AI Demand Is Real, but Productivity Gains Remain Unproven

The AI market’s central tension was visible in Broadcom’s selloff: executives described infrastructure demand as supply-constrained, while public-market pricing left almost no room for disappointment. Caroline Hyde framed the Nasdaq 100’s decline as its worst day since mid-May, with pressure concentrated in chips after a run-up that had left investors anxious that markets had moved “too far, too fast.” Broadcom became the clearest example: Bloomberg’s on-screen chart showed its shares down roughly 15% intraday after the company forecast $16 billion of current-period sales in its custom silicon group, below the $17.2 billion Wall Street had expected.

Ed Ludlow emphasized that the miss was relative to very high expectations, not evidence that demand had disappeared. Broadcom had added $270 billion in market capitalization before the earnings report, he said, and the stock was “priced for perfection.” Hyde later put the market move in starker terms, citing Bloomberg’s “big number” graphic: about $300 billion of Broadcom market value had been erased after the earnings release, with the stock suffering its worst day since January 2025.

The case for continued demand rested on the argument that AI infrastructure remains supply-constrained. Ludlow said Broadcom CEO Hock Tan was expected to address the issue later at the event and that investors would be looking for nuance on the company’s outlook. He also said research he had read that morning pointed to greater visibility into 2028 than Broadcom had described three months earlier. In Ludlow’s reading, the market was punishing the stock for a near-term forecast, while the longer-term signal still depended on demand visibility.

That capacity-shortage argument was expressed most directly by Cerebras CEO Andrew Feldman in a clip from the prior day. Feldman rejected the premise that AI infrastructure demand looked like a bubble. He said builders were far behind customer demand and that the issue was not too much supply chasing speculative use cases, but too little hardware capacity for existing appetite.

Broadcom’s market reaction showed how little room there was for disappointment even when the underlying sales figures remained large. The same executives and investors describing AI demand as constrained by physical capacity were also dealing with stocks and private-company valuations that required continued evidence of that demand.

Tom Giles widened the frame to the private AI companies downstream of the infrastructure buildout. Ludlow said Anthropic had filed confidentially for an IPO that Monday. He also said SpaceX had filed an amended S-1A, priced an IPO at $135 a share, was seeking to raise $75 billion, and was approaching an almost $1.8 trillion valuation. Those were presented on air as Ludlow’s reporting context for a market in which the event included investors with exposure across Anthropic, OpenAI, and SpaceX.

Giles described the AI lab competition as a “horse race” in which OpenAI had long been seen as the company to beat, but Anthropic had become ascendant over roughly the prior six months. He pointed to Anthropic’s cybersecurity tool Mythos and its message that it had captured the business market in a way OpenAI had not. But he cautioned against assuming the lead was fixed: OpenAI could release a new version of ChatGPT and quickly regain momentum. The race, in his account, could produce two potentially back-to-back trillion-dollar IPOs.

Ali Ghodsi offered the most explicit challenge to the prevailing frontier-model narrative. He said he believes artificial general intelligence has already arrived. His evidence was informal but repeated: when he asks audiences whether AGI exists, roughly 10% raise their hands; when he asks whether the frontier AIs they use are smarter than most people they work with most of the time, roughly 90% do. He said he has run that exercise 20 or 30 times.

For Ghodsi, the implication is that the next major bottleneck is not raw model intelligence. It is context. The models are already “plenty smart,” he said, but they lack access to the conversations, processes, business knowledge, and data needed to operate productively inside organizations. That is where he placed Databricks’ role.

Ghodsi connected that thesis to Databricks’ product strategy. He said the company’s focus is on infusing data context into Genie, a product that can answer questions and automate work. The point was not simply that companies own large datasets, but that agents become useful only when connected to the right data in the right operational context.

He also pushed back against the idea that software commoditization makes databases less important. If AI can produce software, he said, that does not eliminate the need for databases; every piece of software uses one. Databricks’ Lakehouse, in his account, is tailored to agents because it gives them access to enterprise data such as KPIs. He cited product groups putting KPIs into Lakehouse so agents can answer questions against those metrics.

Hyde pressed on whether “everyone” is actually using agents and whether productivity is real. Ghodsi’s answer was qualified. People are using agents, but not near their full potential; he suggested perhaps only 1% of the opportunity has been tapped because autonomous agents are not yet broadly collaborating with humans and one another. Again, the constraint was context.

His concrete example was Novo Nordisk. Ghodsi said the company uses Genie to feed clinical-trial information into AI systems, allowing users to ask questions such as how an obesity study is progressing. According to him, the time required to understand those studies fell from weeks to a few minutes. He treated that as an example of agentic work beginning to happen, while still being at an early stage.

