Martin Casado - Entering Uncharted AI Territory - [Invest Like the Best, EP.381]
What this covers
Martin Casado, an a16z partner who co-founded the networking company Nicira, returns to Invest Like the Best to place the current AI cycle against earlier computing shifts. He points to ENIAC making computation about 5,000 times faster than a human computer, and to the Internet pushing distribution costs to zero. In his telling, each shift produced far more software and content rather than wiping out the category. He then turns to why the same logic may not reach machines. After roughly 20 years and $75 billion invested in autonomous vehicles, he notes, the cars on the road still cost far more per ride than Uber, and they draw about 1.3 kilowatts against the brain's 20 watts.
The middle of the conversation is about policy and security. Casado compares today's AI safety rhetoric with the "Digital Pearl Harbor" language of the Internet era. He argues that the Internet's actual threats were specific and emergent, and that openness strengthened security, citing Snort, Sourcefire and Mandiant. He faults the armchair scenarios of Nicholas Bostrom for dampening innovation, and applies the Baptist and Bootlegger framing to regulation, suggesting most advocates are bootleggers. He also describes open source as usually arriving after a closed standard has created a market, as Linux followed UNIX.
The business questions are more contested. Casado argues that model companies need traditional moats because distillation erodes any lead, and he says synthetic data works only in axiomatic domains such as chess. He is technically bearish on agentic systems, citing unsolved memory and long-horizon correctness. He considers current valuations roughly correctly sized, and he describes how new categories find fit through market annealing. He also warns that data-center and GPU cycles tend to overbuild.
Casado argues AI's near-term breakthroughs lie in content, emotional companionship, and language reasoning, where marginal costs collapse, while physical-world robotics stays economically hard because models exploit structure in human-created data that real-world sensor data lacks.
- Marginal cost of digital content creation is heading to zero, as earlier computing and distribution shifts did, producing large expansions in output
- Models reproduce human-created token structure, which is economically viable, but the physical world is chaotic and heavy-tailed, and models have not learned to exploit it
- Model companies face a sawtooth treadmill and must build traditional moats, since distillation erodes model-level advantages
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Open source often democratizes or commoditizes mature businesses after a closed-source standard creates the market, as Linux followed UNIX, MySQL and Postgres followed Oracle and Sybase, and Android followed smartphones, acting as a check on closed-source incumbents.
“one way to view it is almost like this, democratize or commoditize mature businesses.”
Technology booms repeatedly show exuberance about economic benefit outrunning the time needed to build real applications and demand, creating chill periods, and supply cycles for infrastructure such as data centers and GPUs that are out of sync with demand produce overbuild followed by falling prices.
“when the supply cycles are different than demand cycles. So the time it takes to actually build a data center”
Computing and the Internet each changed the world when marginal costs went to zero: ENIAC made computation about 5,000 times faster than a human computer, triggering the compute revolution, and the Internet drove distribution costs to zero, which in turn produced far more software and content than before.
“ENIAC shows up and ENIAC is 5,000x faster than a human, so almost four orders of magnitude”
Durable defensibility for AI companies comes from traditional moats such as API integration stickiness, community and virality, workflow ownership, and brand or styling, since LLMs themselves are unpredictable and hard to make sticky.
“Now interestingly, LLMs have this anti-stickiness property and that they're so unpredictable. You can't rely on them anyways. You have to be ready to swap them out.”
When the cost of creating something falls, the historical pattern is not that the category is wiped out but that far more is produced and far more applications emerge, because access becomes so much broader.
“it's not like it wipes out cost... we end up generating so much more stuff and we enable so many more applications”
Current LLMs may mainly learn the structure embedded in roughly 3,000 years of human-written tokens rather than learning the world itself, so their capabilities reflect human curation of language rather than understanding of physical reality.
“It could be the case that these LLMs have just taken all of those tokens, those 3,000 years of tokens and just learned all of the structure in those tokens”
Robotics companies must be verticalized specialists, since building an agriculture or mining robot requires being an agriculture or mining company, so generalist AI teams should avoid these deals and specialist teams should handle them.
