Dylan Patel, SemiAnalysis, Nebius, Glean, Legora.. 12 Hot Takes From The Biggest Names in AI

1. Key Themes

Enterprise AI Budgets Are Tightening — The Free Spend Era Is Over

The shift from "give me everything" to "prove ROI first" is now the dominant dynamic in enterprise AI. Multiple speakers independently flagged this as the defining friction point of the moment.

"If we were talking 12 months ago, there was way more of give me everything, and I'll take five of it. And now it's like, what are you giving me and how much I'm going to pay?" 00:04:26

— Qasar Younis (Applied Intuition)

"From a business perspective, it is finally businesses are running into this, well, we're going to actually measure our spend on AI. And so that is creating a little bit of friction." 00:30:48

— Arvind Jain (Glean)

Physical AI / Machines Is the Bigger Long-Term Story Than Language Models

Qasar Younis argued forcefully that putting intelligence on physical machines — ships, trucks, construction equipment, defense — is the more durable and historically significant wave, comparable to Amazon and Apple being the real internet winners rather than early static website builders.

"I think, when we look back, I think that's going to be the bigger story. And then look back 25 years from now. I think we'll look back, it's kind of like, if you look at the early days of the internet, there were companies that are really focused on getting websites up, static websites up. And so when you look back over the last 25, 30 years of the internet, you really think about Amazon, you really think about Apple." 00:08:55

— Qasar Younis (Applied Intuition)

Open Source AI Is Dying Faster Than Most Realize

Dylan Patel delivered a blunt, non-consensus data point: Chinese model labs are actively telling inference providers that their next models will NOT be open source — they will shift to licensing.

"The theme here is that all of like, there's multiple Chinese model labs who are telling all the inference guys, hey, our next model is not going to be open source. We're going to license it to you. So open is dying quickly, unfortunately." 00:21:12

— Dylan Patel (SemiAnalysis)

This was immediately contrasted by Arvind Jain's prediction that open source will dominate inference within two years — a direct and unresolved tension between two well-informed observers.

"Open source models are going to dominate AI inferencing. You'll see in the next two years, we'll shift to where it's almost no open source to where it's going to be almost all open source." 00:31:46

— Arvind Jain (Glean)

Infrastructure Builders Are Over-Optimizing for Today's Workloads and Will Waste Billions

Dylan Patel's sharpest warning: companies building data center infrastructure are looking backward at what models look like today, not where workloads are heading — and are customizing and optimizing in ways that will become obsolete.

"A lot of people are trying to optimize on the current rather than think about where the workload is heading and then optimizing. And that's going to lead to a lot of wasted infra spend. Most people don't know what they're doing. They're just buying Nvidia stuff, but a lot of people are also just like, they think they know what they're doing and then they're buying and building and optimizing and customizing. And then ultimately they're gonna waste a bunch of money." 00:19:32

— Dylan Patel (SemiAnalysis)

Token Maxing vs. Token Budgeting Is a Real and Contested Debate

Two knowledgeable operators took diametrically opposite positions. Dylan Patel argued budgeting is "loser mentality" that prevents behavioral transformation; Gil Feig (Merge CTO) argued that more tokens are yielding zero measurable output gains, and cycle time reduction is the only proven lever.

"Token budgeting is a fallacy. The only way to get people to change their behavior is you measure them and you force them to do things differently. And token budgeting is like pulling back on them." 00:22:32

— Dylan Patel (SemiAnalysis)

"Token maxing, not working. You're getting zero results from using more tokens. The only thing that people are seeing a real connection between productivity and usage of AI is how you use AI to bring your cycle times down, get feedback and iterate really quickly." 00:49:54

— Gil Feig (Merge)

Geography Over Vertical as AI Go-to-Market Strategy

Barak Kaufman (Wonderful AI) made the non-consensus argument that geographic expansion — particularly rest-of-world markets — is a more powerful differentiator than vertical specialization for enterprise AI companies. Wonderful expanded to 30+ countries in roughly six months.

"I believe geographies will be even bigger than verticals when it comes to AI. Countries are way more interesting from an expansion perspective and a focus perspective than verticals and that you can go horizontal without a vertical specialization with AI as long as you're able to solve the go-to-market strategy with it." 00:43:12

— Barak Kaufman (Wonderful AI)

Data Is the Unsung Hero of AI — "Data Is Back"

CJ Desai (ServiceNow) argued that the quality of the data layer is the actual gating factor for AI application quality, and that this is broadly underappreciated.

