Apr 14, 2026 · 1h 15m · news
The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of 20VC, AI investor and AMP founder Anj Midha discusses the scaling laws of frontier AI, the geopolitical battlegrounds of data sovereignty, and the early, heavily rejected days of Anthropic. He outlines his vision for treating compute as a standardized utility grid and advocates for a coordinated Western defense system to secure frontier model inference.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 16.4% of the talking time here. How this is scored →
Harry as informed peer 3.1 Guest teaching 6.0 Guest disagreement 3.4 Harry pushing back 2.9
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 21:11 Anj dismisses VC critics of Public Benefit Corporations
When Harry brings up an anonymous VC asking why PBC founders don't 'just win their market first', Anj dismisses the comment, telling the critic to call him when they want to invest in the fastest-growing business of all time.
Hardest push from Harry ▶ 25:21 Harry challenges compute provisioning as a VC loss leaderHarry directly asks Anj whether offering compute at cost is merely a loss leader strategy to force founders into granting venture allocation to his fund.
Biggest teaching moment ▶ 36:04 Anj details non-fungibility and memory bottlenecks of GPUsAnj breaks down the underlying technical reasons why GPU compute cannot easily flow between chip generations like H100s and Blackwell, resulting in massive GPU wastage across data centers.
Harry holds his own ▶ 1:02:24 Harry demonstrates hands-on technical experience with AI toolsHarry counters the perception of non-technical investors by explaining his direct experience using vibe coding platforms and hitting real technical bottlenecks with Supabase integrations.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| LLMs in Science: Debunking Diminishing Returns on Scaling Laws | 3 | 7 | 6 | 3 | Anj immediately and forcefully rejects Demis Hassabis' premise that scaling laws are diminishing, pointing to Periodic Labs' super-exponential gains in material science. He reframes scaling across specific domain modalities and outlines four key bottlenecks in AI: context feedback, compute, capital, and culture. | |
| Vertically Integrated AI and Domain-Specific Superhuman Capabilities | 2 | 6 | 2 | 2 | Anj defines domain-specific superhuman capabilities, explaining why specialized automated R&D labs operate on a completely different execution frontier than general coding models. | |
| Sovereign Data, the Cloud Act, and the Rise of Mistral AI | 3 | 7 | 3 | 3 | Anj educates Harry on the US Cloud Act and how sovereign data regulatory constraints created the strategic necessity for local European AI infrastructure like Mistral AI. | |
| Geopolitics of AI: Anthropic's Mission and Global Security | 3 | 6 | 4 | 3 | Anj recounts the early days of pitching Anthropic to Sand Hill Road VCs, mocking their lack of technical understanding when 21 out of 22 firms rejected the seed round. | |
| Public Benefit Corporations (PBCs) and Aligning Mission with Profit | 4 | 6 | 7 | 6 | Harry presses Anj with harsh criticism from an anonymous VC peer about Public Benefit Corporations (PBCs). Anj hits back forcefully, telling the critic to call him when they want to invest in the fastest-growing business in history. | |
| Securing Compute and Treating It as a Grid | 2 | 6 | 2 | 2 | Anj explains how his early compute procurement efforts at Andreessen Horowitz led to AMP's grid concept, coordinating compute like an independent system operator. | |
| "Back to the Future" Venture Capital and Incubation | 3 | 7 | 5 | 5 | Harry challenges Anj on whether cheap compute is just a loss leader to extract venture equity. Anj rejects this premise and delivers a deep historical overview of early Silicon Valley venture incubation. | |
| Lessons from the Industrial Revolution and Investment Frameworks | 4 | 6 | 4 | 3 | Harry quotes Brian Singerman's philosophy on picking founders rather than predicting the future. Anj counters that the safest way to predict the future is to invent it through systematic experimentation. | |
| Funding European AI and System Design in Infrastructure | 3 | 7 | 2 | 2 | Anj breaks down the capital requirements and financial structuring for European AI infrastructure, comparing continental gigawatt needs to Google's internal compute capacity. | |
| The Pre-Standardization Era and GPU Wastage | 2 | 8 | 3 | 2 | Anj delivers an in-depth technical explanation of why GPU compute lacks fungibility across chip generations, resulting in stranded clusters and infrastructure wastage. | |
| Misaligned Incentives and the Need for Open AI Standards | 3 | 7 | 2 | 2 | Anj differentiates statistical machine learning models from deterministic spreadsheets, calling for open standards and standardized government procurement protocols. | |
