Erik Bernhardsson

1. Key Themes

Inference Alone Is a Commodity; the Platform Is the Business

Erik repeatedly argues that selling tokens for dollars is a race to the bottom and that the durable business is a bundle of adjacent primitives (sandboxes, post-training, RL, batch, low-latency inference) on top of hard-to-copy infrastructure. Modal's second product, sandboxes, is already about a third of revenue.

"If you're just selling like tokens for dollars, there's no switching costs, you know? People can just switch to a different provider. It's just a race to the bottom." [00:14:30]
"In the infrastructure underneath is a lot of differentiation, a lot of, you know, staying power in the market, cuz it's very hard to copy like the infrastructure required to build all of these products." [00:14:59]
On the data: "sandboxes today is a $300 million run-rate business. It's about a third of our revenue." [00:34:33]

The Inference Market Fragments Into Specialized Submarkets

Because the market turned out 10-100x larger than anyone expected, it will split into many specialized products, which Erik sees as a chance to "deconstruct the cloud and rebuild it for this new era."

"It's like 10x or 100x bigger than I ever thought. And that, to me, means there's going to be so much room for specialization. Low latency, high latency, like various types of fine-tuning and post-training." [00:16:41]
"That creates a lot of opportunity to basically kind of rethink what a cloud looks like in the era of AI and kind of deconstruct the cloud and rebuild it for this new era of software." [00:17:43]
"When a market is starting to get 10 billion, 100 billion, whatever, like you're going to have a lot of different submarkets." [00:30:15]

Software Margins Are Dead Because of Layered Stack Economics

AI value chains pass each dollar through applications, models, infrastructure, cloud, and NVIDIA, with a 20-30% gross margin taken at each layer. Modal's own ceiling is 40-60%.

"At every layer, there's going to be, you know, 20%, 30% gross margins." [00:15:46]
"The 90% software margin businesses are gone. I think Modal has an opportunity long term to get to 40, 50, 60% margins, but we're never going to get to software margins." [00:16:15]

Capacity Aggregation Is the Hidden Structural Advantage

Modal controls its own file system, container runtime, and scheduler, and treats 25 neoclouds' capacity as one pool. Pooling many unpredictable customers produces smooth aggregate demand, which makes capacity planning tractable and gives customers instant burst capacity.

"We work with 25 different neo clouds... as we control more and more of the world's compute, you can aggregate that capacity and aggregate demand on the other side and manage that capacity in demand extremely efficiently." [00:18:28]
"If you pool all of those customers together, you end up with a very sort of smooth overall demand." [00:19:20]
On forecasting: "monthly growth rates, which have been about 30, 40% month-on-month for the last 12 months. And you just project that, you know, 3 to 4 months into the future. You end up with like 2 to 3x, you go out and, you know, buy that capacity." [00:20:28]

Infrastructure Success Comes From Building Primitives and Surviving Until Demand Arrives

Modal's serverless function product (2021) found its killer app in Stable Diffusion and GenAI inference (2022-23). Sandboxes (2023) exploded in 2025 when models got good at generating whole apps. Erik admits they did not anticipate RL, vibe coding, or background agents.

"We put out a serverless function product in 2021. It saw success in 2023 with GenAI inference. We put out a sandbox product in 2023... and then it saw success in 2025." [00:35:37]
"Those were things we actually didn't anticipate... we didn't anticipate Stable Diffusion when we started building our serverless function product. It just happened to be the killer app." [00:35:37]

Hardware Prices Fall Eventually, but Demand Has Outrun Supply

Erik declines to forecast GPU pricing and invokes "hardware gravity." Meanwhile, demand is so latent that overbuilding in the next five years seems implausible. The real risk is customers who aren't yet profitable running out of funding.

"Hardware prices always go down in the very long run. Right now, we're in a very, you know, the, there's a supply shortage... They may continue to go up for another year, 2 years, 5 years. I don't know, but eventually it will come down." [00:21:48]
"There's a lot of companies that are not yet making money yet, and they're using a lot of GPUs. They just have to, eventually, start to make money. But like that lag could take a while." [00:23:08]

Post-Training and RL Work Where Feedback Loops Are Short

Erik expects RL to excel in verifiable domains (code, math via Lean, cybersecurity, maybe law) and sees a limit where feedback loops run long.