“if you're building a robot for agriculture, you better be an agriculture company actually. And if you're building one for mining, you're better be a mining company actually.”
Distillation lets competitors train cheaply on a released model to catch up, creating a perverse economy of scale in which the first model's advantage is quickly eroded, so text model companies need a traditional defensibility plan.
“there's a massive perverse economy of scale because distillation works so well”
Nick Bostrom's Platonic armchair thought experiments about superintelligence influenced figures like Demis Hassabis, Elon Musk, Sam Altman, and Ilya Sutskever, and were conflated with actual systems before those systems existed, which Casado says is now dampening innovation.
“this was all before the systems were actually built. And this was a Platonic armchair thought experiment.”
During the Internet's creation there were the same sovereignty and cyber-Pearl Harbor fears that AI regulation now invokes, yet open source and academia produced most top security contributions, including Snort, Sourcefire, and Mandiant, showing that openness strengthened security.
“Many of the top security contributions to the Internet came from academia or open source... the entire industry came out of open source.”
The rhetoric about the Internet, including Digital Pearl Harbor and cyber sovereignty threats, closely parallels current AI safety rhetoric, yet the Internet-era conclusions were to avoid hampering innovation and to embrace open source.
“We were talking about Digital Pearl Harbor or Cyber Pearl Harbor. There's all of this concern.”
It is unknown whether cameras and robots collecting data from the physical world can produce economically viable models, because no one has yet trained a useful model on unfiltered universe data, and it is not even clear the approach is possible in the near term.
“We don't know if that could be economically viable in the near term, we don't even know if it's possible.”
Autonomous vehicle compute can draw around 1.3 kilowatts for GPU planning and inference, while the human brain runs on about 20 watts and is highly optimized for navigating 3D space, so robotics competes against an extremely efficient biological system.
“The human brain consumes about 20 watts. You're competing with the most efficient part of the human brain.”
After roughly 20 years and $75 billion invested in autonomous vehicles, the cars on the road today still have unit economics far more expensive than Uber, even though AI beats humans on certain perception tasks.
“We, as an industry, have invested $75 billion in autonomous vehicles. And if you actually look at the unit economics of the cars that are working on the street today, they're still more expensive than Uber by quite a bit”
Casado is technically bearish on agents because memory, correctness over long horizons, and planning are unsolved, and it is not clear that LLM advances carry over to fully automated agentic systems.
“there are some very key technical problems that we do not track, like, for example, the memory one that you've mentioned, correctness is another one that we haven't done. We don't really know how to do planning very well.”
Many of the top Internet security contributions, including Snort (which became Sourcefire) and Mandiant, came from academia or open source, so the security industry itself grew out of open-source work.
“Many of the top security contributions to the Internet came from academia or open source, STLinux is a very famous example of this. Well listen, Snort which became Sourcefire was a great example, Mandiant was a good example”
Markets are broadly efficient and the current AI valuations are sized correctly, as dot-com overpayment was ultimately justified by growth, but any individual venture bet can look silly because the expected outcome is unknown.
“I think markets are very efficient. I think markets realize that this is disruptive and there's going to be a lot of great companies.”
Once the roughly 3,000 years of human-created data is exhausted, progress requires finding new ways to create data, and thousands of people writing code and annotating sentences will not match the scale of all code and text ever written.
“we've literally exhausted the data created in the last 3,000 years”
Synthetic data helps in axiomatic domains where state spaces can be fully simulated, such as chess or arithmetic, but it does not provide a general solution for language or creativity because those are not axiomatic.
“If you have an axiomatic system where you can fill out the state space computationally like playing chess or arithmetic”
The physical world is chaotic, fractal, sub-similar, and nonlinear, and it is unclear whether current models can handle that complexity, so AI has not yet learned to produce new real-world outputs the way it produces new human-created content.
“In the physical world, as we all know, it's chaotic. It's fractal. It's sub similar, it's nonlinear.”