The software-volume argument came through when Ludlow asked about Jensen Huang’s view that agents would become the ultimate users of software, creating more demand rather than destroying the category. Ghodsi said Huang was “spot on” and added Databricks’ own calculation: within nine months — or, more conservatively, within one or two years — more software will be written than in all prior human history. However that estimate is derived, Ghodsi said, it checks out when backed into from the growth of AI-assisted development.

That explosion in software, in his view, increases demand for the surrounding ecosystem: databases, data infrastructure, and other layers around applications. Databricks, as a Lakehouse provider, would benefit from that expansion.

On capital markets, Ghodsi separated Databricks from the companies rushing toward IPO windows. He said private capital remains available because private investors still need companies in which to deploy money. Databricks will eventually be public, he said, but not because it lacks access to private funding. The more immediate reason is employee liquidity: at Databricks’ scale, with about 10,000 current employees and perhaps 14,000 including former employees, public markets create a transaction mechanism.

He also called the current year a “terrible” one to go public because “there’s so much happening.” Asked whether he studied Rubrik’s S-1 or would track Anthropic’s IPO as a template, Ghodsi said Databricks is not trying to time the market but to win the market over the long run. He had not read the S-1 filing word by word.

Ghodsi’s Databricks argument and Todd McKinnon’s Okta argument met at the same operational bottleneck. Both described AI agents as capable enough to matter but constrained by their access to enterprise systems. Ghodsi emphasized the missing context: the data, processes, conversations, and metrics that make a model useful inside a company. McKinnon emphasized the control layer: the identity, permissions, and shutdown mechanisms required when agents begin touching real systems.

Okta had beaten first-quarter expectations and raised full-year guidance. Bloomberg charts showed its stock up more than 30% over five days, which served as the market backdrop for McKinnon’s argument that software companies can benefit from the agentic shift rather than be destroyed by it. He said Okta was founded during the move to cloud computing, but the shift to AI and AI agents is bigger than anything since the internet, “possibly bigger.”

His argument was that identity becomes more important as applications, services, and agents proliferate. Okta’s new product cycle is “Okta for AI Agents,” which McKinnon described as the identity layer and connectivity layer for agents in the enterprise.

The emphasis on connectivity paralleled Ghodsi’s context argument, but McKinnon approached it through security and control. Agents need connections to databases, data sources, and applications to be useful. Those same connections create risk: information leakage, agents behaving unexpectedly, and the need to shut them down quickly. Okta’s role, as he described it, is to provide a blueprint for secure agent connectivity.

Ludlow asked about Jensen Huang’s argument that the software-doom narrative is “complete nonsense” because agents will become the end users of legacy software and there will be more agents than people. McKinnon did not simply endorse a stock-market reaction. He compared the AI transition with the shift to cloud, where every layer of the stack was reimagined: hardware, infrastructure, platforms, developer tools, and applications. But he said this transition adds something new: a layer of “digital work” on top of the stack.

That layer creates a two-part opportunity for Okta. First, identity must be reinvented and reprioritized as the existing software stack changes. Second, automated digital work becomes a new category in which Okta for AI Agents can operate.

On labor, McKinnon rejected the simple version of AI-driven headcount reduction. Hyde asked whether Okta would need as many employees in the new era. McKinnon said employees must become more productive using tools such as Claude Code, marketing tools, and support tools. But when asked directly whether he would keep the same number of people, he said he disagreed with many observers: in five years, Okta will have far more software engineers than it has now.

The reason was not that engineers would be less efficient. It was that the amount of software to build would expand. Automation increases capability, and capability increases demand. This was the same structural argument Ghodsi made from the data-infrastructure side, but McKinnon’s version put the emphasis on permissioning: AI-written software and AI agents require controlled access to systems of record if they are going to do useful work.

McKinnon’s practical example was a flight cancellation. Ludlow described a world in which an agent can access his credit card, flight information, and internal travel systems, then resolve the problem without a human phone call. McKinnon said those are exactly the connections at issue. The barrier is not that the models lack sufficient raw capability. In his view, models now surpass companies’ ability to connect systems together and build practical agentic applications.

Security remains unavoidable because sensitive data is being connected and automated in new ways. McKinnon argued that customers are responding to Okta’s practical framing: they want more automation and more capable agents, but they need a secure and manageable way to grant access.

That led to his dismissal of the “SaaS-pocalypse” thesis. He called it “completely overblown.” Some SaaS companies will fail to embed AI capabilities and will become legacy vendors, just as some companies were left behind in the move to cloud. But many others, he said, will make the transition and thrive.