"Coding is like a great example, right? Like you can do coding with reinforcement learning pretty well because... you can run code in sandboxes, and a lot of labs are running Modal sandboxes for that exact purpose." [00:41:45]
"There's a lot of tasks out there that like have way longer feedback loops that I don't know if like post-training quite yet are is within reach... I have a feeling like reinforcement learning, we're going to start to run into some limitations around that." [00:41:45]

Brand, Personality, and Speed Substitute for Moats in Early-Stage AI

Erik dislikes the word "moat" for early companies and prefers the "gingerbread man" strategy of running faster. He believes brand value has risen sharply, and that founder-led marketing drove Modal's early growth.

"You never have moats in the early days... It's like none of those things happen until you're like 10 years in." [00:10:24]
"For almost every startup these days, it's like you just have to like move faster than everyone else." [00:10:58]
"In the early days of Modal, like that was a big part of our growth, right? Like me just like tweeting all the time about what we're building... Half of them ended up angel investing." [00:31:54]

Software Engineering Demand Expands, Not Contracts

Erik, 30 years into coding, argues each efficiency leap has increased demand. Modal spends about $500K a month on coding agents across 150 engineers.

"Every single time we make programmers more efficient, it just increases demand for software." [00:43:08]
"One of the biggest misconceptions from the outside is that developer is about like translating from English to code. That has never been... the job of software engineers." [00:43:08]

2. Contrarian Perspectives

Open-Source LLM Inference Has Not Peaked, and the Pie Isn't Fixed

Harry relayed Jason Lemkin's thesis that open-source LLM inference peaked as frontier providers get cheap and US enterprises avoid Chinese models. Erik rejected the zero-sum framing and said both will grow by orders of magnitude.

"I think that's a little bit of a fixed pie size mental... I do think both spaces will grow 5x, 10x, 20x over the next 12 months and and beyond." [00:38:18]
Modal's own experience supports this: after DeepSeek they went "from like 1 or 2 people working on this to like 30 people 12 months later, and now we're one of the top three providers." [00:37:24]

CUDA Itself Is Not the Moat; the Ecosystem Is

Against the conventional NVIDIA bull case, Erik says CUDA is annoying to work with. The real lock-in is kernel optimizations, PyTorch support, and open-source engines like vLLM and SGLang being NVIDIA-first.

"I don't think CUDA itself is the moat. CUDA's actually kind of annoying to work with... But I do think all the NVIDIA ecosystem is." [00:24:32]
"For the last like 15 years, I've been saying... 'next year is the year of non-NVIDIA accelerators.' It never keeps happening, I'm always wrong." [00:24:56]

Neoclouds Moving Up the Stack Will Struggle

Despite Nebius and others explicitly targeting inference, Erik is skeptical that infrastructure-financing businesses can build a complex software platform. He frames neocloud excellence as financial engineering.

"It's very hard to be good at like your core business and also trying to move into different market at the same time." [00:25:42]
"A lot of the constraint for building a neocloud very quickly is on the financial side. You need to... lock in like a long-term customer contract, then you go raise a bunch of money against it, then you go buy the chips... It's a capital game. It's a financial game." [00:26:54]

Go Smaller on Capital and Keep VCs Hungry, Even in AI

Pushed by Harry's thesis that capital is the new moat, Erik says Modal was extremely capital-efficient for most of its life, and that VCs' incentives are misaligned toward over-funding.

"It's often better to raise a lot less money at a lower valuation and then keep the VCs hungry for the next round." [00:08:55]
Capital became more relevant only recently: "Capital helps at this point" once Modal had to buy GPU capacity. [00:09:52]

Owning Your Own Intelligence Is Overrated as a Universal Strategy

Even though Modal could benefit from every enterprise training custom models, Erik says it only makes sense for certain use cases, and cites latency, cost, performance, hackability, and adjacent services as equally important.

"I don't think it makes sense for everyone to own their own intelligence." [00:47:41]
"Owning your intelligence is like just one thing out of many." [00:47:56]

The Data/Environment Business Is a Great Cash Business but Maybe Not a Venture Business

Erik thinks selling RL environments to labs is lucrative now but may not endure.

"If I was 20, and I just wanted to make a lot of money, I would go out and start an environment company... you make a lot of money today, but like is this a venture fund... I don't know if they're going to make a lot of money in the very long run, but like who cares?" [00:45:53]

3. Companies Identified

Modal

Serverless AI cloud infrastructure (functions, sandboxes, training, batch) co-founded and led by Erik. Just announced a $750M round led by Accel at a $15.75B valuation, with $1B in ARR and 15x growth in 12 months. Why mentioned: Core subject. Growing 30-40% month-on-month, profitable on P&L, with negative cash flow due to GPU buildout.