Unlike the Internet, which had specific emergent threats like the Morris worm and a real doctrine shift, current AI safety arguments rest on unprovable science-fiction scenarios such as recursive self-improvement and bioweapons, which have largely been refuted.
“we don't have any of the proof points that we had on the Internet. There is no clear doctrine shift.”
The current AI wave may be economically viable only because humans already did the hard work of creating structured data, so models can make viable pictures, videos, and tech, but training on raw universe data to build viable machines remains unproven.
“the current AI wave is honestly just exploiting human-created content because we did the hard work of creating structure, and that's economically viable”
In category creation, companies find product-market fit through market annealing: iteratively tweaking the product while pushing and educating the market, since new-market customers do not know the value, have no workflow, and interviews produce made-up answers.
“So there's a ton of work take you're actually hammering the market and making it so they can accept it.”
Regulation often coalesces around true believers (Baptists) and profit-seeking opportunists (bootleggers); in the AI safety debate, Casado thinks most advocates are bootleggers protecting big company agendas, though he cannot identify which individuals are which.
“I think the vast majority of those are actual bootleggers, 'Listen, this is good for my big company.'”
The actual bitter lesson in computer science is that progress is a sigmoid: exponentials attract excitement, flatten off, and then another unlock restarts the cycle, and the current wave may also plateau.
“everything is a sigmoid and you have these exponentials and people get super excited and it flattens off”
AI startups historically struggled because traditional AI like perception had diminishing marginal returns and required humans in the loop to handle heavy-tailed real-world cases, so AI investing for startups was a bloodbath while incumbents such as Meta and Google captured value.
“AI investing for startups has been a bloodbath. It's been very difficult. It has been an open question, why is that the case? And the answer is because the economics have been crapped.”
Model companies launch models whose usage spikes and then drops, forming a sawtooth pattern in which they must keep launching new models; without a traditional moat they converge to the same product as they stay on this treadmill.
“it kind of looks like a TCP sawtooth. People use it and then they get bored and it drops off.”
When markets expand they tend to fragment rather than consolidate, creating white space; image models are settling into distinct niches like anime, sci-fi, photoreal, and cartoon-text, so many viable companies can coexist for a long time.
“when markets expand, they tend to fragment, not consolidate. And so white space is created.”
The Internet changed US security posture by creating asymmetry: a more advanced, network-dependent society became more insecure than a less connected adversary, replacing the mutually assured destruction stalemate.
“in the Internet, the more advanced society was the more insecure because you were relying on the infrastructure.”
Early Internet oddities such as the coffee pot webcam and the big red button looked silly but foreshadowed giants like Google, Yahoo, and Amazon, so today's companionship and AI image toys may similarly foreshadow the next major platforms.
“those are the things that grew up to be the Googles and the Yahoo!s and the Amazons and none of us knew what that looked like on the other side.”
Unlike APIs, LLMs are inherently unpredictable and unreliable, so customers must be ready to swap models, which means the model itself provides little lock-in and stickiness must come from the layers around it.
“LLMs have this anti-stickiness property and that they're so unpredictable. You can't rely on them anyways. You have to be ready to swap them out.”
Traditional AI, such as perception and robotics, struggled economically because real-world problems are heavy-tailed and require humans in the loop with diminishing returns, while LLM-based products have better economics because they reproduce human-generated content people already want.
“the economics have been crapped. It just been tough... the universe is very heavy tailed. So you have to bring humans in”
AI training and inference communication patterns differ sharply from cloud workloads, so networking fabric and silicon must be reimagined, and the fabric may not be the same over time, similar to how data centers forced networking innovation.
“The communication patterns for training and the requirements on fabric are very different than cloud workloads”
Asking customers what they want in a new category produces made-up answers because customers feel they must respond and lack knowledge of the new value, so interviews are among the most broken ways to build a company there.
“imagine this, you're going to go interview a customer and ask what they want. They're sitting down there, and they feel like they need to answer something”
An agent is a system with a control loop and state: it takes a high-level objective, performs multiple steps while maintaining state, and uses a feedback loop to decide each next step, a combination Casado thinks has not been solved.