| China's Full-Stack AI Strategy and a Western "Iron Dome" | 4 | 7 | 2 | 3 | Anj details China's full-stack systems co-design and adversarial distillation strategy, advocating for a Western Iron Dome proxy to protect frontier models. | |
| Thiel's "Competition is for Losers" vs. "Optimal Competition" | 3 | 7 | 5 | 3 | Anj reframes Peter Thiel's 'competition is for losers' maxim, introducing 'optimal competition' while criticizing VCs for burning capital on dozens of redundant inference startups. | |
| Model Provisioning, General Products, and Market Segmentation | 5 | 6 | 3 | 4 | Harry cites a partner's thesis on model providers withholding state-of-the-art models for internal products. Anj uses general technology economics to explain product vs enterprise segmentation. | |
| The Transition from "Foundation Models" to "Frontier Systems" | 2 | 7 | 6 | 3 | Anj corrects host terminology, arguing that 'foundation model' is a misnomer created by VC associates and that winning companies build full-stack 'frontier systems'. | |
| Capital Scaling in AI and the Evolution of Venture Capital | 4 | 6 | 5 | 4 | Harry probes whether Anj's capital reserves are sufficient for massive Capex scale. Anj responds while criticizing legacy VC managers for missing key generational checks. | |
| Personal Philanthropy, Stanford Teaching, and the "One-Person Lab" | 6 | 4 | 2 | 2 | Harry demonstrates his own practical technical expertise by discussing vibe coding tools and Supabase integration bottlenecks, finding common ground with Anj. | |
| Key Attributes of Dario Amodei and Life's Scaling Laws | 2 | 3 | 1 | 1 | Anj praises Dario Amodei's empiricist mindset and shares personal reflections on time and family, leading Harry to share a poignant story about funding his mother's medical treatments. | |
| Escaping the Money Treadmill and Defining Success by Independence | 2 | 4 | 2 | 2 | Anj reflects on his path to financial independence from his Singapore scholarship days, describing independence as his primary core motivation over sheer wealth accumulation. | |
| Legacy, Growing Up Screen-Free, and the Future of Technology | 3 | 3 | 2 | 2 | Anj discusses his desire to be remembered as someone who accurately predicted the future, reflecting on his screen-free childhood in India and trading warm banter with Harry. |
Statements from this episode (48)
Midha: Human alignment is a bigger challenge than technical AI alignment
“AI alignment, don't get me wrong, is hard, but not the hardest problem. Human alignment is really the problem right now.”
Midha: West Needs an 'Iron Dome' for AI Inference to Keep Lead
“If we don't secure frontier model inference, or what I call state of the art inference behind a coordinated iron dome, I don't think we have a sustainable shot at staying at the frontier over the next decade.”
Midha: Superconductor discovery shows no signs of saturation
“There's no saturation in superconductor discovery at all.”
Midha: Coding AI benchmarks are super-saturated, yielding diminishing gains
“In, in certain domains that are well explored, like coding, for example, yes, there's an increasing amount of compute required to get an incremental gain in some eval that's super saturated.”
Midha: Periodic Labs operates a 30,000-sq-ft facility in Menlo Park
“This is my incubate, like my latest incubation is called Periodic Labs. I spent three days a week here. In, in Menlo Park, we have a 30,000 square foot facility”
Midha: Compute scaling yields super-exponential gains in superconductor discovery
“And I can tell you throwing more compute at the problem is probably having super exponential gains right now per iteration. So it depends on which domain you're talking about, which modality. There's no saturation in superconductor discovery, for example, at a…”
Midha: Algorithmic innovation in AI is primarily driven by company culture
“Algorithmic innovation, I think, is a function of culture, basically, because if you have the right culture, you get to attract the best researchers, the best research talent, then wants to work on pushing the frontier, and algorithmic innovation just falls ou…”
Midha: Claude and Gemini Failed Badly at Physical Science Benchmarks
“About a year ago, as an example, I realized there was a lot of talk about models being good at physical physics and chemistry, AI for science. And I was a visiting scientist at the applied Physics department at Stanford. And we started benchmarking these model…”
Midha: Scientific AI is bottlenecked by data locked in physical labs
“If you need physics and science, That's a real bottleneck because that data is locked up in national labs and academic labs. It's locked up in physical you know, semiconductor manufacturing plants.”
