders are shifting toward buying GPUs and running open models. A bank executive told Chamath’s team the bank had built six independent data centers and cut AWS and Google Cloud Platform spending.

Nikesh Arora says enterprises favor open-weight models because they can modify weights and inspect reasoning for explainability, auditability, and regulatory review. He expects hybrid infrastructure: hyperscalers or their partners supply bare-metal compute while companies use open models, since building global capacity independently is difficult.

PHP vulnerabilities dubbed “PHP to Shell” triggered a rushed patch release after exploit details leaked. The language underpins roughly 70% of websites; the PHP security team rated the flaws 9.2 out of 10, and companies and governments took systems offline or issued warnings as attacks surged.

Arora says persistent compute helped solve the PHP vulnerability after researchers had considered it too difficult to coordinate around. Powerful models can find and chain vulnerabilities, but attackers can also reverse-engineer patches; defenders need to patch exposed systems and use perimeter products for virtual patches.

Arora separates hacked websites from agents logging into bank accounts with a user’s credentials. Banks struggle to distinguish a person from an agent, and the central liability question is whether the user, agent provider, or bank bears responsibility when an agent makes an unauthorized transaction.

Arora argues that cyber defense cannot yet out-compute attackers because company systems are fragmented and carry technical debt. He recommends aggregating security signals to detect anomalous behavior, accepting that breaches may occur, and responding faster; he says enterprise cyber spending commonly ranges from 4% to 10% of IT budgets.

Arora says the next AI models will expose new weaknesses, creating a continuing patching race; he rejects the claim that frontier labs are deliberately withholding defensive tools. He also expects AI security tools and governance to emerge as companies deploy agents faster than they build safeguards.

France’s protests spread to Italy, Spain, and Albania after beginning with high-school grievances, then widening into looting and unrest. The host reported damage to businesses, injured police, arrests, and the burning of a 143-year-old city hall; Chamath called the violence criminal and argued for a forceful response.

Chamath blamed France’s unrest on government failures and weak responses to violence; Jason emphasized immigration and failed integration, while Friedberg pointed to broad frustration over schools, jobs, healthcare, and government support. Speakers also cited reports of rioters targeting Chinese-owned businesses and described the attacks as driven by resentment toward successful immigrants.

Friedberg argued that France and the United States face similar fiscal pressure, citing high public spending, debt, and deficits; he called France a warning for Western countries. Chamath said bond-market pressure could force a French reset, while Jason warned that a disorderly reset could shift political power and erode rights.

Friedberg proposed winding down federal student loans, making existing student debt dischargeable, reducing housing-credit support, and shifting Social Security and other defined-benefit programs toward defined contributions over a decade. He argued that healthcare and housing rules also constrain prices, while Chamath supported bankruptcy relief for student borrowers.

Chamath and Friedberg argued that immigration policy should prioritize whether newcomers can support themselves and share civic principles; Jason stressed language and cultural integration. Friedberg said government support can weaken incentives, while Jason defended the Second Amendment as a way for business owners to protect themselves during riots.

Chamath said Elon Musk planned to give every American child a SpaceX share, which he framed as a major charitable transfer that could seed long-term investment accounts. The discussion also linked AI abundance to a possible way out of fiscal strain, while acknowledging that people facing immediate hardship may distrust that promise.

Jason argued that President Trump should speak more about housing, inflation, and lower-paid workers alongside data centers and technology; Chamath defended the administration’s AI investment strategy. Jason proposed gradual federal minimum-wage increases, while Friedberg argued tax cuts would help workers without distorting hiring or prices.

Arvind Krishna says quantum computers could break current cryptography around 2030, though the timeline remains uncertain. He expects error-correction claims soon, progress by 2027, and scaling milestones that could make quantum machines capable of surprising applications by 2029.

Arvind Krishna says post-quantum cryptography is already understood and implemented in code, but organizations have delayed adoption. Nikesh Arora warns that stolen encrypted data can be stored and cracked later, so companies should begin replacing vulnerable infrastructure now.

Arvind Krishna says AI could discover new quantum algorithms, but classical hardware cannot deliver their quantum speed advantage. Classical AI might still find new cryptanalysis methods, though Krishna says a method taking ten years to crack one key would not pose a practical threat.

Nikesh Arora says companies should let employees experiment with the strongest available AI models before optimizing cost. For routine cybersecurity work, he favors small, task-specific models; during an active breach, the business stakes can justify frontier models and agent swarms.

Nikesh Arora argues that open-weight models will win much enterprise usage because they cost less, keep model weights under company control, and support auditability. He estimates that open-weight models could handle 75% to 90% of token use, while frontier models remain valuable for specialized, high-stakes work.

Arvind Krishna expects the gap between leading Chinese open-weight models and Western alternatives to remain limited, while Nikesh Arora says companies will keep sensitive model reasoning and data in-house. Both distinguish self-hosted Chinese models from using those models through a cloud provider.

Nikesh Arora says AI vulnerability tools help attackers because they need only one exploitable flaw, while defenders must verify and patch every finding. He reports that these tools still hallucinate, so security teams retest results rather than trust model-generated vulnerability reports or patches.

Arvind Krishna says cyber defense depends on alliances that pool vulnerabilities and distribute patches, making defense cheaper and attackers' work harder. He also argues that defenders need technological superiority and lower costs to avoid losing an asymmetric contest.

Arvind Krishna says coding agents now write about half the code at his company and can improve quality by running evaluations before release. He cautions that agents only find flaws their tests can detect, while fully verifying arbitrary software remains computationally difficult.

Nikesh Arora says AI development will continue because countries will not coordinate to stop it, but companies remain responsible for software they release, including code produced by agents. Arvind Krishna says AI and data-center regulation reflects political pressure alongside concern for long-term economic advantage.

Nikesh Arora sees compute supply, rather than affordability alone, as the main risk for companies outside the United States and China. Arvind Krishna argues that countries willing to invest heavily in capital and accept job displacement will gain an advantage over countries focused on preserving every existing job.

Arvind Krishna credits IBM's vertically integrated mainframe with lasting 65 years, and says IBM once controlled production from raw materials through chips and software. He argues firms should own strategically important layers, while outsourcing commodities or operations that require uneconomic scale.

Nikesh Arora says AI products will increasingly combine models, reasoning, memory, context, and applications into end-to-end stacks. Persistent personal-agent context can create switching costs, while frontier labs are likely to compete for consumer control and push existing apps toward headless operation.

Nikesh Arora says cybersecurity companies are consolidating products so customers do not have to stitch together dozens of vendors. Arvind Krishna expects broader enterprise consolidation as AI replaces minor software functions and distribution costs favor a handful of platforms; both say the timing remains uncertain.