Oct 30, 2025 · 1h 3m · mad
State of AI 2025 with Nathan Benaich: Power Deals, Reasoning Breakthroughs, Real Revenue
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck interviews Nathan Benaich, founder of Air Street Capital, to analyze key findings from the State of AI 2025 report. They explore technical breakthroughs in reasoning and robotics, energy and semiconductor bottlenecks, geopolitical open-source competition, and the evolving economics of generative AI.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 25.7% of the talking time here. How this is scored →
Matt as informed peer 3.9 Guest teaching 3.6 Guest disagreement 1.3 Matt pushing back 1.3
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 31:40 Nathan dismisses national sovereign AI initiatives as 'sovereignty washing'
Nathan forcefully re-frames national compute pledges, calling them marketing tools for re-industrialization and dismissing pure sovereign AI claims as 'sovereignty washing'.
Hardest push from Matt ▶ 52:25 Matt refutes travel booking agents as a classic overhyped demoMatt directly rejects the premise that consumer flight-booking agents represent significant agentic progress, calling it the 'sad canonical use case in San Francisco'.
Biggest teaching moment ▶ 44:00 Nathan illustrates the massive gap between AI safety spend and lab CapExNathan educates the host on the vast financial disparity in AI safety, revealing that all global non-profit safety orgs combined spend $134M annually—equivalent to what frontier labs spend in a single day.
Matt holds his own ▶ 46:00 Matt challenges whether copyright settlements establish pre-training moatsMatt shows strong domain understanding by pressing Nathan on whether $1.5 billion copyright settlements create an impenetrable capital moat that prevents early-stage startups from pre-training.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Welcome & The Scope of the State of AI Report | 3 | 2 | 1 | 0 | Host sets up the interview acknowledging the scope of the 312-slide report and asking how reasoning models evolved. Nathan outlines breakthroughs in verifiable domains like Math Olympiad and biology co-scientists. | |
| Robotics: From Chain of Thought to Chain of Action | 3 | 3 | 1 | 0 | Host introduces the shift in robotics from chain of thought to chain of action. Nathan explains the separation between planning and actuation, while contrasting industrial deployment momentum with humanoid long-tail risks. | |
| Recommended Deep Dive Podcast Episodes | 4 | 3 | 1 | 1 | Host plugs prior podcast episodes before guiding the discussion to business metrics. He constructively synthesizes data comparing business adoption rates against individual employee usage to highlight shadow AI. | |
| The Margin Debate & Economics of AI Models | 4 | 4 | 1 | 1 | Host invites Nathan to unpack the unit economics and margin debate surrounding AI startups. Nathan details gross margin compression caused by flat token pricing relative to reasoning-heavy queries and lab vintage layer cakes. | |
| Evaluating the AI Bubble: Wall Street vs. Silicon Valley | 4 | 3 | 1 | 2 | Host probes the bubble argument by noting how public market constraints distort Wall Street sentiment relative to private Silicon Valley activity. Nathan outlines financial engineering risks, off-balance-sheet debt, and circular deals. | |
| Godfather Skeptics vs. Modern AI Builders | 4 | 5 | 2 | 2 | Host presents the ideological split between pioneer AI skeptics and current lab builders before raising power bottlenecks. Nathan presents massive data center CapEx stats and gas turbine workarounds, while Matt raises under-discussed sustainability concerns. | |
| NVIDIA Dominance, Custom ASICs, and Global Chip Challengers | 4 | 4 | 1 | 1 | Host actively follows report metrics, citing slide 166 regarding Nvidia's stock outperformance against competitors. Nathan outlines Broadcom's ASIC growth around stable transformer architectures and Chinese chip dynamics. | |
| Sovereignty AI Initiatives and National Compute Strategies | 3 | 4 | 2 | 2 | Host introduces national sovereign AI initiatives across various countries. Nathan offers a skeptical reframe, characterising sovereign AI programs as political marketing and sovereignty washing. | |
| Geopolitics of Open Source and US-China AI Model Competition | 5 | 3 | 1 | 1 | Host demonstrates currency on market movements by bringing up Reflection AI's $2B raise alongside Chinese open-source dominance. Nathan notes Qwen's download metrics and predicts US labs will re-embrace open source for political leverage. | |
| AMD-OpenAI Partnership, Market Concentration, and Lab Financialization | 4 | 4 | 2 | 1 | Host prompts a discussion on AI market concentration, leading Nathan to reveal warrant terms in the AMD-OpenAI partnership and describe Anthropic's cultural evolution toward commercial revenue priorities. | |
