The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI

ELI5/TLDR

Anj Midha — founding investor in Anthropic, former a16z GP, now running AMP — walks through the four bottlenecks holding back AI progress: context feedback loops, compute, capital, and culture. He pitched Anthropic to 22 VCs and got 21 rejections because most didn’t even know what GPT-3 was. His big thesis now is that compute infrastructure needs the equivalent of an electrical grid — standardized, fungible, shared — and that without it, billions of dollars of GPU capacity sits stranded and wasted.

The Full Story

The Four Bottlenecks

Midha frames the AI landscape through four constraints: context feedback, compute, capital, and culture. He argues culture might be the most important, because if you get culture right, algorithmic innovation solves itself. Mission-driven teams don’t get attached to one architecture. They just want to push the frontier.

Context feedback is where the real commercial advantage lives. Pre-training data from the internet gets you coding models. But physics, chemistry, materials science — that data is locked in national labs, academic institutions, semiconductor plants. It doesn’t exist on the internet in usable form.

“There’s this disconnect between the marketing hype of AI for science and the reality where these models are terrible.”

His solution: build the feedback loops yourself. His company Periodic Labs has a physical lab in Menlo Park with robots synthesizing new materials, X-ray diffraction machines validating predictions, and the verification data piped back into training runs. He claims compute applied to superconductor discovery is seeing “super-exponential gains per iteration.” The bitter lesson, he says, is alive and well — just not in the saturated domains everyone benchmarks.

The Anthropic Origin Story

Midha knew Tom Brown (lead author on GPT-3) for years. In early 2021, Tom called to say a group was leaving OpenAI to start a new lab. Midha, Dario Amodei, and Tom started weekly sessions to figure out how to turn a research hypothesis — the scaling recipe — into a business.

The plan was simple: raise money, buy compute, collect coding feedback data, deploy models, pipe that feedback into training, repeat. Inference revenue buys more compute while generating the context feedback to keep improving capabilities.

Midha invested his life savings (most of his net worth was tied up in Discord stock) and introduced the founders to 22 VC friends on Sand Hill Road. Twenty-one said no.

“They said, ‘What’s GPT-3?’ I was like, ‘Oh my god.’”

They originally tried to raise $500 million, had to reanchor to a $100 million seed. The VCs who passed didn’t understand the technology. The people who eventually got it were either from the ML/effective altruist community (like SBF) or Amazon, who saw the strategic value of hosting frontier models on AWS as a counterweight to the Azure-OpenAI partnership. That led to the $4 billion Amazon deal.

On Dario specifically, Midha describes three qualities: genuine scientific brilliance (he’s a physicist at heart, not a computer scientist), an obsessive empiricism, and ruthless mission alignment that attracts world-class talent even in the face of public criticism.

Compute as the New Electricity

Midha’s central metaphor: we’re in 1885, the pre-standardization era of compute infrastructure. Factories are running their own generators at half capacity. The answer is a grid.

The core problem is that compute isn’t fungible. Not just across manufacturers (Nvidia vs. AMD), but within Nvidia’s own lineup — H100s, GB200s, and GB300s are completely different chip types. Training runs can’t seamlessly move between them. The result: billions of dollars of stranded compute sitting unutilized while frontier teams can’t get access.

“We are not in an AI bubble. We are in a GPU wastage bubble.”

AMP is building what Midha calls “the AMP Grid” — an independent system operator (like for the electrical grid) that coordinates capacity so teams provision for base load and burst up or down as needed. They’ve secured roughly 1.3 gigawatts of compute infrastructure, about $40 billion in cloud spend over four years, financed 80% with debt and 20% equity.

For European sovereignty, he points to the Cloud Act problem: US law requires access to data on US-managed infrastructure. European military and logistics companies can’t have their AI workloads subject to that. This is Mistral’s real investment thesis in his view — full-stack independence, from land and power to locally-trained open models.

Optimal Competition, Not Monopoly

Midha updates Peter Thiel’s “competition is for losers” to something more precise: perfect competition is for losers, but so are monopolies. Fifty inference companies racing to the bottom is wasteful. But a monopolist stops innovating and starts hoarding resources.