Midha: AI has reached superhuman capabilities in domain-specific tasks
“Within coding, within material science, within each of these domain distributions, we are seeing capabilities that are superhuman.”
Midha: Recursive Self-Improvement in AI Is Already Happening
“Some people would call that recursive self-improvement totally happening.”
Midha: Unique context feedback loops create superior AI moats
“Context feedback loops where you are, you have unique and differentiated access is where progress will be most legible to you. And if there are other teams who don't have access to that context, it'll also be where you have a superior business model.”
Midha: US Cloud Act gives US government access to American-managed data
“The, you know, the US Cloud Act says that, hey, if there's mission, if there's any data workloads, infrastructure, cloud workloads running on infrastructure that is managed by an American company, then the US government has to be able to access that data.”
Midha: Sovereign data needs open cloud market to startups after 15 years
“Because the context, the mission-critical context of those workloads is so important to be run locally that you can't run them on Amazon, AWS, GCP, or Azure, and it's the first time In 15 years that the sort of hyperscaler dominance is up for grabs for startup…”
Midha: Investment thesis for Mistral AI is a fully independent European stack
“Independence at scale, at every part of the AI infrastructure stack, like land, PowerShell in Europe, that's sovereign, it's local. Compute infrastructure, that's local. And models that are trained locally, by the way, fully open, so they can be deployed and c…”
Midha: Anthropic's mission has always been aligned with US national interests
“Anthropic, I will say, you know, the mission and vision has always been very I think it's always been very American aligned, right? They've always said, hey, America is, The crown jewel of the world in terms of innovation. This is where we're located. Anthropi…”
Midha: 21 of 22 Sand Hill Road VCs Rejected Anthropic
“So I introduced them to 22, you know, friends up and down Sandhill road. And so there's some investors there and we got 21 no's, right?”
Midha: Anthropic Cut Initial Seed Target From $500M to $100M
“Remember we originally tried to go out and raise five hundred million and then had to re-anchor to only raising a hundred million dollar seed round”
Midha: Amazon's Initial Deal With Anthropic Was $4 Billion
“And that's why, you know, it resulted in deep compute and capital for equity partnership with Amazon. That was originally four billion dollars.”
Midha: Tech companies will increasingly adopt public benefit charters
“We need more public benefit charters in Silicon Valley and in technology, and I think we will get there.”
Midha: Anthropic is the fastest-growing business of all time
“Tell them to give me a call when they'd like to be investors in the world's fastest growing business of all time, and then they can lecture me about public benefit governance and market share adoptions.”
Midha: AMP provides billions in compute infrastructure at cost
“We're actually giving away most of our compute at cost. Now, if you're a shareholder, you'd go, wait on, do you have billions of dollars of compute infrastructure you're giving away at cost? Yes, because we think that's the right thing for humanity.”
Midha: Andreessen Horowitz used its balance sheet to procure compute via Oxygen
“When I got to A-sixteen Z as a general partner, the first thing I did is I sat down with Mark and Ben and said, we need more compute. We need compute access for these incubations I'm going to do. And they said, no problem, Ansh, let's set up a program. What do…”
Midha: AMP does not own data centers or operate as a traditional VC
“We, we're not a cloud provider. We don't own our own data centers. We're not a traditional venture capital firm either. We see ourselves as an independent system operator, which means our job is to coordinate capacity across the ecosystem in a way that allows …”
Midha: Frontier AI Labs Are Running Compute at Half Capacity
“And many of them are running their own generators in their backyards at half capacity. And I'm going, this makes no sense. Let's all pull our generators so that a shoe factory can spike up during the day, a steel factory can spike up during the night, and then…”
Midha: Traditional VC and Company Incubation Models Cannot Coexist in One Firm
“I think it's very hard for them to coexist inside of one person, and it's very hard to coexist sometimes inside of even one firm”
Anj Midha: Compute infrastructure requires an electrical grid-like coordinating mechanism
“Well, obviously in the world of infrastructure, I think we need something like the grid. For, in the compute infrastructure. So that's what I've spent most of my days on, which is a coordinating mechanism for that, that allowed this, the, not the commoditizati…”
Midha: AMP Is Securing 1.3 Gigawatts of Compute Infrastructure
“We have started securing about 1.3 gigawatts of computer infrastructure. That's roughly forty billion dollars of cloud spend over the next four years. And that is financed roughly, you know, between with about 20% of equity. The remaining is debt. So 20%, that…”
Midha: Google operates 12 to 15 gigawatts of compute infrastructure
“I think Google Is roughly at 12 to 15 gigawatts of that I'm aware of infrastructure for internal and external deployed needs.”