"Data is the unsung hero and data is back. You cannot create an AI application without a great data layer and your AI application is as good as your data." 00:56:13

— CJ Desai (ServiceNow)

The Pricing Model for AI SaaS Is Shifting to Consumption-Based

Legora (legal AI) became the first legal AI company to move to consumption-based pricing, and Max Junestrand predicted the entire category will follow. He framed it as an inevitability driven by the economics of token costs.

"If you have a flat rate and you say, you know, 200, 300 dollars per user per month, and then they use $2,000 worth of tokens, you have a problem... Token consumption have been the business model in coding tools since inception. It's the pricing model of the big labs, and now it's going to be the pricing model in legal. And I expect many of the other companies to follow us into that." 00:47:44

— Max Junestrand (Legora)

Multi-Model Routing Is the Only Rational Enterprise AI Strategy

Apoorv Agrawal (Altimeter) framed the four leading labs as four seasons — unpredictable, cyclical, none permanently dominant — and argued CIOs making hundred-million-dollar bets must architect for a "climate" of multi-model routing and evals rather than committing to one provider.

"There are four seasons in AI. They're called OpenAI, Anthropic, SpaceX, and Google. So if you're a CIO, if you're a CEO, if you're about to make a multimillion, billion dollar decision of which lab to build your intelligence on, plan for the climate, not for the weather." 00:35:43

— Apoorv Agrawal (Altimeter)

Security and Governance Will Bottleneck Enterprise AI Adoption Before Openness Wins

Gil Feig argued that integrations are the danger point where LLMs shift from annoying to genuinely dangerous — and predicted a wave of governance and lockdown before AI gets truly opened up inside enterprises.

"Integrations are the point where an LLM actually becomes dangerous. Something isolated, an LLM, it might insult you, but it's not going to do much worse than that. But the second that thing can send your data elsewhere, that's where all the problems come in." 00:50:32

— Gil Feig (Merge)

---

2. Contrarian Perspectives

Open Source Is Actually Dying, Not Winning

While the conventional narrative is open source AI's triumph, Dylan Patel has direct intelligence from Chinese model labs that their upcoming models will NOT be open — they're moving to licensing. This runs directly counter to Arvind Jain's prediction in the same episode.

"There's multiple Chinese model labs who are telling all the inference guys, hey, our next model is not going to be open source. We're going to license it to you. So open is dying quickly, unfortunately." 00:21:12

— Dylan Patel (SemiAnalysis)

Token Budgeting Is Counterproductive — Organizational Transformation Requires Unconstrained Usage

The dominant enterprise instinct right now is to control token spend. Dylan Patel argued this is backward: forcing people to budget tokens prevents them from learning new workflows and prevents companies from reshaping themselves with AI.

"If you're token budgeting hardcore, then your people are not going to learn the new workflows and then they're going to get... not going to reshape your company and make it more efficient and drive a lot more work with fewer people. Token budgeting is a fallacy." 00:22:04

— Dylan Patel (SemiAnalysis)

The Hidden Risk in AI Is an LTCM-Style Cascade, Not a Gradual Slowdown

Most people model AI risk as a gradual correction. Qasar Younis raised the specific analogy of Long-Term Capital Management — a widely lauded, seemingly untouchable firm that collapsed rapidly due to peripheral global events compounding. He asked whether some external shock could cascade through the AI industry the same way.

"Is there something that happens more peripherally in the AI business? And then it has a domino effect into our businesses. Like, that's what I kind of think about a little bit... Could there be an LTCM situation in our business where you have something that's so elite and hopeful and then suddenly some things don't work, and it stumbles and it has real repercussions." 00:11:57

— Qasar Younis (Applied Intuition)

Geographic Focus (Rest of World) Beats Vertical Specialization as an AI Moat

Most enterprise AI companies are racing to own a vertical. Barak Kaufman argued the actual unlocked arbitrage is geographic — going horizontal across use cases but owning non-US markets where competition is thin and the AI go-to-market playbook hasn't been deployed.

"The non-consensus part was really the geographic focus to go generalized, to go horizontal in terms of the use cases, but to actually focus on rest of world versus US. The rest of it around the challenges with applied AI is very similar to the way most of us probably think about it, but it's the playbook that you bring to it and local delivery that was most non-consensus about Wonderful." 00:44:35

— Barak Kaufman (Wonderful AI)

European Startup Complaints About the AI Gap Are Largely Self-Inflicted

Max Junestrand argued bluntly that European startup underperformance is not structural — it's a work ethic and mindset problem. Companies that adopt a global mindset from day one compete fine; companies that complain about being locked out of US frontier models are rationalizing their own lack of intensity.