The sweet spot is three to four teams per frontier, each making real progress but never comfortable enough to coast. He’s blunt about the current VC ecosystem: too many firms are funding 50+ inference companies, lighting hundreds of millions on fire, while the actual innovators can’t get compute because it’s been hoarded.

“The existential threat to innovation in this category is lack of compute.”

Back to the Future of Venture Capital

Midha sees a return to the founding era of VC — Arthur Rock co-building Intel from the office every day, Bob Swanson co-founding Genentech in Kleiner’s basement. He does daily 8am standups with his Periodic Labs co-founder. His compute team sits upstairs procuring GPUs for the lab.

He’s dismissive of the check-writing, SaaS-era model. Software can now play many coordinating roles of VC firms. And he thinks the wealth creation from frontier AI is being captured too narrowly — no VC firms were in Anthropic’s seed round, meaning the pension funds and endowments behind those firms missed out.

The Personal Stuff

Grew up in India, attended Rishi Valley (a boarding school with no technology — one hour of computer time per week). Government scholarship to Singapore. Stanford for grad school. He credits his older sister for fighting the family battles that gave him the model for financial independence.

On health: recent experiences with family and personal health made him reassess time. His scaling law for Stanford students — take life seriously, but not so seriously you forget what makes it worth living.

What he wants on his tombstone: “He was right.”

Key Takeaways

The four AI bottlenecks are context feedback, compute, capital, and culture — culture may be the most important because it determines whether you attract the researchers who solve the algorithmic problems
AI models are terrible at physics and materials science because the training data for those domains is locked in physical labs, not on the internet — the internet is mostly blogs and code
Periodic Labs closes the loop physically: LLMs predict materials → robots synthesize them → X-ray diffraction validates → verification data feeds back into training
21 of 22 VCs rejected Anthropic’s seed round — most didn’t even know what GPT-3 was in early 2021; the team had to reanchor from $500M to $100M
Amazon’s Anthropic investment was defensive: they watched Azure + OpenAI and saw hosting frontier models on AWS as strategically accretive
Compute is not fungible — even within Nvidia’s lineup (H100 vs GB200 vs GB300), training runs can’t move between chip types, creating billions in stranded capacity
The “AI bubble” is actually a GPU wastage bubble — capabilities aren’t saturating, but infrastructure is badly misallocated
AMP Grid model: independent system operator for compute, like the electrical grid — teams provision for base load and burst as needed; 1.3 GW secured, ~$40B over 4 years, 80% debt-financed
The Cloud Act creates structural demand for sovereign AI infrastructure — European defense and logistics companies can’t run workloads on US-managed cloud
“Perfect competition is for losers, but so are monopolies” — optimal competition means 3-4 teams per frontier, enough to keep innovating but not so many that resources are wasted
Dario Amodei is a physicist, not a computer scientist — his strength is obsessive empiricism combined with ruthless mission alignment
The VC model is reverting to co-founding: Arthur Rock at Intel, Bob Swanson at Genentech — deep partnership, not check-writing

Claude’s Take

This is a solid 75-minute window into how one of AI’s most connected investors thinks about the infrastructure layer. Midha’s grid metaphor for compute is genuinely useful — the comparison to pre-standardization electricity in 1885 clarifies why “AI bubble” talk misses the point. The real problem isn’t that capabilities are overhyped; it’s that the plumbing is badly organized.

The Anthropic founding story is worth hearing firsthand. Twenty-one rejections from Sand Hill Road because VCs didn’t know what GPT-3 was — in 2021 — is a damning data point about how disconnected most of the industry was from the ML research community.

Where this gets a bit self-serving: Midha is essentially pitching AMP throughout. The grid, the incubation model, the public benefit corporation structure — it all loops back to his current venture. That’s fine, but worth noting. The sovereign compute angle with Mistral and Europe is more interesting when you separate the investment thesis from the genuine geopolitical reality.

Harry Stebbings asks decent questions but doesn’t push back much. The conversation could have used more skepticism on whether AMP’s grid model is technically feasible or whether “fungibility” of compute is an achievable goal versus a useful abstraction.

Score: 7/10. Genuinely informative on AI infrastructure economics and the Anthropic backstory, with a clear investor’s lens that you should account for.

The Early Days of Anthropic & How 21 of 22 VCs Rejected It | The Four Bottlenecks in AI | Anj Midha