Anj Midha on Anthropic, Compute Bottlenecks and Building Frontier AI

Anjney “Anj” Midha, founder of AMP and an early Anthropic investor, joins Harry Stebbings to discuss AI research, compute infrastructure and venture capital. Midha has investments and operating interests in several companies discussed, including Anthropic, Mistral and Periodic Labs; his product, market and policy claims are his own views.

Stebbings asks whether increasing compute is producing diminishing returns. Midha says the answer depends on the domain and evaluation: a saturated coding benchmark may demand more compute for an incremental gain, while he believes materials science has much more room to advance. His claims about the pace of progress in superconductor discovery are not a verified benchmark in the interview.

At Periodic Labs, which he is incubating, Midha describes a loop in which models propose materials, robots synthesize them and physical instruments check their properties. Those results are fed back into training. He presents this as an example of how physical verification can generate data that is not readily available on the internet.

Midha identifies four bottlenecks: context and feedback, compute, capital and culture. He argues that a mission-driven culture helps attract researchers willing to change architectures, making it easier to solve algorithmic problems. This is his framework for organizing the problem, not a demonstrated ranking of constraints across all labs.

The feedback loop matters, in his view, because deploying a model in a domain and observing its performance can provide differentiated information. He says he and colleagues tested language models on physics and chemistry around a year earlier and found them weak at the time. The interview does not provide the tests, scores or a current independent comparison.

He attributes part of that weakness to missing scientific data held in laboratories and industrial facilities rather than on the public web. Periodic’s proposed response is to produce and verify data in a physical lab, then fund the compute needed to use it. Midha has a direct operating and investment interest in that approach.

Asked whether more vertically integrated model companies will emerge, Midha distinguishes strong performance on particular tasks from the ability to run a whole laboratory autonomously. He says AI can help analyze large quantities of lab data and automate some coding-related work, but cannot simply be instructed to obtain permits, raise funding and build a physical lab end to end.

Stebbings asks which vertical businesses might be displaced by a foundation-model provider adding similar functionality. Midha says access to a distinctive context-and-feedback loop may make progress and business advantage more visible, but cautions that context alone is not necessarily a moat. The distinction is his investment thesis, not a guarantee of defensibility.

Midha points to sensitive data and European demand for local infrastructure. He gives a broad interpretation of the US CLOUD Act and argues that some organizations may prefer local providers for critical workloads. His account is a commercial and legal view, not legal advice or an exhaustive description of the law’s access conditions or cross-border obligations.

Midha says Mistral’s appeal to him is an independent stack spanning local infrastructure, compute and models that can be deployed and customized. The video describes him as a Mistral board member, so his favorable account carries an investment and governance interest. Stebbings asks how US labs would compete for European government and enterprise workloads.

Midha recalls the early days of Anthropic. He says he knew co-founder Tom Brown and began meeting with Brown and Dario Amodei in early 2021 to help turn a research hypothesis into a business. In his telling, the team explored an AI programming assistant and a loop in which deployment would generate feedback and revenue for further compute.

He says he invested personal savings and introduced the founders to 22 investors, 21 of whom declined. He recalls an initial goal of raising $500 million before a smaller seed target. These figures and the reasons for investor refusals are his recollections, not independently verified records of every discussion.

Midha says the compute needs and investment logic were difficult to explain at the time. He credits later investors and Amazon with recognizing a possible fit between frontier models and cloud infrastructure, and describes an initial $4 billion compute-and-capital partnership. His account is shaped by his own early investment in Anthropic.

Stebbings asks whether successful AI labs inevitably face public conflict. Midha favors public-benefit governance, arguing that it can give leaders room to balance mission and profit over a long horizon. He cites other organizations as analogies; the conversation does not establish that their legal structures or regulatory histories are directly comparable.

For AMP, Midha says the public-benefit mission is to support an independent frontier-technology ecosystem. He says it provides much of its compute at cost because he believes smaller research teams need access. This is his description of AMP’s business and mission, and he has an interest in presenting the arrangement favorably.

Asked how AMP obtains scarce compute, he says earlier procurement work at Andreessen Horowitz and relationships with suppliers helped. He describes AMP Grid as coordinating capacity among independent providers rather than owning data centers. The idea is to let teams provision for ordinary demand while accessing additional capacity for training or inference peaks.

Midha compares today’s fragmented compute supply to factories running separate generators before electricity grids. Pooling capacity, he argues, could improve utilization. Stebbings asks whether inexpensive compute is a loss leader tied to venture investments; Midha denies that this is the deal and says AMP also incubates new companies directly.

He describes incubating one company at a time alongside scientists and engineers, citing historical partnerships such as Intel and Genentech. His preferred role combines capital, operations and company formation. He contrasts this with a model of writing checks across many companies and says the two approaches can be hard for one person to perform simultaneously.

At Periodic, Midha says he spends three days a week working with the team and holds regular planning meetings with co-founder Liam. He says AMP’s compute team helps secure capacity for the lab. This account illustrates the close relationship between the company he is promoting and the infrastructure business he runs.

As an investor, he recommends stating hypotheses, running parallel experiments and being willing to tell limited partners when a thesis is wrong. The future is not predetermined, he says, even when historical analogies offer useful questions. He uses industrial-era infrastructure as an analogy for today’s need to coordinate compute.

Midha says AMP has started securing about 1.3 gigawatts of compute infrastructure and describes it as roughly $40 billion in cloud spending over four years. He mentions approximately 20% equity and a roughly $10 billion equity figure, which are not arithmetically identical; these are his approximate financing claims, not independently verified commitments or realized spending.