"There is a lot of complaining from European startups, rather than just locking in and building for the global stage. I think there's a bit of laziness... If you want to build the biggest companies in the world, you need to look past that. And you need to understand that you're competing with the US, you're competing with China. And if you want to win globally, you need to work as hard as they do." 00:48:49

— Max Junestrand (Legora)

---

3. Companies Identified

Applied Intuition

Physical AI company building intelligence for autonomous machines across automotive, defense, construction, and mining. 1,000+ engineers; has barely touched its funding. Mission: put intelligence on a billion machines.

"We're working with Huntington [Ingalls] to put intelligence on the actual ships... or we're working with Heidelberg Materials to put intelligence in quarries, and in ports and in mines." 00:06:04 — Qasar Younis

SemiAnalysis

Semiconductor and AI infrastructure research firm run by Dylan Patel; built through conference networking and deep supply chain intelligence. Provides forward-looking analysis on infra, chips, and model economics.

"I mean, Paris is beautiful, but it's hot as... I don't know. AC. Paris is meant to be enjoyed in spring and fall, not in summer." 00:17:34 — Dylan Patel

Nebius

AI cloud infrastructure company with 20+ data centers, pursuing a diversified portfolio approach to avoid single-project risk. Growing toward tens of billions in revenue; expanding from model training to inference to agentic workloads.

"We have 20 data centers. We're multiplying that into the future. So the biggest challenge that many of our competitors are facing is being single threaded... Fortunately, at Nebius, we're actually pursuing a broad, diversified portfolio approach." 00:26:09 — Marc Boroditsky

Glean

Enterprise AI platform and leader in context graphs — enabling agents to access the full corpus of corporate knowledge efficiently. Positioned at the intersection of ROI measurement and token cost reduction for enterprises.

"Being the leader in context graphs is actually helping create a massive demand for Glean... with our context graph, when a model is trying to do some complex work, it doesn't have to spend all this time just trying to assemble the raw materials to do that work." 00:29:26 — Arvind Jain

Legora

Legal AI company (European, globally minded); first legal AI company to adopt consumption-based pricing. Raised $600M in two weeks. Powers legal work for law firms and enterprise legal teams. Uses TurboPuffer for search.

"We were actually the first legal AI company to move into consumption based pricing... we've been running at a very positive gross margin for a long time, and this will help us continue to do so." 00:47:44 — Max Junestrand

TurboPuffer

Vector search engine optimized for AI workloads, built on object storage (S3/GCS) with intelligent caching for near-native performance at object-storage economics. Powers search for Cursor, Notion, Legora, and Anthropic.

"The reason folks are able to do vector search over the entire web on TurboPuffer is because we're based on object storage. So that's S3 or Google Cloud Storage, which is rock bottom storage prices... Our largest workloads cost tens of millions of dollars to search over right now." 00:40:06 — Nikhil Benesch

Wonderful AI

Enterprise applied AI partner for large enterprises outside the US. Operating in 30+ countries, expanded most of them in the last six months. Horizontal use case approach with geographic specialization as the moat.

"We're an enterprise applied AI partner to the world's largest enterprises outside of the US. We're both a platform and a partner to them." 00:43:12 — Barak Kaufman

Navan

Business travel AI platform; public for ~9 months. Growing 50% in usage, 40% revenue growth, cash flow positive, profitable. Processing $10B+ in bookings annually. Built a proprietary model to prevent hallucinations in travel use cases.

"We grew last quarter by 50% in terms of usage. We grew our revenue by 40%. We became cash flow positive, became profitable. We have more than $10 billion of bookings a year now." 00:54:07 — Ariel Cohen

Merge

API integration platform (CTO: Gil Feig). Focused on secure integrations between AI systems and enterprise data; flagged integrations as the primary AI security risk vector.

"Integrations are the point where an LLM actually becomes dangerous." 00:50:32 — Gil Feig

Decagon

AI customer support company; cited by Apoorv Agrawal for running ~90% of their model volume through open source — a real-world example of the multi-model routing strategy.

