Chamath and Sacks on AI Bubble Fears, Cyber Risk, and France's Unrest

The episode opens with Ken Griffin's comments at the annual conference on the future of global markets this week. According to the host, Griffin said AI is almost certainly a bubble, while adding that major economic transformations usually come with one.

Griffin is optimistic about the long-term potential of AI, but he is concerned that companies are struggling to turn their AI spending into productivity. The host noted that companies are looking at their bills from Anthropic and OpenAI and asking where the productivity is.

Chamath described a discussion with Ara Khangaldyan, chief economist at Ramp. He said he had posted a rough calculation suggesting that about 1% of companies account for about 90% of AI spending, measured in tokens. He said about 80% of that 90% comes from loss-making, VC-backed startups.

Chamath described the cycle as follows: venture capital firms raise money from pension funds, limited partners, foundations and family offices, then invest it in startups. Part of that money is spent on using AI from other companies, which lets the startups show revenue growth and raise more money from the same investors.

According to Chamath, for AI to scale as a predictable engine of long-term profit, it has to move beyond the startup ecosystem. He said it is a structural problem if 75 cents of every dollar spent on AI comes from startups that are effectively moving money around to raise their own valuations.

Chamath compared this to the early days of cloud computing, when AWS was not tied to a profitable retail business in the same way. He said the current AI pattern is a structural concern rather than a settled fact about the market.

Chamath argued that AI has a diffusion problem because it is too concentrated in a small ecosystem. He said enterprises are increasingly building their own data centers or virtual private clouds and moving toward open-source models.

Nikesh Arora noted that compute is very scarce around the world and that building one's own data centers is difficult, citing power, land and chip delivery timelines. He said Palo Alto Networks runs only about 10 to 20 megawatts of inference and does not operate dozens of data centers worldwide, so a likely outcome is a hybrid approach.

Arora said the bottleneck is becoming less about chips and more about the availability of compute, because he sees demand as insatiable. He said, "If every company in the world today said, 'I'm in, I'm ready, I embrace AI, let's go as fast as possible,' there isn't enough compute in the world to satisfy us."

The conversation then turned to cyber risk. The hosts raised a recent discovery that AI found two major exploits affecting much of the internet. They pointed out that PHP, an open-source programming language, powers roughly 70% of all websites, including WordPress and Wikipedia.

The hosts discussed how traditional bank data access usually ran through APIs connected to verified apps, with intermediaries such as Plaid. Many personal AI agents, however, log in using the user's own credentials, so banks cannot easily tell whether the traffic comes from the user or from an agent.

Arora said the problem has two parts. One is unauthorized access that occurs because websites have been hacked, possibly through PHP vulnerabilities. The other is the agent problem, in which an agent holds a person's credentials and acts on their behalf.

Arora said the central issue is liability. He said that once an agent receives credentials and money disappears, the question becomes who bears the responsibility.

Arora added that banks may simply absorb small disputed charges because the cost of customer service exceeds the amounts involved, and that banks are trying to shut agents out with disclaimers about not being at fault.

Arora said companies now carry technical debt that creates security holes. He said nearly every large enterprise is now using a Metasploit-equivalent tool to break into its own defenses, find the gaps and close them.

He argued that the next generation of models will find more ways into infrastructure, so defenders will be playing a cat-and-mouse game requiring constant patching. He suggested a shift in mindset: assume you will be hacked and focus on how fast you can find and fix vulnerabilities. His advice was to aggregate all signals and data so that AI-driven defensive compute can look for anomalous behavior.

The hosts raised the theory that frontier models are causing much of the chaos by withholding the strongest cyber-defense tools, while open-source projects give white-hat security teams more capability.

Arora said that theory gives the labs too much credit. He said most people at those labs are researchers asking what to do with systems they find extremely powerful. He argued the labs are compute-constrained, which limits even public access to the latest models, and that their main concern is how to monetize the technology and grow revenue.

Friedberg said that if an AI agent acts outside its intended purpose, it becomes a product liability issue. He said that if a product does not behave the way it is supposed to, the platform is responsible, and that this is true for consumer products generally.

The hosts also discussed a company incubated by the All-In team that provides models without guardrails and is the responsibility of the customer who runs a red team. Chamath said he finds the idea of such models frightening and that the risk is currently underestimated.

The hosts argued that cybersecurity budgets often stay low until a major incident occurs. Chamath said that IT budgets typically devote about 4% to 10% to cyber, with financial services and the federal government at around 10% and cloud-heavy firms closer to 4%.

Arora suggested that the CISO could become the next CEO, because every business function now wants AI help. He said his CFO had not wanted AI at first, but now every function asks for AI staff.