| Application Layer Value Creation and High-Impact Enterprise Domains | 4 | 5 | 2 | 2 | Host asks about app layer slop and the weakening regulatory landscape. Nathan dismisses hand-wringing over AI slop and presents a stark comparison between global AI safety budgets and lab CapEx. | |
| Data Copyright Litigation, Licensing Deals, and Pre-Training Moats | 5 | 4 | 2 | 3 | Host presses Nathan on whether $1.5B copyright settlements establish an insuperable financial moat for early-stage pre-training startups. Nathan responds by citing synthetic data generation and DeepSeek architectural efficiency. | |
| AI Cybersecurity Attack Vectors, Model Context Protocol, and M&A | 5 | 4 | 1 | 2 | Host demonstrates sharp domain awareness regarding cybersecurity exit trajectories, predicting rapid M&A by incumbent security vendors. Nathan details complex AI attack vectors including embedded malware LLMs. | |
| Computer Use Agents, Agentic Commerce, and Agent Experience Design | 4 | 4 | 2 | 3 | Host explicitly rejects consumer travel booking agents as an overhyped San Francisco demo. Nathan re-focuses on high-converting agentic commerce and Agent Experience (AX) software design. | |
| AirStreet Capital's Investment Thesis and AI Sectors | 3 | 3 | 0 | 0 | Host asks Nathan to outline AirStreet Capital's core investment thesis. Nathan details portfolio investments spanning tech bio, defense, general-purpose robotics, and voice agents. | |
| State of AI Report Predictions for the Next 12 Months | 4 | 3 | 1 | 0 | Host closes the interview by discussing Nathan's top predictions for the next 12 months. Host offers an insightful quote from Workday's CEO illustrating why legacy enterprise software remains defensible. |
Statements from this episode (34)
OpenAI and DeepMind achieved Math Olympiad gold medals
“And, you know, we saw gold medals on the International Math Olympiad by a couple of labs, including OpenAI and DeepMind.”
DeepMind reasoning models identified disease targets validated in wet labs
“And so DeepMind has shown that you can integrate this kind of reasoning model to sort of decipher new targets for disease, new mechanisms that were actually also proven in a wet lab scenario post facto.”
Humanoid robotics will mirror self-driving car long-tail struggles
“My personal bet is I think it's gonna be, the humanoid space is gonna look much more like self-driving, where we have some very good isolated demos, but the long tail will kill you. And so we're gonna go through many false starts, and I think this is just the …”
Top 20 major AI companies generate tens of billions in revenue
“If you sum sort of the top 20 or so major AI companies from the labs to the most popular kind of vertical applications, you know, across them, they're making tens of billions of dollars of revenue.”
Corporate AI subscription retention rose from 50% to 80% since 2022
“Across a segment of 43,000 or so US customers, we work with ramp to show that retention of subscriptions on AI products across this customer set has really improved markedly since 2022. On 2022 is around the 50% after 12 months. And now in 25, it's hitting aro…”
Average corporate AI spend hits $500,000 and will reach $1M
“The total spend on AI products per customer kind of went up from 35,000 dollars or so maybe two years ago. Now it's around half a million dollars, and it's predicted to hit a million dollars next year.”
76% of surveyed AI practitioners pay for tools out of pocket
“95% of people use AI in their personal life and in their professional life. About 76% of people pay out of their own pocket for it. It's like 10% of people pay more than 200 bucks a month for it. And then looking at the organizations that they work at, it's li…”
Vertical AI application products generate gross margins around 30%
“There are cases where these kind of vertical products are making margin, gross margins of like 30%, and sometimes they get worse with scale.”
Top AI serving companies achieve 70% to 90% gross margins
“There are companies here that are making very, very good margins on serving their AI systems, like, 7080, sometimes 90%, depending on the modality and so, like with everything, the average number sucks but, like, when you look at the best companies, it's reall…”
Wall Street discusses AI bubbles far more than San Francisco
“The finance crowd in New York is definitely talking about bubbles a lot more than what we're talking about in San Francisco, where they're, Their view is like, this is the golden era of AI, and a lot of things are working.”
Meta raises tens of billions in off-balance-sheet debt for data centers
“You have this, sort of, offloading of debt from big companies, for example, Meta, that raises tens of billions of dollars to fuel its data center ambitions, but that doesn't sit on Meta's balance sheet.”
