Erik Bernhardsson
Related people
Episodes
Notes
The Twenty Minute VC: Venture Capital | Startup Funding | The Pitch · 20VC: Who Wins the Inference Market | Has Open-Source LLM…
-
Is Nebius a Better Business Than CoreWeave?
Erik Bernhardsson says he does not know whether Nebius is a better business than CoreWeave. Both, in his view, have executed extremely well. What he respects most is the financial engineering. A neocloud locks in a long-term customer, raises money against that contract, buys the chips, delivers them, and pays the debt down over the following years. He says CoreWeave has done that exceptionally well. The constraint on building a neocloud quickly is that financing, not only the servers.
He treats the top neoclouds, Nebius included, as friends and foes at once. They are trying to move up the stack into inference, and he is somewhat skeptical. Being excellent at the core business and entering a
-
Is CUDA Still a Moat for NVIDIA?
Erik Bernhardsson says CUDA is a little overstated as NVIDIA's moat. He does not think CUDA itself is the moat. He calls it an annoying programming model.
The advantage he names is the NVIDIA ecosystem. Kernel work and Torch support NVIDIA much better than other chips, and open-source engines, including SGLang, do the same, so the stack stays NVIDIA-first. For 15 years he has said that next year would be the year of non-NVIDIA accelerators, and it has not happened. Trainium, TPUs, and AMD look quite good and are catching up. He still thinks they are quite behind.
He says he has been wrong about that call every year. NVIDIA's staying power is what has kept it dominating the market.
-
Will GPU Prices Come Back Down?
Erik Bernhardsson says GPU prices will come back down, though he cannot say when. Hardware prices are like gravity and fall in the very long run. A supply shortage has pushed prices up, so the same number of GPUs costs more, and paying more than other buyers is what gets a large allocation. NVIDIA is shipping an enormous number of GPUs. He does not know how many hundred thousand Blackwells are being installed, and he does not call the market efficient. Prices may keep rising for a year, two years, or five years, until supply catches demand.
Modal buys on a three-to-four-month lead time. He draws a straight line or an exponential line through demand from a diversified set of customers. Growt
-
Can You Raise a Venture Round Without a Pitch Deck?
Erik Bernhardsson says he has never made a pitch deck, and the rounds still closed. He kind of hates fundraising. It is fine when the people are good, and the hard part is saying no to a lot of investors. Every round so far, he got to know people well before asking for money. The Series D announced the day before, with Accel, moved quickly because he had been talking with those investors for at least a year, and only then did they discuss terms.
He does not treat the round as a game of getting the highest price. Later stages can make the money itself more like a commodity. Early on, he does not chase the top price.
-
Who Will Win the AI Inference Market?
Erik Bernhardsson says the AI inference market will be enormous, like a new cloud, and inference alone will not be enough to win it. Selling tokens for dollars is a commodity. There is no switching cost, and customers can move to another provider. That is a race to the bottom.
What lasts, in his view, is a suite. Inference has to come with post-training, sandboxes, reinforcement learning, low-latency inference, and batch inference. Buyers want a set of tools for building AI applications, not one trick. Modal has been building that platform for five years. He does not call the breadth a moat. He calls the infrastructure underneath a very hard product to copy.
-
What Does Erik Bernhardsson Think Makes a Good Dad?
Erik Bernhardsson treats being a good dad as setting his two kids up for success in life, and he always thinks about the right amount to push them. He is not a tiger dad, but kids need a little push so they get exposure.
He makes them play piano. He does not know why piano in particular, but he thinks it is a good way to learn something he hopes they will appreciate at 20 or 30. They do not like it that much. The push should be a little, not too much. He does not want a future where they tell him they wish he had made them do something, and he does not want them at 20 asking why he made them do it.
His parents were very lax. He discovered math and programming, and they asked why he was doi
-
How Much Do AI Labs Pay New Graduates?
Erik Bernhardsson has lost new graduates to some of the big labs for 20-plus million. The people are exceptional: extremely sharp, hungry, and hardworking, and they know everything about technology. They go to the big labs, which pay a lot. A few other employers hire from the same pool, while the big labs are the ones known mainly for paying a lot of money.
The company pays a lot, can go very aggressive, and wants to sit at the 95th percentile in the market. A few companies will probably pay a little more. The equity has way more upside, so the package does not have to match the labs. The pitch is that the compensation is basically better, and they will go very far to hire the best talent.