"We grew 15x in the last 12 months... we're growing 30, 40% every month." [00:39:09]
"I wouldn't do this if I didn't think we could grow and become like a generational company." [00:39:50]

Accel

VC firm leading Modal's most recent round; Erik cultivated the relationship for more than a year before raising. Why mentioned: Lead investor in the round; example of Erik's relationship-first fundraising with no pitch deck.

"I've been speaking to Rich and Gonzalo at Accel for I think at least a year... And then it happened very quickly, cuz we already had that relationship." [00:08:24]

Typeface (Jev)

A specialized model company whose founder, "Jev," has a distinctive and unhinged personality-led brand. The transcript garbles the names ("Typesafe," "Jev," "Jas-"), so the exact company is uncertain. Erik describes a classification/regression-focused model company that launched overnight on Modal. Why mentioned: Erik's example of brand-as-asset and of Modal's burst-capacity value prop. He said he'd have invested at $8.5B and sees a $100B opportunity.

"I think there is an opportunity to create a $100 billion company with a lot of specialized models and inference." [00:00:00]
"They launched overnight on Modal, and we weren't even aware of it. They just like scaled up, and we gave them a ton of GPUs. And it just happened automatically." [00:19:47]
"Everyone wants to clone [him]. No one wants to get up on stage and wear a fucking pink fur coat and just like talk about the... revolution in an unhinged way. That is what brand, to me, is." [00:12:23]

Turbopuffer

Vector database company founded by Simon, a Danish entrepreneur and Erik's angel pick if he could only back one. Why mentioned: His single-pick angel investment; he sees many adjacent expansion paths.

"I love Turbopuffer. I know Simon really well, Danish guy... Just a fantastic entrepreneur, I just love the business." [00:54:44]
"If I ran that company, I would have 500 different ideas of how to expand into many adjacent things around databases, data warehouses, other types of query systems." [00:54:54]

OpenRouter

Model routing marketplace; Erik is a personal investor and said it had a great outcome. Why mentioned: Test case for whether model routing is defensible standalone.

"I invested in OpenRouter personally, and they had a great outcome. I actually think they could have done even better on a standalone basis." [00:50:07]
"The growth rate, and the revenue... And the brand frankly, like they have a really strong brand." [00:50:58]

Cerebras

Specialized AI chip company; Harry noted its stock is down 30%. Why mentioned: Example of the specialized-chip fragmentation alongside NVIDIA.

"I think NVIDIA's going to keep growing, they're going to keep shipping amazing chips. I think there's also going to be like Cerebras and all these Etched and all these other ones like shipping specialized chips, and they're also going to grow." [00:23:55]

Etched

Specialized AI chip startup, mentioned alongside Cerebras as part of chip fragmentation.

"...all these Etched and all these other ones like shipping specialized chips, and they're also going to grow." [00:23:55]

NVIDIA

GPU leader whose ecosystem dominates AI compute. Why mentioned: Erik credits ecosystem lock-in and says NVIDIA is smart to sponsor US open-source models via Poolside and Reflection.

"They've shown to be incredibly good at having that staying power and just dominating the market." [00:24:56]
"I think it's smart of NVIDIA to like buy some of those and sponsor them effectively in order to create like American-built models." [00:29:05]

CoreWeave

Public neocloud. Why mentioned: Erik praises its financial engineering execution.

"CoreWeave has done it exceptionally well. It's a capital game." [00:26:54]

Nebius

Neocloud moving up the stack into inference. Why mentioned: Harry's competitor discussion; Erik is skeptical of its upstack push but respects the execution.

"They've executed extremely well both of them. I have a tremendous respect for what we've done." [00:26:40]

Fireworks and Baseten

Inference platform competitors to Modal. Why mentioned: Erik says he is "paranoid about competition" but believes Modal has stronger community and brand vibes.

"I think we have like really, really good vibes. Like I'm terminally online, like people tweet like way more about Modal." [00:11:55]

Sandbox customers: Suno, Runway

Erik names these as examples of companies running proprietary models on Modal.

"Suno, Runway, Typeface, and many others, like running their own proprietary model on Modal." [00:48:53]

Poolside and Reflection

US open-source model companies that Harry says NVIDIA has acquired or sponsored.

