Investor Anjney Midha: Anthropic's First Round Rejected by 21 VCs
Anjney Midha recalled that in early 2021, he arranged meetings with about 22 venture capital firms for Anthropic on Sand Hill Road, of which 21 rejected the proposal, including those he once viewed as heroes. The team at that time consisted of researchers like Dario Amodei and Daniela Amodei, who had left OpenAI, with backgrounds in GPT-3, scaling laws, and interpretability. The common reason for rejection was, "The direction is good, but where's the evidence?" Midha mentioned that he was almost ready to invest most of his savings but couldn't convince familiar funds.
The company ultimately did not complete its first round through traditional institutions. In May 2021, Anthropic announced it had completed a $124 million Series A round, led by Jaan Tallinn, co-founder of Skype, with participants including Dustin Moskovitz, Eric Schmidt, James McClave, and the Center for Emerging Risk Research. The official purpose was for compute-intensive research to create larger models that are more controllable, interpretable, and reliable. Midha later stated that the founding round was primarily funded by a few individuals with tens of millions of their own capital; the Series A and B rounds were more challenging, and it wasn't until Spark Capital led the Series C round that institutions began to follow suit.
He mentioned in interviews like 20VC that they initially aimed to raise about $500 million but later lowered the target to around $100 million; at that time, most investors were not even aware of what GPT-3 was. Midha himself was one of the earliest individual investors, later leaving Andreessen Horowitz to establish AMP and continued to invest in subsequent rounds of Anthropic. The valuation metrics after the Series A were never officially pinned down by the company, with secondary market and media figures ranging from hundreds of millions to later trillions; what is certain is that early checks came from individuals who understood the safety bets, rather than funds looking at revenue curves.
Traditional due diligence requires products, revenue, PMF, and growth. Cutting-edge models need to burn significant compute resources before any demonstrable "evidence" appears. The rejections in 2021 treated the non-existent loss curves as a veto. Individual checks filled the gaps left by institutions and shaped the early equity structure of the company into effective altruism and founder networks, rather than a standard venture capital syndicate.
In market mechanisms, this is a fill-in funding after cognitive mismatch, not product-market fit. The buyers are individuals and family offices willing to pay for an illiquid ten years for the yet-to-emerge scaling curves; the sellers are research teams blocked by institutional questionnaires. The beneficiaries are the individual checks from the Series A and the institutions that entered in the Series C; the pressured side is the Sand Hill Road funds that review scientific projects using PMF templates—they later used premiums to buy back entry tickets. Funds flowed from personal balance sheets into training clusters, while institutional funds only began to price in after evidence appeared.
Source: Public Information
ABAB AI Insight
Midha's 22 presentations turned Sand Hill Road into a controlled experiment: the same group of people, the same GPT-3 background, and 21 times they wanted to see evidence. In 2021, evidence meant a trained, controllable product with revenue, which was precisely what the team said needed to be spent first to obtain. The logic behind Tallinn's lead was not an ROI model but writing the existing risks into capital use. Moskovitz and Schmidt provided personal balance sheets, not partner meeting minutes. The institutional gap allowed the company to avoid growth clauses when it needed compute the most, and also transferred the bargaining power for subsequent rounds to the individuals who already held shares.
The capital path is that belief capital precedes risk capital. The $124 million from Series A was for the research roadmap, not a sales machine. The difficulty in Series B indicates that institutions still viewed safety labs as scientific projects before the ChatGPT explosion. Spark's entry into Series C marked a template switch: the model began to have a product surface, and funds could then write memos. Midha put his savings on the line because the introducer saw the scaling paper before the funds did. He later used the same judgment to raise AMP, effectively productizing the "betting before evidence appears" concept.
Similar structures include how Google was viewed as a scientific experiment by early portals and how SpaceX relied on Musk's personal guarantees for survival before proving reusability. The best-understood projects often lack counter-consensus space. The industry phase has moved from that failure to over-crowding: today, the same credentials would secure checks before the PPT is finished, and the next failure will occur on another bet that has yet to show a loss curve.
Structural judgment is about the transfer of pricing power. The mechanism is that the due diligence framework determines who has the right to pay for the yet-to-happen training. PMF questionnaires deem companies that prioritize compute as unqualified; individual checks rewrite unqualified into trainable. Once evidence appears, pricing power returns to institutions, but the equity has already been taken by those who believed first. What is scarce is not intelligence, but the ability to recognize the compute that needs to be burned before the indicators appear.
ABAB News · Cognitive Laws
When evidence has not yet been burned out by compute, questionnaires will systematically say no.
What individual checks can buy is the decade not yet covered by institutional templates.
The best-understood projects often no longer have counter-consensus to buy.