AI has shattered the long-held iron rule of the venture capital world, and the $67 billion in hard facts is conclusive proof of this.
The Shen Translation Team is a translation and compilation division under 36Kr, focusing on sectors including technology, business, workplace and life, with a spotlight on introducing new technologies, new perspectives and emerging trends from overseas markets.
Editor's Note: $67 billion realized in just 7 days! An a16z partner reviews the underlying truth behind this staggering profit: the manpower-intensive tactic is obsolete, computing power leverage is rewriting the long-standing iron law of venture capital returns, and a 20-person team is fully capable of leveraging sky-high valuations. This article is adapted from translated content.
As a core partner at a16z, Martin Casado has personally experienced three waves of AI development. None of the previous waves, however, have presented a scenario as dramatic as the one unfolding right now.
The current reality is that within a single week, SpaceX completed the $60 billion acquisition of Cursor, and Stripe also finalized a deal to acquire OpenRouter for more than $7 billion.
Both of these transactions involved projects backed by a16z. Within 7 days, $67 billion of value was realized.
Throughout the history of this industry, we have never been able to "invest 10 dollars and immediately see tangible returns of real money". But now, it is absolutely possible to invest 10 dollars and directly and intuitively recoup a considerable amount of returns.
Casado is no armchair strategist. He joined a16z in 2016 and currently leads its underlying infrastructure investment segment; prior to that, Nicira, the company he founded, was acquired by VMware for $12.6 billion — an extremely shocking blockbuster deal at that time.
To save you time, I have gone through the entire interview on your behalf.
Below are 10 core insights worth paying attention to.
1. The "10-dollar rule" that never came true before has paid off twice in just one week
Every previous AI wave has promised such grand visions of capital returns, but none of them have ever been truly delivered.
Casado joined a16z in 2016, at the time of the AI investment boom centered on drones and autonomous driving that he refers to. Over the following decade, he witnessed the awkward situation of capital being burned recklessly. Ten years ago, if you gave $1 billion to a startup team, what you often got in return was only staff expansion, luxurious offices, and total failure in the end.
Now, if you invest $1 billion in a cutting-edge lab, the fund can be directly converted into GPU computing power hours and tangible products that users are willing to pay for within the same quarter.
The most scarce resource has shifted from manpower to computing power. However, most venture capital firms still follow the old 2015 mindset when it comes to the logic of risk pricing.
The two landmark transactions in one week have fully verified his judgment. Both SpaceX's acquisition of Cursor and Stripe's acquisition of OpenRouter closed the loop of huge capital and turned it into mature products within just a few months — a return cycle that would previously take at least ten years in the models of all traditional funds.
2. A 20-person team built a model worth $2 billion
Casado only used one simple sentence as evidence to prove this "new capital logic".
A well-known large language model in the industry... actually only has a core team of about 20 people. I estimate that its R&D and training costs probably exceed $2 billion.
He did not name the specific model in his public remarks. But this extremely disproportionate input-output ratio speaks for itself: a landmark product serving millions of users has a core team that is small enough to fit around a single conference table.
In the past, the ceiling of ambition was often limited by the size of the workforce. Once the engineer team becomes too bloated, collaboration losses will accelerate cash burn and dilute the product roadmap — this is exactly the "mythical man-month" curse that all software teams will eventually encounter. Now, the new ceiling is capital and computing power, and their expansion rhythm is completely decoupled from the old manpower logic of the past. In the past, 20 people were only enough to support a prototype team at the seed round stage; today, 20 people can manage asset allocation worth up to $2 billion. At the micro end of the same leverage effect, even a super individual (one-person team) can leverage equally astonishing output.
3. His prediction: Top labs take 80% of the capital, but lose 60% of the total API calls
If you ask Casado who will win in this AI ecosystem competition, he will give you two numbers with completely opposite trends.
If I have to make a prediction, the shortage of computing power supply will probably be alleviated around 2028. Weighted by capital value, I think large cutting-edge labs may capture 80% of the market share in the future, because this is the monopoly pattern that traditional giants have always demonstrated throughout historical cycles; but weighted by token consumption, I believe 60% of the actual call volume will eventually flow to long-tail applications and the open source community.
The huge gap between these two numbers is exactly the essence of the whole logic. Calculated by revenue value, top labs take away the vast majority of real money; but calculated by the actual processing scale of tokens, the long-tail forces composed of open source models and small and medium-sized service providers will carry the vast majority of call loads — because their application scenarios are more diverse and penetrate into a wider range of business contexts.