The hosts were asked why Bernie Sanders and others have taken such a strong stance on stopping AI. Friedberg said it is a populist topic, because AI labs warn that half of office jobs will disappear and that many people will die, which makes it an effective issue for populists on the right or the left.

Friedberg argued that such positioning is an effective way to attract attention and votes.

Chamath called participants in the French unrest criminals and terrorists, and urged a harsh response. Friedberg disagreed, arguing that the protests began as legitimate and peaceful, including protests by high school students over grievances, before turning criminal.

Friedberg said the grievances are broad, ranging from schools and Palestine to healthcare, jobs and underfunded public services. He tied the unrest to immigrants who expected the state to provide for them, a promise he said had never been met.

Freeberg argued that a decentralized government has no accountability, no progress and no incentive structure, so the more it tries to take care of people, the more they are left behind. He said health care, housing and education are the three things government tries to deliver in the US, and that spending more on them has made them worse and less affordable.

He pointed to a RAND Group chart of government spending as a share of GDP, comparing France with the Soviet Union at its 1989 peak, communist Romania, Venezuela under Chavez and China. He said US public spending is about 36% of GDP, with federal spending around $7.4 trillion and state and local around $4.6 trillion on a GDP of about $33 trillion.

Chamath noted that in France the cost of university is about 300 euros a year, while the state spends about 12,000 euros per student. He argued that two of the four issues he had raised, free education and free healthcare, are largely solved in France, and a third, student debt, does not exist there, yet riots continue.

He concluded that the problem is not fully solved. In his view, the underlying issue is a feeling of oppression that weighs on the population.

A host noted that it takes weeks to get a doctor's appointment in France, school quality is poor, and some classrooms have 60 children. Chamath said there is a fundamental sense of oppression in the population, in which a higher class does better and people were promised they would do well too.

Chamath also cited a video of rioters attacking Chinese-owned shops in Paris, whom they described as having become too successful. He argued that the pressure of an enormous government and socialism leads people to treat those who appear to be ahead as the problem.

A speaker compared two possible outcomes: a social collapse with riots and murder, as seen in some countries, or an economic collapse like the ones in Argentina, Brazil and, more recently, Greece. The speaker argued that the US cannot sustain a deficit higher than France's at a similar debt-to-GDP ratio with only a modestly higher growth rate.

The speaker suggested the US will need some form of reset and proposed options such as ending federal student loans and cutting federal mortgage support, which could reduce middle-class wealth by 10% to 20%. The speaker also suggested that pensions could move from defined-benefit to defined-contribution plans, as in Australia and Norway.

Jason argued that the 'armchair do-gooder' supports open borders and caring for everyone, but refuses to take migrants into their own home, feed them or employ them. He said that is the same contradiction seen in Europe.

He described an activist in France whose shop was vandalized and who shouted that she was an activist who had stood up for everyone, yet the rioters did not care. He also criticized Kara Swisher's view on Musk's gift of shares to children, arguing that holding shares makes recipients part of capitalism.

A speaker argued that a country can only absorb a limited number of people who do not speak the mother tongue, because it becomes a huge burden that requires a second education and healthcare system. The speaker cited figures for Europe's North African immigrant population, saying about one in 19 people come from North Africa and about one in nine is an immigrant.

The speaker proposed admitting immigrants who can find work and support themselves, and removing welfare benefits for immigrants so that the right population self-selects.

A speaker noted that France has among the highest tax-to-GDP ratios, around 46%, meaning nearly half of each dollar of GDP is collected as tax. The speaker argued that the US should lower taxes and spending as a share of GDP to restore individual opportunity.

Another speaker countered that most Americans would not want to live in the countries ranked above the US, and noted that Brazil's youth unemployment has been around 15% to 20%.

Sacks said the federal minimum wage affects only about 10 to 15 million people, making an increase a small gesture. He suggested that California's attempt to raise the minimum wage to $20 an hour went too fast, and proposed raising it by $1 a year over seven years to reach $15.

Friedberg said that after California's SB 1228, fast-food jobs fell by 18,000, which he attributed to higher prices and lower demand. He argued that government intervention in markets breaks both supply and demand.

Sacks countered that a higher minimum wage increases spending by low-wage workers, which also benefits businesses. The two did not reach agreement and moved on.

Sacks said Trump's approval rating is at a historic low and argued that the administration talks extensively about data centers and crypto while saying little about the minimum wage, housing or inflation. He said failing to communicate with lower-income Americans is the biggest problem for the administration.

The hosts agreed that the administration has communicated poorly, though they disagreed on whether the president's economic record is strong.