All-In Podcast: Ramp Data Reportedly Backs Chamath's Claim That 1% of Companies Drive 90% of AI Spend, While Palo Alto's Arora Says Compute Scarcity Caps Enterprise Self-Build
Episode recorded October 22, 2026, with Palo Alto Networks CEO Nikesh Arora as guest, amid a PHP exploit scare and a sell-off in AI infrastructure names
AI spend is concentrated, circular and leaning on venture money
The most useful new data point on the episode came from Chamath Palihapitiya, who said his Social Capital research team had spoken with Ramp's chief economist, Ara Kharazian, to test a back-of-the-envelope estimate. Chamath's claim is that roughly 1% of companies account for 90% of AI spending, measured in tokens, and that about 80% of that spending comes from unprofitable, venture-backed companies. According to Chamath, Kharazian said the estimate was "pretty much correct." Ramp has not yet published the underlying chart, and the hosts said they were given a first look. Ramp sells cards and banking services mainly to startups, so its visibility into that cohort is strong, though its view of large enterprises is less certain.
Chamath framed the finding as a circularity problem. Venture funds raise money from pensions, foundations and family offices and invest it in startups. The startups spend it on growth, much of which means consuming other companies' AI. That spend shows up as revenue traction, which supports the next round from the same investors. "You can't be in a situation where on the one hand you're talking about breathtaking valuations, but where 75 cents of every dollar is actually just a startup sending it back and forth so that they can mark themselves up," he said. He distinguished this from early AWS, whose startup customers were real but whose parent, Amazon retail, was roughly free-cash-flow breakeven. Here, he argued, money-losing entities are passing money among themselves, and AI has to spread into the broader economy to become a predictable earnings driver. That is the substance of Citadel founder Ken Griffin's warning this week that AI is "almost certainly a bubble" and that companies are "looking at their bills from Anthropic and OpenAI and asking where the productivity is."
The second element of the Ramp chart bears on frontier-lab economics. Chamath said the share of top-spender budgets going to GPU compute, which he described as companies buying their own capacity to run open-source models, is "ticking up" relative to spend on closed frontier models accessed through services such as AWS Bedrock. His reading is that diffusion is slower than the market would like, and that where it is happening, buyers are choosing open weights. He cited his own 8090 business, saying large-company buying patterns changed after Palantir CEO Alex Karp's CNBC commentary on data sovereignty. He also relayed that the CIO of one of the largest US banks said it had cut spend on AWS and GCP and built six independent data centers from scratch. This is a single anecdote, and Chamath declined to name the bank.
David Friedberg's counter, which Chamath did not fully rebut, is that startup stampedes into a technology are historically a positive signal. He said that at his fund, which invests in about 100 companies a year after reviewing more than 10,000, startups migrate toward whatever is "the cheapest, best, fastest edge," as they did with cloud and mispriced Facebook ads. The tension between the two views is unresolved: concentration can be read as a validation signal or as a fragile base of demand.
Arora: enterprises prefer open weights, but the compute is not there to go it alone
Palo Alto Networks CEO Nikesh Arora gave a candid view of the enterprise buyer's position. He said Palo Alto prefers open-weight models because they can be modified and retrained, and because they allow inspection of the model's reasoning. He said the second point "most people don't talk about" matters most in regulated industries: "if I can't validate it, it's not going to be okay to go to your boss and say, 'I don't know what happened, but the IT guys put an AI model in there and it did it.' That's not explainable to the regulator in financial services and healthcare." Arora said enterprises will increasingly demand explainability, verifiability and auditability, which closed models cannot provide.
He was blunt about the constraint on acting on that preference. "It's very hard to find a global resilient highly available inference provider outside of the hyperscaler community right now," he said. A startup can rent 10 megawatts in one location and tolerate an outage, but a company providing 99.9% reliability cannot. Chamath asked whether Palo Alto might demand bare metal from hyperscalers or build its own facilities. Arora said that after closing all of Palo Alto's data centers when he joined, he is now "a taker of terms, not a dictator of terms," because compute is scarce. "You talk to Jensen, he can't give you chips till 2028," he said, adding that land, power and shell are equally hard to secure. Building independently is, in his words, "kind of an empty threat." He expects a hybrid in which Palo Alto takes bare metal from a hyperscaler or a "cousin of a hyperscaler," or stitches together compute from several sources.
Arora also offered a view on where economics are heading. He said the profit pool has so far sat with Nvidia and the memory industry but is shifting toward hyperscalers and others with large compute availability, because the constraint is moving from chips to deployable capacity. He described demand as "insatiable" and said that if every company in the world embraced AI today there would not be enough compute to serve them. That sits uneasily with the bubble argument, and Arora did not address the quality of the demand Chamath described.
The PHP exploits and an acceleration in AI-driven vulnerability discovery
The hosts described a cyber scare tied to two critical PHP vulnerabilities, nicknamed PHP 2 Shell, that were found by AI-assisted researchers in August and rated 9.2 out of 10 in severity by the PHP security team. PHP reportedly powers about 70% of websites. A coordinated patch was planned for October 22, but details leaked, and the release was moved up to October 21. The hosts said AWS reported a jump in hacking activity, Cloudflare reported millions of attempts, and governments across Europe and Wikipedia took systems offline to patch.