This differentiation is not obvious at the moment, because several top labs control the huge centralized procurement advantage of computing power that ordinary people cannot reach. The 2028 node Casado marks is exactly the moment when this computing power barrier begins to collapse: once GPU supply catches up with demand, the AI application ecosystem will, like all technology stack evolutions in history, move towards full fragmentation and decentralization.
4. Why he never mentions "recursive self-improvement"
When you ask about the compound interest effect and self-evolution of AI, he will correct your wording before he starts to answer.
The so-called "recursion" means that you copy something completely and entirely onto itself. "Self-catalysis", on the other hand, means using the thing itself as a tool to help you produce and iterate it faster.
Under his cognitive framework, the so-called "recursive self-improvement" implies a positive feedback loop that replicates itself in full, eventually gets completely out of control and no one can govern. However, the actual dynamics of major labs at present are more pragmatic and focused: using existing AI as a production tool to accelerate the development of the next generation of AI — just as since humans invented the first compiler, software engineers have been using old software to write new software.
He defines this as a "self-catalyzing reaction", and points out that this pattern has played out repeatedly over the past few decades. It is constrained by the physical boundaries of capital and data, but continues to generate compound interest at an extremely steady pace within those boundaries. An out-of-control loop means no ability to intervene, while the compound interest advantage means that a few leading labs can continuously expand their moat, and other players still have space to observe and respond — this is the essential law behind all self-evolution loops that are worth betting on.
5. The "model routing scheduling problem" that even he cannot solve
All developers building upper-layer applications on top of large models will eventually encounter the same ultimate proposition: which model should this specific question be assigned to for an answer?
Model Routing is indeed an extremely challenging technical barrier. I think it belongs to the "AI-complete" problem. Just imagine: to figure out "what questions should the smartest being in the universe answer", you must first have the smartest being in the universe to give you the answer.
Casado splits the "routing scheduling" concept that most developers confuse into two completely different directions. The first is "quality-based routing", which means accurately selecting the model with the best answer quality every time, but this is close to a logical dead loop: you must have a model smart enough to score the outputs of all other candidate models in real time within milliseconds, and there is no model with such capability at present.
The "cost-based routing" has been successfully implemented in practice now. You only need to select the model that meets your quality threshold at the lowest possible price on the Pareto optimal frontier, just like what Cursor's own routing mechanism and OpenRouter are already doing. This is the routing product that is really worth building with all efforts now: find the model that is just good enough at the lowest possible cost.
6. Arbitration tricks used by coupon collectors for $200 subscription plans
The most astonishing reveal Casado dropped during the entire conversation has nothing to do with the model itself, but about how some people exploit loopholes in monthly subscription bills.
In China, some extremely shrewd grey-industry teams specifically use individual service plans to carry out risk-free arbitration. They subscribe to a $200 monthly plan, burn through all the token quota for the month in just three days, and then immediately cancel the subscription; even if they have used up all the quota, the system will still refund most of the payment proportionally based on the remaining 27 days.
Top cutting-edge labs are actually deliberately subsidizing heavy users. The marginal cost of obtaining token calls is relatively controllable, but once a developer's business is deeply tied to a certain API, its churn cost will be extremely high, so this subsidy is essentially a user growth leverage. Casado points out that the vast majority of losses come from the top 5% of users, and major vendors are now continuously tightening account risk control and quota limits to curb this loss loophole.
The arbitration team he mentioned precisely captures the rule loophole between the "fixed subscription system" and the "usage-based billing system", and exploits it extensively: drain the full-month quota at a very fast speed, cancel the subscription immediately before the next billing cycle, and successfully cash out the proportional refund for unused days. Casado jokingly calls it a different kind of "brand new routing mechanism" — instead of scheduling between different models, it arbitrages loopholes in the rules of subscription plans.
7. Why enterprises will not easily replace old models even if better models come out
The traditional consensus is that large models are no different from undifferentiated commodities, and once a stronger competitor comes out, the old model will be ruthlessly abandoned. Casado says bluntly that this view is completely wrong.
Practice has shown us that the user stickiness of these models is far higher than outsiders imagine. Everyone is shouting "one-click replacement is possible at any time", but very few people actually do that in reality. I have already purchased a large amount of computing power quota from OpenAI in advance, why should I replace it?"
In business decision-making, model procurement is often determined by business procurement processes rather than benchmark rankings. Once an enterprise allocates budget to a service provider and completes system engineering docking, huge migration costs will appear in many links that have nothing to do with "model quality": prepaid quota, long-term contract terms, and the surrounding toolchains that have long been polished around the specific API.
Procurement inertia itself is an unbreakable moat.
It is this extremely high stickiness that gives strategic value to aggregation management platforms — a unified dashboard that can integrate Anthropic keys, OpenAI keys and the OpenRouter backend has long exceeded the value of the routing function itself. It holds the distribution channel leading to millions of users on one hand, and forms a hedging card to prevent enterprises from being completely locked in on the other.