"A big portion of their volume goes through open source, about 90% now. And then the frontier stuff, you know, discovering new use cases will always go through the most intelligent models." 00:36:40 — Apoorv Agrawal

MongoDB

Database platform cited as powering 75% of Fortune 100 and leading AI-native startups; integrates vector search and embeddings from Voyage AI natively. Referenced in sponsor segment.

Brex

Intelligent finance platform combining cards, expenses, and banking with agentic finance capabilities. Used by Vercel, OpenAI, Anthropic, Granola, and Deepgram. Referenced in sponsor segment.

---

4. People Identified

Qasar Younis

CEO/Co-Founder of Applied Intuition. Former Y Combinator partner. Building physical AI for autonomous machines across automotive, defense, and industrial sectors. Known for financial discipline — company has barely touched its funding.

"We're still in that phase, which is great... we have a lot of resources to put towards any opportunity that we think is worth deploying a billion dollars for or deploying multiple billion dollars for." 00:07:25

Dylan Patel

Founder of SemiAnalysis. Deep supply-chain and infrastructure intelligence on AI chips, data centers, and model economics. Has direct access to lab and hardware OEM pricing intelligence.

"We've seen next generation hardware receive price increases before they've even started production. Like there was a quoted price and now the quoted price is higher." 00:21:36

Arvind Jain

Co-Founder/CEO of Glean. Former Google engineer. Building the context graph layer for enterprise AI. Has strong conviction that open source will dominate inference within two years.

"Open source models are going to dominate AI inferencing. You'll see in the next two years, we'll shift to where it's almost no open source to where it's going to be almost all open source." 00:31:46

Marc Boroditsky

President of Nebius. Overseeing a 20+ data center portfolio with a diversified, multi-geography build-out strategy. Focused on expanding from training infrastructure into inference and agentic workloads.

"We're always looking at the new and innovative things that are taking place to be able to improve the quality, the capabilities, the reliability of the solutions that we're delivering." 00:27:06

Max Junestrand

CEO/Co-Founder of Legora. European-founded legal AI company with global ambitions. Raised $600M in two weeks. Pioneer of consumption-based pricing in legal AI.

"Typically you don't raise 600 million dollars in two weeks. But we got some great advice when we were in Y Combinator, which is if you build a great company, it's easy to fundraise." 00:47:02

Nikhil Benesch

Co-Founder/CEO of TurboPuffer. Building the foundational search layer for AI agents on object-storage economics. Serves Cursor, Notion, Legora, and Anthropic.

"If you're spending $5 on TurboPuffer per user, but you're only charging your users $5 a month, the economics just don't work. But if we can bring that cost down by an order of magnitude, so you're only spending 50 cents a user on searches, suddenly you're able to build a product where you couldn't before." 00:40:42

Barak Kaufman

Chief Strategy Officer and first institutional investor in Wonderful AI. Architect of the geographic-first, horizontal enterprise AI go-to-market thesis.

"Geographies will be even bigger than verticals when it comes to AI... you can go horizontal without a vertical specialization with AI as long as you're able to solve the go-to-market strategy with it." 00:43:12

Ariel Cohen

Co-Founder/CEO of Navan. Built a hallucination-prevention platform for travel AI; took the company public; growing at 40-50% with $10B+ in bookings and profitability.

"If I send you to the wrong flight... People are so emotional when it comes to travel. You cannot have any fuck ups. Hallucination is a huge, huge, huge fuck up. We've built our own platform, our own model to prevent that." 00:52:43

Gil Feig

CTO of Merge. Skeptic on token maxing; strong advocate for integration security and AI sovereignty as the primary enterprise governance concerns.

"There's never been a good position in business to hand over the fate of your company, your future, your development to another company. So I think people are going to want sovereignty." 00:51:11

Apoorv Agrawal

Partner at Altimeter Capital. Teaches Stanford course on AI stack; has hosted NVIDIA, Grok, Databricks, BaseStand, OpenAI, and Anthropic speakers. Coined the "four seasons of AI" mental model for lab selection.

"There are four seasons in AI. They're called OpenAI, Anthropic, SpaceX, and Google." 00:35:43

CJ Desai

President & COO of ServiceNow. Argues data infrastructure is the true gating factor for AI application quality.

"Data is the unsung hero and data is back. You cannot create an AI application without a great data layer." 00:56:13

Jesse (Decagon)

CEO of Decagon (mentioned by name by Apoorv Agrawal); wrote publicly about routing 90% of their AI volume through open source models — cited as a real-world template for multi-model architecture.