“An agent to me has a control loop and state.”
AI is being crammed into old enterprise use cases like chatbots and automation, but its real near-term value lies in creativity, emotional connection, and language reasoning, which computers previously could not solve, producing a creative social explosion rather than enterprise automation.
“it's not clear to me that's actually the best use of it. A lot of the problems that it solved are problems that haven't solved on the computers like creativity and the emotional connection”
A $200 million custom ASIC could be economically justified for a single $1 billion training run if it yields a 50 percent improvement, since the $500 million saved exceeds the cost, which is why silicon investment is returning.
“So for $200 million, you can get a 50% improvement, a 50% improvement of $1 billion is $500 million.”
The access layer to the Internet may shift from search engines and banking apps to AI characters, as Casado's 13-year-old daughter uses Character AI characters in group chats with friends and he expects she will open a bank account through a character.
“It's like almost like the access layer to the Internet is changing, right?”
Because the marginal cost of digital creation has gone to zero, AI-driven content generation will naturally proliferate, and the question is one of time scales rather than whether it happens.
“When it comes to content creation like any sort of digital creation, the marginal cost has gone to zero”
First movers gain a large early advantage from creating user behavior, brand, and feedback, which breaks Pareto early with roughly 80 percent of value going to them, but that advantage erodes over time.
“first movers tend to have a massive advantage here. And it's because they're creating a new user behavior and they have a brand”
Developers spend about 90 percent of their time maintaining code, so correctness and readability matter, humans must still understand the system, and AI will increase the amount of coding rather than eliminate programmers, just as low-code promised and failed.
“I think people spend about 90% of their time when they develop maintaining code.”
Any claim that a task will be done soon is a cliche because the universe is heavy-tailed with many exception cases, as Tesla's autonomous driving difficulties illustrate, so timelines for real-world AI should be treated skeptically.
“It just feels like the universe is very heavy tailed. There's lots of exception cases, and it takes a long time.”
The more an AI product reproduces what humans already create, the more economically viable it is, because human demand is clear and humans have already done the work of structuring the data.
“The closer you get to a model that's just trained on data further by human, the more economically viable the company is”
A AAA video game costing around $500 million could have a cost of goods sold near $10 if its music, sound, textures, meshes, characters, cutscenes, and images were generated by models, since nearly every asset can already be created with them.
“rather they'd be $500 million be actual COGS cost should be something like $10”
Rohan's experiment showed that data corpora that compress well with gzip produce more learning in models, which suggests models exploit structure already present in human-created data.
“if you have a data corpus and you gzip it, use compression, if it compresses a lot, then the model learns a lot on the same data”
Academia has become largely muzzled on AI open source, unlike its role in the Internet era when Berkeley's BSD and MIT's EFF championed open infrastructure, and a California AI bill reportedly came from a Berkeley professor and her student.
“academia is totally complicit, the California bill, this horrible bill in California”
Because enough human content has been created, we are close to typing a prompt that auto-generates a complete, deep virtual world with every desired asset, eliminating the need for humans to hand-construct metaverse environments.
“we're very close to being able to type a prompt into a computer and end up with an entire virtual world that you can enter”
Casado believes at least one-third of dollars spent on generative models go to emotional connection products such as virtual friends, interactive games, and virtual romantic partners, which he treats as a distinct category from content creation.
“at least 1/3 of that are people that are just playing an emotional connection”
Language reasoning tasks such as reviewing a legal PDF, which takes a $700-an-hour lawyer three days and yields partly wrong answers, are likely to shift to LLMs, because models do a good enough job on language reasoning at far lower cost.
“the lawyer spends 3 days reading it. The lawyer gives me answer some questions based on it. Some of those questions are right, some of those questions are wrong”
The worst thinking in AI is zero-sum thinking, because open and closed source models and multiple competitors are not mutually exclusive in a market that is growing rapidly, so the market is abundant rather than winner-take-all.
“I think actually the worst thinking in this AI way is zero sum thinking.”
