Ali Ghodsi (Databricks CEO) · 02:17 "We already have AGI. The problem isn't AI capability — it's that the organization's context hasn't reached the AI."
The guest is Ali Ghodsi, co-founder and CEO of Databricks. Iranian-Swedish, with a PhD in distributed systems from KTH Royal Institute of Technology. He joined UC Berkeley's AmpLab in 2009 (then the most active AI lab in the world, where Michael Jordan was based). He commercialized Apache Spark, which originated there, and founded Databricks in 2013. Over 13 years, he grew the lakehouse + AI platform to over $100 billion in valuation, with more than 20,000 customers. The host is Apoorv Agrawal — Stanford GSB instructor and partner at Altimeter Capital (an investor in Databricks).
The central thesis of this session is captured in a single line that directly contradicts the prevailing Silicon Valley consensus: "we already have AGI." Ali runs a show-of-hands poll with the audience. "Who thinks AGI is already here?" — 10%. "Who thinks most of the people you interact with are smarter than the latest AI model?" — almost everyone raises a hand. "Now again: who thinks AGI hasn't arrived yet?" — many hands again (laughter). "See, this is mass hypnosis. You answered with your own hand, but you move the goalposts."
Ali argues that today's frontier models comfortably exceed the AGI definition the research community used back in his 2009 AmpLab days. He has checked with people from that era, and they all say "by that definition, we passed it long ago." Yet the industry keeps pointing to edge cases — "it can't count the r's in 'strawberry'" — and insisting "so it isn't AGI." That, he argues, is what's being used to justify hundreds of billions of dollars of GPU / data center investment aimed at "superintelligence."
Key Observations
The real reason 95% of POCs fail isn't AI — it's "the John or Jane problem" (05:03)
The MIT Tech Report figure — "95% of AI POCs are failing" — is directionally right, Ali grants. But the cause isn't AI capability. Every organization has one person about whom everyone says, "ask John (or Jane), they know everything." A 10-, 15-, or 30-year veteran who holds all the context in their head. Lose that person and the company stops.
None of what's in John or Jane's head has been transferred to the model. So AI operates in a context-starved state and produces "stupid mistakes" at scale. "Even if AGI arrives, even if it solves hard math problems, it's useless inside organizations until we download that context into silicon." Ali's prescription: get carbon (humans) talking to silicon (AI). That, in his view, is the rewiring work organizations everywhere will spend the next decade on.
40 years from the dynamo — adoption lag the size of the electric motor (16:54)
Ali recommends a 1990 paper by Stanford professor Paul David, "The Dynamo and the Computer." The electric motor (the dynamo) was invented in 1880, but it didn't show up as a productivity gain in factories until 1920 — a 40-year lag. Early on, factory owners just swapped a steam engine for an electric motor, keeping the same dense, vertical factory layout (they simply rewired the steam-driven lineshaft to the electric motor).
The productivity explosion only happened once factory owners realized "electricity can be distributed" and moved factories out of cities into the suburbs, spreading them out horizontally and running multiple motors independently. The PC followed the same arc — in the 1990s, PCs were used as "electric typewriters," printing on paper, filing, handing to an assistant, all preserving the old workflow. That created the Solow paradox: "you see the computer age everywhere except in the productivity statistics." Today's AI, Ali argues, is at the same stage.
Inside Databricks: from 3 quarters to 1 quarter × 7x — but not because of AI (19:58)
Databricks used to take three quarters to build a connector to Salesforce / Workday / NetSuite. When LLMs got fast, Ali tried it himself: "I wrote one in two days." He asked the team, "why does this take nine months?" Two weeks later, the answer came back: "we've analyzed it, and we can shorten the cycle by about 1.5 months." A flat anticlimax.
He gave the same problem to another employee — someone who thinks from first principles. Their answer: "I can build seven of these in one quarter." What was different? Not the model — they rewrote the process from first principles. A Stanford-trained PM visiting each customer and writing a 60–80 page requirements doc (a full quarter of work) was replaced by a one-week rough draft, then AI-driven rapid rewriting. Building test environments for Salesforce and others (the part the team was bad at) got parallelized to external contractors. The "one person per connector" model became "seven people across seven connectors" as a group effort (eliminating the bus factor).
"GPT-7 or Opus 6 wasn't going to help here. What was needed was human refactoring and process change." This, he says, is a microcosm of the problem organizations everywhere are facing.
Pushing back on "software is dead" — Seven Powers as the surviving moats (07:16)
"Software is dead" is the meme of the moment, but Ali pushes back immediately. "If all software dies, then OpenAI and Anthropic should die too — they're software companies. NVIDIA too — they're smart people writing chip-design software. They should die. But they're the most valuable companies in the world. So software doesn't die."
