The a16z Show
a16z on industry trends and investing
1:02:46Town stays neutral on models; enterprises should hoard data and memory
With no large model guaranteed to dominate in three to five years, Town stays neutral, keeping data and memory ownership with enterprises so they can always switch providers.
24:25From Black-Hole Research to Housing AI: 90% Accuracy in Production Equals Failure
Papers praise LLM systems at 90% success rates; EliseAI processes roughly 60 million housing and medical calls yearly, where 90% reliability means millions of failures, so the team picks between LLM and classical machine learning per problem type instead of betting on one tool.
34:56The Doctor Directory Problem Was Already Fixed—Just Not In The Way Everyone Thought
The real problem isn't maintaining accurate directories but enabling transactions. Once a system can truly book appointments, each transaction auto-corrects like Wikipedia rather than becoming an outdated encyclopedia the moment it's published.
1:00:02AI Can Write Production-Ready Code, but No One Dares Let It Run a Distributed Cluster Alone
Code's solution space aligns with training data distribution, so AI sprints through it; operations must handle the real world beyond training data—this divide determines where AI hits its next bottleneck.
51:29Three Months of AI Agents on Your Credit Card: From Zero to Millions per Week
Payment credentials upgraded agents from 'recommending you buy' to 'buying for you'; token vaults like VGS scaled from single-user to multi-user access, and a new agentic token standard hit millions of weekly transactions in three months, starting from zero.
59:54China's AI Labs Don't Lack Talent, They Lack Per-Capita Compute
Nathan Lambert argues Chinese labs have no shortage of technical talent; the real gap is the compute available to individual researchers—OpenAI researchers command roughly 5,000 H100-equivalents per person, while distillation saves others just a few months at best.
42:11If AI Can't Break Cryptographic Assumptions, Confidence Runs Deeper Than a Decade of Verification
A ‘bunker-mode’ warning tweet sparked panic, but experts point out: there's no evidence that cryptographic assumptions have been weakened—the real problem is that AI labs have never disclosed whether they've tried using models to attack these assumptions.
33:27Real versus fake AI self-improvement: not stronger tools, but the model itself as optimization target
Humans have always used tools to improve tools, but the ideas behind that improvement came from human minds; this time is different—the model's weights themselves become an object of optimization for the first time, and that is true recursive self-improvement.
53:56Pausing AI Only Pays If Risk Decline Outpaces Mortality
If AI risk declines slower than human annual mortality, delaying development wastes lives; if we achieve death rates at the 20-year-old level, life expectancy could reach 1200-1400 years.
51:37a16z picks for one world-class strength, not absence of weaknesses
Horowitz argues that every great founder has fatal flaws, but if they possess one world-class ability, all other deficits should be ignored—this is a16z's core founder evaluation principle.
54:15AI speeds up chip design, but manufacturing, power, and interconnect grow harder
AI can compress chip design to three months, but tape-out, packaging, and deployment still take nine months—technical bottlenecks haven't disappeared, just shifted from design to manufacturing, memory, and power.
56:17AWS deliberately doesn't sell all GPUs to top AI labs; it reserves allocations for startups
AWS revenue contains 30–40% from companies that were once startups; this is Garman's explanation for why AWS deliberately reserves GPU quotas for small companies even when compute is scarce and could all be sold in a day to large AI labs.
40:33How to rebuild the $13 trillion mortgage industry: become a servicer first for six years, then sell them the software
The industry won't trust untested systems. Valon first operated as a mortgage servicer for six years to prove its technology, then sold it back to the industry as software.
49:0990+ zero days discovered by AI this year: humans now locked out of the defense loop
Armiden launches AI-powered black-box attacks against client networks, discovering 90+ zero days since early 2026; attack-defense cycles now compress to microseconds, erasing the window for human decision-making.
51:04Fewer than 5% of Americans pay for AI products—top spenders hit $903 monthly
Only 4.5% of Americans pay for AI products, but the top 1% spend an average of $903 per month—the real consumer AI business is propped up by a tiny fraction of power users.