"Does it matter if NVIDIA owns the open-source model layer in the US with first Poolside and now Reflection?" [00:28:57] (Harry)

Anthropic and OpenAI

Frontier labs; Erik says they pay new grads $20M+ packages, and that Anthropic's cheap, performant models create competition for open-source.

"We've lost new grads going to, you know, some of the labs for like... talking, you know, 20-plus million." [00:00:14]

Harvey and Legora

Legal AI companies with opposing views on specialized models. Harvey says specialization has been great; Legora says the underlying models move too fast.

"Harvey, who says that it's fantastic... and then you have Legora, who say, 'Ridiculous, the underlying models move so fast, it's the wrong strategy.'" [00:47:13] (Harry)

Mercor

Data/expert-labor marketplace; Harry cited Brandon (the CEO) saying they spend 3x human salaries on inference.

"Brandon at Mercor, um they spend 3x their human salaries on inference." [00:44:27]

Spotify

Erik's former employer; he attributes Modal's consumer-grade product thinking to his time there.

"The consumer background of coming from Spotify was very important in how I thought about building what's now a B2B company." [00:13:04]

Snap

Referenced via Ben Thompson's "gingerbread man" strategy of running faster rather than having moats.

"Snap doesn't have a moat, they just run faster. Like they're the gingerbread man." [00:10:35]

Vendors named: DeepSeek, vLLM, SGLang, PyTorch

DeepSeek's release forced Modal to staff up LLM inference quickly; vLLM/SGLang/PyTorch illustrate NVIDIA-first ecosystem lock-in.

"Then DeepSeek came out and we're like, 'Fuck, like we got to move very fast.'" [00:37:24]
"PyTorch always supports NVIDIA much better than anything else and the open-source engines like vLLM and SGLang because of that, they're always supported much better." [00:24:32]

Alternative accelerators: Trainium, TPUs, AMD

Erik says these look good but remain behind NVIDIA.

"There's like Trainium, there's TPUs, there's AMD, there's a lot of other stuff that does look quite good and they're catching up, but, you know, I still think they're quite behind." [00:24:56]

Sponsors: Blitzy, Flex, Vanta

Podcast sponsors (not discussed as excellence by Erik): enterprise autonomous software development, financial platform for owners, and compliance automation.

Lovable

Mentioned by Harry as part of Erik's angel portfolio.

"You have a great angel portfolio: OpenRouter, you mentioned Lovable before." [00:54:33] (Harry)

Claude Code

Harry cites it as a mother lode, "$60 billion of revenue" in his framing, and a key driver of Anthropic's valuation; Erik expects spend on it to keep rising.

"It's just going to eat more." [00:45:22]

4. People Identified

Founder and CEO of Modal, Swedish-born (Örebro), started coding at eight, competed in programming competitions at 16-17, former Spotify engineer. Why mentioned: Guest; built Modal from a 2021 start into a $1B ARR company in five years.

"I'm just like the biggest like optimist that there is in the world about software, and computers, and like I just love this shit." [00:45:22]

Jev (Typeface founder)

Founder of a specialized-model company with a charismatic, unconventional personality-led brand. The transcript's name and company are garbled, so the precise identity is uncertain. Why mentioned: Erik's archetype of brand-as-moat and a potential $100B outcome.

"His charisma, his like you have to create like a personality like around your brand." [00:31:26]

Simon (Turbopuffer founder)

Danish entrepreneur building a vector database.

"Just a fantastic entrepreneur, I just love the business." [00:54:44]

Rich and Gonzalo (Accel)

Accel partners who led Modal's Series D, with whom Erik built a relationship over about a year.

"I've been speaking to Rich and Gonzalo at Accel for I think at least a year." [00:08:24]

Jensen Huang

NVIDIA CEO. Erik cites his recent London remarks on pricing and supply.

"He said exactly the same thing I always tell people. Like hardware prices, like it's like gravity." [00:21:48]

Ben Thompson

Stratechery author whose "gingerbread man" Snap piece shapes Erik's view of moats.

"There's an old blog post I always come back to it by Ben Thompson." [00:10:35]

Eric Vishria

Benchmark partner; Erik echoes his view that NVIDIA and specialized chips will all grow.

"I think Eric Vishria was actually... the same thing. Like all of these things are going to grow." [00:23:55]

Zach Frankel

Mentioned by Harry as someone who described Erik's childhood and exceptional ability.