Apoorv Agrawal described AI’s shift in financial terms: it had moved from a technology product cycle to “one of the largest capital formation cycles.” The clearest split, in his view, is between companies receiving AI capital expenditure and those spending it.

The receivers are the “picks and shovels” providers: compute, memory, optics, networking, energy, and related infrastructure. The spenders are the labs and platform companies trying to scale AI capabilities. Ludlow introduced the discussion by referring to Alphabet having “come out and upsized an $84 billion, now equity offering.” Agrawal then said Google had “just raised $84 billion,” and separately cited OpenAI’s $122 billion raise, Anthropic’s financing activity, Nvidia’s $80 billion stock buyback, and SK Hynix retiring $8 billion of stock with more dividends coming. In the article’s terms, these remain claims made in the exchange, used by Agrawal to describe the scale and asymmetry of AI capital formation.

Bloomberg showed a chart, attributed on-screen to Bloomberg consensus estimates and reported financials, titled “Hyperscalers’ AI capex surge, free cash flow fades.” It compared Amazon, Alphabet, Microsoft, Meta, and Oracle from 2022 through 2027 estimates, with stacked bars rising toward the later years and a total free-cash-flow line weakening. The visual reinforced Agrawal’s divide: AI infrastructure spending is not only a product story, but a cash-flow and capital-allocation story for the largest technology companies.

That distinction also framed the concern about whether the world can support multiple large AI labs and model companies. Hyde asked whether there was room for all of the large language models, given the persistent narrative that models will be commoditized. Agrawal rejected a binary classification of companies as consumer AI, enterprise AI, or space AI. The reality, he said, is more multi-faceted.

OpenAI, in his description, is not only a consumer product company. It has just under a billion ChatGPT users, 5 million Codex users in enterprise, and additional bets in hardware, robotics, and frontier enterprise offerings. SpaceX, meanwhile, sits across “the two largest markets on the planet, space and AI,” and is now selling compute. When Ludlow characterized SpaceX as stealthily becoming a hyperscaler, Agrawal called the approach “Switzerland”: sell compute, sell models, build coding and consumer products, and maintain optionality across the stack.

Bloomberg also showed a graphic describing a “2026 IPO boom,” with reported estimates for SpaceX, Anthropic, and OpenAI. The on-screen table listed SpaceX at $75 billion with a June date; Anthropic at $60 billion and fourth quarter 2026; and OpenAI with the amount listed as TBA, also fourth quarter 2026. The source line attributed the estimates to Bloomberg News and The Information.

Asked whether there would be sufficient public-market demand for a wall of offerings from companies such as Anthropic, OpenAI, and SpaceX, Agrawal said investors had long been looking for ways to get access to the AI supercycle and that “that moment has arrived.” He characterized IPO markets as cyclical, with the last major cycle four to five years earlier in 2021. His claim about demand sat across two objects: demand from investors for access to these companies as they come to market, and demand for the AI capabilities the companies are building.

He also distinguished proven and unproven capability. In coding models, he said, the market crossed a threshold in December of the previous year, including with Opus 4.5. Consumer demand is already widely experienced. Still to be proven are robotics, manufacturing, OpenAI’s hardware device, and other adjacent opportunities.

Agrawal’s “long/short” teaching exercise at Stanford clarified where investors and founders are concentrating their optimism and caution. More than half of the speakers in his class were long the picks-and-shovels providers. Compute was the most common long theme; energy was second, because energy becomes the bottleneck at scale. On the short side, the most common theme was incumbents that are not innovating.

His own bearish filter was not a sector label but a relationship to scaling laws. If a business is harmed by the empirical pattern that more compute, resources, data, and algorithms drive intelligence higher, he said, it is a difficult place to be. Altimeter believes intelligence is going to rise substantially, so businesses that suffer as intelligence improves are structurally exposed.

Mary Daly separated enthusiasm from measured economic impact. Asked where she goes for evidence on whether AI is improving productivity, Daly said she goes to businesses actually using the technology, not enthusiasts or doomsayers. Over the prior year, she said, she had seen interest turn into investment: firms training workforces to be AI-ready and asking how AI can affect front-office operations, not just back-office tasks.

The activity is broad. Daly said she sees it in small, medium, and large companies; global and regional firms; agriculture, machining, manufacturing, and services. But she was explicit that widespread productivity gains have not yet appeared. The return on investment is still developing.

Ludlow pressed the point, noting that he had previously asked her repeatedly to show the productivity gains. Daly said productivity growth has been outside historical norms, which is positive for the U.S. economy, but she would not simply attribute it to AI. It is possible, she said, that businesses are using LLM assistants to do more with fewer workers. But businesses are not yet reporting transformative ongoing productivity gains.