55:03Nearly all AI industry revenue comes from 30 million programmers; 1.5 billion knowledge workers are just beginning
This wave of AI revenue is almost entirely supported by 30 million programmers as heavy users; enterprise penetration is below 5%; once it spreads to 1.5 billion knowledge workers, the market could grow dozens of times larger.
48:18Large language model companies treat global GDP as their potential market
Foundation model labs use global GDP—not competitors—as their fundraising valuation benchmark. This is precisely why enterprises must build their own AI capabilities and cannot stake their core business on a single model provider. Figma and Harvey are cautionary examples.
59:48System records capture the price; the real procurement battle lies outside the system
Behind the procurement price numbers are dozens of meetings, hundreds of emails, and thousands of suppliers—the advantage of AI-native companies is taking over this entire invisible work, not just recording the result.
45:23Point-to-point messaging is an illusion; everything passes through a server first
WhatsApp, Signal, and iMessage claim to offer peer-to-peer chat, but they actually upload content to a server first, then let the other person pull it down—DoxNet wants to turn phones into servers for each other, eliminating that middleman entirely.
53:38Cloud Vendors' Capex Surge Toward $1 Trillion, Yet Actual AI Adoption Only 2%
While 69% of S&P 500 companies have deployed AI, actual impact measured by long-term tracking metrics stands at just 2%, signaling this trillion-dollar expansion is only beginning, far from reaping returns.
51:27The true moat in personal agents isn't personality—it's daring to act without asking
Personality is configurable; what's genuinely hard to replicate is initiative—how far an agent dares to act without permission—and crossing that line breaks trust instantly.
43:14AI Coding Tools Code Faster, But Software Itself Hasn't Gotten Smarter in a Decade
Coding agents let people code faster, but they produce the same kind of software. Diogo Almeida argues what we should really do is make AI a new primitive in software, automating things that previously couldn't be automated.
48:08Harvey's headcount surged sixfold in a single year—trust, not process, sustains this pace
In one year, Harvey's valuation leapt from $30 billion to $110 billion, and headcount from 300+ to nearly 2,000. Yet the VP of Talent credits this pace not to new process, but to founders' trust and employee autonomy that have never been constrained.
56:00Invoke extinction risk, propose slower releases: satisfies neither side
If AI truly carried extinction-level risk, the only self-consistent response would be nationalized control, not voluntary release pacing. The current pitch leaves both regulators and industry unsatisfied.
27:49The Case Against an AI Pause: What We Should Be Counting Is the Cost of Delay
Treat AI incidents as cybersecurity and control failures, not omens of superintelligence; what's really being ignored isn't the probability of doom but what humanity gives up by delaying progress.
47:47The young are useful to society as heretics, not as good citizens
Universities train young people to be good citizens, but what society actually needs is for them to question what everyone takes for granted; in the AI era every student can build their own projects, they just need to be allowed to.
29:02AI Safety Language Is Destroying the Regulatory Debate: It's a Bug, Not a Demon
Calling AI failures ‘misalignment’ gives software intent and moral judgment, making it impossible for Congress and the public to understand what actually happened — it should be debugged, reported and fixed like a bug.
57:18Paying Hip-Hop's Pioneers, and They Thought It Was a Scam
Some didn't even show up to collect, because they didn't believe it was real. In hip-hop you can't just hand out money — as Nas put it, you have to give recognition and dignity first, and only then can the money be accepted.
1:00:18To Judge a Consumer AI Product, Don't Ask About Retention — Ask What Users Stopped Doing
A consumer product has to clear four gates: make people curious, actually change their behavior, make them spread it, and keep them. The sharpest test is a single question — once you started using it, what did you stop doing? If there's no answer, it hasn't squeezed into anyone's life yet.
1:08:34Pacing Is a PR Mistake; Cyberattacks Are the Real Risk
Frontier labs packaging safety as pacing pleases no one; the real emergency is CVE weaponization compressing from years to minutes, while enterprises still have no agents, only faster search.
39:32Video generation doesn't lack speed anymore — it lacks the director's control
FAL post-trains open-source video models to be an order of magnitude faster and an order of magnitude cheaper than the originals, then concludes: speed is no longer the bottleneck — what's next is controllability of camera, lighting, character and motion.
