AGI Is Already Here — But Your Org Chart Isn't Ready for It
Ali Ghodsi opens with a contested but confident claim: AGI has already arrived by the definitions researchers were using in 2009. Yet he immediately undercuts any triumphalism by noting that 95% of enterprise POCs are failing and that even AI-native companies are still hiring salespeople the old-fashioned way. The gap between what the technology can do and what organizations are actually doing with it is the central problem he returns to throughout.
His core framework is that the blocker is organizational context, not model capability. Models and agents don't know what Jane learned in 30 years at a company, and without that knowledge they make costly mistakes. He illustrates this with a detailed case study from inside Databricks: a production connector that used to take nine months to ship barely moved to 7.5 months when engineers simply applied AI to their existing workflow. Only after management intervened to redesign the entire process — parallel teams, outsourced test instances, compressed requirements cycles — did seven connectors ship in a single quarter. His pointed conclusion: a better model (GPT-7, whatever comes next) would not have helped; the bottleneck was human process.
On competitive dynamics, Ghodsi offers a nuanced take on software moats. Barriers to entry and switching costs are collapsing — he calls it a potential "SaaS apocalypse" — but incumbents can survive through economies of scale, brand, trust certifications, patents, and proprietary data. He flags a clear warning sign: any software company whose product looks identical to ten years ago but whose revenue kept climbing should be worried. He also uses the electric dynamo analogy (1880–1920, a 40-year lag before productivity showed up in statistics) to caution against expecting AI's economic impact to appear quickly.
On where to invest, Ghodsi is unambiguous: applications will win, not infrastructure or frontier models. He sees the model layer becoming a thin-margin, economies-of-scale commodity game, with open-source closing the gap to roughly one month behind frontier proprietary models. He singles out healthcare (17% of US GDP) and — more speculatively — AI-powered education as sectors capable of producing trillion-dollar companies, provided someone builds a genuine data moat and demonstrates measurable outcomes.
His closing advice is deliberately anti-hype: take a long-term view, resist tunnel vision on whatever problem feels loudest today (he uses 2000-era multicast research as a cautionary tale), and look for the secular trend the way Bezos looked at internet retail. Good ideas are rare and hard to see; the 22-year-old interns panicking that a six-month delay will cost them the AI era are, in his view, asking exactly the wrong question.
Databricks CEO Ali Ghodsi argues we've crossed the AGI threshold but enterprise AI adoption is failing because companies haven't rewired their processes, not because the models aren't good enough. The real bottleneck is getting decades of organizational context into AI agents. Long-term, applications — not models — will capture most of AI's value.
figure out how to get that context into the AIs inside of an organization, how do you transform how old school business is happening and how do you get those processes into the agents then you will have massive impact because AGI is already here
Good ideas are very hard to come by; humans are bad at identifying them and tend toward tunnel vision on the wrong problems, as illustrated by the multicast example.
We already have AGI so we already have artificial general intelligence
the AI researchers at AMPLab in 2009 agreed that by their definition of AGI at the time, we have already hit that
95% of the POCs are failing
if you don't get all the context that exists inside of these organizations and how humans work and everything, all the context we have in our heads, if you don't get that to the models and the agents, they're going to do lots of stupid mistakes and they're useless
barriers to entry have significantly gone down and switching costs have significantly gone down for software
if in the future you're just talking to an agent, that switching cost gets eliminated because you're just talking to an agent
Databricks has approximately 20,000 customers
building a production connector from Databricks to Salesforce took three quarters (9 months) before AI assistance
the team that builds connectors could only compress the timeline from 9 months to 7.5 months using AI, despite Ali claiming he could write a connector in two days
if a company has been around for 10 years and they have not innovated, if their software looks the same as 10 years ago but the revenue has been going up, they should be worried
Ghodsi uses Databricks' Genie product daily for numerical and quantitative internal decisions such as ROI and cost analysis.
it's not an AI problem, it's a human problem — you have to kind of rewire all your processes in the organization to be able to do it
Ali Ghodsi was late to the meeting because he was meeting with the CEO of one of the big banks who reported seeing no productivity gains from AI in their organization
if OpenAI and Anthropic are just software companies with researchers writing software, they would be dead too under the 'software is dead' thesis, yet they have trillion dollar valuations
The product requirements process for connectors was reduced from one quarter to one week by writing down requirements quickly and iterating faster.
Standing up Salesforce, Workday, and NetSuite test instances was outsourced to external firms working in parallel, shrinking that bottleneck.
The team moved from one person per connector to seven people working across all seven connectors simultaneously, eliminating bus-factor-one risk.
Seven connectors shipped in one quarter as a result of those process changes.
Having a better AI model—up to GPT-7 or a future frontier model—would not have helped ship those seven connectors faster; the bottleneck was process, not model capability.
All value in technology consistently moves up the stack over time—from IBM PCs, to operating systems (Microsoft), to virtualization (VMware), and now toward applications.
Applications will be the winners in the five-layer AI stack; that is where Ghodsi would allocate investment.
In the early 2000s the consensus among networking researchers was that the multicast problem—efficient one-to-many broadcast—was the most important problem on the internet, yet it turned out to be irrelevant because bandwidth costs collapsed.
Healthcare is approximately 17% of US GDP, making it a candidate for a trillion-dollar AI company.
Education has historically been a poor VC investment—no public-market company has reached even $100 billion in market cap in education.
An AI-powered education company with proven learning outcomes could become a trillion-dollar company with data moat and winner-take-all dynamics within geographies.
Moonshot's Kimi 2.6, released Tuesday (two days before recording), is the best model ever produced in history—better than any frontier or non-frontier model—but only relative to what existed in January 2025.
Open source models are closing the gap with proprietary frontier models; the gap that was three to four months is now approximately one month.
The frontier model business will become an economies-of-scale game with thin gross and operating margins, analogous to Amazon's original book-selling business.
Declining cost of software reduces barriers to entry and switching costs, creating a 'SaaS apocalypse.'
Airbnb could have been started in 2001—no additional technological development was required—yet it took until 2009 for Brian Chesky to have the idea, triggered by a personal need at a conference.
Ghodsi stopped using Twitter/X after Elon Musk acquired it.
even inside AI companies, they're hiring salespeople from old school companies and running things in old school ways — the futuristic agentic co-worker vision is not happening
22-year-old interns at Databricks are asking whether delaying starting their own company by six months means their career is ruined because AGI is coming and they'll miss the boat
Students and practitioners risk focusing on the loudest current trend (e.g., frontier AI models, AGI) the way 2000-era researchers focused on multicast—which may prove irrelevant.
The key blocker to enterprise AI adoption is that models and agents lack the organizational context that experienced employees ('John or Jane') carry in their heads after 10-20-30 years at a company
Software moats that survive AI disruption include economies of scale, brand, trust/security certifications, patents, switching costs, and proprietary data
Historical analogy: the electric dynamo took 40 years (1880 to 1920) to show productivity gains because factories had to redesign their entire floor plans, not just swap steam engines for electric ones — AI adoption will follow the same long diffusion curve
The connector shipping improvement was a human refactoring and process-change problem, representative of what the whole world needs to do to succeed with AI.
Ghodsi's investment strategy for AI applications: early-stage seed strategy, invest in many startups, expect most to fail, but a few will become the next Google.
Jeff Bezos identified the internet as the secular trend, made a long-term bet on online purchasing, and started with books—the least differentiated commodity—as the entry point.
Shared with Earmark · earmark-ai.com