For Europe, he suggests comparing infrastructure needs with the scale of a major cloud provider and says local capacity could support an independent ecosystem. He offers estimates of Google’s capacity and possible European needs, but the interview supplies no underlying capacity inventory. These figures are his directional views, not established requirements.

On financing, Midha says matching long-term equity, debt and the needs of institutions was harder than finding capital in isolation. He describes the arrangement as a systems-design problem. The comments reflect his fundraising experience and should not be read as proof that capital is readily available to every project.

Midha argues that the industry faces a “GPU wastage bubble” rather than an AI capabilities bubble. He claims significant pockets of compute sit underused, an assessment consistent with AMP’s business of coordinating capacity. He does not provide an independent utilization survey or establish the scale of any market-wide bubble.

His technical explanation is that compute is not fully fungible: different chip generations, memory capacities and software environments can constrain which workloads run on which clusters. An older cluster may not be interchangeable with a newer one for a particular training job, even if it remains useful for other tasks. He hopes common standards and coordination can reduce stranded capacity.

Stebbings asks whether standardization could remove boom-and-bust cycles. Midha thinks it could help substantially, citing earlier infrastructure industries, but that is a prediction rather than proof that standards would eliminate cycles. He believes disagreement over how to procure and govern statistical AI systems is a barrier to coordination.

Midha calls for technologists and standards bodies to propose and evaluate common procurement rules. He distinguishes statistical models from deterministic software and argues that institutions should resolve the implications together. His preference for industry and government coordination does not specify a complete standard or settle the policy debate.

Asked about Chinese models, Midha praises engineering across chips, systems and model deployment. He also alleges adversarial distillation of Western models and predicts that open-model strategies could change as capabilities improve. The discussion does not establish that any named model or organization used an illicit method; his broad geopolitical claims are his assessment.

He proposes a coordinated defensive layer for frontier inference, with providers sharing signals when they suspect an attack. He says his board relationships already give him informal opportunities to connect founders. Claims about state-sponsored activity and insider threats are not independently substantiated in the interview, and the proposed system raises implementation questions the conversation does not resolve.

On inference demand, Midha expects growth if capabilities continue improving, but says access to secure compute and capital could constrain it. He argues that labs might miss revenue targets if capacity is unavailable. These are conditional projections, not evidence that a particular company will miss its targets.

Midha revisits Peter Thiel’s phrase about competition and proposes “optimal competition”: several strong teams in a field, rather than a single monopoly or a crowded market with no differentiation. He argues that monopoly can weaken innovation, while too many competitors can waste scarce resources. This is his preferred industrial structure, not an established optimum.

He says numerous inference providers compete for the same scarce compute and argues that a smaller group with reliable supply could be healthier. AMP is actively procuring capacity for such teams, giving him a commercial interest in that diagnosis. He names access to supply as a key factor in which providers succeed.

Stebbings asks whether model creators will reserve their best systems for their own applications. Midha expects general-purpose products to be broadly distributed when providers can spread development costs, while specialized products may be segmented by customer. He sees an ecosystem of different models rather than a single universal model; this remains a prediction.

Midha prefers the term “frontier systems companies” to “foundation-model companies.” He argues that models, infrastructure, products and feedback loops form a connected system, citing Anthropic’s coding product and Mistral’s compute efforts. His investment in the companies and AMP’s infrastructure strategy are relevant to that full-stack thesis.

Asked whether AMP has enough capital, he says scaling beyond the 1.3-gigawatt project will require more funding and does not give a finite target. He expects continued fundraising while advances continue and an independent ecosystem needs support. The project scale is his description of capacity being secured, not proof that all infrastructure is operational.

Midha expects some venture investors to build institutions around close company formation rather than simply act as intermediaries for limited partners. He worries the public is not sharing enough in frontier-technology value creation. Stebbings counters that pension funds and endowments invest in venture funds, while Midha argues that access to particular early opportunities remains uneven.

Midha says he invested many hundreds of millions of dollars in Anthropic across several rounds, from its first to its most recent, and intends to give much of the proceeds to public-benefit causes. The amount is his account of investments across rounds, not a realized return or a single investment. His continuing stake is material context for his optimism about the company.

His advice to limited partners is to do the reading and understand the technical bottlenecks before choosing managers. Stebbings agrees that investors should build with AI themselves. Midha describes a course for government officials that combines technical study with building and deploying agents, as an example of the practical learning he favors.

Asked what makes Dario Amodei effective, Midha praises scientific ability, empirical reasoning and commitment to mission. As an early investor and longtime associate, he is speaking from a close relationship with Anthropic, not offering an independent evaluation. He believes those qualities help attract researchers and guide difficult trade-offs.

Midha says health experiences in his family changed how he values time and relationships. He urges students to take their work seriously without forgetting friends and family. The conversation becomes personal as he reflects on living far from his parents and Stebbings shares his own family experience.

Asked how to escape the pursuit of ever more money, Midha describes a Singapore scholarship and his wish for financial independence. He says watching his sister navigate family expectations helped him understand the freedom independence can provide. For him, knowing a mission makes it easier to turn down opportunities that do not fit.

In the closing exchange, Midha reflects on wanting to understand and communicate where technology is going. He describes attending a school in rural India with limited computer access and says he would allow children to use social media in moderation. His preference is to teach habits that make technology useful without allowing dependence.