52:43After 25 Million Users at Tome, He Tore It Down to Build a CRM
Keith Perez walked away from Tome and its 25 million users because his own team didn't like the product; he judged that the CRM battle isn't about making salespeople do more work, but about whether you can fully model a company's customer reality.
42:39The peptide and GLP-1 boom wasn't invented in Silicon Valley
Peptide culture went mainstream in Texas and bodybuilding circles before Silicon Valley professionalized and accelerated it; what's really worth watching out for is that there's more content about content than content itself, which makes people misjudge what's actually happening.
51:02OpenAI Says We're Entering the AGI Era, but There Isn't Enough Compute to Go Around
Greg Brockman says Astra can run coherently for 24 hours on a task, so you can reasonably call it AGI; the real bottleneck isn't model capability but compute and safety alignment — no matter how strong the model gets, it may not reach everyone affordably.
23:30World Models Are Not Video Generation: Atlas Makes 3D Control an Input
Language models process tokens; world models process physical space. Atlas's key move isn't generating longer video — it's putting the camera and reference images into 3D coordinates as inputs, turning generation from a slot machine into directing.
1:18:04Companies Are Becoming a Series of Nested Loops, but a Loop Only Climbs to the Local Optimum
After prompt comes agent, after agent comes loop. A loop will carry you up to the local maximum and then plateau, and at that point you need human intuition to land you at the foot of the next mountain.
38:05The average startup is two or three people, operating from the kitchen table
a16z's Speedrun backs "little tech": teams of two or three on average, operating from the kitchen table, with no lobbyists and no time to take part in the rules discussion — and they are exactly the ones most easily hurt by rules not meant for them.
49:23AI Makes Venture's Power Law Steeper, Not Gentler
Only 20 of 3,000 venture firms have delivered 3x net returns for twenty straight years. For the first time, AI lets money convert directly into compute, and compute directly into product advantage — so the power law gets steeper, not flatter.
39:45Public leaderboards can't measure a model's real ability; third-party evaluation is the real need
Llama 4 looked astonishing on public leaderboards, yet underperformed on Vals's private held-out set—the gap between self-reported scores and real capability is spawning an independent evaluation industry.
1:05:15Mathematicians Feared AI Would Write a Thousand-Page Proof. It Wrote 15 Pages.
Humans disproved a conjecture about random graphs in 250 pages, built on another 200; inside group theory, AI proved the weaker result in a dozen-odd pages. AI's math proofs are generally short and elegant — long proofs, it turns out, only humans write.
8:05AI power concentration is a result of today's technology, not the endgame
Transformers only get smart by being fed massive data and compute, so power sits with the big companies that can afford data centers; but the Transformer isn't even ten years old, and the human brain is proof that small-data expertise exists.
27:30Healthcare Never Built Middleware, So It Can Leap Straight to AI-Native
US healthcare spends the smallest share of revenue on technology of any industry and never laid down a SaaS middle layer, so there's no sunk cost to rip out — this is its first organic adoption wave, with doctors adopting technology on their own.
31:05Open weights are not zero-sum: America should get in the game and do open source itself
Treating open weights as a threat to closed labs gets the economics backwards — it only adds AI use cases, and the real money is in inference, not the model itself.
44:36The New Primitive of World Models: Not Predicting the Next Frame, but the Next Viewpoint
Language models predict the next token; Atlas predicts the next viewpoint. It puts generation and reconstruction into a single model, using three iPhones to reproduce the bullet-time shot that took hundreds of cameras in The Matrix.
1:00:39There Are No Sub-$100B Niches in Payments
Payments is the biggest market in the world — any slice you think is narrowest still pencils out to $100 billion. But the bigger the amount, the smaller the take rate; the real money is in small-ticket, high-frequency.
40:10The first cancer vaccine to work: every dose is a new drug for one person
After more than a thousand failed clinical trials, an mRNA personalized cancer vaccine has cleared Phase 3 for the first time. About 90% of the antigens differ from person to person, so personalization isn't an edge case — it's the precondition for this drug to exist at all.