That said, two large changes are real: barriers to entry have collapsed, and switching costs have collapsed. Writing software has gotten dramatically cheaper (though not zero). UI familiarity will fade too, once everyone talks through agents (users won't care which Salesforce / CRM is running under the hood). But the moats from Hamilton Helmer's 7 Powers — scale economies (AWS), brand (Ferrari / Rolex), counter-positioning (trust, patents, data exclusivity), switching costs, network economies, process power — still apply. Companies that haven't innovated in 10 years will be wiped out by the SaaS apocalypse; companies that do innovate can pivot and survive.
Value moves to the application layer — don't repeat the Multicast mistake (25:08)
Apoorv's question: "If you had $100 to invest across Jensen's 5-layer stack (energy / chips / infrastructure / models / apps), where would you put it?" Ali's answer: "Obviously the application layer. But I'm not an investor — this isn't financial advice (laughs)."
His evidence: a bitter personal experience. During his PhD in the 2000s, the best mathematical minds in networking were working on the "Multicast problem" — how to efficiently broadcast from one source to the entire world (e.g., live sports). Ali started a company. While he was working on it, fiber-optic capacity was being laid in huge quantities, bandwidth costs collapsed, and the multicast problem itself disappeared. A complete waste of time.
At the time, nobody serious paid attention to the "weird ideas" that turned out to be the real internet winners: the taxi business (Uber), selling books online (Amazon), renting your bedroom (Airbnb), short-form posts (Twitter). In 2000, saying any of these were the future would have sounded crazy. Today, the entire AI industry has tunnel vision on AGI / superintelligence from NVIDIA / OpenAI / Anthropic / DeepMind. But the real winners may be "left field" applications like education and healthcare, Ali predicts. Healthcare is 17% of U.S. GDP, and humans will pay infinitely for an AI company that can say "I've seen a million patients — given your genetic makeup, here are the risks." Education is the same — a sector VCs have traditionally avoided, but parents will pay seriously for their own child's education.
Frontier models converge to "Amazon book-selling" margins (33:01)
Moonshot (China) shipped Kimi 2.6 on Tuesday — "if this had been released in January, it would have been the best model in human history" — that level of performance. But now it gets surpassed by the frontier within months. The gap to open source has shrunk from 3–4 months to one month.
Ali's prediction: frontier model provision becomes a "token factory" business (centralized GPU data centers like AWS) and converges to a scale-economies game. "It'll settle into gross margins like Amazon.com's book-selling business. The number of providers will be tiny, and gross and operating margins will be small." Meanwhile, the AI model itself gets commoditized (both open-source pressure and scale economies push that way). Value shifts to the application layer — the same pattern IT history has repeated for decades.
Advice to students: take secular long bets, like Bezos (35:44)
"Don't rush, don't get whipsawed by fear-mongering" is Ali's message to students. Airbnb could have launched in 2001, but it took Brian Chesky nine years to get to the idea (a designer conference where they needed futons; that's where the spark came from). Good ideas are rare, humans are fundamentally bad at this, Ali says.
The model is Bezos: in his investment banking days, he saw the secular trend that "the internet keeps getting bigger," started from the most boring, undifferentiated category (books), and grew it into the "everything store." "Don't bet on what everyone is shouting about on Twitter right now (the multicast-style stuff). Bet on what you think is true over the long term."
Video Outline
- (00:00) State of the union — the current state of the AI industry
- (02:17) "We already have AGI" declaration + audience show-of-hands experiment
- (02:45) The 2009 AmpLab definition of AGI has already been surpassed
- (04:17) Why 95% of POCs fail — lack of context
- (05:49) The "John or Jane" problem — tacit organizational knowledge hasn't reached the AI
- (07:16) Pushing back on "software is dead"
- (08:54) Two changes — barriers to entry and switching costs both collapsing
- (10:53) Hamilton Helmer's 7 Powers — the surviving moats
- (13:11) Ethan Mollick's jagged frontier diagram / the reality across 20,000 customers
- (16:54) Dynamo to computer — 40 years of adoption lag from 1880 to 1920
- (19:58) Inside Databricks: how connector development went from 3Q to 1Q × 7x
- (24:21) Hamilton Helmer + Jensen's 5-layer stack investment allocation
- (25:16) The application layer wins — avoiding the Multicast trap
- (26:04) Ali's PhD-era failure — the Multicast startup
- (28:11) The real winners of the internet were Uber / Amazon / Airbnb / Twitter
- (28:42) Promising application-layer candidates — healthcare (17% of U.S. GDP) and education
- (32:15) Value keeps moving up the stack (IBM → Microsoft → VMware)
- (33:01) Moonshot Kimi 2.6 — open source catching up to the frontier
- (34:33) Frontier model provision converges to the token-factory business
- (34:51) Favorite AI product — Cursor (before Elon bought it)
- (35:44) Databricks' 10-year vision — riding the SaaS apocalypse
- (35:44) Advice to students — take secular long bets, like Bezos
Sources
Class #4 | MS&E 435: Economics of the AI Supercycle, Stanford University Spring '26, Apoorv Agrawal