Midha: Europe needs Google-scale compute infrastructure for AI sovereignty
“Like that's roughly what the continent needs for full sovereignty, right? To have as, at least as much infrastructure locally as there is within the alphabet holdings sort of pool. Over the next four years.”
Midha: AI ecosystem is deeply underinvested in secure compute
“We are deeply under invested in security and secure compute.”
Midha: Tech Is in a GPU Wastage Bubble, Not an AI Bubble
“We're not in an AI bubble for sure. I'll tell you that, which is the question I keep getting asked. We are definitely in a GPU wastage bubble where there are stranded pockets of compute, like billions of dollars of compute that are sitting unutilized.”
Midha: AI Compute Is Non-Fungible Across Nvidia Chip Generations
“Computer is not fungible today. So forget fungibility of compute across different manufacturers, like NVIDIA and AMD. Within a manufacturer, NVIDIA chips, for example. The H 100, the GB 200, the GB 300, these are all completely different chip types. So if you …”
Midha: Trump Is Providing Necessary Regulatory Freedom for US AI Innovation
“Now, president Trump is actually, I think trying to do his best from what I can tell in at least giving America enough freedom to innovate that these standards can even be discovered in our labs here. Because first you need somebody to actually pioneer and fig…”
Midha: Integrated Huawei Chips Rival Top US AI Chips
“Huawei chips are able to produce capabilities improvements today in China that rival some of the best chips here when integrated up and down the stack.”
Midha: Compute shortages will prevent top AI labs from hitting revenue targets
“I think if there's any reason why OpenAI, Anthropic, Gemini, and so on don't hit their revenue targets over the next few years, it's because they won't have access to enough compute.”
Midha: Optimal tech market competition consists of three or four key players
“The optimal competition set set up is you have three or four teams. In every frontier that are making extraordinary progress. And so if you invest in them, you get extraordinary returns, but they're not so comfortable as to be a monopoly such that they can sto…”
Midha: VCs Are Burning Hundreds of Millions on Redundant Inference Startups
“It's not clear to me that we need 50 inference companies, and it's not clear to me that VCs are smart enough to realize that they're just lighting hundreds of millions of dollars on fire in a category where having four or five really good inference trusted pro…”
Midha: General-purpose AI models will be made available to everyone
“So if you have a general model that's good for everybody, it will be available to everyone.”
Midha: Many Future Frontier AI Companies Will Pass $100B Valuations
“Oh, so many. I'm saying periodic is one.”
Midha: Venture Capital Community Has Forgotten How to Build Businesses
“The commercial community has forgotten how to build businesses, and they've forgotten the difference between first principles and marketing.”
Midha: AI infrastructure projects will continuously raise capital without end
“As long as the capabilities frontier keep moving and we want a healthy, independent ecosystem, we'll just keep Raising more capital. There's no end to that. I don't really, the day machine learning stops working as a systematic way to give humanity more capabi…”
Midha: VC managers are misallocating public capital into non-viable startups
“There's a huge misallocation of public capital into venture managers who didn't, are not capturing enough value in Frontier AI. Instead, they're investing in a bunch of stuff that's not going to exist, and the public's going to be mad.”
Midha: Invested hundreds of millions into Anthropic across multiple rounds
“I've had the privilege to invest many hundreds of millions of dollars into Anthropic across several rounds from the first to the most recent one.”
Anj Midha intends to donate most of his Anthropic returns to education
“I intend to give most of that away to public benefit causes, public benefit education programs.”
Shorts cut from this episode
▶ China’s Secret to Winning in AI · 20VC with Harry Stebbings (@42:46)
▶ Anthropic's 1st Round: We Got 21 No's · 20VC with Harry Steb (@17:15)
▶ What makes Anthropic's Dario so special · 20VC with Harry St (@1:04:33)
▶ Is this what Europe needs to be competitive? · 20VC with Har (@34:06)
▶ "AI Alignment is NOT the hardest problem..." · 20VC with Har (@0:00)
▶ We're in 1885 Industrial Revolution England Right Now · 20VC (@24:56)