8. Marketing has completely become a pure financial decision
In the past, marketing expenses were always the most difficult item to justify and defend in board meeting presentation materials. Casado says AI is completely reshaping this work into a mathematical problem.
Today's CMOs have basically become CFOs wearing marketing vests. In the past, people would endlessly discuss the so-called "art of marketing": paid traffic placement, content construction, offline summits, social media matrices. Now, the core topic has directly become whether we should increase subsidies or reduce subsidies?
Offline customer acquisition, content marketing and brand activities all have a long-standing pain point — when you invest one dollar, no one can accurately tell exactly how much return it brings. AI completely eliminates this ambiguity, because by directly subsidizing token call volume, you can convert it directly into real registered paying users with a certainty that no traditional channel can match.
Every new position created in the future will essentially be a CFO wearing a trench coat.
He adds that technology R&D is also being rapidly flattened. In the past, to build a technology company, you had to go through a long list of extremely tedious technical decisions; now, it is increasingly simplified to one core proposition: can you raise enough real money to buy the GPU computing power you want.
9. What exactly did SpaceX spend $60 billion to buy? The answer is by no means just the large model itself
Casado distills this largest unlisted enterprise acquisition in tech history into four core elements, and "money" is only one of them.
To simplify the logic of this transaction to the extreme, one party holds proprietary data, one party has massive computing power, one party owns distribution channels, and the other has abundant strategic resources enough to support staying at the cutting edge.
Cursor itself is already a highly explosive business model. It brings high-value real programming behavior data, and an enterprise engineering culture that values "delivering products quickly" more than "writing and publishing papers": Casado mentions that the founders invested 30% to 40% of their energy in recruitment and cultural construction, and even with top researchers in the team, they always firmly adhere to the product-first route.
One of my favorite cases is their design lead Ryo Lu. He independently developed a project called Ryo OS, a simulated operating system full of classic retro Mac style.
SpaceX completes the other half of the puzzle — a supercomputing scale that a startup can almost never reach on its own, and a strong balance sheet that can continuously spend heavily to acquire resources. This is the core buyer logic in all acquisition deal structures.
10. He will never dictate what you should do, and that is exactly the key point
If you go to an average venture capitalist for advice on emerging opportunities, most of them will sell you a pretentious industry track theory; but Casado only gives you one code of conduct.
I am a hill climber (a hill-climbing optimization seeker in algorithms). My greatest passion has always been cutting-edge technology, startup ecosystems, creative destruction, and pure innovation. I am extremely bullish on Silicon Valley in the long run. As long as there are still mountains worth climbing in front of me, and as long as I can keep climbing on this land I love, I will feel extremely satisfied.
He never bothers to predict the so-called "next trend", which means that if a founder asks him "what should I do", they are asking the wrong person from the very beginning. His real screening mechanism is "Founder-Market Fit": whether the growth trajectory of a specific founder highly resonates with the real pain points of the target market, which he considers far more important than pure personal intelligence. He would rather place big bets on three or four extremely strong founders in vertical tracks he fully understands, than blindly chase those exaggerated track narratives that he cannot even justify himself — looking back at every winning business plan he has overseen and finalized, all of them fit this underlying logic without exception.
Practical Guide
The dramatic change that capital can now be directly materialized into launched products within a few months is completely rewriting the winning rules of the business landscape.
For founders: Stop obsessing over model parameters or benchmarking quality. Casado emphasizes that even if you are only slightly better than competitors by a tiny epsilon margin, you have already secured pricing power. Therefore, focus your energy on the product interaction layer built around the model, distribution and customer acquisition networks, or the business models of underlying scheduling routing.
For investors: Abandon the black-and-white zero-sum game perspective. Before hastily concluding that a company is overvalued, carefully evaluate the core strategic hubs it firmly occupies, rather than only focusing on the current financial statements — this restraint and insight in due diligence dimensions is the key line that distinguishes true conviction from blind follow-up.
For business operators: Conduct a comprehensive inventory of your team's token consumption, to verify the stickiness logic Casado pointed out. If you have not re-evaluated your model suppliers for reasons other than "model quality differences" for more than 6 months, it means you are already deeply trapped in procurement inertia, and you must include this in your optimization agenda for the next quarter.
For everyone else: Pay close attention to the "80/20 and 60/40 market differentiation pattern". If his prediction holds, top labs will take away the vast majority of revenue, while the huge peripheral ecosystem will occupy most of the call volume — this huge market gap is the perfect hotbed for nurturing new unicorns.