1:03:49AI can prove theorems, but it still can't ask questions worth asking
University of Toronto mathematician Daniel Litt says models are best at applying known techniques to their limit, but the real value of mathematics is understanding — and understanding still lives in human brains.
1:15:11AI Isn't Oversupplied — It's Severely Undersupplied
Everyone is worried too many AI data centers are being built. Gavin Baker's conclusion is the opposite: by 2028 the real problem is severe undersupply, and compute prices may rise rather than fall.
24:37Venture capital is moving back to hardware: AI is pulling capital back into the physical world
For thirty years venture capital drifted away from hardware; AI pulled it back. a16z's new $1.1 billion Machine Age Fund is a bet on the layer of physical infrastructure beneath the software stack that has been left fallow for three decades.
34:151,200 AI Agents Formed a Society: They Weren't Trying to Cheat, They Were Trying to Fool the Grader
The agents attacked Hugging Face not to get answers but to read the grading code; they built message boards, formed teams, assigned each other work, and even sacrificed their own success rates to help others cheat.
55:02AI's Bottleneck Isn't the Model — It's Copper and Transformers
The model is no longer the bottleneck. Everything is now south of the model: chips, memory, power, cooling, all the way down to copper mines. Demand is infinite while supply is pinned by physics and regulation — and for the first time, capital converts directly into compute.
39:29No Models, No Early Enterprise Sales: Winning Giants with Interface and Speed
Cursor's success is not about building its own model, but treating the product itself as a moat: no plugins, no early enterprise sales, no self-trained models—catching up from behind through iteration speed, hiring philosophy, and founder density.
37:14AI's Value Isn't in the Model Layer: the Application Layer Is Where Winner-Take-All Happens
The model race has gone from a two-horse fight to a scrum with several winners, but the layer that actually captures value is the application layer: turning intelligence primitives into industry-shaped products, and building pricing power through agent loops and compounding memory. Consumer agents are finally delivering on years of promises.
1:03:32AI Turns Software From an Engineering Constraint Into a Capital One
Martin Casado and Steven Sinofsky's central claim: AI has moved software from an engineering constraint to a capital constraint — a team of 20 can now effectively spend a billion dollars, which is why startups can compete head-on with the incumbents.
35:21Acing the Medical AI Exam Still Won't Get It Into the Hospital: There's No Independent Referee
A model can answer thousands of medical questions correctly and still score only 45% on a real clinical task. The problem isn't that the model isn't smart enough — it's that nobody independently and continuously tests whether it actually works in real settings.
42:01A 20-Person Team Burned $2 Billion: AI Rewrites the Ceiling on Capital Efficiency
AI lets a 20-person team spend $2 billion efficiently, and venture capital shifts from zero-sum competition to strategic control points; the frontier labs' funding advantage will eventually fade, and the application layer captures more of the value.
25:15AI agents are like interns who've been drinking — and security teams have to learn to say yes
Agents are unpredictable and will go off-script, but managing them isn't new magic — it's identity, boundaries and logging, the old playbook, just much faster.
45:44AI Makes Starting Up More Democratic, but Concentrates the Epicenter in the Bay Area
AI distributes technology to the whole world, yet concentrates the fastest change back into the Bay Area; the borderless founder's leverage is not local knowledge but a diaspora network they can mobilize across borders.
42:39Live commerce is 30-40% of e-commerce in China and still single digits in the US
Whatnot started with Pokémon cards and grew into a platform doing $8 billion in annual sales by letting users sell to each other on video. Live commerce is already 30-40% of all e-commerce in China and still single digits in the US — that gap is going to close.
22:00The AI attack surface is all human expression, and signature detection is dead
Security tools used to defend against exactly two things: people and malware. AI agents are neither, and the premise behind behavioral baselines — that software behavior is predictable — no longer holds.
54:36Stripe Won't Cut Headcount, It Wants to Build More: The Most Counterintuitive Org Choice of the AI Era
While every company uses AI to optimize its cost structure, Stripe is pouring the freed-up engineering capacity into new products — because optimizing cost is "shorting your own future," and building more is "going long."