-
Are Software Engineers Going to Be Fine?
Erik Bernhardsson says software engineers are going to be fine for the next 10 to 20 years. Work an agent can finish within 24 hours or a week will probably get automated. Engineers work on a longer timescale, with longer feedback loops, and reinforcement learning will start to run into limits around that. He has been wrong for so long, and reinforcement learning may eventually reach tasks that take a human six months.
He has been coding for 30 years, through new programming languages, the cloud, and many frameworks. Every time programmers become more efficient, demand for software increases, because many parts of the world desperately need software. Translating English into code has never
-
Has Open-source LLM Inference Peaked?
Erik Bernhardsson says open-source LLM inference has not peaked. Both open-source inference and frontier providers will grow 5x, 10x, or 20x over the next 12 months and beyond, and both will grow by orders of magnitude over the next few years. Which side ends up larger is hard to call.
More resources should have gone to language-model inference earlier. The work sat on the sidelines because the team did not realize how big open-source inference would be and expected a race to the bottom in a commodity space. Staffing then jumped from one or two people to 30 people 12 months later, and Modal is now one of the top three providers.
He calls the peak argument a fixed-pie view: frontier provide
-
How Did Harry Stebbings Thank New Followers?
Harry Stebbings sent a personal thank-you to every follower who followed him before 50,000.
He typed each note himself. Every night from 1 till 2.30 a.m. he sent 50 to 100 messages. One note said he saw the person was from Sweden, that he loved the Grand Hotel, and that they should let him know when they were in London.
He treated the messages as a customer acquisition flywheel. The next day a recipient would walk into the office, mention that Harry Stebbings had written, and tell coworkers the show was pretty good and worth a listen.
-
Is Buying GPUs Like an Open-air Drug Market?
Erik Bernhardsson says buying GPUs at NVIDIA's GTC conference in San Jose felt like an open-air drug market, with GPUs instead of drugs. He ran around talking to about 500 cloud vendors, asking who had GPUs, and came out with a lot of deals. Getting that capacity is pretty crazy, so Modal does the deals under the hood and customers do not have to chase them.
Modal works with 25 different neoclouds. He puts Modal's share of the world's compute at 0.1% or whatever, and he does not know the exact number. The long-term goal is to control more of that compute, aggregate capacity and demand, and manage both efficiently. Modal built its own file system, container runtime, and scheduler, and does n
-
Do Companies Still Have Software Margins?
Erik Bernhardsson says almost no company today has software margins anymore. The 90% margins plus that Silicon Valley showed 10 years ago are basically gone. Modal has a long-term opportunity to get to 40%, 50%, 60% margins, and Modal is never going to get to software margins.
The reason is topological. The stack of vendors runs through applications, models, infrastructure, cloud, and Nvidia. When a customer pays a dollar at the top of the stack, the dollar has to flow through a lot of different layers, and at every layer gross margins are going to be 20%, 30%.
There is a lot of margin opportunity and a lot of opportunity to make a lot of money, because the space is so big. The dollar is s
-
How Much Does Modal Burn Each Month?
Erik Bernhardsson says Modal burned only a few million a month up until last year, and monthly burn is now a lot. Modal is profitable on a profit and loss basis, while cash flow is quite negative. Heavy spending on GPUs and on building out capacity is what pulls cash flow down.
Modal has not been capital constrained for most of the life of the company, and the business has been extremely capital efficient. A few million a month already counts as a lot of money. Last year the spending pattern changed, because capacity now takes a lot of money. Profit on the books and negative cash flow describe the same period: the company makes money on a profit and loss basis and still sends a lot of cash
-
Why Did Erik Bernhardsson Wait Until His Mid-30s to Start Modal?
Erik Bernhardsson started Modal, his first company, at 36 or 37 because he never had quite enough confidence in himself, and a little of the money was missing too. He does not think he had ever necessarily wanted to start a company. He saw a big opportunity in the market, asked himself why no one else was doing such an obvious idea, and that led him to start Modal. Five years later, it has gone pretty well.
He looks at Stanford kids who are 21 and feels they maybe have too much confidence, already sure they will take over the world. The ideal age, in his view, is maybe 32, with a bit more experience in how to navigate people, hire people, and fire people. He does not know whether he would h