The important word, for Daly, was “yet.” When she asks companies for a timeframe, they point to next year or the year after. The reason is that productivity gains require business-process transformation, not simply adopting a model or agent. Companies must rework how work is done around possibilities they may not yet fully understand.

Daly’s bullishness came less from technology companies’ own investment plans and more from adoption by companies outside the technology core. She described a machine-making business scanning 50 years of plans, then using those plans and a model to generate faster, better, cheaper product ideas. She also described touring a robotics company building tools that help manufacturers improve shipping and distribution. These examples, she said, have the capacity to change the economy because they show AI being applied beyond software companies and AI labs.

On the question of a market bubble, Daly declined to equate AI with the dot-com period even though she sees some similarities in San Francisco: productivity growth, enthusiasm, and pressure on housing. The dot-com boom was different, she said, because AI is already being put into businesses and is more pervasive. She did not jump from 1990s echoes to a conclusion that this must end like the 1990s.

Financial stability was another area where Daly distinguished market exuberance from systemic risk. Hyde asked whether high valuations and upcoming public listings could themselves become a financial stability issue. Daly said rising markets are not enough. A financial stability issue would mean stress spreading to banks, consumers, or businesses. She said she was not seeing evidence of that, while adding that the Fed watches it carefully.

Daly treated AI infrastructure as a timing problem for inflation. Ed Ludlow asked whether massive data-center capex, bottlenecks in areas such as memory, and utility-scale power demand are inflationary or disinflationary. Daly’s answer was that both can be true at different points.

In the beginning, large construction projects and electricity demand create competition for limited services. Areas supplying those services may see price pressure. But the infrastructure being built — data centers, and potentially new electricity generation supported by large tech buyers — can eventually help prices. “But not here and now,” Daly said. Policymaking, in her view, requires looking at current prices, forecasts, and evidence rather than assuming the eventual outcome.

At that moment, she said, she was focused more on energy and food prices. Oil prices and fertilizer prices had pushed into food inflation, connected to uncertainty around a war in Iran. Futures markets suggested oil around $80 a barrel by year-end, but Daly emphasized uncertainty. The Fed’s policy stance, she said, was in a good place and prepared to respond either way.

That uncertainty made her resistant to forward guidance on rate cuts in 2026. The economy could develop with inflation risks persisting, or the war could end, oil prices could fall, and underlying dynamics — including possible positives from AI — could reassert themselves. Trying to resolve the uncertainty immediately, she said, would be a mistake because it could close policymakers’ minds to incoming evidence.

Her labor-market comments tied AI directly to hiring caution. Daly said the labor market had stabilized relative to the end of the prior year, when she had been among policymakers worried about it and supportive of cuts to stabilize conditions. Businesses now feel more cautiously optimistic, which could feed into hiring, but they are not rushing. They are interrogating how much AI can do before they hire.

The reason is partly skills uncertainty and partly risk management. Businesses do not want to hire many people, discover that AI can handle some of the work, and then need a different skill mix. They also want to avoid over-hiring followed by layoffs, which are painful for workers and companies. Daly expects that caution to remain for a while.

In San Francisco, she acknowledged echoes of the late 1990s. She moved to the city in 1996 and remembered the affordability pressure of the dot-com era. A new wave of IPOs and employee wealth could pressure housing and labor costs again. But she framed housing inflation primarily as a supply problem. The region wants investment, growth, and employment, but limited housing supply means increased demand pushes prices up. Those questions, she said, fall to federal and local policymakers outside the Fed, including the mayor.

Daly also addressed AI inside the Federal Reserve, but with the same caution she applied to businesses. The Fed wants to adopt new technology to work more efficiently, she said, while preserving public trust and public funds. People must know they can access their money, that banks are well supervised, and that monetary policy is made by people using judgment — not by machines.

Her example of modernization was check processing. When she joined the Fed in the 1990s, check processing existed everywhere the Fed had locations. As check demand declined, those activities were consolidated into fewer sites. She used that history to describe the continuing obligation to modernize and improve efficiency. Asked about Chairman Warsh, she said he had just joined and should have time to announce his own agenda, but she had heard him emphasize the same compass she had seen from prior chairs: doing the Fed’s best work for the American people.

On credit and data-center financing, Daly placed the issue below inflation in her hierarchy of immediate risks. The Fed is watching carefully, and she noted that many companies are investing their own resources. But if she had to stack-rank current concerns, getting inflation back to target and giving Americans relief remained her top priority.