If small tech and academia do not mobilize, AI regulation driven by doom narratives will hurt the United States, academia, researchers, and startups, causing the country to regulate itself out of leadership.
“we don't regulate ourselves out of our leadership position.”
Creativity and language reasoning come from the prefrontal cortex, which Casado describes as roughly 250,000 years old and not highly optimized, so computers can surpass humans there more easily than in physical navigation.
“It turns out that computers are much, much better at the newer part of the brain solving that.”
Generating a Pixar-style likeness of oneself costs roughly $200 with a designer versus about a hundredth of a penny in model inference, a gap of roughly four orders of magnitude similar to ENIAC versus a human computer.
“the inference cost to create a Martin as a Pixar character is like hundredth of a penny”
For generative products, the model itself is the app because it is magical enough that users find value directly, so investors should back the best team building the best model in music, speech, video, and images.
“I'm a firm believer that these models train on human data, training stuff, the model is the app.”
Venture investing should be judged on expected value (probability times payout), so a bet like solving free energy has infinite expected value despite near-zero probability, and portfolio-level thinking matters more than individual outcomes.
“The expected value of solving free energy is infinite, but the chance is that it happened is almost zero.”
The first popular video on the Internet was a Cambridge researcher's coffee-pot camera, which went viral because it sparked the idea that many such feeds could be built.
“It was a coffee pot. I think this is guy in Cambridge, some researcher.”
Venture investors cannot know the expected outcome of any given bet, so even in efficient markets the way to place bets is across a portfolio, since any independent bet may look silly.
“The question is how do you extensively place those bets when you don't know for any given bet, what the expected outcome is going to be.”
Data pipelines are shifting from SQL-friendly structured warehouses to ingesting large volumes of unstructured video, images, and music that must be processed reproducibly, so unstructured and structured data workflows are converging.
“So it's almost like the unstructured role and the structural world are converging”
Although dot-com valuations were widely called crazy, the Internet's growth became the growth driver of the world economy for twenty years, more than making up for overvaluation, and AI will likely follow the same pattern.
“everybody said that dot com was crazy and they overpaid, except for because literally the growth driver of the world economy for 20 years”
Whether to back deeply technical AI founding teams or commercial teams that can apply a modality is an open debate at a16z; Casado leans toward deep technical teams from his own research background, though he admits it is not always right.
“Do you back the core technical team or do you back the commercial team and then get the modality?”
Querying large pools of unstructured video in a way comparable to SQL is not yet solved, so the industry must mature its data tooling for labeling, collection, and querying unstructured data.
“they're not solved in the sense that somebody can't just take a big pool of video and then be able to query it just like SQL right away”
Paying thousands of people to write code snippets or annotate sentences in post-training is already happening, but a thousand people writing code cannot compare to the volume of all code ever written, so this data source will slow the scaling curve.
“you can't compare like a 1,000 people writing code with all the code ever written before”
Releasing open-source models such as Facebook's Llama without an enterprise sales force means the model lands with users but is not pushed and pulled toward what enterprise buyers need, while Mistral pairs open source with customer engagement.
“when you release open source without actually having an enterprise sales force and go-to-market engine around it, I think it kind of lands people to do whatever they want”
AI is producing a rising number of PhD-led CEOs with deep technical backgrounds, such as George Fraser of Fivetran, suggesting AI is closer to science than to pure engineering or business.
“you are seeing a horizon of PhD CEOs. It's actually pretty rare.”
The key takeaway from Bostrom is not that AI takes over, but that we probably live in a simulation and are about to create the next one, which Casado offers as a closing reflection.
“we probably are living in a simulation, and we're probably about to create the next one.”
The California AI bill that Casado calls horrible was authored by a Berkeley professor and her student, showing that academia is complicit in pushing regulation that protects against models that do not yet exist.
“the California bill, this horrible bill in California... it literally came from a Berkeley professor and her student.”
Ideogram was an obvious investment because its team, which included Google Imagen pioneers of diffusion models, built an image model that quickly reached very high quality.
“We invested in Ideogram. This was like the Google Imagen team is one of the pioneers of diffusion models.”