A 1GW AI data center requires $50B CapEx and $8B-$11B annually
“The biggest step for me is one gigawatt of a data center for AI basically costs fifty billion dollars in capex. And on an annual running basis, it costs between, like, another eight to nine eight to nine to maybe even eleven billion dollars to run.”
Major tech companies abandon green commitments to secure AI power
“Well, a year or two ago, big companies did make commitments to be green as of, you know, as of, you know, as soon as they started inking deals with you know, nuclear companies and, Various energy providers for data centers, all those commitments basically got,…”
NVIDIA chips power 90% of open-source AI research papers
“If you do that analysis, you basically find that 90% of all papers make use of a NVIDIA chip.”
Custom AI chips become viable only after model architectures stabilize
“As soon as you get to a point where there's some convergence on an architecture that's looks like it's stable and is revenue generating and developers are coming to sort of work on it and confirm that it is like the thing, then you can flip towards doing a cus…”
NVIDIA stock outperformed private AI chip startup investments 12x to 2x
“If you bought NVIDIA stock on the day of the announcement of all the, like, private rounds in these companies, what would the value of your stock be in NVIDIA versus these companies? And if I recall correctly, it's basically 12 X in NVIDIA versus two X in in t…”
NVIDIA's sovereign AI business generates around $20 billion in revenue
“Nvidia has even started marketing this as like a new kind of product line basically for its business that currently generates, I think, around twenty billion dollars worth so it's real money”
Sovereign AI initiatives are mostly marketing and sovereignty washing
“I think it's more marketing than it is like a real policy, because at the end of the day, if you buy your stock from the U S and you're not an ally of the U S at some point, then they'll just switch it off. And so part of this is like sovereignty washing, I th…”
50% of Hugging Face model derivatives are now based on Qwen
“I think 50% of all model derivatives being downloaded from Hugging Face or Quen base now.”
OpenAI holds AMD stock warrants tied to a $600 share price
“There's like another financial sweetener in the deal where OpenAI has warrants in AMD if the stock price hits 600.”
A large portion of CoreWeave's revenue comes directly from Microsoft
“A lot of CoreWeave's revenue also comes from Microsoft on the way back.”
AI slop is acceptable if consumers actually enjoy engaging with it
“And I think like the slop thing, I mean, if it's bad, people won't look at it. And if it, if people look at it and they enjoy it, then, you know, good, good for them. Like, I don't necessarily have like a huge problem with that.”
Only three EU member states have implemented the EU AI Act
“The EU AI act has had delays in implementations, only three member states that have actually implemented it.”
AI model capabilities on cybersecurity tasks double every six months
“The capabilities on cyber tasks of models are doubling every six months.”
Independent AI safety organizations spend an average of $134 million annually
“Independent safety organizations. There's maybe like five or six. These are usually nonprofits that are still nonprofits or private companies. They spend on average, one hundred and thirty four million dollars a year in total.”
Anthropic agreed to a $1.5 billion training data copyright settlement
“And then there was a biggest settlement that happened in the last few months with Anthropic that agreed to pay out one and a half billion.”
Smarter pre-training methods will reduce massive AI data spending requirements
“So I think we're just getting smarter about how to do pre-training rather than shoving everything we have into a bucket and like seeing what happens. And so as a result of that, you might not necessarily have to spend the exact same amount of money to get a ca…”
SaaS software will eventually be replaced by autonomous AI agents
“At some point, basically, just as desktop software became SaaS, at some point, SaaS will just become an agent because it's no longer really like a human that's It's actually doing everything in the software product, but a software that's running the software p…”
AI search traffic converts at higher rates than direct e-commerce traffic
“Traffic that's generated through conversations and AI search onto a commerce platform converts high to higher level than direct traffic.”
Google Search survives because ChatGPT relies heavily on referencing it
“People say, oh, Google search is dead. I think that's like probably completely wrong because ChatGPT references Google a ton.”
ProFluent developed the first AI language model-created CRISPR genome editor
“ProFluent, which is kind of leading the charge for these language models in protein design, developing the first, like CRISPR genome editor that an AI has created.”
Data center NIMBYism will feature prominently in 2026 political campaigns
“We predicted this kind of nimbyism, not, not in your backyard will kind of take precedence in in major political campaigns in 20, 26.”