"As Jesse from Decagon wrote very eloquently. A big portion of their volume goes through open source, about 90% now." 00:36:40 — Apoorv Agrawal

---

5. Operating Insights

Consumption-Based Pricing Is the Forcing Function That Fixes AI Unit Economics

Legora's move to consumption-based pricing solves a structural problem that is quietly destroying margins across SaaS AI companies: flat-rate subscriptions create an existential mismatch when users consume $2,000 of tokens on a $200/month seat. Max Junestrand identified this as the reason the model will spread across categories.

"If you have a flat rate and you say, you know, 200, 300 dollars per user per month, and then they use $2,000 worth of tokens, you have a problem. And so we've been running at a very positive gross margin for a long time, and this will help us continue to do so." 00:47:44 — Max Junestrand (Legora)

Cycle Time Reduction, Not Token Volume, Is the Measurable AI Productivity Lever

For operators trying to assess whether their AI investment is generating returns, Gil Feig offered a precise diagnostic: ignore token consumption as a proxy metric. The only causal link to productivity that holds up empirically is reduced iteration cycle time — faster feedback loops.

"The only thing that people are seeing a real connection between productivity and usage of AI is how you use AI to bring your cycle times down, get feedback and iterate really quickly." 00:49:54 — Gil Feig (Merge)

Build a Diversified Portfolio of Infrastructure Projects, Not a Single Bet

Marc Boroditsky identified the core operational failure of competing data center builders: single-threading. In an environment with local regulatory, community, and power challenges in every geography, a portfolio of projects is structurally more resilient than any optimized single build.

"The biggest challenge that many of our competitors are facing is being single threaded, which is strange when you're in a multi-threaded, parallel processing industry and getting stuck with a single project." 00:26:09 — Marc Boroditsky (Nebius)

Always Be Actively Hunting for Hidden Risk — In Your Company, Team, and Market

Qasar Younis articulated a mental model he applies continuously: paranoid risk scanning across leadership, technical strategy, and external market conditions. He specifically flagged the LTCM analog as the type of peripheral shock most operators ignore.

"You always want to be thinking about hidden risk in your company, in your leadership team, in your technical strategy, in the market at large, and how you play it." 00:12:25 — Qasar Younis (Applied Intuition)

---

6. Overlooked Insights

TurboPuffer's "Search Is Still an Order of Magnitude Too Expensive" Is a Massive Latent Market Signal

Nikhil Benesch made a throwaway comment that deserves far more attention: TurboPuffer was founded because search was an order of magnitude too expensive, they've already brought it down by an order of magnitude, and they still think it needs to come down another order of magnitude before certain product categories become economically viable. This is not just a company story — it's a map of entirely new product categories that do not yet exist because search economics don't support them. The $5/user/month search cost vs. $5/month subscription example is a specific, testable heuristic for identifying markets that will open as costs fall.

"It kills us that there are still products out there that are limited in their ambition because search is still too expensive. What product isn't currently in the market because search is too expensive." 00:40:42 — Nikhil Benesch (TurboPuffer)

"If you're spending $5 on TurboPuffer per user, but you're only charging your users $5 a month, the economics just don't work. But if we can bring that cost down by an order of magnitude, so you're only spending 50 cents a user on searches, suddenly you're able to build a product where you couldn't before. Your margins work, you get crazy growth, you explode." 00:40:42 — Nikhil Benesch (TurboPuffer)

France's Nuclear Power Export Role Is Quietly Becoming a Strategic AI Infrastructure Constraint Across the EU

Dylan Patel dropped a single observation that has enormous geopolitical and infrastructure investment implications: France's surplus nuclear power has been acting as a grid stabilizer for Spain, Italy, Belgium, the Netherlands, and Germany. The moment France redirects that surplus to domestic data centers, multiple EU national grids become fragile. This means French data center expansion faces a hard political ceiling imposed by its neighbors — a constraint that will shape where European AI infrastructure actually gets built over the next decade.

"They have all this spare power. They're like, oh, okay, let's build a bunch of data centers. But then Spain, Italy, Belgium, the Netherlands and Germany all freaked out because it's like, oh wait, all this spare power is actually like keeping our grid alive. If you stop selling us the spare power and you build data centers, our grids are screwed. And so there's been a lot of pushback from the other EU countries — don't build data centers in France because we need the nuclear power." 00:17:51 — Dylan Patel (SemiAnalysis)