"I was talking to Zach Frankel last night, and he said about your childhood and like early indicators of success." [00:04:30] (Harry)

Jason Lemkin

SaaStr founder and Harry's weekly co-host; holds the thesis that open-source LLM inference has peaked.

"He always says that now we've reached peak open-source LLM inference usage." [00:37:49] (Harry)

Lin (Fireworks)

Referenced for the "own your own intelligence" thesis.

"Lin has said too from Fireworks, about the importance of owning your own intelligence." [00:47:13] (Harry)

Marc Benioff

Salesforce CEO; cited for spending $300M, about 5% of engineer salaries, on Claude Code.

"Marc Benioff, who says that they spend 5% of engineers' salaries, 300 million, on Claude Code today." [00:44:39] (Harry)

Brandon (Mercor CEO)

"Brandon at Mercor, um they spend 3x their human salaries on inference." [00:44:27] (Harry)

Neil Mehta

Greenoaks founder; Harry credits him with the "10x the customer experience" prompt.

"Neil Mehta has a great thing, which is like, 'How can I 10x the customer experience?'" [00:32:57] (Harry)

Hamilton Helmer (implied via 7 Powers)

Erik references the book but not the author by name; included for completeness.

"You go read like 7 Powers or like one of those things." [00:10:19]

5. Operating Insights

Build Relationships With Investors a Year Before You Need Capital

Erik has never made a pitch deck. He builds relationships long before raising, so when he needs capital, terms are negotiated rather than pitched, and speed is a byproduct.

"I like to get to know people well in advance of raising money... we're like, 'Actually, you know what? We could take some money. Like, what are the terms?'... And then it happened very quickly." [00:08:24]

Forecast Capacity With Simple Extrapolation on Diversified Demand

Modal projects aggregate customer growth, assumes 3-4 month GPU lead times, and buys a 2-3x multiple of capacity ahead. Diversification across thousands of customers makes aggregate demand predictable.

"You can actually look at the demand in aggregate pretty well and predict based on monthly growth rates... and you just project that, you know, 3 to 4 months into the future." [00:20:28]

Run Employee Secondaries With Guardrails to Make Equity a Hard Currency

Erik takes few secondaries himself but champions them for employees, capped at roughly 10%, as a recruiting weapon against labs paying $20M+ to new grads.

"Have a precedent of doing employee secondaries... Up to 10%. Like, I think you have to have guardrails." [00:52:21]
"It's a way to like create a very liquid, hard currency out of the equity." [00:52:21]

Staff Up Within Months When a Market Flips, as Modal Did After DeepSeek

Modal misjudged open-source LLM inference as a commodity. When DeepSeek landed, they reallocated from 1-2 people to about 30 in 12 months and became a top-three provider.

"We didn't realize how big open-source inference would be, and we thought it would be kind of a race to the bottom... And so we staffed up a team immediately." [00:37:24]

Hire More Forward-Deployed Engineers and Sales to Personalize Onboarding

Erik's 10x customer experience answer: more products plus FDEs and sales reps who proactively reach out to self-serve signups with domain-specific examples.

"We should hire like way more people that reach out to you personally and say, 'Hey, I saw you get started on Modal. I saw you're working in the bio space. How can we help you?'" [00:33:07]

6. Overlooked Insights

Modal Is "Long" GPUs Because Price Increases Pass Through to Customers

Erik says that in a tight market Modal is effectively more long than short GPUs. Its model locks long-term reservations and rents them on shorter durations, so rising underlying prices flow through to customers, and the company benefits from the spread. This reveals a structural hedge most people assume doesn't exist for neoclouds' customers, and the mirror-image risk if prices fall.

"We're actually more long than we're short GPUs cuz, you know, we can charge a higher price to our customers... we lock up longer term reservations and rent them out on slightly shorter duration." [00:21:20]

The Real AI Financing Risk Is Customer Funding Dependency, Not Compute Overbuild

Erik says overbuilding is unlikely in the next five years, but he flags the actual fragility: customers burning GPUs without revenue who rely on continual equity infusions. A venture pullback would hit GPU demand through this channel, and he predicts the pullback explicitly while staying bullish long term.

"Companies that are dependent on continual infusion of equity capital, I would not invest in them... if there's a venture capital pullback, they will struggle to raise more money." [00:55:39]
"There's going to be a pullback, you know, 20%, 30% in the stock market, everyone's going to freak out... I'm almost certain it's going to happen, but like I'm so bullish." [00:00:06]