Asked whether such episodes will recur every 10 days, Arora said, "I think 10 days is too long." He attributed the severity less to new attack technique than to accumulated technical debt: customers running unpatchable legacy versions that were never upgraded because it "wasn't important." He said the PHP flaw was "solved by persistent compute," in that a problem identified years ago but considered too complex to coordinate was cracked once enough compute was applied. Once a patch is published, he noted, attackers reverse-engineer it to target unpatched systems. Palo Alto scanned its own infrastructure and issued virtual patches at the perimeter, a mitigation also offered by Cloudflare and other firewall vendors.
Arora rejected the thesis that defenders will simply out-compute attackers, at least for now. The infrastructure is not organized to launch defensive agent swarms, he said, citing a company with 5,000 websites that has never inventoried them. He said nearly every large enterprise is using Anthropic's Mythos or an equivalent to attack its own defenses, but each new model will find more holes, so patching alone is a cat-and-mouse game. His advice is to assume breach and invest in aggregating signals to detect anomalous behavior and respond quickly. He also dismissed the theory that frontier labs are gatekeeping defensive capability, saying they are rate-limited by compute and "just trying to go after profit pools," and that even he is rate-limited on the newest publicly available model.
On budgets, Arora said cyber typically consumes 4% to 10% of IT spend. Financial services and the federal government sit near 10%, while cloud-heavy companies sit closer to 4% to 5% because hyperscalers handle infrastructure security. He said that figure "is going to have to go up." He described an AI security market that barely exists, noting that companies are deploying agents without a way to find all the agents running in their environment. He pointed to a researcher finding that sharing a Muse conversation could expose private information. "We haven't built the AI security harnesses that we need," he said.
Agent access to bank accounts raises an unresolved liability question
A second issue emerged alongside the PHP scare: banks struggling to distinguish a customer from a personal agent logging in with the customer's credentials. Arora said the unauthorized access banks are seeing is probably tied to hackable websites, while agent traffic is a separate problem. He said banks cannot tell human and agent traffic apart, and that captcha is no longer a reliable barrier because AI can solve it. The main obstacle is liability, and secondarily the cost of support calls demanding explanations. He expects banks to prohibit agent access.
On who bears responsibility if an agent drains an account, Arora and Friedberg both argued the user ultimately does. "There's a word in agent. It's called agency," Arora said, adding that whoever grants it owns the liability. He carved out a "rogue agent" exception where the product fails to behave as advertised, which would be a product-liability matter for the platform. Friedberg noted that Amazon's blocking has already changed behavior: Grok's shopping agent will now fill a cart but not click buy.
The hosts also discussed Obliteration AI, a company they incubated that offers models without guardrails to red teams and corporations. Chamath called the category "unbelievably scary" and said it is underappreciated as a risk vector. Arora called it "the dark side of AI" and compared unconstrained models to handing out "a nuclear bomb." Jason Calacanis countered that access is restricted to vetted white-hat customers. The exchange did not resolve how such a service would be controlled.
Market sentiment: infrastructure names sold off
The hosts noted that AI infrastructure stocks fell on Thursday. CoreWeave was down about 10%, Meta fell 6% after earnings on capex concerns, and AMD and SpaceX each fell about 5%. They described the moves as part of a choppy but generally upward trend. Jason pointed to Jensen Huang's view that financing will be available for the buildout. The episode offered no new evidence on the capex payback question beyond the Ramp spending data.
France: yields at a 2002 high and a macro-political debate
The remainder of the episode covered riots in France that the hosts said had drawn more than half a million protesters, damaged over 3,000 businesses and injured more than 1,000 police officers, with copycat unrest in Italy, Spain and elsewhere. On the market side, France's 10-year yield passed 5% this week, the highest since 2002, while the US 10-year slipped just below 5%. The Wall Street Journal attributed that to flows from EU debt into Treasuries. Polymarket, an All-In partner, was cited as giving 91% odds of a Moody's or S&P downgrade of France by October 31 and 18% odds of a state of emergency by year-end. Chamath noted France has already proposed EUR 50 billion of austerity and said it will have to go higher.
Most of the discussion was political opinion rather than analysis. Chamath called the rioters "domestic terrorists" and said "bond vigilantes" should push French yields toward 8% to force reform. Friedberg argued that public spending near communist-era highs as a share of GDP shows state-provided services degrade outcomes, and he said the US faces the same fiscal arithmetic. He cited a US deficit of 6.2% of GDP against France's 5.4%, and growth of 2.3% against 0.5%. His prescriptions included winding down federal student loans, shifting defined-benefit retirement systems to defined-contribution accounts, and reforming housing and healthcare. Jason pushed back that France already provides low-cost education and healthcare, so the grievances are not simply unmet services. Several of the statistics cited were off the cuff, and the hosts disagreed with one another on immigration and on whether Americans would prefer European-style services.
The hosts also discussed Elon Musk's pledge to give every American child a share of SpaceX, which Jason valued at about $10 billion at a price of $150 to $170 per share, and argued over the political case for a federal minimum wage increase and President Trump's messaging ahead of the November 3 election. Jason proposed phasing in $1 a year to $15, while Friedberg and Chamath argued that wage mandates destroy jobs, citing California's fast-food wage increase and a claimed loss of 18,000 jobs. They agreed to disagree.